CN116049764A - Cross-scale time sequence data fusion method and system for Internet of things - Google Patents
Cross-scale time sequence data fusion method and system for Internet of things Download PDFInfo
- Publication number
- CN116049764A CN116049764A CN202310077575.2A CN202310077575A CN116049764A CN 116049764 A CN116049764 A CN 116049764A CN 202310077575 A CN202310077575 A CN 202310077575A CN 116049764 A CN116049764 A CN 116049764A
- Authority
- CN
- China
- Prior art keywords
- data
- model
- time sequence
- sequence
- feature
- Prior art date
- Legal status (The legal status 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 status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/901—Indexing; Data structures therefor; Storage structures
- G06F16/9024—Graphs; Linked lists
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/901—Indexing; Data structures therefor; Storage structures
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Databases & Information Systems (AREA)
- Physics & Mathematics (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a trans-scale time sequence data fusion method and system for the Internet of things. The method comprises the following steps: acquiring cross-scale time sequence data from multiple subsystems and multiple types of sensors; performing time window sliding treatment on the cross-scale time sequence data, wherein the treated result comprises two types of data of equal interval multivariable time sequence and multielement association relation; modeling feature fusion is carried out on two types of data of the multivariate time sequence and the multivariate association relation at equal intervals, and parameters are dynamically loaded and operated along with the transaction processing period synchronously, so that fused implicit feature data are obtained; and processing the fused implicit characteristic data according to specific steps, and applying the processed implicit characteristic data to safety monitoring, system situation assessment and inter-subsystem linkage operation decision. The invention can effectively solve the problems that a large amount of sensor data and subsystem data in certain environments cannot be uniformly managed, fused and shared, and logic linkage control cannot be performed among subsystems.
Description
Technical Field
The invention relates to a trans-scale time sequence data fusion method and system for the Internet of things, and belongs to the technical field of electronic information.
Background
In recent years, with the development of semiconductor technology, internet of things technology and the like, intelligent sensors are widely applied to the fields of large-scale buildings, traffic, power grids, petrochemical industry and the like, and the widely-existing terminal information carriers realize interconnection and intercommunication through communication modes such as Ethernet, twisted pair, optical fibers, 4G, 5G and the like.
At present, various intelligent sensors exist in the internet of things, the adopted communication protocols and interface standards are different, the trans-scale time sequence data from each sensor is lack of unified storage and management, and analysis and application of the time sequence data are limited in an independent (autonomous) subsystem. The subsystems have no data sharing, so that the whole Internet of things system cannot realize cooperative operation.
Taking a large building as an example, a large number of subsystems exist inside the whole building, such as a lighting system, an elevator system, a water supply system, a parking management system, an intrusion alarm system, an access control system, an electric power monitoring system, a fire automatic alarm system and the like. The subsystems do not realize information interconnection and intercommunication, and data fusion and sharing, and the systems are difficult to realize linkage cooperation.
Disclosure of Invention
The invention mainly aims to provide a cross-scale time sequence data fusion method and system for the Internet of things, which can effectively solve the problems that a large amount of sensor data and subsystem data in certain environments cannot be uniformly managed, fused and shared and logic linkage control cannot be performed among subsystems.
In order to solve the problems, the invention adopts the following technical scheme:
a cross-scale time sequence data fusion method oriented to the Internet of things comprises the following steps:
and (3) data acquisition: acquiring cross-scale time sequence data from multiple subsystems and multiple types of sensors;
and (3) fusion of sensing data: performing time window sliding treatment on the cross-scale time sequence data, wherein the treated result comprises two types of data of equal interval multivariable time sequence and multielement association relation;
modeling feature fusion: modeling feature fusion is carried out on two types of data of the multivariate time sequence and the multivariate association relation at equal intervals, and parameters are dynamically loaded and operated along with the transaction processing period synchronously, so that fused implicit feature data are obtained;
transaction processing: and processing the fused implicit characteristic data according to specific steps, and applying the processed implicit characteristic data to safety monitoring, system situation assessment and inter-subsystem linkage operation decision.
