WO2019218395A1 - 一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法 - Google Patents

一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法 Download PDF

Info

Publication number
WO2019218395A1
WO2019218395A1 PCT/CN2018/088913 CN2018088913W WO2019218395A1 WO 2019218395 A1 WO2019218395 A1 WO 2019218395A1 CN 2018088913 W CN2018088913 W CN 2018088913W WO 2019218395 A1 WO2019218395 A1 WO 2019218395A1
Authority
WO
WIPO (PCT)
Prior art keywords
gas
malodorous
gas sensor
sensor array
monitoring
Prior art date
Application number
PCT/CN2018/088913
Other languages
English (en)
French (fr)
Inventor
高大启
张小勤
王泽建
赵黎明
Original Assignee
华东理工大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority claimed from CN201810471708.3A external-priority patent/CN108896706B/zh
Priority claimed from CN201810471613.1A external-priority patent/CN108709955B/zh
Application filed by 华东理工大学 filed Critical 华东理工大学
Priority to US16/642,531 priority Critical patent/US10948467B2/en
Publication of WO2019218395A1 publication Critical patent/WO2019218395A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
    • G01N33/0004—Gaseous mixtures, e.g. polluted air
    • G01N33/0009—General constructional details of gas analysers, e.g. portable test equipment
    • G01N33/0011—Sample conditioning
    • G01N33/0016—Sample conditioning by regulating a physical variable, e.g. pressure or temperature
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N1/00—Sampling; Preparing specimens for investigation
    • G01N1/02—Devices for withdrawing samples
    • G01N1/22—Devices for withdrawing samples in the gaseous state
    • G01N1/24—Suction devices
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N27/00—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means
    • G01N27/02—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating impedance
    • G01N27/04—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating impedance by investigating resistance
    • G01N27/12—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating impedance by investigating resistance of a solid body in dependence upon absorption of a fluid; of a solid body in dependence upon reaction with a fluid, for detecting components in the fluid
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N27/00—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means
    • G01N27/26—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating electrochemical variables; by using electrolysis or electrophoresis
    • G01N27/403—Cells and electrode assemblies
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N27/00—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means
    • G01N27/62—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating the ionisation of gases, e.g. aerosols; by investigating electric discharges, e.g. emission of cathode
    • G01N27/64—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating the ionisation of gases, e.g. aerosols; by investigating electric discharges, e.g. emission of cathode using wave or particle radiation to ionise a gas, e.g. in an ionisation chamber
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
    • G01N33/0004—Gaseous mixtures, e.g. polluted air
    • G01N33/0006—Calibrating gas analysers
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
    • G01N33/0004—Gaseous mixtures, e.g. polluted air
    • G01N33/0009—General constructional details of gas analysers, e.g. portable test equipment
    • G01N33/0027—General constructional details of gas analysers, e.g. portable test equipment concerning the detector
    • G01N33/0031—General constructional details of gas analysers, e.g. portable test equipment concerning the detector comprising two or more sensors, e.g. a sensor array
    • G01N33/0032—General constructional details of gas analysers, e.g. portable test equipment concerning the detector comprising two or more sensors, e.g. a sensor array using two or more different physical functioning modes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N1/00—Sampling; Preparing specimens for investigation
    • G01N1/02—Devices for withdrawing samples
    • G01N1/22—Devices for withdrawing samples in the gaseous state
    • G01N1/26—Devices for withdrawing samples in the gaseous state with provision for intake from several spaces

Definitions

  • the invention relates to an online monitoring and analysis method for a multi-point centralized electronic nose instrument for malodorous gas, which is oriented to the market supervision requirement of the environmental protection and management department, and is oriented to (a) petrochemical, rubber, perfume flavor, pharmaceutical, paint, brewing, papermaking Industrial parks; (b) waste and sewage treatment areas such as landfill, landfill and incineration, sewage treatment; (c) farms; (d) online monitoring and analysis requirements for odorous areas such as neighbouring residential areas, involving the environment Protection, analytical chemistry, computer, artificial intelligence, big data, automatic control, precision measurement and other technical fields, mainly to solve the automation, integration and miniaturization of electronic nose instruments, online monitoring of various odor pollutants in the odorous pollution area and on-line online prediction. And the problem of trolling pollution source tracking and control.
  • Odor pollutants refer to all odorous gas substances, and generally refer to all substances that emit a foul odor. Odorous pollutants are widely distributed in all petrochemical enterprises, such as petrochemicals, garbage and sewage treatment, pharmaceuticals, and aquaculture, and neighboring residential areas. They are widely distributed and have a wide range of influence. The current situation of odor pollution in China is that there are many sources of emissions, the odor components are complex, the national standards are lagging behind, and complaints from residents are frequent.
  • the evaluation of malodorous pollution is a malodorous gas, and the evaluation method is divided into an olfactory method and an instrumental analysis method.
  • GB14554-93 "Emission Standards for Odor Pollutants" stipulates that the emission control indicators for odor pollutants include a qualitative dimensionless odor concentration and eight quantitative single component concentrations, namely trimethylamine (C 3 H 9 N), styrene. (C 8 H 8 ), hydrogen sulfide (H 2 S), methyl mercaptan (CH 4 S), methyl sulfide (C 2 H 6 S), dimethyl disulfide (C 2 H 6 S 2 ), ammonia ( NH 3 ), carbon disulfide (CS 2 ).
  • GB/T18883-2002 Indoor Air Quality Standards
  • GB14554 stipulates that the concentration of odor is determined by the three-point comparative odor bag method, and the concentration of C 3 H 9 N, C 8 H 8 , H 2 S, CH 4 S, C 2 H 6 S and C 2 H 6 S 2 is determined.
  • “Odor concentration” refers to the odor samples collected at the scene of successive dilutions in the laboratory with odorless clean air to the dilution factor sniff member of the olfactory threshold, a measure of the value of EU standards EN17325-2003 with OU (odor unit).
  • the standard identification method of odor concentration mainly relies on the nose of the sniffer [2]! China, Europe, America, Japan and South Korea and other countries and regions are the same.
  • the implementation of the national standard GB/T14675-93 for 25 years, "Determination of ambient air - malodor - three-point comparison odor bag method" regulates the selection of olfactory identification, odorous gas sample collection, sample manual dilution and olfactory determination .
  • countries in Europe, America, Australia, New Zealand and other countries use dynamic olfactometer to dilute odor samples [3].
  • GB/T14675 and HJ905 stipulate that the malodorous gas sample is collected by the staff on the spot with a sampling bottle or an odorless air bag (for example, 10L).
  • odorless air bag for example, 3L
  • a certain amount of syringe and diluted with odorless clean air and finally sniffed by the sniffing group members.
  • a odorless air bag for example, 3L
  • One of the cores of the three-point comparison odor bag method is that after the odor sample is diluted once, one sniffer must sniff three 3L air bags, one of which is a diluted odor bag, and the other two are odorless. Air bag, and can identify the odor bag.
  • the NH 3 olfactory threshold is 5 times different from Japan, the H 2 S is nearly 3 times different, the trimethylamine is 28.12 times different, the positive valeric acid is 65.67 times, and so on.
  • the above results at least indicate two problems: (1) the olfactory process for determining the odor concentration is complicated, and the odor evaluation is very costly; (2) the olfactory threshold given by the various units in each country is not objective and not reproducible. .
  • the three-point comparative odor bag method specified in GB/T14675 can reflect the ordinary people's feelings, but the operability is very poor.
  • To do a sniffer test requires a large number of sampling and sniffing personnel, which is costly, especially unsuitable for low concentration and toxic.
  • the smell of matter [2, 4-5].
  • the odor evaluation result of the three-point comparative odor bag method is selected by 1 on-site sampling point; 2 sampling device; 3 laboratory conditions; 4 sniffer ability and state; 5 odor concentration and initial dilution factor; 6 sniffer time
  • Olfactory is the complex feeling of a large number of olfactory cells in the nasal cavity.
  • the olfactory simulation method uses an array of multiple gas sensors with overlapping performances to achieve rapid odor detection and qualitative and quantitative analysis, which has attracted people's attention [10].
  • electronic nose technology can determine the taste class, intensity, quality level, true and false, freshness, control production process, formulation and production process, etc. through the multi-sensory response of the gas sensor array to the odor.
  • the electronic nose method is now mainly used for qualitative and quantitative analysis of complex odors, for example, wine [11], tea [12], milk [13], grain and oil quality [14]; fruit maturity [15]; freshness of fish products [16] Water and ambient air monitoring [17-18]; disease diagnosis [19-20]; bacterial odor perception [21];
  • the application of electronic nose technology has broad prospects.
  • One of the development trends is to develop high sensitivity and high selectivity gas sensing devices to achieve qualitative and quantitative detection and analysis of odor.
  • the sensitivity of SnO 2 semiconductor gas sensors has reached the order of 10 -9 V/V (ppb) [10], which directly produces a V-level voltage response to odor without the need for secondary amplification, which is a odorous pollutant.
  • Online monitoring is very attractive.
  • the second trend of electronic nose technology is to form an array with a plurality of different types of gas sensors with the necessary sensitivity, focusing on the use of data analysis methods to improve the selectivity of the detected objects, to achieve odor recognition, strength estimation and key component prediction.
  • the existing electronic nose monitoring system adopts a monitoring point, an electronic nose, that is, a "one-on-one nose” layout method, which is obviously affected by the installation mode of the surveillance camera.
  • the electronic nose and the camera have similarities, but the conditions of use are completely different.
  • the installation of the electronic nose directly on the monitoring site will cause a series of problems: for example, the electronic nose instrument, especially the core component—the gas sensor has long been subjected to wind, sun and rain, which affects the life; the gas sensor is prone to fatigue when exposed to malodorous gases for a long time.
  • malodorous gases are that (1) the composition is numerous and complex. Except for a few inorganic substances such as H 2 S, NH 3 , CS 2 , SO 2 , etc., most of them are organic substances, so-called “volatile organic compounds”; (2) some odorous substances have low olfactory threshold, but contribute to odor concentration. The degree is very large; vice versa; (3) some substances are non-toxic and harmless, contributing little to the odor concentration, and the gas sensor is very sensitive.
  • Gas sensor selection should consider factors such as sensitivity, selectivity, response speed, stability, commercialization, miniaturization, life, and cost.
  • the MOS type gas sensor has high sensitivity and obvious advantages.
  • the disadvantage is that the selectivity is not ideal and should be the first choice for the array components.
  • the EC type gas sensor has the advantages of better selectivity, the disadvantage is that the sensitivity is 1-2 orders of magnitude lower than the former, the life is short (generally 1-2 years), the size is large, and the stability is poor. It is mainly used for the detection of toxic gases such as H 2 S, NH 3 , CS 2 and SO 2 [ 10 , 22 ].
  • the PID type gas sensor is characterized by being sensitive to VOCs between n-hexane and n-hexadecane, but it is not unique, and has disadvantages such as large size, short life, and high price.
  • the Airsense and alpha MOS electronic noses are based on the European standard EN13725.
  • EN13725 stipulates that the odor can be continuously diluted by any multiple (maximum 2 ⁇ 10 6 times) with a dynamic sniffer, and 0.04 ⁇ 10 -6 (V / V) n-butanol is used as a standard substance to specify the standard gas pair.
  • the human olfactory organ acts as the basic unit of odor concentration (OU E ), "the olfactory threshold of any mixed odor is equivalent to the stimulating amount of the standard gas to humans", and the quantity conforms to the ISO dimension system and has an automatic test characteristic.
  • GB/T14675 belongs to the olfactory static measurement method, which is a manual test method, and its operability and scientificity are worse than EN17325 [26].
  • a malodorous electronic nose instrument can realize simultaneous online monitoring of multiple observation points in a specific area (for example, within an area of 4km 2 ), that is, fixed point monitoring or moving point monitoring, of course, monthly or even
  • the unit continuously monitors 24 hours a day; proposes a simple and effective machine learning model and algorithm to achieve 24-hour continuous estimation and prediction of the above 10+1 odor pollutant concentrations, and uses wireless WIFI technology to transmit monitoring data and analysis results in real time.
  • To the monitoring center and various terminals realize remote control of malodorous pollution based on the Internet.
  • RVD Goor MV Hooren
  • AMD AMDingemans
  • B. Kremer K. Kross
  • Training and validating a portable electronic nose for lung cancer screening Journal of Thoracic Oncology, 2018 (In press), doi.org/10.1016/j .jtho.2018.01.024.
  • the invention is a prior invention patent "a machine olfactory device and its olfactory simulation test method” (application number: 02111046.8), "a method for identifying a olfactory odor based on a modular combined neural network” (application number: 03141537.7) , "A olfactory simulation instrument and qualitative and quantitative analysis methods for multiple odors” (application number: 201010115026.2), and "a multi-channel integrated olfactory simulation instrument and online analysis method for biological fermentation process” (application number: 201310405315.X)
  • a online monitoring and analysis method for multi-point centralized electronic nose instrument of malodorous gas was invented to solve the problem of long-term online monitoring of multiple monitoring points in odorous pollution areas and on-line prediction of various odor gas concentration control indicators.
  • the odorous gas multi-point centralized online monitoring and analysis system of the present invention comprises a malodorous electronic nose instrument I, a gas sampling probe II, an external vacuum pump III, an ambient air purifying device IV, a clean air bottle V, a gas pipeline,
  • the electronic thermometer and hygrometer VI, the central control room VII and a plurality of fixed/mobile terminals VIII realize long-term online monitoring of 10 monitoring points in the malodorous pollution area and online estimation and prediction of various odor pollutant concentration control index values.
  • the malodor electronic nose instrument I includes a gas sensor array and its constant temperature studio I (a), multi-point centralized malodorous gas automatic sample introduction system I (b), computer control and data analysis system I (c) three major components.
  • the gas sensor array thermostat chamber I(a) is composed of a gas sensor array and its annular working chamber I-1, a heat insulating layer I-2, a resistance heating wire I-3, and a fan I-4.
  • the gas sensor array I-1 is composed of 16 gas sensors, and is evenly distributed in a sealed cavity with a diameter of ⁇ 140 mm and a section size of 21 mm ⁇ 17 mm to form an annular working chamber of the gas sensor array, which is in a constant temperature chamber of 55 ⁇ 0.1 °C.
  • Multi-point centralized malodorous gas automatic sample introduction system I(b) includes built-in micro vacuum pump I-14 and 14 two-position two-way solenoid valves (I-5, I-6-1 ⁇ I-6-10, I-8 , I-10, I-13), throttle valve I-11, flow meter I-12, vacuum pressure gauge I-7, gas buffer chamber I-9, located at the lower right of the malodorous electronic nose instrument I.
  • Computer control and data analysis system I(c) includes computer motherboard I-15, data acquisition card I-16, display I-17, drive and control circuit module I-18, precision linear and switching power supply module I-19, hard disk, Network card, graphics card, located on the left side of the malodorous electronic nose instrument I.
  • the computer control and data analysis system I(c) uses the machine learning cascade model to determine the odor odor concentration of the monitoring point, the ammonia NH 3 specified by GB14554, the hydrogen sulfide H 2 S, the carbon disulfide CS 2 , 8 compounds such as trimethylamine C 3 H 9 N, methyl mercaptan CH 4 S, methyl sulfide C 2 H 6 S, dimethyl disulfide C 2 H 6 S 2 , styrene C 8 H 8 , GB/T18883
  • the specified sulfur dioxide SO 2 and total volatile organic compounds total 10+1 odor pollutant concentration control index values are analyzed and predicted in real time, and the monitoring data and prediction results are transmitted remotely to the central control room VII and designated through the wireless Internet.
  • Fixed/mobile terminal VIII The specified sulfur dioxide SO 2 and total volatile organic compounds total 10+1 odor pollutant concentration control index values are analyzed and predicted in real time, and the monitoring data and prediction results are transmitted remotely to the central control room VII and designated through the wireless Internet
  • the malodorous electronic nose instrument I obtains a 16-dimensional response vector every single cycle T 0 and stores it in a data file on the computer's hard disk.
  • Use 10 two-position two-way solenoid valves I-6-1 ⁇ I-6-10 to sequentially control the on and off of the malodorous gas in 10 monitoring points in the 4km 2 area, and realize the cycle of the odorous gas cycle of T 10T 0
  • the monitoring points of the malodorous gas are monitored online, and the monitoring data is sequentially stored in 10 data files. These data are the numerical basis for predicting the various concentrations of malodorous pollutants by the malodorous electronic nasal instrument I, thereby realizing the cyclic online prediction of the 10+1 odor pollutant concentration control index value.
  • the single sampling period T 0 of the malodorous gas includes: initial recovery of the gas sensor array I-1 (95-215 seconds), accurate calibration of clean air (30 seconds), balance (5 seconds), headspace sampling of malodorous gas (30 seconds), Purify ambient air flush (20 seconds) for a total of 5 stages.
  • the solenoid valve is disconnected, and the built-in micro vacuum pump I-14 sucks the malodorous gas in the gas buffer chamber I-8 at a flow rate of 1,000 ml/min, and flows through the annular working chamber of the gas sensor array to pass over the surface of the gas sensor sensitive membrane.
  • the gas sensor array I-1 thus produces a sensitive response for 30 seconds.
  • the computer control and data analysis system I(c) continuously records sensitive response data, including balance (5 seconds), malodorous gas headspace sampling (30 seconds), and clean ambient air flushing (first 10 seconds).
  • the gas sensor array I-1 which has a total of 45 seconds in these three stages, responds to the data and is temporarily stored in a text file.
  • the response data of the other period T 0 other time is not recorded.
  • the computer control and data analysis system I(c) predicts the 10+1 odor pollutant concentration control index value based on this response vector.
  • the gas sensor array I-1 is composed of 11 metal oxide semiconductor types, 4 electrochemical types, and 1 photoion type gas sensor. Among them, 11 metal oxide gas sensors are used to detect various organic/inorganic compounds; 4 electrochemical gas sensors are used to detect 4 inorganic compounds such as NH 3 , H 2 S, CS 2 and SO 2 ; A photoionizing gas sensor is used to detect total volatile organic compounds.
  • the odorous gas multi-point centralized online monitoring and analysis system realizes a specific area of up to several square kilometers, and can also realize online monitoring and analysis of as little as 1 production workshop or 1 building, or even 1 point.
  • 10 monitoring points can be set.
  • the malodorous electronic nose instrument I is located indoors and is connected to each monitoring point through a stainless steel pipe with an inner diameter of ⁇ 10 mm.
  • the gas sampling probe is in the form of a faucet and is connected to a commercial dust removal and dehumidification unit for replacement. To change the position of the monitoring point, you only need to re-lay the stainless steel pipe, install and move the gas sampling probe to the designated position, upstairs and downstairs, high or low, just as easy to lay water pipes or cables.
  • Eight or more monitoring points are set around the boundary of the monitoring area, and the position of the malodorous electronic nose instrument I and the 10 monitoring points is set to target the shortest stainless steel pipe.
  • the malodorous electronic nose instrument I is arranged in a certain room in the regional center; for the landfills and sewage treatment plants where there is no road, the malodorous electronic nose instrument I is arranged. In a certain room at the border of the area.
  • the malodorous gas sucked into the malodorous electronic nose instrument I does not flow through the annular working chamber of the gas sensor array I- 1, but was directly discharged to the outside.
  • the malodor electronic nose instrument I is internally provided with a gas buffer chamber I-8 having a size of ⁇ 40 mm*5 mm, and the flow rate of the malodorous gas here is 16 times lower than that of the inner diameter ⁇ 10 mm stainless steel pipe. Only in the stage of the top air sampling (30 seconds) of the malodorous gas, the built-in micro vacuum pump I-14 sucks the malodorous gas in the gas buffer chamber I-8 into the annular working chamber of the gas sensor array, the gas sensor array I- 1 therefore produces a sensitive response. The built-in micro vacuum pump I-14 draws fresh malodorous gas.
  • the 1,000 ml/min clean air calibration step (30) caused multiple sensing of the malodorous gas by the gas sensor array I-1 on the same baseline.
  • the standard volume of the 12 to 15 Mpa compressed gas cylinder V is 40 L, and the conversion to normal temperature and normal pressure is 6 m 3 .
  • the ambient air outside the malodorous electronic nose instrument I is first cleaned with a commercial air purifier and then used to flush the gas sensor array I-1 to initially return to the baseline state to reduce operating costs.
  • the big data set of malodorous gas includes: (1) Gas sensor array I-1 for a large number of odor pollutants in chemical parks (including flavor and fragrance plants), pharmaceutical plants, landfills, sewage treatment plants, farms, neighboring residential areas, etc. On-site inspection data on site.
  • the malodorous electronic nose instrument I uses the machine learning cascade model to predict the odor smell concentration and various odor pollutant concentration control index values at the t+1, t+2 and t+3 moments in the future.
  • the first stage of the machine learning cascade model the Convolutional Neural Network (CNN) layer is responsible for predicting the response of the gas sensor array I-1 to a malodorous gas at a monitoring point at times t+1, t+2, and t+3. According to the current time t and the recently occurring gas sensor array I-1 time series response.
  • the second level of the machine learning cascade model the deep neural network (DNN) layer further predicts the odor odor concentration and various odor pollutant concentration control index values at t+1, t+2, and t+3. It is based on the long-term accumulation of malodorous gas big data and the predicted value of the first-level convolutional neural network layer of the cascade model.
  • Convolutional neural network CNN i1 online learning preprocessed gas sensor i time series response data set X i1 is:
  • the target output is:
  • d i1 (x i (t) x i (t-1) x i (t-2) x i (t-3) x i (t-4) x i (t-5) x i (t-6 ) x i (t-7) x i (t-8) x i (t-9)) T ⁇ R 10 ,
  • the convolutional neural network CNN i1 responds to a 9-dimensional time series based on the most recent time period after the online learning is completed within 10 seconds:
  • x i1 (x i (t-8) x i (t-7) x i (t-6) x i (t-5) x i (t-4) x i (t-3) x i (t -2) x i (t-1) x i (t)) T ⁇ R 9
  • the response x i (t+1) of the gas sensor i at time t+1 is predicted.
  • T 0 40 minutes, it is equivalent to predicting the response of the gas sensor i in the 40th minute in the future.
  • the online learning preprocessed data sets X i2 and X i3 are:
  • the time series response of the convolutional neural networks CNN i2 and CNN i3 in the target output and prediction of the learning phase is the same as CNN i1 .
  • T 0 40 minutes
  • predicting the response of the gas sensor i at time t+2 and t+3 x i (t+ 2) and x i (t+3), respectively, are equivalent to predicting the response of the gas sensor i in the next 80 minutes and 120 minutes.
  • the overall prediction problem of the organic compound and odor odor concentration total 10+1 odor pollutant concentration control index value is decomposed into 11 single concentration values one-to-one prediction problem, and the machine learning cascading model second level uses 10+1 singles
  • the output three hidden layer depth neural network module respectively predicts the value of the 10+1 odor pollutant control index.
  • the single-output deep neural network training set is the gas-sensitive sensor array I-1 of the malodorous electronic nose instrument I.
  • the malodorous gas big data obtained from the standard odor/gas sample and the large-scale pollution field online detection, the target output is the odor odor value and Color spectrum and spectrophotometric routine instrument offline measurements, as well as resident complaint data.
  • a single single output three hidden layer deep neural network DNN j uses a bottom-up layer-by-layer offline learning method.
  • the first and second hidden layer learning adopts a single hidden layer peer-to-peer neural network structure, that is, the hidden layer of the single hidden layer peer-to-peer neural network—the output layer weight is directly equal to its input layer—the hidden layer weight, and the target output is directly equal to Its input, input component and output component are scaled to the range [0, 3] according to the size of the feature component.
  • the jth concentration value y j (t+1) of the malodorous gas at time t+1 is predicted, and the jth single output depth neural network DNN j is based on 16 convolutional neural networks for the t+1 moment gas sensor
  • the predicted response of array I-1 ⁇ x 1 (t+1), x 2 (t+1),..., x 16 (t+1) ⁇ predicts y j (t+2) and y j (t+3 ) based on the predicted responses of 16 convolutional neural networks for t+2 and t+3 times (x 1 (t+2), x 2 (t+2),..., x 16 (t+2)) T and (x 1 (t+3), x 2 (t+3), ..., x 16 (t+3)) T .
  • the actual output of DNN j is an estimate of the current concentration value y j (t) of the malodorous gas component j.
  • the odorous electronic nose instrument I predicts the long-term online monitoring of multiple monitoring points in the odor-contaminated area and the online prediction of various odor pollutant concentration control index values, including the following steps:
  • the purified ambient air flows through the 2/2-way solenoid valve I-5, the gas sensor array annular working chamber I-1, and the second position at a flow rate of 6,500 ml/min.
  • the solenoid valve I-10 is passed through and then discharged to the outside.
  • the temperature in the annular working chamber I-1 of the gas sensor array reaches a constant 55 ⁇ 0.1 ° C from room temperature.
  • the external vacuum pump III sucks the malodorous gas of a certain monitoring point with a straight line distance of up to 2.5km in a minute by a pumping rate of 250-280L/min and an ultimate vacuum of 100-120 mbar through a stainless steel pipe with an inner diameter of ⁇ 10 mm.
  • the external vacuum pump III continues to pump malodorous gas until the operator clicks the "External Vacuum Pump Off" option on the on-screen menu.
  • the two-position two-way solenoid valve I-6-k+1 is turned on, and the remaining nine of the two two-position two-way solenoid valves I-6-1 to I-6-10 are disconnected, including two bits.
  • the solenoid valve k is opened, and the external vacuum pump III is turned to suck the malodorous gas of the monitoring point k+1. Due to the action of purifying the ambient air, the heat accumulated in the annular working chamber of the gas sensor array is taken away, and the malodorous gas molecules adhering to the surface of the sensitive membrane of the gas sensor and the inner wall of the pipeline are initially washed away, and the gas sensor array I- 1 gradually restored to the baseline state; lasted 20 seconds. among them:
  • (b2) Gas Sensing Sensor Array Response Prediction The first level of the machine learning cascade model—16*3 convolutional neural networks are based on the current time t before [t-18, t], [t-19, t-1] and The gas-sensing sensor array time series response vector that has occurred in the [t-20, t-2] time period, realizes online self-learning, and predicts the future T 0 , 2T 0 and 3T 0 moments of the gas sensor array I-1 the response to.
  • the malodor electronic nose instrument I realizes the online monitoring, identification and prediction of the 10+1 odor pollutant control index value of 10 monitoring points.
  • Figure 1 is a diagram of the present invention - an online monitoring and analysis method for a multi-point centralized electronic nose instrument for malodorous gas - the development of a malodorous electronic nose instrument, a machine learning cascade model and algorithm, and an online detection and prediction of malodorous pollutants. block diagram.
  • FIG. 2 is a schematic diagram of the present invention - an online monitoring and analysis method for a multi-point centralized electronic nose instrument for malodorous gas - a working principle of a multi-point centralized monitoring and analysis system for malodorous electronic nose instruments and malodorous pollution areas.
  • FIG. 3 is a schematic diagram of the working principle of the odorous electronic nose instrument (headspace sampling state) of the present invention, a method for online monitoring and analysis of a odorous gas multi-point centralized electronic nose instrument.
  • FIG. 4 is a schematic diagram of the present invention, a method for online monitoring and analysis of a multi-point centralized electronic nose instrument for malodorous gas, a malodorous electronic nose instrument and a plurality of monitoring points.
  • Figure 5 is a schematic diagram of the present invention - an online monitoring and analysis method for a multi-point centralized electronic nose instrument for malodorous gas - a gas sensor array arrangement and a ring working chamber.
  • a gas sensor array component unit (a) a gas sensor array component unit; (b) an annular working chamber cover; (c) a sectional view of the annular working chamber.
  • Figure 6 is a schematic diagram of the present invention - a method for online monitoring and analysis of a malodorous gas multi-point centralized electronic nose instrument - a malodorous gas buffer chamber.
  • Fig. 8 is a perspective view showing the appearance of a odorous electronic nose instrument according to the present invention, a method for online monitoring and analysis of a multi-point centralized electronic nose instrument for malodorous gas.
  • Figure 9 is a schematic view of the present invention - a method for online monitoring and analysis of a multi-point centralized electronic nose instrument for malodorous gas - a back view of a malodorous electronic nose instrument.
  • Figure 10 is a method for online monitoring and analysis of a multi-point centralized electronic nose instrument for malodorous gas - Convolutional neural network CNN i1 predicts t+1 time (for example, the 40th minute in the future) gas sensor i response x i (t +1) Schematic.
  • Figure 11 is a schematic diagram of the present invention - an online monitoring and analysis method for a multi-point centralized electronic nose instrument for malodorous gas - a deep neural network DNN j layer k learning process.
  • Figure 12 is a schematic diagram of the present invention - an online monitoring and analysis method for a multi-point centralized electronic nose instrument for malodorous gases - a machine learning cascade model predicting the concentration of various malodorous pollutants at time t+1 (e.g., 40 minutes in the future).
  • t+1 e.g. 40 minutes in the future.
  • (a) first level - modular convolutional neural network layer
  • (b) second level - modular deep neural network layer.
  • FIG. 1 is a block diagram showing the relationship between the online monitoring and analysis method of a malodorous gas multi-point centralized electronic nose instrument, the development of a malodorous electronic nose instrument, a machine learning model and algorithm, and online detection and prediction of malodorous pollutants.
  • the invention first analyzes the characteristics of malodorous pollutants and gas sensors from a chemical and physical point of view.
  • the malodorous gas is composed of many and complex components, often containing dozens or even hundreds of odorous components, both organic and inorganic; some malodorous components contribute to the odor concentration but the true concentration may be low, so the gas sensor response is very Small; some malodorous ingredients contribute little to the odor concentration but the true concentration may be high, and the gas sensor is therefore large; and vice versa.
  • the present invention selects a small gas sensor array module composed of MOS type, EC type and PID type gas sensor.
  • the present invention proposes a multi-point centralized monitoring mode of the malodorous gas in which the key components are located indoors, and develops the malodorous electronic nose instrument accordingly.
  • the present invention further proposes to establish a large odor data, and proposes a new machine learning cascade model to realize on-line monitoring and prediction of various odor pollutants.
  • the malodorous gas big data includes: (1) the laboratory off-line detection data of the gas-sensitive sensor array I-1 of the malodor electronic nose instrument I for a large amount of malodorous standard sample headspace volatile gas, including ⁇ -phenylethyl alcohol, 5 standard odors such as isovaleric acid, methylcyclopentanone, ⁇ -undecanolactone, ⁇ -methyl hydrazine and C 3 H 9 N, C 8 H 8 , H 2 S, CH 4 S
  • Different concentrations of single-component standard malodor samples prepared from 9 kinds of malodorous compounds such as C 2 H 6 S, C 2 H 6 S 2 , NH 3 , CS 2 , SO 2 , etc., also include mixed component standards prepared from a plurality of different concentrations of compounds Odor samples; (2) on-line detection data of a large number of odorous pollutants on the gas sensor array I-1; (3) off-line olfactory data of a large amount of odor pollutants in the laboratory; (4) a large
  • FIG. 2 is a schematic diagram of the working principle of a multi-point centralized monitoring and analysis system for malodorous electronic nose instruments and malodorous pollution areas.
  • the multi-point centralized monitoring and analysis system for the odorous pollution area includes the odor electronic nose instrument I, the gas sampling probes II-1 ⁇ II-10 of the 10 outdoor monitoring points, the external vacuum pump III, the ambient air purification device IV, the clean air V,
  • the electronic thermometer and hygrometer VI, the central control room VII and its plurality of fixed/mobile terminals VIII realize long-term online monitoring of 10 monitoring points in the malodorous pollution area and online prediction of various concentration control index values of malodorous gases.
  • the position of the gas path and the solenoid valve is the first monitoring point II-1.
  • the malodorous gas is sucked to the malodorous electronic nose instrument I, and the gas sensor array I-1 thus generates a sensitive response working state.
  • FIG. 3 is a schematic diagram of the working principle of the malodorous electronic nose instrument 1. Its constituent units include:
  • Gas sensor array thermostat studio I (a): gas sensor array and its annular working chamber I-1, insulation layer I-2, resistance heating wire I-3, fan I-4, located in malodorous electronics The upper right side of the nose instrument.
  • Valve I-8 gas buffer chamber I-9, control of gas-sensitive sensor array annular working chamber I-1 odorous gas and clean air 6,500ml/min and 1,000ml/min flow conversion two-position two-way solenoid valve I-10 Throttle I-11, flow meter I-12, two-position two-way solenoid valve I-13 for controlling clean air on and off, built-in micro vacuum pump I-14, located at the lower right of the malodorous electronic nose instrument.
  • FIG 4 is a schematic diagram showing the mutual position of the malodorous electronic nose instrument I and the 10 monitoring points II-1 to II-10.
  • the malodorous electronic nose instrument I For the path-reachable areas shown in Figure 4(a), such as chemical parks and residential areas, the malodorous electronic nose instrument I should be placed in a certain room in the center of the monitoring area; the roadless reachable as shown in Figure 4(b) In the area, the malodorous electronic nose instrument I should be placed in a certain room at the boundary of the monitoring area.
  • the position of the malodorous electronic nose instrument I is determined by the principle that the straight line distance from each monitoring point is the shortest.
  • the external vacuum pump III, the ambient air purification device IV, the clean air V, and the electronic thermometer and hygrometer VI are disposed near the malodorous electronic nose instrument I.
  • Most of the contaminated areas such as industrial parks, garbage and sewage treatment areas, farms, and neighboring residential areas that need to be monitored are within 1 km 2 .
  • the gas pipeline connecting the malodorous electronic nose instrument I and each monitoring point is arranged around the boundary.
  • the external vacuum pump III can pump the monitoring point gas into the malodorous electronic nose instrument 1 within 1 min.
  • the odor electronic nose instrument and the odor pollution area multi-point centralized monitoring and analysis system of the invention are particularly suitable for application in a production workshop, a sewage pool, a farm, etc., and can realize a specific area as large as several km2, as small as one. Online monitoring and analysis of production workshops or 1 building, or even 1 point.
  • Figure 5 is a schematic view showing the arrangement of the gas sensor array I-1 of the present invention and its annular working chamber.
  • Figure 5(a) shows a specific example: the gas sensor array consists of three types of 16 types of gas sensors, including 11 MOS models (4 TGS2000 series I-1-1, 3 plastic cases). TGS800 series I-1-2, 4 stainless steel housings TGS800 series I-1-3), 4 EC type I-1-4 and 1 PID type I-1-5.
  • MOS type gas sensor has high sensitivity, long life and sensitivity to organic and inorganic components.
  • EC type gas sensor has good selectivity and is mainly used for detecting inorganic gas.
  • PID type gas sensor is for n-hexane to n-hexadecane. The VOCs are more sensitive.
  • the machine learning cascade model determines the concentration of inorganic components such as H 2 S, NH 3 , SO 2 and CS 2 according to the responses of 11 MOS types and 4 EC type gas sensors; according to 11 MOS types and 1 PID type gas
  • the response of the sensitive sensor collectively determines the TVOC concentration of malodorous gas and the concentration of organic components such as C 3 H 9 N, C 8 H 8 , CH 4 S, C 2 H 6 S, C 2 H 6 S 2 ; according to all 16 gas sensors
  • the responses together determine the olfactory concentration OU value.
  • the gas sensor array annular working chamber I-1 is made of stainless steel base I-1-6, sealing ring I-1-7, stainless steel cover I-1- 8.
  • the partition plate I-1-9 and the gas sensor socket I-1-10, the sealing material I-1-11, and the screw I-1-12 are formed to form a sealed annular working chamber.
  • the malodorous gas is sucked from the air intake hole, and then the TGS2000 series gas sensor I-1-1 is swept through the ring type working chamber, and the three plastic casings TGS800 series gas sensor I-1-2 4 stainless steel casing TGS800 series gas sensor I-1-3, 4 EC type gas sensor I-1-4 and 1 PID type gas sensor I-1-5, finally flowing out from the air outlet, gas sensitive
  • the sensor array thus produces a sensitive response.
  • FIG. 6 is a schematic view of the malodorous gas buffer chamber I-9 of the present invention.
  • the gas buffer chamber is located in the malodorous electronic nose instrument I, having an inner diameter of ⁇ 40 mm and a net depth of 5-10 mm. Since the ratio of the inner diameter of the gas pipe connecting the malodorous electronic nose instrument I and the 10 monitoring points II-1 to II-10 is 4:1, the gas flow rate drops 16 times in the buffer chamber, and the built-in micro vacuum pump I-14 can A sufficient amount of malodorous gas is drawn from here.
  • 10 two-position two-way solenoid valves I-6-1 to I-6-10 that control the on-off of the odorous gas at 10 monitoring points are only turned on and off once, in any single cycle T 0 At any one time, there is one and only one is turned on, and the other nine are disconnected.
  • the two-position two-way solenoid valve I-5 for controlling the air-to-opening of the ambient air is controlled, and the two-position two-way solenoid valve I-8 for controlling the flow of malodorous gas into the annular working chamber of the gas sensor array is controlled.
  • the two-position two-way solenoid valve I-13 for controlling the clean air on and off is turned on and off 10 times, and the two-position two-way solenoid valve I-10 for controlling the flow conversion is turned on and off 20 times.
  • the two-position two-way solenoid valve I-13 is disconnected, and the two-position two-way solenoid valves I-5 and I-10 are turned on, and are cleaned by the device IV under the suction of the built-in micro vacuum pump I-14.
  • the ambient air flows through the two-position two-way solenoid valve I-5, the gas pipeline, the gas sensor array annular working chamber I-1, the two-position two-way solenoid valve I-10, and the micro vacuum pump I at a flow rate of 6,500 ml/min. -14, then discharged to the outside.
  • FIG 8 is a schematic perspective view of a malodorous electronic nose instrument.
  • the gas sensor array I-1 is located at the upper right portion of the malodorous electronic nose instrument I; we can see the display I-17, the vacuum pressure gauge I-7, and the flow meter I-12 from the front view.
  • FIG. 9 is a schematic view of the back of the malodorous electronic nose instrument 1.
  • the stinky electronic nose instrument I has an external display interface, 2 USB interfaces, a mouse interface, a keyboard interface, an Internet interface, a clean ambient air and a clean air inlet, 10 monitoring points, a malodorous gas inlet, an external vacuum pump III outlet, and an exhaust gas exhaust. Export.
  • the machine learning cascade model of the computer control and data analysis system I(c) predicts 10+1 odor pollutants based on the response vector x(t). Concentration control indicator value.
  • FIG. 10 is a schematic diagram showing the structure of a convolutional neural network CNN i1 of a gas sensor i response x i (t+1) at a time t+1 (for example, the 40th minute in the future).
  • Table 2(a) shows the time series response training set X i1 ⁇ R 10 ⁇ 9 of CNN i1 with a total of 10 samples and a dimension of 9.
  • the input and output components of Tables 2(a) and 2(b) are all scaled to the range [0, 3].
  • the present invention predicts the response x i (t+2) of the gas sensor i at time t+2 (e.g., the 80th minute in the future) and t+3 (e.g., the 120th minute in the future) using the convolutional neural networks CNN i2 and CNN i3 , respectively. ) and x i (t+3).
  • the CNN i2 and CNN i3 structures and learning parameters are the same as NN i1 .
  • Tables 3 and 4 show the time series response training sets X i2 ⁇ R 10 ⁇ 9 and X i3 ⁇ R 10 ⁇ 9 of the two convolutional neural networks.
  • the two convolutional neural networks still use the same time series response sample x i (t) as CNN i1 as shown in 2(b) to predict the response x i of the gas sensor i at times t+2 and t+3 ( t+2), x i (t+3).
  • the time spans of X i2 and X i3 are [t-19, t-2] and [t-20, t-3], respectively.
  • t is farther away, so the predictive value of CNN i2 and CNI i3 is less reliable than CNN i1 .
  • Both CNN i1 , CNN i2 and CNN i3 complete online learning and prediction within 10 seconds of the air-conditioning phase of the gas sensor array environment. Therefore, when all 16 response curves of the gas sensor array are subjected to t+1, t+2, and t+3 time response predictions, the present invention uses a 3*16 convolutional neural network; if only t+1 is predicted At the moment of response, only 16 single output convolutional neural networks are needed.
  • the invention decomposes the overall prediction problem of the multiple concentration values of the malodorous gas into a plurality of single concentration values one-to-one prediction problem according to the “divide and conquer” strategy, and uses the machine learning cascade model second level—multiple single output deep neural networks to one by one. Predict multiple single concentration values to effectively reduce the machine
  • the number of single output DNNs is equal to the number of malodorous gas concentration control indicators to be predicted, one-to-one correspondence. For example, to predict the dimensionless odor concentration OU, NH 3 , H 2 S, CS 2 , C 3 H 9 N, CH 4 S, C 2 H 6 S, C 2 H 6 S 2 , C 8 H 8 For 9 odor pollutant concentrations and TVOC concentrations such as SO 2 , 10+1 single output DNNs are required.
  • a single output DNN learns the big data of malodorous gas.
  • the input value is the data of the gas sensor array and the temperature and humidity data of the malodorous electronic nose instrument.
  • the target output is the offline measurement value of the conventional instrument such as olfactory value and color spectrum, and the resident complaint data.
  • Some samples in the big data of malodorous gas only have gas sensor array response and no offline measurement values such as olfactory value and color spectrum, and residents' complaint data, and do not participate in learning.
  • a single output DNN j has 3 hidden layers, and the hidden layer and output layer use Sigmoid to modify the activation function.
  • the gas sensor array response data and the target output component are each proportionally transformed to a range [0, 3].
  • the first and second hidden layers are feature transform (coding) layers, which are learned offline from bottom to top, and the structure and weight parameters are determined by a single hidden layer peer-to-peer neural network.
  • FIG. 11 is a schematic diagram of a peer-to-peer neural network learning process for determining the kth layer-k+1th hidden layer weight and threshold of DNN j .
  • Figure 11 (a) shows that the number of output nodes of a peer-to-peer neural network is equal to the number of input nodes, both of which are linear activation functions.
  • the weight and threshold of the hidden layer-output layer are directly equal to the input layer-hidden layer, and the target output. Directly equal to its actual input.
  • Figure 11 (b) shows that after the peer-to-peer neural network learning is completed, the number of hidden nodes of the k+1th layer of DNN j is equal to the number of hidden nodes of the peer-to-peer neural network, and the k- th layer-k+1th hidden layer weight
  • the threshold is equal to the input layer of the peer-to-peer neural network - the hidden layer.
  • the number of learning samples of DNN j be N
  • the peer-to-peer neural network learning factor ⁇ 2/N
  • the maximum iteration step number ⁇ max 10,000.
  • the third hidden layer of DNN j is a nonlinear mapping layer, which is fitted with a single output unit j to fit the jth concentration control index value of the malodorous gas.
  • Figure 12 is a schematic diagram of the machine learning cascade model predicting the concentration of various malodorous pollutants at time t+1 (e.g., the 40th minute in the future).
  • the first stage of the machine learning cascade model first learns the time series response generated by 16 gas sensors by using 16*3 sets of single output single hidden layer convolutional neural networks, and then Table 2(b)
  • the second stage of the machine learning cascade model uses 10+1 single output three hidden layer depth neural network modules to predict the above 10+1 odor pollutant control index values.
  • the actual output of DNN j is an estimate of the current concentration value y j (t) of the malodorous gas component j.