The cross-scale time sequence data fusion system for the Internet of things realizes all functions from data acquisition to transaction processing and mainly comprises the following modules:
1) And a data acquisition module: the system comprises a sensing data fusion module, a sensor data acquisition module, a data processing module and a data processing module, wherein the sensing data fusion module is used for acquiring cross-scale time sequence data from various subsystems and various types of sensors, denoising and filtering the acquired cross-scale time sequence data and then sending the acquired cross-scale time sequence data to the sensing data fusion module;
2) And a sensing data fusion module: the method comprises the steps that cross-scale time sequence data received from a data acquisition module are subjected to time window sliding processing, a processed result comprises two types of data of equal interval multi-variable time sequence and multi-element association relation, and the two types of data are provided for a modeling feature fusion module through an internal interface;
3) And a modeling feature fusion module: the modeling feature fusion module is used for dynamically loading parameters along with the transaction processing period and running the parameters, and providing fused implicit feature data for the transaction processing module through an internal interface;
4) A transaction processing module: the method is used for processing the implicit characteristic data provided by the modeling characteristic fusion module according to specific steps and then applying the implicit characteristic data to safety monitoring, system situation assessment and inter-subsystem linkage operation decision.
Further, the cross-scale time sequence data involved in the data acquisition module of the present invention includes data with different time granularity (if millisecond, second, minute), different communication modes (such as ethernet, twisted pair, wifi, 5G), different communication protocols (such as KNX, BACnet, OPC, modBus, HTTP, 104, DDE, etc.), and different data types (floating point type, boolean type, long form, character string).
Further, the multi-element association relation referred to in the invention refers to establishing a causal relation between the cross-scale time series data and other relations (including but not limited to a spatial relation, a device topological relation and the like) between the cross-scale time series data established by using a map model.
Furthermore, the invention adopts the time window to realize the sensing data fusion by the sliding processing of the time window in the sensing data fusion module, and combines the following two aspects: a) Converting the cross-scale time sequence data into equidistant time sequence data, namely an equidistant multivariable time sequence by algebraic interpolation, machine learning fitting, spline function or numerical integration equivalent method; b) And dynamically extracting the multi-element association relation among the data in the set time window along with the running process of the system.
Further, the modeling feature fusion related in the modeling feature fusion module comprises the following three steps:
a) Establishing a multivariate time sequence prediction model by taking historical data of the equal interval multivariate time sequence received from the sensing data fusion module as a reference sample and taking current and subsequent time sequence data as learning targets;
b) Taking the attribute of the measured entity in the Internet of things as a node, taking the multi-element association relationship as an edge, establishing a multi-element feature graph model, and mining potential association relationship or association relationship change;
c) Model parameters of the multivariate time sequence prediction model and the multivariate feature map model are saved as readable data files and updated regularly through a programmed automatic optimization solver.
Further, the dynamic loading parameters related in the modeling feature fusion module comprise the following four types of parameters:
a) Time attribute of data sampling moment in period set by time sequence signal, according to characteristics of application scene, phase in period
Or the following (time-sequential) position-coded expression:
where i represents the sampling time, d is a coefficient set in the model, and T represents the period set by the timing signal.
The dynamic updating mode of the parameters is as follows: the ratio i/T of the sampling moment i relative to the periodic dividing starting position is updated along with the system operation.
b) Modeling feature fusion focuses on loading data in time sequence, including sampling data X of each sensor i at each moment j i, Because the physical data sizes represented by various sensors are different, parameter normalization processing is often required:
the dynamic updating mode of the parameters is as follows: when the update time of equal interval is triggered, each data item loading history is shifted left by one bit value along the time sequence, namely X i, ←X i, ，j＝2,3,...,t。
c) Modeling the original data e with feature fusion focusing on the association relation ij E {0,1 time window [ t ] 0 ,t 1 ,...,t N ) Sliding the generated data updates. Dynamic updating mode for time-varying stability determination and parameterization determination of association heterogeneity of association relationship between entity nodes (i, j) is referred to the determination method in sub-step 1) and sub-step 2 in the multi-feature graph model construction) Is transformed by the medium code.