Landscapes

  • Chemical & Material Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Immunology (AREA)
  • Pathology (AREA)
  • Biochemistry (AREA)
  • Analytical Chemistry (AREA)
  • Theoretical Computer Science (AREA)
  • Food Science & Technology (AREA)
  • Medicinal Chemistry (AREA)
  • Combustion & Propulsion (AREA)
  • Molecular Biology (AREA)
  • Biomedical Technology (AREA)
  • Chemical Kinetics & Catalysis (AREA)
  • Electrochemistry (AREA)
  • Data Mining & Analysis (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Evolutionary Computation (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Artificial Intelligence (AREA)
  • Toxicology (AREA)
  • Sampling And Sample Adjustment (AREA)
  • Investigating Or Analyzing Materials By The Use Of Electric Means (AREA)
  • Investigating Or Analyzing Materials By The Use Of Fluid Adsorption Or Reactions (AREA)

Abstract

一种恶臭电子鼻仪器(Ⅰ)和恶臭气体多点集中式在线监测方法：其中，室内的恶臭电子鼻仪器(Ⅰ)与多个监测点(II-1～II-10)之间通过不锈钢管道相连接，气体采样探头(Ⅱ)可固定或移动布置，可高可低；外置真空泵(III)可在1min内将直线距离达2.5km的现场恶臭气体抽吸到恶臭电子鼻仪器(Ⅰ)内；内置微型真空泵(I-14)只在30秒顶空采样阶段持续抽吸气体缓冲室(I-9)的新的恶臭气体，使在顶空采样阶段持续抽吸气体缓冲室(I-9)的新的恶臭气体流经气敏传感器阵列(I-1)环形工作腔；模块化卷积神经网络(CNN)在线学习气敏传感器阵列(I-1)近期时间序列响应并预测即将发生的响应，模块化深度神经网络(DNN)依据恶臭大数据离线建立气敏传感器阵列响应与多种恶臭污染物浓度值之间的非线性关系；在建筑物、工业园区、垃圾与污水处理区、居民生活区等污染现场，恶臭电子鼻仪器(Ⅰ)可循环在线监测多达10个点，并用机器学习级联模型实现臭气嗅感浓度OU值和10种恶臭污染物浓度控制指标值的循环在线估计与预测。

Description

一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法 技术领域
本发明—一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法，面向环境保护与管理部门的市场监管需求，面向(a)石油化工、橡胶、香料香精、制药、涂料、酿造、造纸等工业园区；(b)垃圾转运、填埋与焚烧、污水处理等垃圾与污水处理区；(c)养殖场；(d)邻近居民生活区等恶臭污染区域的在线监测与分析需求，涉及环境保护、分析化学、计算机、人工智能、大数据、自动控制、精密测量等技术领域，主要解决电子鼻仪器自动化、集成化与小型化，恶臭污染区域多种恶臭污染物在线监测和浓度在线预测，以及恶臭污染源寻踪与控制问题。
背景技术
我国社会现阶段主要矛盾为人民日益增长的美好生活需要和不平衡不充分发展之间的矛盾，环境大气污染与美好生活需要之间当然是矛盾的。2015.01.01日起施行的《中华人民共和国环境保护法》第四十二条将恶臭气体与废气、废水等并列为环境污染物和公害。
“恶臭”特指难闻的臭味，是一切刺激人的嗅觉感官，损害人类生活环境、令人难以忍受或不愉快的气味的通称，有时称“异味”。恶臭污染物特指一切恶臭气体物质，泛指一切散发恶臭气味的物质。恶臭污染物广泛存在于石油化工、垃圾与污水处理、制药、养殖等一切有废气排放的企业及邻近居民区，分布很广，影响范围很大。我国恶臭污染现状是，排放源众多，臭气成分复杂，国家标准滞后，居民投诉频发。
近几年，恶臭作为一种扰民和危害人体健康的污染已成为比较突出的环境公害问题。随着人们生活水平的提高和环保意识的增强，一些国家或地区的恶臭污染投诉占环境投诉的比例越来越高。据不完全统计，恶臭投诉占环境投诉的比例，美国在50％以上，澳大利亚高达91.3％，日本每年达数万件。我国也不例外，据***报道，2017年全国环保举报平台共接到618,856件举报[1]。其中，涉大气污染举报最多，占56.7％；而恶臭/异味污染举报占涉气举报的30.6％，又是最多！这就是说，恶臭/异味污染举报占2017年环保举报数的17.35％。
恶臭污染评价对象是恶臭气体，评价方法分为嗅辨法和仪器分析法。GB14554-93《恶臭污染物排放标准》规定，恶臭污染物排放控制指标包括1种定性的无量纲臭气浓度和8种定量的单一成分浓度，即三甲胺(C 3H 9N)、苯乙烯(C 8H 8)、硫化氢(H 2S)、甲硫醇(CH 4S)、甲硫醚(C 2H 6S)、二甲二硫(C 2H 6S 2)、氨(NH 3)、二硫化碳(CS 2)。此外，GB/T18883-2002《室内空气质量标准》特别推荐了二氧化硫(SO 2)和总挥发性有机化合物(Total volatile organic compound,TVOC)浓度这2种定量控制指标。现阶段，恶臭污染评价指标体系主要由上述1种定性指标和10种定量指标构成。GB14554规定，测定臭气浓度用三点比较式臭袋法，测定C 3H 9N、C 8H 8、H 2S、CH 4S、C 2H 6S、C 2H 6S 2浓度用气相色谱法，测定NH 3和CS 2浓度采用分光光度法；GB/T18883规定，测定TVOC浓度采用气相色谱法；GB/T15262-94规定，测定SO 2采用分光光度法。
“臭气浓度”是指 现场采集的臭气样品在 实验室用无臭清洁空气连续稀释至 嗅辨员嗅觉阈值的稀释倍数，欧盟标准EN17325-2003用OU(odor unit)值度量。目前，臭气浓度的标准鉴别方法主要靠嗅辨员的鼻子[2]！我国、欧美、日韩等国家和地区均是如此。实施已25年的国标GB/T14675-93《环境空气-恶臭的测定-三点比较式臭袋法》规范了嗅辨员选拔、恶臭气体样品采集和样品人工稀释与嗅辨测定等三个环节。欧美、澳大利亚、新西兰等国家用动态嗅觉仪稀释臭气样品[3]。
GB/T14675和HJ905规定，恶臭气体样品先由工作人员在现场用采样瓶或无臭气袋(例如10L)采集，
表1,中国和日本对几种典型恶臭物质的嗅觉阈值测定结果(V/V，ppm)[4]
Figure PCTCN2018088913-appb-000001
然后运回到嗅辨室，再用注射器按一定比例抽吸移至无臭气袋(例如3L)并用无臭清洁空气稀释，最后由嗅辨小组成员嗅辨。三点比较式臭袋法核心之一是：臭气样品稀释一次后，一个嗅辨员需嗅闻3只3L气袋，其中1只为稀释后的有臭气袋，另2只为无臭气袋，并能从中鉴别出有臭气袋。
“选对选错全靠嗅辨员嗅闻后的主观判断”。尽管GB/T14675已施行25年，但现状是，许多恶臭物质要么没有嗅阈值，要么不同国家或组织给出的嗅阈值差别很大。2015年，天津环科院国家环境保护恶臭污染控制重点实验室从更具有统计意义的期望出发，组织30名嗅辨员(男13人，女17人)对40种恶臭物质进行了嗅觉阈值测定，表1为其中10种恶臭物质嗅辨结果与日本嗅辨值的比较[4]。根据表1，NH 3嗅觉阈值与日本相差5倍，H 2S相差近3倍，三甲胺相差28.12倍，正戊酸相差65.67倍，等等。上述结果至少说明两个问题：(1)确定臭气浓度的嗅辨过程很复杂，嗅评一次代价很大；(2)各国各单位给出的恶臭物质嗅觉阈值本身不客观，不具备重复性。
GB/T14675规定的三点比较式臭袋法尽管可体现普通人感受，但可操作性极差，做一次嗅辨测试需要大量采样和嗅辨人员，成本很高，特别不适于低浓度和有毒物质的嗅辨[2,4-5]。三点比较式臭袋法的嗅评结果好坏受①现场采样点选择；②采样装置；③实验室条件；④嗅辨员能力与状态；⑤臭气浓度与初始稀释倍数；⑥嗅辨时间与疲劳等诸多因素影响，其中的人工采样、人工稀释和人工嗅辨方法存在很多局限性。
广东省环境监测中心高级工程师肖文深有感触地说[2]：“我本人和监测中心都是恶臭的受害者，深圳、惠州交界处的垃圾填埋场臭味扰民遭到周边老百姓频繁投诉，去年(注：2016)开始每个月我们都要监测一次，每次都很头大，采样要5个人，实验室分析要8个人。但结果很难说清楚，有时现场很臭，但实验室测不出来；有时反之。现在，老百姓把我们告上法庭，政府部门对我们也不满意，真是两头受气。”国家环保部前总工程师万本太呼吁[2]，“监测恶臭气体不能再靠鼻子去闻，就像靠舌头舔鉴别有无毒害一样，这不是要命吗？要研究在线自动仪器监测方法，…。”
人的嗅觉能感觉到的致臭成分有4,000多种，其中对人体健康危害较大的有几十种，实际生活中闻到的臭味往往含有数十、数百种致臭成分[4]。例如，H 2S散发臭鸡蛋味；胺(Amines)类物质散发腐臭鱼味；氨类和醛类成分散发刺鼻味，等等[5]。
面对人们过美好生活的期望，世界各国普遍重视环境污染治理。近年来，我国环保部门尤其高度重视恶臭气体污染的在线监测与治理[1,6-9]，这是建设智慧城市的要求。根据《国家环境保护标准“十三五”发展规划》，2019年将发布GB14554和GB/T14675修订版，亦将发布《恶臭污染物环境监测技术规范》和《环境空气和废气恶臭气体在线监测技术规范》。我们注意到，行业标准HJ905-2017《恶臭污染环境监测技术规范》已于2017.12.29发布，于2018.03.01日正式实施。
由于嗅辨法和常规仪器分析法时效性差，代价高；还由于嗅辨法对人体有害，嗅辨结果不客观，嗅觉模拟—电子鼻技术与仪器因此特别引人注目[2,9]。
嗅觉是鼻腔大量嗅细胞的复杂感觉。嗅觉模拟方法利用性能重叠的多个气敏传感器组成阵列，实现气味快速检测和定性定量分析，引起人们的高度重视[10]。例如，电子鼻技术可以通过气敏传感器阵列对气味的多元感知响应来确定呈味物质类别、强度、质量等级、真假、新鲜程度，控制生产过程，调整配方与生产工艺，等等。电子鼻方法现在主要用于复杂气味定性定量分析，例如，酒[11]、茶[12]、牛奶[13]、粮油质量[14]；水果成熟度[15]；鱼肉制品新鲜程度[16]；水与环境空气监测[17-18]；疾病诊断[19-20]；细菌气味感知[21]；等等。
电子鼻技术应用前景广阔，发展趋势之一是，发展高灵敏度、高选择性的气敏器件，以实现气味的定性定量检测与分析。令人鼓舞的是，SnO 2半导体气敏器件灵敏度已达10 -9V/V(ppb)数量级[10]，对气味直接产生V级电压响应，不需二次放大，这对恶臭污染物的在线监测是很有吸引力的。电子鼻技术发展趋势之二是，以具有必要灵敏度的多个不同类型气敏元件组成阵列，着重利用数据分析方法来提高对检测对象的选择性，实现气味的识别、强度估计和关键成分预测。
电子鼻理论与应用研究相关检索结果如下：(1)文献。1990年以前仅60多篇，2000年前累计500多篇，现在累计已达6,000余篇，说明电子鼻研究近几年广泛展开。(2)专利。500余项国际发明专利和100余项国内发明专利大多是近5年公开和授权的，显示嗅觉模拟知识产权保护已受到重视。(3)技术标准。国际标准数据库HIS尚无与嗅觉模拟有关的产品技术标准。(4)应用。国内绝大多数工作以国外商品化电子鼻进行实验室研究[12,15]。法国FOX型电子鼻2007年进入我国市场，价格昂贵(100多万元RMB)，主要用于实验室离线检测，不可能用于恶臭污染过程在线检测。上述结果说明，嗅觉模拟—电子鼻理论与应用研究亟待深入。
ISI数据库查询结果表明，电子鼻方法应用于环境恶臭气体过程检测与分析的文献不多，仅130余篇，不到电子鼻文献总数的2％，且大多为室内空气、水、土气味的离线检测和实验室数据处理；尚未发现恶臭污染物现场电子鼻在线监测报道，尚无成熟的恶臭电子鼻仪器商品[17-18]。
我国的恶臭污染电子鼻监测应用工作走在世界前面。在政府环保管理部门主导下，国内一些化工园区、垃圾填埋场、污水处理厂等污染源排放单位通过招标采用了德国Airsense公司和法国alpha MOS公司的商品化电子鼻[22]。这两款产品由4个金属氧化物半导体(Metal Oxide Semiconductor,MOS)、4个电化学(Electrochemical,EC)、1个光离子(Photoionization Detector,PID)气敏元件为阵列，是专门针对中国市场开发的，实际应用过程中存在监测标准不一致、分析模型不适用、稳定性与一致性差、设备和运行费用高昂等一系列问题。国内拓扑智鑫公司的恶臭监测***用1个PID和8个EC气敏元件组成阵列，关注重点放在简单的偏最小二乘(partial least squares,PLS)算法和数据云端传输，企图依据被测样品与标准样品的比较来做判断，没有考虑恶臭气体成分复杂性和环境多变性[23]。
现有电子鼻监测***都采用一个监测点一台电子鼻即“一点一鼻”的布点方式，这显然受监控摄像头安装方式的影响。电子鼻与摄像头二者作用有相似之处，但使用条件完全不同。电子鼻直接安装在监控现场会产生一系列问题：例如，电子鼻仪器尤其是核心元件—气敏传感器长期经受风吹日晒雨淋而影响寿命；气敏传感器长时间接触恶臭气体极易产生疲劳与“中毒”效应；不同监测点同型号气敏传感器会产生一致性问题；电子鼻仪器一经安装将很难再移动；不能随风向变化而变化；布点过密不仅影响市容，而且成本高；配备“零气”会导致电子鼻仪器体积庞大和安全问题；与嗅辨和色质谱结果差别大；等等。
安装了4台alpha MOS电子鼻的上海老港垃圾填埋场和耗资1483.5万元安装了115台Airsense电子鼻的天津大港工业区仍分列2017年环保举报次数第1和第3位[24]。这一事实充分说明，现在所谓的“臭气电子鼻”只是传感器级的初级应用。
为了将电子鼻技术与仪器用于恶臭气体在线监测与分析，我们必须解决以下问题：
1,气敏传感器阵列设计问题
恶臭气体的特点是，(1)组成成分众多且复杂。除H 2S、NH 3、CS 2、SO 2等少数无机物外，大多数为有机物，即所谓“挥发性有机化合物”；(2)有些恶臭物质嗅觉阈值很低，但对臭气浓度贡献度却很大；反之亦然；(3)有些物质无毒无害，对臭气浓度贡献度很小，气敏传感器却很敏感。
气敏传感器选择要综合考虑灵敏度、选择性、响应速度、稳定性、商品化、小型化、寿命、成本等因素。MOS型气敏传感器灵敏度高，优点明显，缺点是选择性不够理想，应成为阵列组成单元首选。与MOS型相比，EC型气敏元件的优点是选择性较好，缺点是灵敏度比前者低1-2个数量级，寿命较短(一般为1-2年)，尺寸大，稳定性较差，主要用于H 2S、NH 3、CS 2、SO 2等有毒气体检测[10,22]。PID型气敏元件的特点是对正己烷～正十六烷之间的VOCs较敏感，但并不唯一，还存在尺寸大、寿命短、价格高等缺点。我们 应深入理解不同敏感元件的性能和特点，设计出小型气敏传感器阵列模块，解决其长期稳定性差、噪声消除、温湿度补偿、性能无差别更换等问题[25]。
2,标准适用性问题和与感官、常规仪器分析结果一致性问题
Airsense和alpha MOS电子鼻依据的是欧盟标准EN13725。该标准规定，用动态嗅辨仪对臭气进行任意倍数的连续稀释(最大2×10 6倍)，以0.04×10 -6(V/V)正丁醇作为标准物质，指定该标准气体对人的嗅器官作用为臭气浓度基本单位(OU E)，“任何混合气味的嗅觉阈值等同于标准气体对人的刺激量”，量值符合ISO量纲体系，具有自动测试特征。比较而言，GB/T14675属于嗅觉静态测定法，是手工测试方法，可操作性和科学性要比EN17325差[26]。
臭气组成成分及浓度是时刻变化的。在单个气敏元件选择性较差的情况下，如何将气敏传感器阵列响应转化为与臭气嗅辨浓度、色质谱等常规仪器分析相一致的结果，这既是一个涉及计算机和分析化学的理论问题，更是一个涉及恶臭污染源类型的实践问题[5]。
3,基于大数据和人工智能的臭气浓度及其关键成分浓度预测问题
人类社会处于大数据和人工智能时代，健康大数据、金融大数据、交通大数据、商业大数据、基因大数据等正在深刻地改变人们的生活和工作方式。在我国，环境大数据已提上议事日程，政府环保管理部门正在大力推动中[7-8]。
由于恶臭气味复杂性和环境多变性，小数据和常规分析方法不足以有效建立估计和预测恶臭气体多种成分的数学模型。没有恶臭电子鼻仪器对大量恶臭污染现场测试产生的气敏传感器阵列响应数据，没有嗅辨人员对大量恶臭样品的实验室嗅辨数据，没有色质谱等常规仪器对大量恶臭样品的离线检测数据，企图单纯靠气敏传感器阵列和简单的数学模型来估计臭气浓度与多种污染物成分是不可能的。德、法电子鼻正是这样做的，由此产生的监测数据的作用十分有限，甚至可以说是不可信的。
我们应以气敏传感器阵列响应数据、嗅辨数据、色质谱与分光光度等常规仪器分析数据为基础，建立恶臭气体大数据，深入研究人工智能理论与算法[27-29]，从恶臭大数据中挖掘出关键成分浓度等有用信息，以实现电子鼻仪器对上述10+1种主要恶臭污染物浓度控制指标的实时预测。
4,恶臭电子鼻仪器集成化、自动化与智能化问题
恶臭污染源众多，恶臭气体组成成分众多，环境变化多端，恶臭污染物排放形式众多。我们应摒弃“一点一鼻”的分散式监测方式，研究气敏传感器阵列优化与融合，多点集中式精密自动进样***模块化与小型化；将气敏传感器阵列模块、恶臭气体自动进样模块、驱动与控制电路模块、计算机等集成在一个测试箱内，发明和研制尺寸小、重量轻、操作简便的新型恶臭电子鼻仪器；优化仪器内部工作条件，以内部的“不变”应对外部的“万变”。理想情况是，一台恶臭电子鼻仪器能实现特定区域(例如，面积4km 2以内)多个观测点的同时在线监测，即可固定点监测，也可移动点监测，当然是以月乃至年为单位的每天24小时连续监测；提出简单有效的机器学习模型与算法实现对前述10+1种恶臭污染物浓度的24小时连续估计和预测，并利用无线WIFI技术，实时把监控数据和分析结果传输到监控中心及各种终端，实现基于Internet网的恶臭污染远程控制。
参考文献
[1]赵艳红，环保部：2017年接到环保举报60余万件，大气举报占近六成，人民网，2018.01.23,politics.people.com.cn/n1/2018/0123/c1001-29782203.html
[2]杜娟，面对恶臭污染只靠鼻子闻？专家建议尽早实现24小时监测，广州日报，2016.07.06.
[3]李伟芳，耿静，翟增秀，闫凤越，恶臭物质的嗅觉阈值与致臭机理研究概况与展望，安全与环境学报，2015,15(3):327–330.
[4]王亘，翟增秀，耿静，韩萌，鲁富蕾，40种典型恶臭物质嗅阈值测定，安全与环境学报，2015,15(6):348–351.
[5]王亘，翟增秀，韩萌，荆博宇，鲁富蕾，孟洁，恶臭污染排放源指纹谱编制及初步应用，环境科学研究，2017,30(12): 1944–1953.
[6]王金南，蒋洪强，刘年磊，关于国家环境保护“十三五”规划的战略思考，中国环境管理，2015,7(2):1–8.
[7]詹志明，尹文君，环保大数据及其在环境污染防治管理创新中的应用，环境保护，2016,06:44–48.
[8]程春明，李蔚，宋旭，生态环境大数据建设的思考，中国环境管理，2015,7(6):9–13.
[9]滕建礼，王玉红，我国环境监测仪器行业发展现状及展望，中国环保产业，2017,12:19–24.
[10]P.Boeker,On'Electronic Nose'methodology,Sensors&Actuators B:Chemical,2014,204:2–17.
[11]M.Paknahad,A.Ahmadi,J.Rousseau,H.R.Nejad,M.Hoorfar,On-chip electronic nose for wine tasting:a digital microfluidic approach,IEEE Sensors Journal,2017,17(14):4322–4329.
[12]J.C.Zhu,F.Chen,L.Y.Wang,Y.Niu,Z.B Xiao,Evaluation of the synergism among volatile compounds in Oolong tea infusion by odour threshold with sensory analysis and E-nose,Food Chemistry,2017,221:1484–1490.
[13]V.F.Pais,M.I.S.Veríssimo,J.A.B.P.Oliveira,et al.,Using acoustic wave sensors to follow milk coagulation and to separate the cheeses according to the milk origin,Sensors&Actuators B:Chemical,2015,207:1121–1128.
[14]R.Upadhyay,S.Sehwag,H.N.Mishra,Electronic nose guided determination of frying disposal time of sunflower oil using fuzzy logic analysis,Food Chemistry,2017,221:379–385.
[15]S.Jiang,J.Wang,Internal quality detection of Chinese pecans(Carya cathayensis)during storage,Postharvest Biology&Technology,2016,118:17–25.
[16]W.Wojnowski,T.Majchrzak,T.Dymerski,J.Gebicki,J.Namiesnik,Electronic noses:Powerful tools in meat quality assessment,Meat Science,2017,131:119–131.
[17]P.Schneider,N.Castell,M.Vogt,F.R.Dauge,W.A.Lahoz,A.Bartonova,Mapping urban air quality in near real-time using observations from low-cost sensors and model information,Environment International,2017,106:234–247.
[18]A.Blanco-Rodríguez,V.F.Camara,F.Campo,L.Becherán,A.R.Garcia-Ramirez,Development of an electronic nose to characterize odors emitted from different stages in a wastewater treatment plant,Water Research,2018,134:92–100.
[19]R.V.D Goor,M.V.Hooren,A.M.Dingemans,B.Kremer,K.Kross,Training and validating a portable electronic nose for lung cancer screening,Journal of Thoracic Oncology,2018(In press),doi.org/10.1016/j.jtho.2018.01.024.
[20]G.Suarez-Cuartin,J.Giner,J.L.Merino,A.Rodrigo-Troyano,O.Sibila,Identification of pseudomonas aeruginosa and airway bacterial colonization by an electronic nose in bronchiectasis,Respiratory Medicine,2018,136:111–117.
[21]E.N.Carmona,V.Sberveglieri,A.Ponzoni,et al.,Detection of food and skin pathogen microbiota by means of an electronic nose based on metal oxide chemiresistors,Sensors&Actuators B:Chemical,2017,238:1224–1230.
[22]苏爱华，周俊杰，电子鼻在垃圾填埋场的恶臭在线监测***中的应用，环境与可持续发展，2017,42(4):129–131.
[23]郭勇辉、樊喜来、霍金芳、耿利华，环境恶臭污染定量监测方法，发明专利，2013100985244,2016.03.16授权.