d) The variable parameters related to the model structure and model training (i.e. parameter learning) in the modeling feature fusion module comprise linear mapping parameters in vector conversion in a multivariate time sequence prediction model and a multivariate feature map model, model parameters in nonlinear units (Transformer, RELU/GELU, MLP) in a neural network, and super parameters (such as learning rate of the neural network training, training data batch and training iteration epoch number) set in the model training. The dynamic updating mode of the parameters is as follows: 1) Collecting time sequence data and data association relation in a new time window along with system operation, so as to update training set data in batches at regular intervals; 2) If the total number of the data batch for model training is regulated to be kept unchanged, mixing the historical training data and the newly collected training data according to different weight proportions, for example, under the condition that the total number of the model training batch is unchanged, lifting a new training set epoch, and reducing the historical training set epoch or randomly discarding part of the historical data; 3) The multivariate time sequence prediction model and the multivariate feature map model adopt respective optimization solvers, wherein the neural network training learning rate can be set in the optimizers for self-adaptive dynamic adjustment; and performing iterative optimization by taking the currently used model parameters as initial values, and training through a new data set to obtain optimized model parameter values.
Further, the multivariate timing prediction model (Multivariate Time Series model, hereinafter referred to as MTS-R1 module in the drawings) referred to in the modeling feature fusion module of the present invention is an integrated model composed of three modules, namely a single-step prediction target fit (hereinafter referred to as MTS-R1 module in the drawings), a cross-period prediction target fit (hereinafter referred to as MTS-R2 module in the drawings), and a sequence-to-sequence model (hereinafter referred to as MTS-S2S module in the drawings). The multivariate time sequence prediction model is shown in fig. 3, and follows the following steps to learn the prediction task of the system situation:
1) Data acquisition
And receiving a multi-element time sequence formed by collecting the multiple types of internet of things data integrated by the sensing data fusion module along with time, wherein the multi-element time sequence is represented by the following two-dimensional data:
wherein X is i,j Sample data representing the ith sensor at the jth time; the length of a processing time window of the sampling data is t, and the total number of the sensors of the Internet of things incorporated into the model is M. The cross-scale nature of the data X is embodied for different sensors i, X i,j The dimensions and ranges of (2) may be different; the timing of the data X is embodied for different sampling instants j, X i,j May be different and exhibit dynamic characteristics.
2) Data dimension reduction
In order to overcome the interference of irrelevant or secondary information in mass data of the Internet of things, importance ranking is carried out by using characteristic importance coefficients through a single-step prediction target fitting module and a cross-period prediction target fitting module, and the steps are as follows:
the single-step prediction target fitting model receives multiple time sequence data X as input, and k prediction targets in the next step (time t+1) are used as model learning tasks:
wherein a single target prediction task is learned when k=1, such as predicting only the amount of electricity used in the next period of time of the system; and learning a multi-target prediction task when k >1, and synchronously predicting the deviation of the system power consumption and the environment monitoring index from the ideal state.
Outputting importance coefficients of corresponding features of each dimension sensor after the single-step prediction target fitting model finishes training:
the cross-period prediction target fitting model focuses on system signals with periodic fluctuation, such as days, weeks, quarters and the like; receiving multiple time sequence data X as input, and taking the following k prediction targets after multiple periods (t+s time) as model learning tasks:
wherein a single target prediction task is learned when k=1, and a multi-target prediction task is learned when k > 1.