[24]中华环保联合会VOCs污染防治专业委员会，11月全国“12369”共接环保举报5万余件其中涉及大气、噪声污染的举报最多，www.acef-vocs.com.cn/sv_view.aspx？FId＝t8:36:8&Id＝521&TypeId＝36.
[25]J.Fonollosa,L.Fernandez,A.Gutierrez-Galvez,R.Huerta,S.Marco,Calibration transfer and drift counteraction in chemical sensor arrays using Direct Standardization,Sensors&Actuators B:Chemical,2016,236:1044–1053.
[26]王同健，田秀华，王琳，恶臭监测嗅觉测试方法的比较分析，中国环境监测，2013,29(5):169–172.
[27]S.D.Vito,E.Esposito,M.Salvato,O.Popoola,F.Formisano,R.Jones,G.D.Francia,Calibrating chemical multisensory devices for real world applications:An in-depth comparison of quantitative machine learning approaches,Sensors&Actuators B:Chemical,2018,255(2):1191–1210.
[28]A.Peter,High-dimensional function approximation with neural networks for large volumes of data,IEEE Transactions on Neural Networks and Learning Systems,2018,29(2):500–508.
[29]Y.LeCun,Y.Bengio,G.Hinton,Deep learning,Nature,2015,521:436–444.
发明内容
本发明是在现有发明专利《一种机器嗅觉装置及其嗅觉模拟测试方法》(申请号：02111046.8)、《一种基于模块化组合神经网络的机器嗅觉气味识别方法》(申请号：03141537.7)、《一种嗅觉模拟仪器与多种气味定性定量分析方法》(申请号：201010115026.2)、和《一种多通道集成嗅觉模拟仪器和生物发酵过程在线分析方法》(申请号：201310405315.X)的基础上，发明一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法，以解决恶臭污染区域多个监测点的长期在线监测和多种恶臭气体浓度控制指标的在线预测问题。
为了实现上述目的，本发明的恶臭气体多点集中式在线监测与分析***包括恶臭电子鼻仪器I、气体采样探头II、外置真空泵III、环境空气净化装置IV、洁净空气瓶V、气体管道、电子温湿度计VI、中央控制室VII以及多个固定/移动终端VIII，实现恶臭污染区域10个监测点的长期在线监测和多种恶臭污染物浓度控制指标值的在线估计与预测。
恶臭电子鼻仪器I包括气敏传感器阵列及其恒温工作室I(a)、多点集中式恶臭气体自动进样***I(b)、计算机控制与数据分析***I(c)三大组成部分。气敏传感器阵列恒温工作室I(a)由气敏传感器阵列及其环形工作腔I-1，隔热层I-2，电阻加热丝I-3，风扇I-4组成。气敏传感器阵列I-1由16个气敏元件构成，均布于中径φ140mm、断面尺寸21mm×17mm的密封腔内，形成气敏传感器阵列环形工作腔，处于55±0.1℃的恒温室内，位于恶臭电子鼻仪器I右上方。多点集中式恶臭气体自动进样***I(b)包括内置微型真空泵I-14、14个二位二通电磁阀(I-5,I-6-1～I-6-10,I-8,I-10,I-13)、节流阀I-11、流量计I-12、真空压力表I-7、气体缓冲室I-9，位于恶臭电子鼻仪器I右下方。计算机控制与数据分析***I(c)包括计算机主板I-15、数据采集卡I-16、显示器I-17、驱动与控制电路模块I-18、精密线性与开关电源模块I-19、硬盘、网卡、显卡，位于恶臭电子鼻仪器I左侧。
多点集中式恶臭气体自动进样***I(b)对单个监测点恶臭气体采样周期为T 0＝180-300秒钟，默认值T 0＝240秒钟，气敏传感器阵列I-1因此对该监测点产生一个16维响应向量。计算机控制与数据分析***I(c)依据这一响应向量，用机器学习级联模型对该监测点的臭气嗅感浓度、GB14554指定的氨NH 3、硫化氢H 2S、二硫化碳CS 2、三甲胺C 3H 9N、甲硫醇CH 4S、甲硫醚C 2H 6S、二甲二硫醚C 2H 6S 2、苯乙烯C 8H 8等8种化合物，GB/T18883指定的二氧化硫SO 2与总挥发性有机化合物共10+1项恶臭污染物浓度控制指标值进行实时分析和预测，并将监测数据和预测结果通过无线Internet网远程传送到中央控制室VII和指定的固定/移动终端VIII。
恶臭电子鼻仪器I每单周期T 0得到一个16维的响应向量，储存在计算机硬盘的一个数据文件里。用10个二位二通电磁阀I-6-1～I-6-10依次控制4km 2区域内10个监测点恶臭气体的通与断，以T＝10T 0的恶臭气体循环采样周期实现10个监测点恶臭气体的循环在线监测，并将监测数据依次储存在10个数据文件里。这些数据是恶臭电子鼻仪器I预测恶臭污染物多种浓度的数值基础，据此实现对10+1项恶臭污染物浓度控制指标值的循环在线预测。
恶臭气体单采样周期T 0包括：气敏传感器阵列I-1初步恢复(95-215秒)、洁净空气精确标定(30秒)、平衡(5秒)、恶臭气体顶空采样(30秒)、净化环境空气冲洗(20秒)共5个阶段。在单周期T 0内，在计算机控制下，对应监测点的二位二通电磁阀I-6-k(＝1,2,…,10)导通，其余9个监测点的二位二通电磁阀断开，内置微型真空泵I-14以流量1,000ml/min抽吸气体缓冲室I-8内的恶臭气体，使之流经气敏传感器阵列环形工作腔，掠过气敏传感器敏感膜表面，气敏传感器阵列I-1因此产生敏感响应，持续30秒。自平衡状态开始之刻起，计算机控制与数据分析***I(c)持续记录敏感响应数据，包括平衡(5秒)、恶臭气体顶空采样(30秒)、净化环境空气冲洗(前10秒)这3个阶段共45秒的气敏传感器阵列I-1响应数据，并临时存储在一个文本文件里。单周期T 0其它时间的响应数据不记录。
在时长45秒的响应数据内，单个气敏传感器响应曲线的稳态最大值和最小值之差值被提取为响应分量，气敏传感器阵列I-1因此产生一个16维的响应向量。在数据记录结束后的10秒内，即净化环境空气冲洗阶段的后10秒，计算机控制与数据分析***I(c)依据这一响应向量预测10+1项恶臭污染物浓度控制 指标值。
气敏传感器阵列I-1由11个金属氧化物半导体型、4个电化学型和1个光离子型气敏元件组成。其中，11个金属氧化物型气敏元件用于检测多种有机/无机化合物；4个电化学型气敏元件用于检测NH 3、H 2S、CS 2、SO 2等4种无机化合物；1个光离子型气敏元件用于检测总挥发性有机化合物。
恶臭气体多点集中式在线监测与分析***即实现大至数平方千米的1个特定区域，也可实现小至1个生产车间或1栋建筑物，乃至1个点的在线监测与分析。对特定区域、生产车间和建筑物，可设置10个监测点，最大监测区域为2km*2km＝4km 2，其中，9个固定监测点，1个移动监测点。恶臭电子鼻仪器I位于室内，通过内径φ10mm不锈钢管道与各监测点相连接。气体采样探头采用水龙头形式，与商用除尘去湿净化部件连接，随用随换。改变监测点位置只需重新铺设不锈钢管道，安装和移动气体采样探头到指定位置即可，楼上楼下，可高可低，就像铺设水管或电缆线一样简便。
8个或更多监测点绕监测区域边界设置，恶臭电子鼻仪器I和10个监测点位置设置以不锈钢管道最短为目标。对路径可达的化工园区、居民小区等区域，恶臭电子鼻仪器I布置在区域中心的某个室内；对无路可达的垃圾填埋场、污水处理厂等区域，恶臭电子鼻仪器I布置在区域边界的某个室内。
外置真空泵III抽气速率250-280L/min，极限真空度100-120mbar，长期连续工作，可通过内径φ10mm不锈钢管道在1分钟之内将直线距离达2.5km的一个监测点的恶臭气体抽吸到恶臭电子鼻仪器I内。在单周期T 0内，除恶臭气体顶空采样(30秒)这一阶段外，其余阶段被抽吸到恶臭电子鼻仪器I内的恶臭气体并不流经气敏传感器阵列环形工作腔I-1，而是被直接排出到室外。
恶臭电子鼻仪器I内部设置有一个尺寸为φ40mm*5mm的气体缓冲室I-8，恶臭气体在此处的流速较内径φ10mm不锈钢管道骤降16倍。只有在恶臭气体顶空采样(30秒)这一阶段，内置微型真空泵I-14才将气体缓冲室I-8内的恶臭气体抽吸到气敏传感器阵列环形工作腔，气敏传感器阵列I-1因此产生敏感响应。内置微型真空泵I-14抽吸到的都是新鲜恶臭气体。
在恶臭气体顶空采样前，1,000ml/min洁净空气精确标定环节(30)使得气敏传感器阵列I-1对恶臭气体的多次感知在同一基线上进行。12～15Mpa压缩气体钢瓶V标准容积为40L，转换到常温常压为6m 3。当单周期T 0＝3、4和5分钟时，这样1瓶40L压缩洁净空气分别可用25、33和41天。恶臭电子鼻仪器I所处室外的环境空气先用商品化空气净化器净化，然后被用来冲洗气敏传感器阵列I-1，使之初步恢复到基准状态，以降低运行费用。
恶臭气体大数据集包括：(1)气敏传感器阵列I-1对化工园区(包括香精香料厂)、制药厂、垃圾填埋场、污水处理厂、养殖场、邻近居民区等大量恶臭污染物现场的在线检测数据。(2)气敏传感器阵列I-1对大量恶臭标准样品顶空挥发气的实验室离线检测数据，其中包括GB/T14675指定的β-苯乙醇、异戊酸、甲基环戊酮、γ-十一烷酸内酯、β-甲基吲哚这5种标准臭液；GB14554指定的C 3H 9N、C 8H 8、H 2S、CH 4S、C 2H 6S、C 2H 6S 2、NH 3、CS 2与GB/T18883指定的SO 2共9种单一成分恶臭污染物配制的不同浓度标准恶臭样品，还包括不同浓度多种单一化合物配制的混合成分标准恶臭样品。(3)GB/T14675和HJ 905-2017规定的真空瓶或臭气袋在大量恶臭污染物现场采样，并立即运回嗅辨室而得到的无量纲臭气浓度离线嗅辨数据。(4)GB/T18883规定的Tenax GC/TA吸附管恶臭污染物现场采样，气相色谱仪实验室离线检测得到的总挥发性有机化合物数据和分光光度仪实验室离线检测得到的SO 2数据。(5)GB/T14676-14680规定的恶臭污染物现场采样，8种恶臭成分的气相色谱仪、质谱仪和分光光度仪实验室离线检测数据。(6)恶臭污染源邻近区域居民投诉数据。
恶臭电子鼻仪器I用机器学习级联模型预测未来t+1、t+2和t+3时刻臭气嗅感浓度和多种恶臭污染物浓度控制指标值。机器学习级联模型第一级—卷积神经网络(Convolutional neural network,CNN)层负责预测t+1、t+2和t+3时刻气敏传感器阵列I-1对一个监测点恶臭气体的响应，依据的是当前时刻t和近期已发生的气敏传感器阵列I-1时间序列响应。机器学习级联模型第二级—深度神经网络(Deep neural network,DNN)层进一步预测t+1、t+2和t+3时刻臭气嗅感浓度和多种恶臭污染物浓度控制指标值，依据的是长期 积累的恶臭气体大数据和级联模型第一级—卷积神经网络层的预测值。
依据“分而治之”策略，机器学习级联模型第一级用16*3组单输出单隐层卷积神经网络一一预测t+1、t+2和t+3时刻各个气敏传感器的响应。对T 0＝40分钟而言，相当于从当前时刻t算起，预测未来第40、80和120分钟时刻的响应。
以单周期T 0＝40分钟，3个单输出单隐层卷积神经网络模块分别预测t+1、t+2和t+3时刻气敏传感器i的响应为例：
(a)单输出单隐层卷积神经网络CNN i1预测t+1时刻气敏传感器i的响应：
设卷积神经网络CNN i1学习气敏传感器i在t时刻之前已发生的18个时刻时间序列响应，时延长度Δt＝9，则输入节点数m i＝9，取隐节点数h i＝5，输出节点数n i＝1。卷积神经网络CNN i1在线学习经预处理的气敏传感器i时间序列响应数据集X i1为：
Figure PCTCN2018088913-appb-000002
目标输出为：
d i1＝(x i(t) x i(t-1) x i(t-2) x i(t-3) x i(t-4) x i(t-5) x i(t-6) x i(t-7) x i(t-8) x i(t-9)) T∈R 10，
这种方式相当于卷积神经网络CNN i1学习气敏传感器i最近12小时已发生的1个18维时间序列响应，产生10个9维时间序列响应，即样本数为N i1＝10。卷积神经网络CNN i1的隐层和输出层活化函数为Sigmoid修正函数
Figure PCTCN2018088913-appb-000003
采用误差反传算法学习，学习因子为η i＝5/N i1＝0.2。数据集X i1和目标输出d i1均成比例变换到范围[0,3]。卷积神经网络CNN i1在10秒钟内在线学习结束后，依据最近时间段的一个9维时间序列响应：
x i1＝(x i(t-8) x i(t-7) x i(t-6) x i(t-5) x i(t-4) x i(t-3) x i(t-2) x i(t-1) x i(t)) T∈R 9
预测t+1时刻气敏传感器i的响应x i(t+1)。当T 0＝40分钟时，相当于预测未来第40分钟气敏传感器i的响应。
(b)单输出单隐层卷积神经网络CNN i2与CNN i3预测t+2和t+3时刻气敏传感器i的响应
卷积神经网络CNN i2和CNN i3结构仍为：m i＝9，h i＝5，n i＝1。在线学习经预处理的数据集X i2和X i3分别为：
Figure PCTCN2018088913-appb-000004
和
Figure PCTCN2018088913-appb-000005
即X i2和X i3同样有10个9维时间序列响应，样本数均为N i1＝10。卷积神经网络CNN i2与CNN i3在学习阶段的目标输出和预测时依据的时间序列响应与CNN i1相同。当T 0＝40分钟时，相当于学习气敏传感器i在40分钟和80分钟之前的12小时已发生的响应，预测t+2和t+3时刻气敏传感器i的响应x i(t+2) 和x i(t+3)，分别相当于预测气敏传感器i未来第80分钟和120分钟的响应。
依据“分而治之”策略，NH 3、H 2S、CS 2、C 3H 9N、CH 4S、C 2H 6S、C 2H 6S 2、C 8H 8、SO 2、总挥发性有机化合物和臭气嗅感浓度共10+1项恶臭污染物浓度控制指标值整体预测问题被分解为11个单浓度值一一预测问题，机器学习级联模型第二级用10+1个单输出三隐层深度神经网络模块分别预测这10+1项恶臭污染物控制指标值。单输出深度神经网络训练集为恶臭电子鼻仪器I的气敏传感器阵列I-1对标准臭液/气样品和大量污染现场在线检测得到的恶臭气体大数据，目标输出为臭气嗅辨值和色质谱与分光光度常规仪器离线测量值，以及居民投诉数据。
单个单输出三隐层深度神经网络DNN j采用自下而上的逐层离线学习方式。第一和第二隐层学习时采用单隐层对等神经网络结构，即单隐层对等神经网络的隐层—输出层权值直接等于其输入层—隐层权值，目标输出直接等于其输入，输入分量和输出分量依据特征分量大小成比例变换到范围[0,3]。单隐层对等神经网络的隐层活化函数为Sigmoid修正函数
Figure PCTCN2018088913-appb-000006
采用误差反传算法学习，学习因子为η j＝1/N j，学习结束后丢弃隐层—输出层；N j为恶臭气体大数据样本数。
假设对t+1时刻恶臭气体第j个浓度值y j(t+1)进行预测，第j个单输出深度神经网络DNN j依据的是16个卷积神经网络对t+1时刻气敏传感器阵列I-1的预测响应{x 1(t+1),x 2(t+1),…,x 16(t+1)}，预测y j(t+2)和y j(t+3)分别依据的是16个卷积神经网络对t+2和t+3时刻的预测响应(x 1(t+2),x 2(t+2),…,x 16(t+2)) T与(x 1(t+3),x 2(t+3),…,x 16(t+3)) T。
若实际输入是气敏传感器阵列当前响应向量(x 1(t),x 2(t),…,x 16(t)) T，必要时可再加上t时刻温湿度值，则深度神经网络DNN j的实际输出是对恶臭气体成分j当前浓度值y j(t)的估计。
恶臭电子鼻仪器I对恶臭污染区域多个监测点长期在线监测和多种恶臭污染物浓度控制指标值的在线预测，包括以下步骤：
(1)开机：仪器预热30分钟。单击屏幕菜单的“空气净化器开”选项，空气净化器IV开始对恶臭电子鼻仪器I所处的室内空气净化，长期持续工作直至操作人员单击“空气净化器关”选项为止。
在内置微型真空泵I-14的抽吸作用下，净化环境空气以6,500ml/min的流量依次流经二位二通电磁阀I-5、气敏传感器阵列环形工作腔I-1、二位二通电磁阀I-10，然后被排出到室外。气敏传感器阵列环形工作腔I-1内的温度从室温达到恒定的55±0.1℃。
单击屏幕菜单的“外置真空泵开”选项。外置真空泵III以250-280L/min的抽气速率和100-120mbar的极限真空度，通过内径φ10mm不锈钢管道在1分钟内将直线距离最大达2.5km的某个监测点恶臭气体抽吸到恶臭电子鼻仪器I内，依次流过对应的二位二通电磁阀I-6-k(＝1,2,…,10)、真空压力表I-7和气体缓冲室I-8，然后直接排出到室外。外置真空泵III持续抽吸恶臭气体，直到操作人员单击屏幕菜单的“外置真空泵关”选项为止。
修改屏幕菜单恶臭气体“单采样周期T 0”设置，默认值T 0＝40分钟；10个监测点恶臭气体循环采样周期为T＝10T 0。
(2)恶臭气体循环采样周期开始：点击屏幕菜单的“开始检测”按钮，恶臭电子鼻仪器I依次对10个监测点进行循环监测，计算机控制与数据分析***I(c)在指定文件夹自动生成10个文本文件，以存储气敏传感器阵列I-1对10个监测点恶臭气体的响应数据。
(3)监测点k(＝1,2,…,10)恶臭气体单采样周期开始。以T 0＝4分钟为例：
(3.1)气敏传感器阵列初步恢复：单周期T 0第0-155秒，在内置微型真空泵I-14的抽吸作用下，净化环境空气以6,500ml/min的流量依次流经二位二通电磁阀I-5、气敏传感器阵列环形工作腔I-1、二位二通电磁阀I-10，然后被排出到室外。在6,500ml/min净化环境空气的作用下，气敏传感器阵列环型工作腔I-1内积聚的热量被带走，粘附在气敏传感器敏感膜表面和管道内壁的恶臭气体分子被初步冲走，气敏传感器阵列I-1初步恢复到基准状态，历时155秒。
10个二位二通电磁阀I-6-1～I-6-10只有I-6-k导通，其余9个断开，外置真空泵III将监测点k(＝1, 2,…,10)的恶臭气体抽吸到恶臭电子鼻仪器I内。
(3.2)洁净空气精确标定：在单周期T 0第156-185秒，二位二通电磁阀I-13导通，二位二通电磁阀I-5、I-8和I-10断开，二位二通电磁阀I-6-1～I-6-10保持步骤(3.1)的状态。在内置微型真空泵I-14的抽吸作用下，洁净空气以1,000ml/min的流量依次流经二位二通电磁阀I-13、气体管道、气敏传感器阵列环形工作腔I-1、节流阀I-11、流量计I-12、微型真空泵I-14，然后被排出到室外。洁净空气使气敏传感器阵列I-1精确恢复到基准状态；历时30秒。外置真空泵III持续抽吸。
(3.3)平衡：在单周期T 0第186-190秒，二位二通电磁阀I-5、I-8、I-10、I-13断开，二位二通电磁阀I-6-1～I-6-10保持步骤(3.1)的状态；气敏传感器阵列环形工作腔I-1内无气体流动。自单周期T 0第186秒即平衡状态开始之刻起，计算机控制与数据分析***I(c)开始记录气敏传感器阵列I-1实时响应数据，并存储在指定的临时文本文件“temp.txt”里；历时5秒。外置真空泵III持续抽吸。
(3.4)监测点k恶臭气体顶空采样：在单周期T 0第190-220秒，二位二通电磁阀I-8导通，3个二位二通电磁阀I-5、I-13和I-10断开，二位二通电磁阀I-6-1～I-6-10保持步骤(3.1)的状态。在内置微型真空泵I-14抽吸作用下，气体缓冲室I-8内的恶臭气体以流量1,000ml/min依次流过气敏传感器阵列环形工作腔I-1、节流阀I-11、流量计I-12、内置微型真空泵I-14，最后排出到室外。气敏传感器阵列I-1因此产生的敏感响应继续记录在临时文件“temp”里，历时30秒。外置真空泵III持续抽吸。
(3.5)气敏传感器阵列冲洗：在单周期T 0第221-230秒，二位二通电磁阀I-5，二位二通电磁阀I-8、I-10和I-13断开，在微型真空泵I-14抽吸作用下，流量6,500ml/min的净化环境空气以依次流经二位二通电磁阀I-5、气敏传感器阵列环形工作腔I-1、二位二通电磁阀I-10，然后被排出到室外。与此同时，二位二通电磁阀I-6-k+1导通，10个二位二通电磁阀I-6-1～I-6-10的其余9个断开，包括二位二通电磁阀k断开，外置真空泵III转而抽吸监测点k+1的恶臭气体。由于净化环境空气的作用，气敏传感器阵列环型工作腔内积聚的热量被带走，粘附在气敏传感器敏感膜表面和管道内壁的恶臭气体分子被初步冲走，气敏传感器阵列I-1逐步恢复到基准状态；历时20秒。其中：
(a)在单周期T 0第221-230秒，气敏传感器阵列响应数据继续记录在临时文件“temp”里，历时10秒；至第230秒末，计算机控制与数据分析***I(c)停止记录气敏传感器阵列响应数据。
(b)在单周期T 0第231-240秒，计算机控制与数据分析***I(c)与此进行以下三项操作：
(b1)特征提取：自第231秒之刻起，并从时长45秒的临时文件“temp”里提取各个气敏传感器的最大和最小稳态响应值，以最大响应值与最小响应值之差作为各个气敏传感器当前时刻t对监测点k恶臭气体的响应特征分量x i(t)(i＝1,2,…,16)，并记录在对应的数据文件里。