After the cross-period prediction target fitting model finishes training, outputting importance coefficients of corresponding features of each dimension of sensor:
And integrating the evaluation results of the single-step prediction target fitting model and the cross-period prediction target fitting model, and converting the feature importance corresponding to the 1 st to M th sensors into the feature importance:
in order to facilitate quantitative comparison standard, further normalization is carried out to obtain the importance coefficient of each dimension of final characteristics:
will v (1) ～v (M) In descending order, directly selecting top m dimension features according to system characteristics and actual demands or screening features by threshold value (e.g. retaining all importance coefficients)>0.05). The final retained feature dimension is noted as M, i.e., a data record of M sensors is retained from among the M sensors, where M < M.
3) Sequence-to-sequence modeling
M×t two-dimensional data retained after data dimension reduction
Seen as a sequence of column vectors
Wherein the method comprises the steps of
i＝1,2,...,t。
Will be serialized
As t inputs of the sequence to the sequence model, obtain +.>
R sequential output targets as sequence-to-sequence model according to +.>
A number of samples are constructed from different starting time points and model parameters within the neural network structure of the sequence-to-sequence model are learned by training.
4) System situation prediction
Two situations are distinguished: (A) If the predicted target is completely contained in the sequence-to-sequence model 
Within the output (i.e. the prediction target is the direct amount of sensor data within r future cycles), the prediction value is directly from +.>
And acquiring a target value of the sensor corresponding to the corresponding moment in the formed matrix. (B) Otherwise, the predicted target comes from->
Indirect integration of such known information (e.g., predicting the duration of time that the system is synthetically maintained in current mode operation), at which time the sequence is fetched into the implicit direction of the output layer within the sequence model structureThe amount is recorded as
A decoding model with a transducer mechanism is added downstream, and the decoding model is specifically expressed as follows:
the three key parameters Q, K, V of the transducer are calculated by the following formula:
wherein the method comprises the steps of
Is a model parameter; d in the above k The dimension of the vector after conversion for the multi-head attention mechanism of the transducer is as follows: hd k =m. The multi-head attention calculating mode is as follows:
selecting a nonlinear unit RELU or GELU according to the predicted target property, and finally predicting and outputting the system situation:
or (b)
Wherein W is Y Is a model parameter of the output layer.
Optimization objective (loss function): model output for each sample
True "predicted" values from samples derived from historical data +.>
The construction of the objective function is as follows: />
Furthermore, the Multi-Attribute Graph model (hereinafter, and in the drawings, the number MAG is used for fusing the correlation knowledge between the measurement points to further implement more accurate anomaly detection or automatically identify the data linkage relation which is not marked manually) related to the modeling feature fusion module. And taking the fixed-length window time sequence data acquired by each sensor as a multi-element input, and carrying out relation learning by using a correlation graph formed by existing measurement points to construct a characteristic representation model of each node in the correlation graph. The multi-element feature map model is as shown in fig. 4, and the identification task of the multi-element feature map model is completed by following steps:
1) Entity association stability determination
Compared with the internet with frequently changed links, the internet of things is generally more stable in intra-network linkage rules due to the fact that nodes of the internet identify entities in the physical world, such as production and manufacturing circulation in the process industry, topological connection of devices in intelligent buildings, spatial relation of sensors in a geographic information system and the like. To keep the comprehensiveness, the sampling time window t is checked first 0 ,t 1 ,...,t N ) Whether the internal Internet of things is internally provided with entity links or deleted is checked in a way of checking whether the key value pairs/the relation database/the graph database records are provided with link addition and deletion according to the actual storage form.
A. Such as time window t 0 ,t 1 ,...,t N ) No link is added and deleted, the linkage in the Internet of things is stable, and the step 2) is directly skipped;
B. such as time window t 0 ,t 1 ,...,t N ) With link additions or deletions therein, e.g. links varying ij For example, when e is within window period ij The occupied time of=0 (1) is e ij K times the time taken by =1 (0), and k is larger (e.g. k>2) Then consider e ij =0 (1) is [ t ] 0 ,t 1 ,...,t N ) Normal association within, jump to step 2), and e ij A sample corresponding to=1 (0) is taken as an abnormal sample identifiable by the model;
C. otherwise, the two conditions do not meet the requirement that the description does not accord with the characteristics of the Internet of things facing the patent, and the data sampling period t needs to be reselected i -t i-1 Or give up a time window [ t ] 0 ,t 1 ,...,t N ) All models within the model train samples.