(b2)气敏传感器阵列响应预测：机器学习级联模型第一级—16*3个卷积神经网络依据当前时刻t以前[t-18,t]、[t-19,t-1]和[t-20,t-2]时间段内已发生的气敏传感器阵列时间序列响应向量，实现在线自学习，并据此预测未来T 0、2T 0和3T 0时刻气敏传感器阵列I-1的响应。
(b3)恶臭气体浓度控制指标值预测：机器学习级联模型第二级—10+1个深度神经网络依据级联模型第一级的16*3个卷积神经网络预测的气敏传感器阵列响应值，进一步预测监测点k的10+1项恶臭污染物浓度控制指标值，通过显示器显示出来，并将监测和预测结果通过Internet网络传送到中央控制室VII和多个固定/移动终端VIII。
(3.6)监测点k恶臭气体单采样周期结束：k←k+1，回到步骤(3.1)，监测点k+1恶臭气体单采样周期开始。
重复步骤(3.1)～(3.6)，恶臭电子鼻仪器I实现对10个监测点恶臭气体的循环在线监测、识别和10+1项恶臭污染物控制指标值的预测。
附图说明
图1是本发明—一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法—恶臭电子鼻仪器研制、机器学习级联模型与算法和恶臭污染物在线检测与预测三者之间的关系框图。
图2是本发明—一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法—恶臭电子鼻仪器和恶臭污染区域多点集中式监测与分析***工作原理示意图。
图3是本发明—一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法—恶臭电子鼻仪器工作原理示意图(顶空采样状态)。
图4是本发明—一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法—恶臭电子鼻仪器和多个监测点相互位置示意图。(a)有路可达区域；(b)无路可达区域。
图5是本发明—一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法—气敏传感器阵列布置及其环形工作腔示意图。(a)气敏传感器阵列组成单元；(b)环形工作腔盖；(c)环形工作腔断面图。
图6是本发明—一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法—恶臭气体缓冲室示意图。
图7是本发明—一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法—恶臭气体单采样周期为T 0(＝240秒)，循环采样周期为T＝10T 0时，14个二位二通电磁阀I-6-k(k＝1,2,…,10)、I-5、I-13、I-8和I-10通断变化情况示意图(单位：秒)。
图8是本发明—一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法—恶臭电子鼻仪器外观立体示意图。
图9是本发明—一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法—恶臭电子鼻仪器背面示意图。
图10是本发明—一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法—卷积神经网络CNN i1预测t+1时刻(例如未来第40分钟)气敏传感器i响应x i(t+1)示意图。
图11是本发明—一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法—深度神经网络DNN j第k层学习过程示意图。(a),对等神经网络结构与参数；(b),对等神经网络学习结束后的保留结构。
图12是本发明—一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法—机器学习级联模型预测t+1时刻(例如未来第40分钟)多种恶臭污染物浓度示意图。(a),第一级—模块化卷积神经网络层；(b),第二级—模块化深度神经网络层。
具体实施方式
下面结合附图对本发明作进一步的详细描述。
图1是本发明—一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法—恶臭电子鼻仪器研制、机器学习模型与算法和恶臭污染物在线检测与预测三者之间的关系框图。
本发明首先从化学、物理角度对恶臭污染物和气敏传感器的特点进行分析。恶臭气体组成成分众多且复杂，往往含有数十乃至数百种致臭成分，既有有机成分也有无机成分；有些恶臭成分对臭气浓度贡献大但真实浓度可能很低，气敏传感器响应因此很小；有些恶臭成分对臭气浓度贡献很小但真实浓度可能很高，气敏传感器因此很大；反之亦然。综合考虑灵敏度、选择性、响应速度、稳定性、商品化、小型化、寿命、成本等因素，本发明选择由MOS型、EC型和PID型气敏元件组成小型气敏传感器阵列模块。为避免监测区域室外的风吹日晒雨淋，本发明提出关键部件位于室内的恶臭气体多点集中式监测方式并据此研制恶臭电子鼻仪器。考虑到恶臭污染物成分复杂且监测现场环境变化多端等因素，本发明进而提出建立恶臭大数 据，并提出新的机器学习级联模型来实现对多种恶臭污染物的在线监测与预测。
根据图1，恶臭气体大数据包括：(1)恶臭电子鼻仪器I的气敏传感器阵列I-1对大量恶臭标准样品顶空挥发气的实验室离线检测数据，其中包括，β-苯乙醇、异戊酸、甲基环戊酮、γ-十一烷酸内酯、β-甲基吲哚等5种标准臭液和C 3H 9N、C 8H 8、H 2S、CH 4S、C 2H 6S、C 2H 6S 2、NH 3、CS 2、SO 2等9种恶臭化合物配制的不同浓度单一成分标准恶臭样品，还包括由多种不同浓度化合物配制的混合成分标准恶臭样品；(2)气敏传感器阵列I-1对大量恶臭污染物现场的在线检测数据；(3)大量恶臭污染物的臭气浓度实验室离线嗅辨数据；(4)大量恶臭污染物的气相色谱仪、质谱仪和分光光度仪实验室离线检测得到的TVOC和上述9种恶臭成分检测数据；(5)恶臭污染源邻近区域居民投诉数据。
图2是恶臭电子鼻仪器和恶臭污染区域多点集中式监测与分析***工作原理示意图。恶臭污染区域多点集中式监测与分析***包括恶臭电子鼻仪器I、10个室外监测点的气体采样探头II-1～II-10、外置真空泵III、环境空气净化装置IV、洁净空气V、电子温湿度计VI、中央控制室VII及其多个固定/移动终端VIII，实现恶臭污染区域10个监测点的长期在线监测和恶臭气体多种浓度控制指标值的在线预测。此时的气路和电磁阀的位置为第一个监测点II-1恶臭气体被抽吸到恶臭电子鼻仪器I，气敏传感器阵列I-1因此产生敏感响应的工作状态。
图3是恶臭电子鼻仪器I工作原理示意图。其组成单元包括：
(a)气敏传感器阵列恒温工作室I(a)：气敏传感器阵列及其环形工作腔I-1，隔热层I-2，电阻加热丝I-3，风扇I-4，位于恶臭电子鼻仪器右上方。
(b)多点集中式恶臭气体自动进样***I(b)：控制净化环境空气通断的二位二通电磁阀I-5，控制10个监测点恶臭气体通断的10个二位二通电磁阀I-6-1～I-6-10，显示外置真空泵III工作状态的真空压力表I-7，控制恶臭气体流入气敏传感器阵列环形工作腔I-1的二位二通电磁阀I-8，气体缓冲室I-9，控制气敏传感器阵列环形工作腔I-1内恶臭气体和洁净空气6,500ml/min与1,000ml/min流量转换的二位二通电磁阀I-10，节流阀I-11，流量表I-12，控制洁净空气通断的二位二通电磁阀I-13，内置微型真空泵I-14，位于恶臭电子鼻仪器右下方。
(c)计算机控制与数据分析***I(c)：计算机主板I-15，数据采集卡即A/D板I-16，显示器I-17，驱动与控制电路模块I-18，多路直流电源I-19，位于恶臭电子鼻仪器左侧。
图4是恶臭电子鼻仪器I和10个监测点II-1～II-10相互位置示意图。对图4(a)所示的路径可达区域，例如化工园区和居民小区，恶臭电子鼻仪器I应布置在监测区域中央的某个室内；对图4(b)所示的无路可达区域，恶臭电子鼻仪器I应布置在监测区域边界的某个室内。恶臭电子鼻仪器I位置确定以其与各监测点的直线距离最短为原则。外置真空泵III、环境空气净化装置IV、洁净空气V、电子温湿度计VI布置在恶臭电子鼻仪器I附近。
假设最大监测区域为2km*2km＝4km 2，连接恶臭电子鼻仪器I与各个监测点的不锈钢管道内径为d＝φ10mm。考虑最不利情况：气体最长管道出现在图4(b)所示的无路可达区域，最大直线长度为
Figure PCTCN2018088913-appb-000007
Figure PCTCN2018088913-appb-000008
再假设外置真空泵流量Q＝280L/min，最大真空压力为p＝-100mbar，则恶臭气体流速υ＝4Q/(πd 2)＝59.42m/s＝3.57km/min，恶臭气体从最远采样点II-3或II-9被抽吸到恶臭电子鼻仪器I只需t max＝l max/υ＝37.65s。注意到l max＝2.24km且d＝φ10mm时，管道最大容积约为U max＝l max*πd 2/4＝176L，小于外置真空泵1min可抽吸的最大气体容积Q＝250-280L，再考虑泄漏因素，意味着监测点气体被抽吸到恶臭电子鼻仪器I的时间仅为1min左右。在如此短的时间内，恶臭气体来不及发生变质和吸附效应。
绝大多数需要监测的工业园区、垃圾与污水处理区、养殖场、邻近居民生活区等污染区域面积在1km 2之内。假设最大监测区域为1km*1km＝1km 2，连接恶臭电子鼻仪器I与各个监测点的气体管道绕边界布置。仍考虑最不利情况：气体最长管道为l max＝0.5+1+0.5＝2km。外置真空泵III可在1min内将监测点气体抽吸到恶臭电子鼻仪器I内。本发明的恶臭电子鼻仪器和恶臭污染区域多点集中式监测与分析***特别适合应用于生产车间、污水池、养殖场等场合，可实现大至数km2的1个特定区域，小至1个生产车间或1栋建 筑物，乃至1个点的在线监测与分析。
图5是本发明的气敏传感器阵列I-1布置及其环形工作腔示意图。图5(a)表明了一个具体实例：气敏传感器阵列共由3种类型16个型号的气敏元件组成，包括MOS型的11个(4个TGS2000系列I-1-1，3个塑料外壳TGS800系列I-1-2，4个不锈钢外壳TGS800系列I-1-3)，4个EC型I-1-4和1个PID型I-1-5。MOS型气敏传感器灵敏度高，寿命长，对有机和无机成分均敏感；EC型气敏元件选择性较好，主要用于检测无机气体；PID型气敏传感器对正己烷～正十六烷之间的VOCs较敏感。机器学习级联模型依据11个MOS型与4个EC型气敏传感器的响应共同确定H 2S、NH 3、SO 2、CS 2等无机成分浓度；依据11个MOS型与1个PID型气敏传感器的响应共同确定恶臭气体TVOC浓度和C 3H 9N、C 8H 8、CH 4S、C 2H 6S、C 2H 6S 2等有机成分浓度；依据所有16个气敏元件的响应共同确定嗅感浓度OU值。
根据图5(a)、5(b)和5(c)，气敏传感器阵列环形工作腔I-1由不锈钢底座I-1-6、密封圈I-1-7、不锈钢盖I-1-8、隔板I-1-9以及气敏传感器插座I-1-10、密封材料I-1-11、螺钉I-1-12组成，形成密封的环形工作腔。顶空采样时，恶臭气体从进气孔被吸入，然后绕环型工作腔依次掠过4个TGS2000系列气敏传感器I-1-1，3个塑料外壳TGS800系列气敏传感器I-1-2，4个不锈钢外壳TGS800系列气敏传感器I-1-3，4个EC型气敏传感器I-1-4和1个PID型气敏传感器I-1-5，最后从出气孔流出，气敏传感器阵列因此产生敏感响应。
图6是本发明的恶臭气体缓冲室I-9示意图。该气体缓冲室位于恶臭电子鼻仪器I内，内径φ40mm，净深度可为5-10mm。由于与连接恶臭电子鼻仪器I和10个监测点II-1～II-10的气体管道内径之比为4:1，气体流速在此缓冲室内骤降16倍，内置微型真空泵I-14因此能从这里抽吸到足够量的恶臭气体。
图7是恶臭气体单采样周期为T 0，循环采样周期为T＝10T 0时，多点集中式恶臭气体自动进样***的14个二位二通电磁阀通断变化情况及其相互关系。在循环采样周期T内，控制10个监测点恶臭气体通断的10个二位二通电磁阀I-6-1～I-6-10只通断1次，在任一单周期T 0内的任一时刻，有且仅有一个导通，其余9个均断开。
与此相比较，在循环采样周期T内，控制净化环境空气通断的二位二通电磁阀I-5，控制恶臭气体流入气敏传感器阵列环形工作腔的二位二通电磁阀I-8和控制洁净空气通断的二位二通电磁阀I-13均通断10次，控制流量转换的二位二通电磁阀I-10通断20次。
请参见图7，以循环采样周期T＝10T 0内的第一个单采样周期T 0＝240秒为例，有下列几种情况：
(a)整个单采样周期T 0＝240秒。二位二通电磁阀I-6-1一直导通，其余9个二位二通电磁阀I-6-2～I-6-10断开，在外置真空泵III的抽吸作用下，恶臭气体以250-280L/min的流量依次流过第一个监测点气体采样探头II-1、气体管道、二位二通电磁阀I-6-1、气体缓冲室I-9、外置真空泵III，最后被排除到室外。
(b)单采样周期T 0第0-175秒。尽管外置真空泵III将第一个监测点II-1的恶臭气体抽吸到恶臭电子鼻仪器I内，但由于二位二通电磁阀I-8断开，此时的恶臭气体并不流经气敏传感器阵列环形工作腔I-1，而是由外置真空泵III直接排出到室外。这175秒时间段又可分为2小段：(b1)气敏传感器阵列初步恢复的155秒；(b2)气敏传感器阵列冲洗的20秒。在这2小段，二位二通电磁阀I-13断开，二位二通电磁阀I-5和I-10导通，在内置微型真空泵I-14的抽吸作用下，经装置IV净化的环境空气以6,500ml/min的流量依次流经二位二通电磁阀I-5、气体管道、气敏传感器阵列环形工作腔I-1、二位二通电磁阀I-10、微型真空泵I-14，然后被排出到室外。
(c)单采样周期T 0＝240秒第176-205秒。3个二位二通电磁阀I-5、I-8和I-10断开，二位二通电磁阀I-13导通，在洁净空气瓶V的气体压力作用下，洁净空气以1,000ml/min的流量依次流经二位二通电磁阀I-13、气体管道、气敏传感器阵列环形工作腔I-1、流量计I-12、微型真空泵I-14，然后被排出到室外。
(d)单采样周期T 0＝240秒第211-240秒。3个二位二通电磁阀I-5、I-13和I-10断开，二位二通电磁阀I-8导通，在内置微型真空泵I-14的抽吸作用下，恶臭气体以1,000ml/min的流量依次流经二位二通电磁阀I-13、气体管道、气敏传感器阵列环形工作腔I-1、流量计I-12、微型真空泵I-14，然后被排出到室外。
图8是恶臭电子鼻仪器I立体示意图。气敏传感器阵列I-1位于恶臭电子鼻仪器I右上部；我们由正面外观图可以看到显示器I-17、真空压力表I-7和流量表I-12。
图9是恶臭电子鼻仪器I背面示意图。恶臭电子鼻仪器I设置了外接显示器接口、2个USB接口、鼠标接口、键盘接口、Internet接口、净化环境空气和洁净空气入口、10个监测点恶臭气体入口、外置真空泵III出口、以及废气排出口。
在单采样周期T 0记录时长45秒的响应数据内，单个气敏传感器i响应曲线的稳态最大值U imax(t)和最小值U imin(t)之差值被提取为特征分量x i(t)＝U imax(t)-U imin(t)，气敏传感器阵列因此产生一个16维的响应向量x(t)＝(x 1(t),…,x i(t),…,x 16(t)) T∈R 16。在数据记录结束后的10秒内，即环境空气冲洗阶段后10秒，计算机控制与数据分析***I(c)的机器学习级联模型依据响应向量x(t)预测10+1项恶臭污染物浓度控制指标值。
依据“分而治之”策略，机器学习级联模型第一级—卷积神经网络层采用多个单输出单隐层卷积神经网络一一预测各个气敏传感器的响应。图10为预测t+1时刻(例如未来第40分钟)气敏传感器i响应x i(t+1)的卷积神经网络CNN i1结构示意图。表2(a)为CNN i1的时间序列响应训练集X i1∈R 10×9，共有10个样本，维数为9。训练集X i1时间序列跨度为[t-18,t-1]，在T 0＝240秒且T＝10T 0的情况下，相当于CNN i1学习气敏传感器i从12小时前当前到当前时刻已发生的响应。根据表2(a)，CNN i1的一个学习样本相当于气敏传感器i长度Δt＝9的一个时间序列响应。表2(b)给出CNN i1预测时采用的时间序列响应样本x 1＝(x i(t-8),…,x i(t)) T∈R 9。
卷积神经网络CNN i1学习气敏传感器i在t时刻之前已发生的18个时刻的时间序列响应，时延长度Δt＝9，则输入节点数m i＝9，取隐节点数h i＝5，输出节点数n i＝1；卷积神经网络CNN i1在线学习气敏传感器i经预处理的时间序列响应数据集X i1，如表2(a)所示。CNN i1隐节点和输出节点活化函数为sigmoid修正函数
Figure PCTCN2018088913-appb-000009
采用误差反传算法进行学习，学习因子为η i1＝5/N i1＝0.5，最大迭代次数10,000。表2(a)和2(b)的输入输出分量均成比例变换到范围[0,3]。
卷积神经网络CNN i1在气敏传感器阵列环境空气冲洗阶段后10秒时间内完成在线学习，并依据表2(b)给出的时间序列响应样本x i(t)＝(x i(t-8),…,x i(t)) T来预测t+1时刻气敏传感器i的响应x i(t+1)。
本发明用卷积神经网络CNN i2和CNN i3分别预测气敏传感器i在t+2时刻(例如未来第80分钟)和t+3时刻(例如未来第120分钟)的响应x i(t+2)和x i(t+3)。CNN i2和CNN i3结构与学习参数与NN i1相同。表3和表4给出了这2个卷积神经网络的时间序列响应训练集X i2∈R 10×9和X i3∈R 10×9。这2个卷积神经网络仍采用如2(b)所示的与CNN i1相同的时间序列响应样本x i(t)来预测t+2和t+3时刻气敏传感器i的响应x i(t+2)、x i(t+3)。与X i1的时间跨度为[t-18,t-1]相比，X i2和X i3的时间跨度分别为[t-19,t-2]与[t-20,t-3]，距t要远一些，所以，CNN i2和CNN i3的预测值可信度较CNN i1低。
CNN i1、CNN i2和CNN i3均在气敏传感器阵列环境空气冲洗阶段的后10秒内完成在线学习和预测。因此，当对气敏传感器阵列所有16条响应曲线一一进行t+1、t+2和t+3时刻响应预测时，本发明采用了3*16卷积神经网络；若只预测t+1时刻响应，则只需要16单输出卷积神经网络即可。
本发明依据“分而治之”策略，将恶臭气体多个浓度值整体预测问题分解为多个单一浓度值一一预测问题，用机器学习级联模型第二级—多个单输出深度神经网络来一一预测多个单一浓度值，以有效降低机器
表2(a),卷积神经网络CNN i1的时间序列训练集X i1
Figure PCTCN2018088913-appb-000010
表2(b),卷积神经网络CNN i1预测t+1时刻气敏传感器阵列响应x i(t+1)的时间序列响应样本x i(t)
Figure PCTCN2018088913-appb-000011
表3,卷积神经网络CNN i2的时间序列训练集X i2
Figure PCTCN2018088913-appb-000012
表4,卷积神经网络CNN i3的时间序列训练集X i3
Figure PCTCN2018088913-appb-000013
学习模型与算法的复杂程度。单输出DNN个数等于要预测的恶臭气体浓度控制指标数，一一对应。例如，若要预测无量纲臭气浓度OU值、NH 3、H 2S、CS 2、C 3H 9N、CH 4S、C 2H 6S、C 2H 6S 2、C 8H 8、SO 2等9种恶臭污染物浓度和TVOC浓度，则需要10+1个单输出DNNs。一个单输出DNN学习的是恶臭气体大数据，输入值为气敏传感器阵列检测数据与恶臭电子鼻仪器现场温湿度数据，目标输出为嗅辨值、色质谱等常规仪器离线测量值及居民投诉数据。恶臭气体大数据中有些样本只有气敏传感器阵列响应而没有嗅辨值、色质谱等离线测量值及居民投诉数据的，不参加学习。
一个单输出DNN j有3个隐层，隐层和输出层采用Sigmoid修正活化函数
Figure PCTCN2018088913-appb-000014
气敏传感器阵列响应数据和目标输出分量各自成比例变换到范围[0,3]。第一和第二隐层为特征变换(编码)层，从下向上逐层离线学习方式，结构和权值参数用单隐层对等神经网络确定。图11为确定DNN j的第k层—第k+1隐层权值与阈值的对等神经网络学习过程示意图。图11(a)表明，一个对等神经网络的输出节点数与输入节点数相等，均为线性活化函数，隐层—输出层的权值与阈值直接等于其输入层—隐层的，目标输出直接等于其实际输入。图11(b)表明，该对等神经网络学习结束后，DNN j的第k+1层隐节点数等于该对等神经网络的隐节点数，第k层—第k+1隐层权值与阈值等于该对等神经网络输入层—隐层的。设DNN j的学习样本数为N，则对等神经网络学习因子η＝2/N，最大迭代步数τ max＝10,000。DNN j的第三隐层为非线性映射层，与单个输出单元j一起拟合恶臭气体第j个浓度控制指标值。
图12为机器学习级联模型预测t+1时刻(例如未来第40分钟)多种恶臭污染物浓度示意图。根据图12(a)，机器学习级联模型第一级先用16*3组单输出单隐层卷积神经网络一一学习16个气敏传感器产生的时间序列响应，然后表2(b)的时间序列响应样本x i(t)＝(x i(t-8),…,x i(t)) T分别预测t+1、t+2和t+3时刻各个气敏传感器的响应。机器学习级联模型第二级用10+1个单输出三隐层深度神经网络模块分别预测前述10+1项恶臭污染物控制指标值。
假设DNN j对t+1时刻一种恶臭气体浓度值y j(t+1)进行预测，则依据的是16个CNN i1(i＝1,2,…,16)对气敏传感器阵列t+1时刻预测响应(x 1(t+1),x 2(t+1),…,x 16(t+1)) T以及当前时刻温湿度值；DNN j预测y j(t+2)依据的是16个CNN i2对t+2时刻预测响应(x 1(t+2),x 2(t+2),…,x 16(t+2)) T以及当前时刻温湿度值，等等。
若实际输入是气敏传感器阵列当前响应向量(x 1(t),x 2(t),…,x 16(t)) T，必要时可再加上t时刻温湿度值，则深度神经网络DNN j的实际输出是对恶臭气体成分j当前浓度值y j(t)的估计。