2) Node feature encoding (hereinafter referred to as MAG-Encode in the figures) incorporating stable association information
The data of the Internet of things processed in the step 1) are organized into a static structure heterogeneous network in any sliding time window, and each node in the network comprises the time-varying characteristics.
In order to fuse the existing associated information and enrich the feature expression of the nodes, a TransH or TransR algorithm suitable for the knowledge graph is applied to the heterogeneous network in the window period. Considering that entities contained in the internet of things also have heterogeneity (such as different equipment types), splicing the one-hot coding of the characterization type and the acquired time sequence data to serve as original multi-element coding of the node i:
wherein the method comprises the steps of
Indicating that node i belongs to the kth type of device, +.>
Indicating that node i is in window period t 0 ,t 1 ,...,t N ) And acquiring time sequence data. Taking the node i as the original code, and obtaining the node i feature code after the association information is fused by using the TransH or the TransR:
wherein the method comprises the steps of
Representing the characteristic codes of the fused nodes i, and sampling the characteristic codes in a time window [ t ] 0 ,t 1 ,...,t N ) Is internally effective. The adoption of TransH or TransR identifies smaller measured errors algorithms based on known links in the heterogeneous network.
3) Multi-element characteristic diagram model structure
To be without loss of generality
And the feature vector of the ith entity obtained by projection/rotation of the TransH/TransR algorithm on the corresponding semantic space is represented. In order to expand relation judgment in a TransH or TransR algorithm, model input is defined as follows for anomaly detection of heterogeneous association between an entity i and an entity j in the Internet of things:
the design model structure is a multi-layer perceptron (hereinafter referred to as MAG-MLP in the figures) that outputs K dimensions:
wherein the method comprises the steps of
K represents the total number of association types. Outputting the corresponding r-th dimension element Y ij (r) →1 indicates that there is a high probability that the time window [ t ] is between entity i and entity j 0 ,t 1 ,...,t N ) The interior has the r-type linkage relation, otherwise Y ij And (r) →0 is considered to have no such linkage relationship.
To enhance model robustness in actual model construction, the method is used for matching in input dimension
Directly associatedAnd a regular term is additionally added to the model parameters, so that overfitting caused by overutilization of multidimensional time sequence information in a window period is prevented.
4) Training and application
The training samples of the model are from the relationship samples which are not sampled by the TransH or TransR algorithm in the relationship which is stably existed in the sub-step A of the step 1), and the samples which are not involved in the operation of the step 2) due to the relationship change in the sub-step B of the step 1) are taken as abnormal association, and are taken as negative sampling in the model training process when the abnormal association change which can be recognized by the model is taken as the abnormal association change.
Training samples are in the form of
Each dimension element Y in (3) ij (r) ∈ {0, 1. Based on such samples, neural network training is performed on the multi-layer perceptron MAG-MLP. The actual training process adopts a random gradient descent method and a batch gradient descent method to optimize model parameters; new samples are generated as the system is run and collected, and thus the model is typically regularly updated by training.
Application: after the model is put into use, according to the acquired multi-element data
Processing the model according to the steps described in this section to form the required input format, generating a determination of the K-dimensional relationship from the model>
For r=1, 2, …, K, when Y ij (r) Window period [ t ] with the actual System 0 ,t 1 ,...,t N ) When the linkage rule marks set in the system are inconsistent, the system operation is considered to have abnormal association between the entities i and j in the time.
Further, the implicit feature data related in the modeling feature fusion module is a feature vector which is obtained by splicing two vectors respectively extracted from the output layers of the multivariate time sequence prediction model MTS-S2S module and the multivariate feature map model MAG-MLP module, and the fusion processing result of the feature vector does not have explicit physical significance.
Further, the specific steps involved in the transaction module of the present invention include the following three steps:
a) Carrying out different relation association on data received from the modeling characteristic fusion module or data (data after sliding treatment of a time window) of the sensing data fusion module, and deducing a result of another variable by inputting different values of single or multiple variables to form a fuzzy rule;
b) Establishing a data classifier for the implicit features received from the modeling feature fusion module; the data acquired by the data acquisition module is mapped by a two-layer fusion module (namely a modeling feature fusion module and a sensing data fusion module) and finally transmitted to a link formed by the output of a data classifier as a rule formed by data mining;
c) On the basis of the mining rules of running data dynamic verification and expert approval, creating transaction-based ECA rules, namely ECA system cascade operation rules.
Furthermore, the transaction processing module establishes a transaction relation among all subsystems based on a new ECA (Event-Condition-Action) rule of the transaction, and forms a linkage control logic relation among all subsystems, so that the transaction processing module is used for safety monitoring, system situation assessment and inter-subsystem linkage operation decision.
The beneficial effects of the invention are as follows:
according to the cross-scale time sequence data fusion method and system for the Internet of things, data acquisition from all subsystems and sensors is completed, the data are filtered and uniformly stored, and through sensor data fusion, modeling feature fusion and ECA transaction processing rule establishment, correlation or dependency is extracted from a large amount of cross-scale data, so that data fusion sharing and logic linkage among all subsystems are realized.
Drawings
Fig. 1 is a cross-scale time sequence data fusion system architecture diagram for the internet of things, which is provided by the invention;
fig. 2 is a schematic diagram of cross-scale data acquisition for the internet of things according to the present invention;
FIG. 3 is a multivariate timing prediction Model (MTS) according to the present invention;
FIG. 4 is a diagram of a multivariate feature map Model (MAG) proposed by the present invention;
FIG. 5 is a diagram of a subsystem linkage architecture in accordance with the present invention;
fig. 6 is a flowchart of a modeling feature fusion implementation for a large intelligent building application scenario provided by the invention.
Detailed Description
The invention will be described in further detail with reference to the drawings and the specific embodiments. It is understood that the described embodiments are only a part of the present invention, but not all embodiments, and that based on the embodiments in the present invention, a person skilled in the art shall make corresponding changes and modifications according to the present invention, but all shall fall within the protection scope of the appended claims.
The invention will be described in connection with application in large intelligent buildings.
Referring to an architecture diagram of a cross-scale time sequence data fusion system oriented to the internet of things in fig. 1, the embodiment realizes the content of four large modules shown in fig. 1, completes cross-scale data acquisition of a large intelligent building subsystem, performs time window sliding processing on the data, adopts a modeling feature fusion technology, realizes transaction processing of real-time rule reasoning, and completes system operation decision.
The invention discloses a trans-scale time sequence data fusion method oriented to the Internet of things, which specifically comprises the following steps:
and step 1, cross-scale time sequence data acquisition. As shown in fig. 2, in this case, data of more than 30 subsystems such as an intelligent lighting system, a power monitoring system, a fire alarm system, a pneumatic window system, etc. are included, and the protocol types include KNX, OPC, modbus, BACnet, etc. The data acquisition program analyzes the data transmission protocol of the subsystem terminal equipment, unifies the cross-scale time sequence data, and completes cross-node acquisition of the cross-scale data. And meanwhile, the data acquisition program performs denoising and filtering on the collected data, then sends the data to the safe and fast real-time database, and uniformly manages all subsystem data to form a data management platform.
And 2, fusing the sensing data. In the embodiment, the fusion of the sensing data is to preprocess the data collected in the step 1, namely, to convert the time series data of a cross scale into time series data with equal interval by algebraic interpolation, machine learning fitting and other methods; integrating the map level associated data of the Internet of things equipment by dynamically extracting data in a set time window in real time; integrating the data nodes through ECA rules, completing the first data fusion of the acquired 30 subsystem data, and outputting the multivariate time sequence data and the multivariate map data to a modeling feature fusion module.