Claims (13)

  1. 一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法，其特征是，恶臭气体多点集中式在线监测与分析***包括恶臭电子鼻仪器I、气体采样探头II、外置真空泵III、环境空气净化装置IV、洁净空气瓶V、气体管道、电子温湿度计VI、中央控制室VII以及多个固定/移动终端VIII，实现恶臭污染区域10个监测点的长期在线监测和多种恶臭污染物浓度控制指标值的在线估计与预测；
    所述的恶臭电子鼻仪器I包括气敏传感器阵列及其恒温工作室I(a)、多点集中式恶臭气体自动进样***I(b)、计算机控制与数据分析***I(c)三大组成部分；气敏传感器阵列恒温工作室I(a)由气敏传感器阵列及其环形工作腔I-1，隔热层I-2，电阻加热丝I-3，风扇I-4组成；气敏传感器阵列I-1由16个气敏元件构成，均布于中径φ140mm、断面尺寸21mm×17mm的密封腔内，形成气敏传感器阵列环形工作腔，处于55±0.1℃的恒温室内，位于恶臭电子鼻仪器I右上方；多点集中式恶臭气体自动进样***I(b)包括内置微型真空泵I-14、14个二位二通电磁阀(I-5,I-6-1～I-6-10,I-8,I-10,I-13)、节流阀I-11、流量计I-12、真空压力表I-7、气体缓冲室I-9，位于恶臭电子鼻仪器I右下方；计算机控制与数据分析***I(c)包括计算机主板I-15、数据采集卡I-16、显示器I-17、驱动与控制电路模块I-18、精密线性与开关电源模块I-19、硬盘、网卡、显卡，位于恶臭电子鼻仪器I左侧；
    多点集中式恶臭气体自动进样***I(b)对单个监测点恶臭气体采样周期为T 0＝180-300秒钟，默认值T 0＝240秒钟，气敏传感器阵列I-1因此对该监测点产生一个16维响应向量；计算机控制与数据分析***I(c)依据这一响应向量，用机器学习级联模型对该监测点的臭气嗅感浓度、GB14554指定的氨NH 3、硫化氢H 2S、二硫化碳CS 2、三甲胺C 3H 9N、甲硫醇CH 4S、甲硫醚C 2H 6S、二甲二硫醚C 2H 6S 2、苯乙烯C 8H 8等8种化合物，GB/T18883指定的二氧化硫SO 2与总挥发性有机化合物共10+1项恶臭污染物浓度控制指标值进行实时分析和预测，并将监测数据和预测结果通过无线Internet网远程传送到中央控制室VII和指定的固定/移动终端VIII；
    恶臭电子鼻仪器I每单周期T 0得到一个16维的响应向量，储存在计算机硬盘的一个数据文件里；用10个二位二通电磁阀I-6-1～I-6-10依次控制4km 2区域内10个监测点恶臭气体的通与断，以T＝10T 0的恶臭气体循环采样周期实现10个监测点恶臭气体的循环在线监测，并将监测数据依次储存在10个数据文件里；这些数据是恶臭电子鼻仪器I预测恶臭污染物多种浓度的数值基础，据此实现对10+1项恶臭污染物浓度控制指标值的循环在线预测。
  2. 根据权利要求1所述的恶臭气体多点集中式电子鼻仪器在线监测与分析方法，其特征是，恶臭气体单采样周期T 0包括：气敏传感器阵列I-1初步恢复(95-215秒)、洁净空气精确标定(30秒)、平衡(5秒)、恶臭气体顶空采样(30秒)、净化环境空气冲洗(20秒)共5个阶段；在单周期T 0内，在计算机控制下，对应监测点的二位二通电磁阀I-6-k(＝1,2,…,10)导通，其余9个监测点的二位二通电磁阀断开，内置微型真空泵I-14以流量1,000ml/min抽吸气体缓冲室I-8内的恶臭气体，使之流经气敏传感器阵列环形工作腔，掠过气敏传感器敏感膜表面，气敏传感器阵列I-1因此产生敏感响应，持续30秒；自平衡状态开始之刻起，计算机控制与数据分析***I(c)持续记录敏感响应数据，包括平衡(5秒)、恶臭气体顶空采样(30秒)、净化环境空气冲洗(前10秒)这3个阶段共45秒的气敏传感器阵列I-1响应数据，并临时存储在一个文本文件里；单周期T 0其它时间的响应数据不记录；
    在时长45秒的响应数据内，单个气敏传感器响应曲线的稳态最大值和最小值之差值被提取为响应分量，气敏传感器阵列I-1因此产生一个16维的响应向量；在数据记录结束后的10秒内，即净化环境空气冲洗阶段的后10秒，计算机控制与数据分析***I(c)依据这一响应向量预测10+1项恶臭污染物浓度控制指标值。
  3. 根据权利要求1所述的恶臭气体多点集中式电子鼻仪器在线监测与分析方法，其特征是，气敏传感器阵列I-1由11个金属氧化物半导体型、4个电化学型和1个光离子型气敏元件组成；其中，11个金属氧化物型气敏元件用于检测多种有机/无机化合物；4个电化学型气敏元件用于检测NH 3、H 2S、CS 2、SO 2等4种无机化合物；1个光离子型气敏元件用于检测总挥发性有机化合物。
  4. 根据权利要求1所述的恶臭气体多点集中式电子鼻仪器在线监测与分析方法，其特征是，恶臭气体多点集中式在线监测与分析***实现大至数km 2的特定区域，小至生产车间或建筑物，乃至1个点的在线监测与分析；对特定区域、生产车间和建筑物，可设置10个监测点，最大监测区域为2km*2km＝4km 2，其中，9个固定监测点，1个移动监测点；恶臭电子鼻仪器I位于室内，通过内径φ10mm不锈钢管道与各监测点相连接；气体采样探头采用水龙头形式，与商用除尘去湿净化部件连接，随用随换；改变监测点位置只需重新铺设不锈钢管道，安装和移动气体采样探头到指定位置即可，楼上楼下，可高可低，就像铺设水管或电缆线一样简便。
  5. 根据权利要求1所述的恶臭气体多点集中式电子鼻仪器在线监测与分析方法，其特征是，8个或更多监测点绕监测区域边界设置，恶臭电子鼻仪器I和10个监测点位置设置以不锈钢管道最短为目标；对路径可达的化工园区、居民小区等区域，恶臭电子鼻仪器I布置在区域中心的某个室内；对无路可达的垃圾填埋场、污水处理厂等区域，恶臭电子鼻仪器I布置在区域边界的某个室内。
  6. 根据权利要求1所述的恶臭气体多点集中式电子鼻仪器在线监测与分析方法，其特征是，外置真空泵III抽气速率250-280L/min，极限真空度100-120mbar，长期连续工作，可通过内径φ10mm不锈钢管道在1分钟之内将直线距离达2.5km的一个监测点的恶臭气体抽吸到恶臭电子鼻仪器I内；在单周期T 0内，除恶臭气体顶空采样(30秒)这一阶段外，其余阶段被抽吸到恶臭电子鼻仪器I内的恶臭气体并不流经气敏传感器阵列环形工作腔I-1，而是被直接排出到室外。
  7. 根据权利要求1所述的恶臭气体多点集中式电子鼻仪器在线监测与分析方法，其特征是，恶臭电子鼻仪器I内部设置有一个尺寸为φ40mm*5mm的气体缓冲室I-8，恶臭气体在此处的流速较内径φ10mm不锈钢管道骤降16倍；只有在恶臭气体顶空采样(30秒)这一阶段，内置微型真空泵I-14才将气体缓冲室I-8内的恶臭气体抽吸到气敏传感器阵列环形工作腔，气敏传感器阵列I-1因此产生敏感响应；内置微型真空泵I-14抽吸到的都是新鲜恶臭气体。
  8. 根据权利要求1所述的恶臭气体多点集中式电子鼻仪器在线监测与分析方法，其特征是，在恶臭气体顶空采样前，1,000ml/min洁净空气精确标定环节(30)使得气敏传感器阵列I-1对恶臭气体的多次感知在同一基线上进行；12～15Mpa压缩气体钢瓶V标准容积为40L，转换到常温常压为6m 3；当单周期T 0＝3、4和5分钟时，这样1瓶40L压缩洁净空气分别可用25、33和41天；恶臭电子鼻仪器I所处室外的环境空气先用商品化空气净化器净化，然后被用来冲洗气敏传感器阵列I-1，使之初步恢复到基准状态，以降低运行费用。
  9. 根据权利要求1所述的恶臭气体多点集中式电子鼻仪器在线监测与分析方法，其特征是，恶臭气体大数据集包括：(1)气敏传感器阵列I-1对化工园区(包括香精香料厂)、制药厂、垃圾填埋场、污水处理厂、养殖场、邻近居民区等大量恶臭污染物现场的在线检测数据；(2)气敏传感器阵列I-1对大量恶臭标准样品顶空挥发气的实验室离线检测数据，其中包括GB/T14675指定的β-苯乙醇、异戊酸、甲基环戊酮、γ-十一烷酸内酯、β-甲基吲哚这5种标准臭液；GB14554指定的C 3H 9N、C 8H 8、H 2S、CH 4S、C 2H 6S、C 2H 6S 2、NH 3、CS 2与GB/T18883指定的SO 2共9种单一成分恶臭污染物配制的不同浓度标准恶臭样品，还包括不同浓度多种单一化合物配制的混合成分标准恶臭样品；(3)GB/T14675和HJ 905-2017规定的真空瓶或臭气袋在大量恶臭污染物现场采样，并立即运回嗅辨室而得到的无量纲臭气浓度离线嗅辨数据；(4)GB/T18883规定的Tenax GC/TA吸附管恶臭污染物现场采样，气相色谱仪实验室离线检测得到的总挥发性有机化合物数据和分光光度仪实验室离线检测得到的SO 2数据；(5)GB/T14676-14680规定的恶臭污染物现场采样，8种恶臭成分的气相色谱仪、质谱仪和分光光度仪实验室离线检测数据；(6)恶臭污染源邻近区域居民投诉数据。
  10. 根据权利要求1所述的恶臭气体多点集中式电子鼻仪器在线监测与分析方法，其特征是，恶臭电子鼻仪器I用机器学习级联模型预测未来t+1、t+2和t+3时刻臭气嗅感浓度和多种恶臭污染物浓度控制指标值；机器学习级联模型第一级—卷积神经网络(Convolutional neural network,CNN)层负责预测t+1、 t+2和t+3时刻气敏传感器阵列I-1对一个监测点恶臭气体的响应，依据的是当前时刻t和近期已发生的气敏传感器阵列I-1时间序列响应；机器学习级联模型第二级—深度神经网络(Deep neural network,DNN)层进一步预测t+1、t+2和t+3时刻臭气嗅感浓度和多种恶臭污染物浓度控制指标值，依据的是长期积累的恶臭气体大数据和级联模型第一级—卷积神经网络层的预测值。
  11. 根据权利要求1所述的恶臭气体多点集中式电子鼻仪器在线监测与分析方法，其特征是，依据“分而治之”策略，机器学习级联模型第一级用16*3组单输出单隐层卷积神经网络一一预测t+1、t+2和t+3时刻各个气敏传感器的响应；对T 0＝40分钟而言，相当于从当前时刻t算起，预测未来第40、80和120分钟时刻的响应；
    以单周期T 0＝40分钟，3个单输出单隐层卷积神经网络模块分别预测t+1、t+2和t+3时刻气敏传感器i的响应为例：
    a)单输出单隐层卷积神经网络CNN i1预测t+1时刻气敏传感器i的响应：
    设卷积神经网络CNN i1学习气敏传感器i在t时刻之前已发生的18个时刻时间序列响应，时延长度Δt＝9，则输入节点数m i＝9，取隐节点数h i＝5，输出节点数n i＝1；卷积神经网络CNN i1在线学习经预处理的气敏传感器i时间序列响应数据集X i1为：
    Figure PCTCN2018088913-appb-100001
    目标输出为：
    d i1＝(x i(t) x i(t-1) x i(t-2) x i(t-3) x i(t-4) x i(t-5) x i(t-6) x i(t-7) x i(t-8) x i(t-9)) T∈R 10，这种方式相当于卷积神经网络CNN i1学习气敏传感器i最近12小时已发生的1个18维时间序列响应，产生10个9维时间序列响应，即样本数为N i1＝10；卷积神经网络CNN i1的隐层和输出层活化函数为Sigmoid修正函数
    Figure PCTCN2018088913-appb-100002
    采用误差反传算法学习，学习因子为η i＝5/N i1＝0.2；数据集X i1和目标输出d i1均成比例变换到范围[0,3]；卷积神经网络CNN i1在10秒钟内在线学习结束后，依据最近时间段的一个9维时间序列响应：
    x i1＝(x i(t-8) x i(t-7) x i(t-6) x i(t-5) x i(t-4) x i(t-3) x i(t-2) x i(t-1) x i(t)) T∈R 9
    预测t+1时刻气敏传感器i的响应x i(t+1)；当T 0＝40分钟时，相当于预测未来第40分钟气敏传感器i的响应；
    b)单输出单隐层卷积神经网络CNN i2与CNN i3预测t+2和t+3时刻气敏传感器i的响应：
    卷积神经网络CNN i2和CNN i3结构仍为：m i＝9，h i＝5，n i＝1；在线学习经预处理的数据集X i2和X i3分别为：
    Figure PCTCN2018088913-appb-100003
    和
    Figure PCTCN2018088913-appb-100004
    即X i2和X i3同样有10个9维时间序列响应，样本数均为N i1＝10；卷积神经网络CNN i2与CNN i3在学习阶段的目标输出和预测时依据的时间序列响应与CNN i1相同；当T 0＝40分钟时，相当于学习气敏传感器i在40分钟和80分钟之前的12小时已发生的响应，预测t+2和t+3时刻气敏传感器i的响应x i(t+2)和x i(t+3)，分别相当于预测气敏传感器i未来第80分钟和120分钟的响应。
  12. 根据权利要求1所述的恶臭气体多点集中式电子鼻仪器在线监测与分析方法，其特征是，依据“分而治之”策略，NH 3、H 2S、CS 2、C 3H 9N、CH 4S、C 2H 6S、C 2H 6S 2、C 8H 8、SO 2、总挥发性有机化合物和臭气嗅感浓度共10+1项恶臭污染物浓度控制指标值整体预测问题被分解为11个单浓度值一一预测问题，机器学习级联模型第二级用10+1个单输出三隐层深度神经网络模块分别预测这10+1项恶臭污染物控制指标值；单输出深度神经网络训练集为恶臭电子鼻仪器I的气敏传感器阵列I-1对标准臭液/气样品和大量污染现场在线检测得到的恶臭气体大数据，目标输出为臭气嗅辨值和色质谱与分光光度常规仪器离线测量值，以及居民投诉数据；
    单个单输出三隐层深度神经网络DNN j采用自下而上的逐层离线学习方式；第一和第二隐层学习时采用单隐层对等神经网络结构，即单隐层对等神经网络的隐层—输出层权值直接等于其输入层—隐层权值，目标输出直接等于其输入，输入分量和输出分量依据特征分量大小成比例变换到范围[0,3]；单隐层对等神经网络的隐层活化函数为Sigmoid修正函数
    Figure PCTCN2018088913-appb-100005
    采用误差反传算法学习，学习因子为η j＝1/N j，学习结束后丢弃隐层—输出层；N j为恶臭气体大数据样本数；
    假设对t+1时刻恶臭气体第j个浓度值y j(t+1)进行预测，第j个单输出深度神经网络DNN j依据的是16个卷积神经网络对t+1时刻气敏传感器阵列I-1的预测响应{x 1(t+1),x 2(t+1),…,x 16(t+1)}，预测y j(t+2)和y j(t+3)分别依据的是16个卷积神经网络对t+2和t+3时刻的预测响应(x 1(t+2),x 2(t+2),…,x 16(t+2)) T与(x 1(t+3),x 2(t+3),…,x 16(t+3)) T；
    若实际输入是气敏传感器阵列当前响应向量(x 1(t),x 2(t),…,x 16(t)) T，必要时可再加上t时刻温湿度值，则深度神经网络DNN j的实际输出是对恶臭气体成分j当前浓度值y j(t)的估计。
  13. 根据权利要求1所述的恶臭气体多点集中式电子鼻仪器在线监测与分析方法，其特征是，恶臭电子鼻仪器I对恶臭污染区域多个监测点长期在线监测和多种恶臭污染物浓度控制指标值的在线预测，包括以下步骤：
    (1)开机：仪器预热30分钟；单击屏幕菜单的“空气净化器开”选项，空气净化器IV开始对恶臭电子鼻仪器I所处的室内空气净化，长期持续工作直至操作人员单击“空气净化器关”选项为止；
    在内置微型真空泵I-14的抽吸作用下，净化环境空气以6,500ml/min的流量依次流经二位二通电磁阀I-5、气敏传感器阵列环形工作腔I-1、二位二通电磁阀I-10，然后被排出到室外；气敏传感器阵列环形工作腔I-1内的温度从室温达到恒定的55±0.1℃；
    单击屏幕菜单的“外置真空泵开”选项；外置真空泵III以250-280L/min的抽气速率和100-120mbar的极限真空度，通过内径φ10mm不锈钢管道在1分钟内将直线距离最大达2.5km的某个监测点恶臭气体抽吸到恶臭电子鼻仪器I内，依次流过对应的二位二通电磁阀I-6-k(＝1,2,…,10)、真空压力表I-7和气体缓冲室I-8，然后直接排出到室外；外置真空泵III持续抽吸恶臭气体，直到操作人员单击屏幕菜单的“外置真空泵关”选项为止；
    修改屏幕菜单恶臭气体“单采样周期T 0”设置，默认值T 0＝40分钟；10个监测点恶臭气体循环采样周期为T＝10T 0；
    (2)恶臭气体循环采样周期开始：点击屏幕菜单的“开始检测”按钮，恶臭电子鼻仪器I依次对10个监测点进行循环监测，计算机控制与数据分析***I(c)在指定文件夹自动生成10个文本文件，以存储气敏传感器阵列I-1对10个监测点恶臭气体的响应数据；
    (3)监测点k(＝1,2,…,10)恶臭气体单采样周期开始；以T 0＝4分钟为例：
    (3.1)气敏传感器阵列初步恢复：单周期T 0第0-155秒，在内置微型真空泵I-14的抽吸作用下，净化环境空气以6,500ml/min的流量依次流经二位二通电磁阀I-5、气敏传感器阵列环形工作腔I-1、二位 二通电磁阀I-10，然后被排出到室外；在6,500ml/min净化环境空气的作用下，气敏传感器阵列环型工作腔I-1内积聚的热量被带走，粘附在气敏传感器敏感膜表面和管道内壁的恶臭气体分子被初步冲走，气敏传感器阵列I-1初步恢复到基准状态，历时155秒；
    10个二位二通电磁阀I-6-1～I-6-10只有I-6-k导通，其余9个断开，外置真空泵III将监测点k(＝1,2,…,10)的恶臭气体抽吸到恶臭电子鼻仪器I内；
    (3.2)洁净空气精确标定：在单周期T 0第156-185秒，二位二通电磁阀I-13导通，二位二通电磁阀I-5、I-8和I-10断开，二位二通电磁阀I-6-1～I-6-10保持步骤(3.1)的状态；在内置微型真空泵I-14的抽吸作用下，洁净空气以1,000ml/min的流量依次流经二位二通电磁阀I-13、气体管道、气敏传感器阵列环形工作腔I-1、节流阀I-11、流量计I-12、微型真空泵I-14，然后被排出到室外；洁净空气使气敏传感器阵列I-1精确恢复到基准状态；历时30秒；外置真空泵III持续抽吸；
    (3.3)平衡：在单周期T 0第186-190秒，二位二通电磁阀I-5、I-8、I-10、I-13断开，二位二通电磁阀I-6-1～I-6-10保持步骤(3.1)的状态；气敏传感器阵列环形工作腔I-1内无气体流动；自单周期T 0第186秒即平衡状态开始之刻起，计算机控制与数据分析***I(c)开始记录气敏传感器阵列I-1实时响应数据，并存储在指定的临时文本文件“temp.txt”里；历时5秒；外置真空泵III持续抽吸；
    (3.4)监测点k恶臭气体顶空采样：在单周期T 0第190-220秒，二位二通电磁阀I-8导通，3个二位二通电磁阀I-5、I-13和I-10断开，二位二通电磁阀I-6-1～I-6-10保持步骤(3.1)的状态；在内置微型真空泵I-14抽吸作用下，气体缓冲室I-8内的恶臭气体以流量1,000ml/min依次流过气敏传感器阵列环形工作腔I-1、节流阀I-11、流量计I-12、内置微型真空泵I-14，最后排出到室外；气敏传感器阵列I-1因此产生的敏感响应继续记录在临时文件“temp”里，历时30秒；外置真空泵III持续抽吸；
    (3.5)气敏传感器阵列冲洗：在单周期T 0第221-230秒，二位二通电磁阀I-5，二位二通电磁阀I-8、I-10和I-13断开，在微型真空泵I-14抽吸作用下，流量6,500ml/min的净化环境空气以依次流经二位二通电磁阀I-5、气敏传感器阵列环形工作腔I-1、二位二通电磁阀I-10，然后被排出到室外；与此同时，二位二通电磁阀I-6-k+1导通，10个二位二通电磁阀I-6-1～I-6-10的其余9个断开，包括二位二通电磁阀k断开，外置真空泵III转而抽吸监测点k+1的恶臭气体；由于净化环境空气的作用，气敏传感器阵列环型工作腔内积聚的热量被带走，粘附在气敏传感器敏感膜表面和管道内壁的恶臭气体分子被初步冲走，气敏传感器阵列I-1逐步恢复到基准状态；历时20秒；其中：
    (a)在单周期T 0第221-230秒，气敏传感器阵列响应数据继续记录在临时文件“temp”里，历时10秒；至第230秒末，计算机控制与数据分析***I(c)停止记录气敏传感器阵列响应数据；
    (b)在单周期T 0第231-240秒，计算机控制与数据分析***I(c)与此进行以下三项操作：
    (b1)特征提取：自第231秒之刻起，并从时长45秒的临时文件“temp”里提取各个气敏传感器的最大和最小稳态响应值，以最大响应值与最小响应值之差作为各个气敏传感器当前时刻t对监测点k恶臭气体的响应特征分量x i(t)(i＝1,2,…,16)，并记录在对应的数据文件里；
    (b2)气敏传感器阵列响应预测：机器学习级联模型第一级—16*3个卷积神经网络依据当前时刻t以前[t-18,t]、[t-19,t-1]和[t-20,t-2]时间段内已发生的气敏传感器阵列时间序列响应向量，实现在线自学习，并据此预测未来T 0、2T 0和3T 0时刻气敏传感器阵列I-1的响应；
    (b3)恶臭气体浓度控制指标值预测：机器学习级联模型第二级—10+1个深度神经网络依据级联模型第一级的16*3个卷积神经网络预测的气敏传感器阵列响应值，进一步预测监测点k的10+1项恶臭污染物浓度控制指标值，通过显示器显示出来，并将监测和预测结果通过Internet网络传送到中央控制室VII和多个固定/移动终端VIII；
    (3.6)监测点k恶臭气体单采样周期结束：k←k+1，回到步骤(3.1)，监测点k+1恶臭气体单采样周期开始；
    重复步骤(3.1)～(3.6)，恶臭电子鼻仪器I实现对10个监测点恶臭气体的循环在线监测、识别和10+1项恶臭污染物控制指标值的预测。
PCT/CN2018/088913 2018-05-17 2018-05-30 一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法 WO2019218395A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US16/642,531 US10948467B2 (en) 2018-05-17 2018-05-30 Online centralized monitoring and analysis method for multi-point malodorous gases using electronic nose instrument

Applications Claiming Priority (4)

Application Number Priority Date Filing Date Title
CN201810471613.1 2018-05-17
CN201810471708.3 2018-05-17
CN201810471708.3A CN108896706B (zh) 2018-05-17 2018-05-17 大数据驱动的恶臭气体多点集中式电子鼻仪器在线分析方法
CN201810471613.1A CN108709955B (zh) 2018-05-17 2018-05-17 一种恶臭气体多点集中式在线监测与分析***及方法

Publications (1)

Publication Number Publication Date
WO2019218395A1 true WO2019218395A1 (zh) 2019-11-21

Family

ID=68539330

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/088913 WO2019218395A1 (zh) 2018-05-17 2018-05-30 一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法

Country Status (2)

Country Link
US (1) US10948467B2 (zh)
WO (1) WO2019218395A1 (zh)

Cited By (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111289314A (zh) * 2020-04-16 2020-06-16 吕梁学院 一种针对恶劣环境使用的检测采样装置
CN111596006A (zh) * 2020-04-29 2020-08-28 北京雪迪龙科技股份有限公司 一种大气在线监测仪的校准方法及监测仪
CN111950936A (zh) * 2020-08-31 2020-11-17 平安国际智慧城市科技股份有限公司 空气质量监测方法、装置、设备及介质
CN112819158A (zh) * 2021-02-05 2021-05-18 凌坤(南通)智能科技有限公司 一种基于优化的bp神经网络的气体辨识方法
CN113125509A (zh) * 2021-04-15 2021-07-16 浙江大学 基于生物混合***的人体臭味检测与掩蔽效果评价方法
CN113447618A (zh) * 2021-06-29 2021-09-28 重庆大学 多传感器电子鼻的阶梯形气室、双进样检测***及方法
WO2021216172A1 (en) * 2020-04-22 2021-10-28 Evoqua Water Technologies Llc Electrochemical sensor assembly and system
WO2022051518A1 (en) * 2020-09-02 2022-03-10 University Of Southern California Blocking neural networks for high capacity
CN114923896A (zh) * 2022-03-31 2022-08-19 清华大学 光电子鼻的控制***和光电子鼻
CN114958587A (zh) * 2022-06-16 2022-08-30 深圳兆源环境科技有限公司 微生物特异性反应监测气体恶臭ou值的自动化设备
CN115090042A (zh) * 2022-06-29 2022-09-23 国能(山东)能源环境有限公司 基于滤筒除尘器和物联网的室内离散粉尘过滤***及方法
US11636870B2 (en) 2020-08-20 2023-04-25 Denso International America, Inc. Smoking cessation systems and methods
US11760170B2 (en) 2020-08-20 2023-09-19 Denso International America, Inc. Olfaction sensor preservation systems and methods
US11760169B2 (en) 2020-08-20 2023-09-19 Denso International America, Inc. Particulate control systems and methods for olfaction sensors
US11813926B2 (en) 2020-08-20 2023-11-14 Denso International America, Inc. Binding agent and olfaction sensor
US11828210B2 (en) 2020-08-20 2023-11-28 Denso International America, Inc. Diagnostic systems and methods of vehicles using olfaction
US11881093B2 (en) 2020-08-20 2024-01-23 Denso International America, Inc. Systems and methods for identifying smoking in vehicles
EP4118439A4 (en) * 2020-03-10 2024-03-13 Stratuscent Inc. SYSTEMS AND METHODS FOR DETECTING A SENSOR EVENT IN A STREAM OF CHEMICAL SENSOR MEASUREMENT DATA
US11932080B2 (en) 2020-08-20 2024-03-19 Denso International America, Inc. Diagnostic and recirculation control systems and methods
US12017506B2 (en) 2020-08-20 2024-06-25 Denso International America, Inc. Passenger cabin air control systems and methods

Families Citing this family (29)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP7074194B2 (ja) * 2018-07-31 2022-05-24 日本電気株式会社 情報処理装置、制御方法、及びプログラム
KR102607881B1 (ko) * 2018-12-19 2023-11-29 삼성전자주식회사 공기 조화 장치 및 그 제어 방법
TWI721579B (zh) * 2019-09-27 2021-03-11 研能科技股份有限公司 人工智能物聯網處理空氣品質系統
CN113128013A (zh) * 2019-12-30 2021-07-16 鸿富锦精密电子(天津)有限公司 环境状态分析方法、装置、计算机装置及存储介质
WO2021147274A1 (zh) * 2020-01-23 2021-07-29 华东理工大学 一种气敏-气相电子鼻仪器和发酵-恶臭多状态参数在线分析方法
US11742901B2 (en) * 2020-07-27 2023-08-29 Electronics And Telecommunications Research Institute Deep learning based beamforming method and apparatus
US11409797B2 (en) * 2020-08-31 2022-08-09 MS∧2 Inc. Systems and methods for generating smell data and a database thereof
CN112666241B (zh) * 2020-12-28 2023-08-22 北京首创大气环境科技股份有限公司 一种区域tvoc监测方法
CN112798673A (zh) * 2020-12-28 2021-05-14 汉威科技集团股份有限公司 一种多参数恶臭气体浓度在线监测方法
CN112870999B (zh) * 2020-12-29 2022-08-26 国网浙江省电力有限公司电力科学研究院 一种判断气室内c4f7n和co2混合气体中杂质气体的方法
CN112782233B (zh) * 2020-12-30 2023-05-09 江苏智闻智能传感科技有限公司 一种基于阵列气体传感器的气体识别方法
JP2022169273A (ja) * 2021-04-27 2022-11-09 株式会社日立製作所 ガス検出システム及びガス検出方法
CN113238003B (zh) * 2021-05-07 2023-07-14 山东工商学院 一种电子鼻***基础数据集获取及验证平台
CN113341066B (zh) * 2021-05-24 2022-04-08 西南石油大学 基于多传感器融合技术的四氢噻吩浓度的在线检测方法及***
CN113447624B (zh) * 2021-06-29 2022-06-21 重庆大学 一种用于检测茶叶品质的电子鼻及其控制方法和管理***
CN114113491B (zh) * 2021-12-10 2023-03-07 中国科学技术大学 一种利用卷积神经网络计算的集成式机器嗅觉芯片
CN114487314A (zh) * 2021-12-16 2022-05-13 山东省食品药品检验研究院 一种现场快速检测海鲜产品残留违禁药物的装置及方法
CN114368795B (zh) * 2021-12-31 2023-01-17 天健创新（北京）监测仪表股份有限公司 一种在线式黑臭水体的多模态辨识方法及***
CN114646680B (zh) * 2022-05-18 2022-09-13 深圳市瑞达同生科技发展有限公司 一种气体传感器自动测试***
CN115174622A (zh) * 2022-07-05 2022-10-11 中国地质大学（武汉） 一种多路物联网气体监测控制***和方法
CN115356434B (zh) * 2022-07-14 2023-06-02 福建省杭氟电子材料有限公司 六氟丁二烯储放场所的气体监测***及其监测方法
CN115290784A (zh) * 2022-08-03 2022-11-04 南京中车浦镇城轨车辆有限责任公司 一种车辆内厌恶型气味物质确定方法
CN115860056B (zh) * 2023-02-17 2023-05-26 电子科技大学 一种用于混合气体浓度预测的传感器阵列神经网络方法
CN116146309B (zh) * 2023-04-23 2023-07-07 浙江浙能迈领环境科技有限公司 船舶尾气处理***及其方法
CN116764414A (zh) * 2023-06-12 2023-09-19 广东安源鼎盛检测评价技术服务有限公司 一种作业场所空气有毒物质检测及处理装置
CN116577473B (zh) * 2023-07-14 2023-11-17 北京市农林科学院 一种草莓机械损伤发生时间的检测方法及装置
CN116990088B (zh) * 2023-07-31 2024-02-09 珠海深圳清华大学研究院创新中心 城镇污水***温室气体释放监测装置及方法
CN117491565B (zh) * 2023-10-13 2024-05-14 江苏云创环境技术研究院有限公司 一种室外空气环境质量监测装置及其方法
CN117391374B (zh) * 2023-10-30 2024-03-08 江苏国贸酝领智能科技股份有限公司 一种基于物联网技术的智慧水务监管方法及***

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102353798A (zh) * 2011-07-01 2012-02-15 华东理工大学 一种面向生物发酵过程的嗅觉模拟仪器与在线分析方法
CN103454335A (zh) * 2013-09-09 2013-12-18 华东理工大学 一种多通道集成嗅觉模拟仪器和生物发酵过程在线分析方法
CN103472094A (zh) * 2013-07-24 2013-12-25 华东理工大学 一种嗅觉模拟仪器和特定物质气（嗅）味等级现场分析方法
US20140332994A1 (en) * 2013-05-13 2014-11-13 Jeffrey E. Danes Detection in thermoplastics
CN105486812A (zh) * 2015-12-28 2016-04-13 周俊杰 连续环境空气质量监测中电子鼻恶臭等级赋值方法及应用
CN107807199A (zh) * 2017-10-31 2018-03-16 华东理工大学 一种电子鼻仪器和烟草与烟草制品感官质量检测方法
CN107843695A (zh) * 2017-10-31 2018-03-27 华东理工大学 烟草与烟草制品感官质量的电子鼻仪器评价方法

Family Cites Families (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5766551A (en) * 1995-06-29 1998-06-16 Iowa State University Research Foundation, Inc. Device for quantitation of odors from liquid livestock wastes
GB9604627D0 (en) * 1996-03-04 1996-05-01 Neotronics Ltd Sensor for volatile materials and gases
US5756879A (en) * 1996-07-25 1998-05-26 Hughes Electronics Volatile organic compound sensors
US6672129B1 (en) * 1997-10-22 2004-01-06 Microfab Technologies, Inc. Method for calibrating a sensor for measuring concentration of odors
JP3932653B2 (ja) 1998-03-10 2007-06-20 ソニー株式会社 非水電解液二次電池
US8394330B1 (en) * 1998-10-02 2013-03-12 The California Institute Of Technology Conductive organic sensors, arrays and methods of use
US7183115B1 (en) * 2001-11-09 2007-02-27 Lawrence Factor, Inc. Gas sample analysis system
CN1173171C (zh) * 2002-03-15 2004-10-27 上海慧比智能科技有限公司 一种嗅觉模拟装置及其嗅觉模拟测试方法
CN1228625C (zh) * 2003-07-11 2005-11-23 华东理工大学 一种基于模块化组合神经网络的机器嗅觉气味识别方法
US7905154B2 (en) * 2004-11-29 2011-03-15 Jones Jr Arthur T Apparatus and method of contaminant detection for food industry
WO2008024138A1 (en) * 2006-08-23 2008-02-28 Georgia Tech Research Corporation Fluidically-assisted sensor systems for fast sensing of chemical and biological substances
US20080236249A1 (en) * 2007-01-09 2008-10-02 Fernandez De La Mora Gonzalo Method for identification of persons based on the analysis of volatile substances
US7971470B2 (en) * 2007-04-13 2011-07-05 Madison Avenue Management Company, Inc. Method for detecting chemical substances in whole, closed and/or sealed containers
US8312759B2 (en) * 2009-02-17 2012-11-20 Mcalister Technologies, Llc Methods, devices, and systems for detecting properties of target samples
CN101806763B (zh) * 2010-02-26 2012-11-21 华东理工大学 一种嗅觉模拟仪器和多种气味定性定量同时分析方法
US20130298664A1 (en) * 2012-05-08 2013-11-14 Logimesh IP, LLC Pipe with vibrational analytics
CN103196830B (zh) 2013-03-26 2016-03-16 北京拓扑智鑫科技有限公司 环境恶臭污染定量监测方法
US9400766B2 (en) * 2013-04-26 2016-07-26 International Business Machines Corporation Determining combinations of odors to produce a target olfactory pleasantness
US10078044B2 (en) * 2013-10-08 2018-09-18 Nanyang Technological University Method and arrangement for determining at least one pore-related parameter of a porous material

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102353798A (zh) * 2011-07-01 2012-02-15 华东理工大学 一种面向生物发酵过程的嗅觉模拟仪器与在线分析方法
US20140332994A1 (en) * 2013-05-13 2014-11-13 Jeffrey E. Danes Detection in thermoplastics
CN103472094A (zh) * 2013-07-24 2013-12-25 华东理工大学 一种嗅觉模拟仪器和特定物质气（嗅）味等级现场分析方法
CN103454335A (zh) * 2013-09-09 2013-12-18 华东理工大学 一种多通道集成嗅觉模拟仪器和生物发酵过程在线分析方法
CN105486812A (zh) * 2015-12-28 2016-04-13 周俊杰 连续环境空气质量监测中电子鼻恶臭等级赋值方法及应用
CN107807199A (zh) * 2017-10-31 2018-03-16 华东理工大学 一种电子鼻仪器和烟草与烟草制品感官质量检测方法
CN107843695A (zh) * 2017-10-31 2018-03-27 华东理工大学 烟草与烟草制品感官质量的电子鼻仪器评价方法

Cited By (24)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP4118439A4 (en) * 2020-03-10 2024-03-13 Stratuscent Inc. SYSTEMS AND METHODS FOR DETECTING A SENSOR EVENT IN A STREAM OF CHEMICAL SENSOR MEASUREMENT DATA
CN111289314A (zh) * 2020-04-16 2020-06-16 吕梁学院 一种针对恶劣环境使用的检测采样装置
WO2021216172A1 (en) * 2020-04-22 2021-10-28 Evoqua Water Technologies Llc Electrochemical sensor assembly and system
CN111596006A (zh) * 2020-04-29 2020-08-28 北京雪迪龙科技股份有限公司 一种大气在线监测仪的校准方法及监测仪
US11881093B2 (en) 2020-08-20 2024-01-23 Denso International America, Inc. Systems and methods for identifying smoking in vehicles
US11828210B2 (en) 2020-08-20 2023-11-28 Denso International America, Inc. Diagnostic systems and methods of vehicles using olfaction
US12017506B2 (en) 2020-08-20 2024-06-25 Denso International America, Inc. Passenger cabin air control systems and methods
US11932080B2 (en) 2020-08-20 2024-03-19 Denso International America, Inc. Diagnostic and recirculation control systems and methods
US11813926B2 (en) 2020-08-20 2023-11-14 Denso International America, Inc. Binding agent and olfaction sensor
US11760169B2 (en) 2020-08-20 2023-09-19 Denso International America, Inc. Particulate control systems and methods for olfaction sensors
US11760170B2 (en) 2020-08-20 2023-09-19 Denso International America, Inc. Olfaction sensor preservation systems and methods
US11636870B2 (en) 2020-08-20 2023-04-25 Denso International America, Inc. Smoking cessation systems and methods
CN111950936B (zh) * 2020-08-31 2023-08-18 平安国际智慧城市科技股份有限公司 空气质量监测方法、装置、设备及介质
CN111950936A (zh) * 2020-08-31 2020-11-17 平安国际智慧城市科技股份有限公司 空气质量监测方法、装置、设备及介质
WO2022051518A1 (en) * 2020-09-02 2022-03-10 University Of Southern California Blocking neural networks for high capacity
CN112819158A (zh) * 2021-02-05 2021-05-18 凌坤(南通)智能科技有限公司 一种基于优化的bp神经网络的气体辨识方法
CN112819158B (zh) * 2021-02-05 2024-02-27 凌坤(南通)智能科技有限公司 一种基于优化的bp神经网络的气体辨识方法
CN113125509A (zh) * 2021-04-15 2021-07-16 浙江大学 基于生物混合***的人体臭味检测与掩蔽效果评价方法
CN113447618B (zh) * 2021-06-29 2023-11-03 重庆大学 多传感器电子鼻的阶梯形气室、双进样检测***及方法
CN113447618A (zh) * 2021-06-29 2021-09-28 重庆大学 多传感器电子鼻的阶梯形气室、双进样检测***及方法
CN114923896A (zh) * 2022-03-31 2022-08-19 清华大学 光电子鼻的控制***和光电子鼻
CN114958587A (zh) * 2022-06-16 2022-08-30 深圳兆源环境科技有限公司 微生物特异性反应监测气体恶臭ou值的自动化设备
CN115090042A (zh) * 2022-06-29 2022-09-23 国能(山东)能源环境有限公司 基于滤筒除尘器和物联网的室内离散粉尘过滤***及方法
CN115090042B (zh) * 2022-06-29 2024-04-02 国能(山东)能源环境有限公司 基于滤筒除尘器和物联网的室内离散粉尘过滤***的工作方法

Also Published As

Publication number Publication date
US10948467B2 (en) 2021-03-16
US20200400631A1 (en) 2020-12-24

Similar Documents

Publication Publication Date Title
WO2019218395A1 (zh) 一种恶臭气体多点集中式电子鼻仪器在线监测与分析方法
CN108709955B (zh) 一种恶臭气体多点集中式在线监测与分析***及方法
CN108896706B (zh) 大数据驱动的恶臭气体多点集中式电子鼻仪器在线分析方法
Naumova et al. Gas/particle distribution of polycyclic aromatic hydrocarbons in coupled outdoor/indoor atmospheres
Monn Exposure assessment of air pollutants: a review on spatial heterogeneity and indoor/outdoor/personal exposure to suspended particulate matter, nitrogen dioxide and ozone
Kim et al. Monitoring and prediction of indoor air quality (IAQ) in subway or metro systems using season dependent models
Sironi et al. Continuous monitoring of odours from a composting plant using electronic noses
Brown et al. Relationships between socioeconomic and lifestyle factors and indoor air quality in French dwellings
Price et al. Budgets of organic carbon composition and oxidation in indoor air
Zhao et al. Operating behavior and corresponding performance of mechanical ventilation systems in Chinese residential buildings
Orecchio et al. The distribution of phthalate esters in indoor dust of Palermo (Italy)
Huang et al. Land use patterns and SO2 and NO2 pollution in Ulaanbaatar, Mongolia
Han et al. Characterizing oxygenated volatile organic compounds and their sources in rural atmospheres in China
Li et al. A practical framework for predicting residential indoor PM2. 5 concentration using land-use regression and machine learning methods
Pavilonis et al. Indoor and outdoor particulate matter and endotoxin concentrations in an intensely agricultural county
CN102621272A (zh) 一种大气碳氢化合物在线分析仪
Hänninen et al. School environment: policies and current status
Harrison et al. Measurement and modeling of exposure to selected air toxics for health effects studies and verification by biomarkers
CN110187055A (zh) 一种用于不同恶臭源厂家之间的厂界型监测仪
Ramasubramanian et al. Pilot study on the impact of green roofs on ozone levels near building ventilation air supply
CN111443159B (zh) 一种气敏-气相色谱多源感知和电子鼻仪器在线检测方法
CN209841839U (zh) 一种用于不同恶臭源厂家之间的厂界型监测仪
Morgenstern et al. Lack of spatial variation of endotoxin in ambient particulate matter across a German metropolitan area
Qu et al. Rapid determination of chemical concentration and odor concentration of paint-emitted pollutants using an electronic nose
CN205384218U (zh) 受控堆肥条件下高分子材料释放的二氧化碳采集分析装置

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18918451

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18918451

Country of ref document: EP

Kind code of ref document: A1