And 3, modeling feature fusion. In the case, a multivariate time sequence prediction model (shown in figure 3), a multivariate feature graph model (shown in figure 4) and the like are used for researching a modeling feature fusion technology aiming at the reasoning decision requirement applicable to intelligent operation regulation of a large building, and finally high-quality knowledge characterization and reasoning rules applicable to the operation regulation in the building are generated in a converging mode.
Aiming at the characteristic of multi-system and multi-equipment operation dynamic representation time-space association coupling in a large intelligent building, a modeling feature fusion implementation scheme (shown in fig. 6) for jointly deploying a multi-variable time sequence prediction model MTS and a multi-feature graph model MAG is designed in the scheme, namely the MTS and the MAG keep respective complete mechanisms in the scheme, but are not independently applied any more, and the extracted intermediate features are multiplexed with each other. In general, five of the branches 1.1-2.2 in FIG. 6 cover the sum of MTS and MAG modeled feature extraction mechanisms. The following stages illustrate the implementation of MTS, MAG and directed extraction tasks in modeling feature fusion.
1) In the first stage, the MTS-dominated multi-channel signal feature extraction step, in this case, finds new feature combinations that may exist. In order to overcome the interference of irrelevant or secondary information in mass data of the Internet of things, importance ranking is carried out by characteristic importance coefficients through a single-step prediction target fitting module and a cross-period prediction target fitting module (branch 1.1 of fig. 6), and the specific steps are as described in the multivariate time sequence prediction modules MTS-R1 and MTS-R2. In addition, through visual analysis of real data, a large number of equipment measurement time sequence signals in the system have multi-scale cycle approximation (such as mixed fluctuation of a day cycle, a season cycle and even a potential year cycle of an air conditioning system highly related to outdoor weather, the unit in an automatic control mode shows reciprocating signal oscillation between a high setting threshold and a low setting threshold, and the intelligent lighting system shows data fluctuation with the day cycle. Based on the phenomenon, the information combination of the time domain channel and the frequency domain channel is designed in the case to extract and transform the system characteristics rich in information.
Specifically, in the present case, a time domain feature extraction channel (branch 1.2 of fig. 6) is adopted, on the basis of retaining time sequence signals in each dimension of the system, performance improvement brought by a convolution unit in a convolution neural network for a downstream recognition task is referred to, and a one-dimensional window convolution (also called 1-Dimensional Convolution or 1×d Conv) unit is added in a data model in the present case to extract time domain features in a local time slice; according to the different convolution parameters actually set, the potential effects also include the functions of local time domain signal smoothing (overcoming local high-frequency interference) and measuring and calculating the coincidence degree of relative reference signal fragments (shape).
Specifically, in this case, a frequency domain feature extraction channel (branch 1.3 of fig. 6) is also adopted, and since the time sequence signal monitored by the device in the large building does not show the periodicity of a single scale, the multi-scale frequency domain feature shown by the multi-dimensional data needs to keep the data feature of a wider frequency spectrum range. Based on the requirement, a branch channel for frequency domain feature extraction is designed in the scheme, frequency domain conversion is respectively carried out on time sequence data in a window period by using discrete Fourier transform, and besides filtering each single-dimensional signal frequency band abnormality which obviously exceeds a threshold value, the signal quantity in the frequency spectrum of other windows is reserved; further imitates the time domain channel design, adding a 1 x D convolution unit to identify the local frequency domain signal pattern.
The output features of the final first stage are jointly composed of the automatically estimated m-dimensional important original features, the time sequence convolution features and the frequency domain convolution features.
2) And in the second stage, the multi-element characteristic diagram model association extraction flow and the MTS-S2S built-in time sequence coding flow are implemented in parallel. In the associated feature coding flow (branch 2.1 of fig. 6) dominated by the multivariate feature map model MAG, the entity associated stability decision content for a specific building application case includes: element associations that may come from programmatic extraction in Building Information Models (BIM) or from manual arrangement/rule-defined equipment associations, including topological associations between equipment devices or collection points in a building (e.g., pipeline connections, heat exchange channels, power monitoring total-branch tables, etc.), location associations located in adjacent or the same space, logical associations in system operation dynamics; the latter fully reflects the necessity of dynamic determination of entity association stability in MAG along with time window, for example, indoor illumination automatic regulation and control in non-working period transits from strong association with outdoor weather condition to strong association of human body detection sensing in office area, and cold and heat source operation of central air conditioner unit is controlled by indoor CO 2 Content-dynamic switching of associated characteristics transmitted by multiple modes of fresh air switch regulation, and the like.
Specifically, the raw multi-element encoding input received in this column for MAG-Encode specific flow dominated by the multi-element feature map model in Branch 2.1 of FIG. 6
Each of which
The integrated information of m-dimensional important original features, time sequence convolution features and frequency domain convolution features extracted by MTS dominance in multiplexing branches 1.1-1.3 is adjusted, so that MAG utilizes system dynamic features from time-frequency domain outside a model independent mechanism. The MAG-Encode module converts the original code into node code +.>
Then, in order to extract core features from the code for the task facing the subsequent orientation (such as multidimensional time sequence prediction and abnormal data error correction), a module adopting a graph attention (Graph Attention Network, GAT) mechanism is connected in series in the branch 2.1, and the capability of directionally conveying important information to the next stage is adaptively improved by automatically quantifying the importance degree of each association relation between nodes.
In parallel with this, the flow shown by branch 2.2 in fig. 6 focuses on integrating information in terms of timing and encoding. The process is governed by MTS, which aligns the original codes received from the first stage in the time-sequential direction
Encoding the equally spaced multivariate time series into a low dimensional embedded vector via an MTS-S2S built-in encoding module 
In this case, according to the characteristics that the directional tasks in the next stage are different in emphasis, the MTS-S2S built-in coding module selects a coding module with a time sequence Self-Attention (Temporal Self-Attention) mechanism, and adaptively transmits important information in the time sequence direction by capturing the importance degree of the time sequence relationship in the time window.
Based on the implementation steps, the output characteristics of the stage are converted from the high-dimensional mixed information of the first stage into low-dimensional and dense vector characteristics, and meanwhile, the important information of the multi-element association relation and the time sequence correlation is contained.
3) And thirdly, model feature fusion for the directional task. Aiming at the characteristics that multidimensional data in a large building data management platform has time correlation and obvious inter-characteristic correlation (such as the regulation of an air conditioning system in a building, the operation of a weather station system associated with a pneumatic window system and the like) due to the fact that fresh air temperature from outdoors, two types of oriented tasks are designed to serve as evaluation references for joint training of MTS and MAG models, namely an abnormal data error correction task and a multidimensional time sequence data prediction task (shown in a third stage of fig. 6); in the stage, the gating units are placed on the respective task branches to avoid coupling interference among characteristic components, model training optimization is carried out according to a loss function jointly formulated by two types of tasks, and the problem that the overall process of model characteristic fusion falls into a local optimal solution is avoided.
In the two types of directional tasks, the error correction of abnormal data not only comprises the recovery of the original correct distribution of the collected data, but also comprises the identification of abnormal characteristic association among the data, and can be regarded as a pre-training basic task related to the system abnormal detection in the subsequent transaction processing; the multi-dimensional time sequence data prediction task is divided into short-term and long-term data prediction according to the application requirements of building operation and maintenance management, and in the step 3, the fusion result of the MTS-R1 module and the MTS-R2 module can be effectively responded, and the multi-dimensional time sequence data prediction task can also be regarded as a pre-training basic task related to system situation analysis and prediction in subsequent transaction processing. The modeled feature fusion case of the present embodiment can thus provide strong support for the transaction processing of the system (step 4).
And 4, transaction processing. As shown in fig. 5, the present case uses a dynamic linkage scheme iteration and rule reasoning engine technology, and combines with an interactive customization scene to implement intelligent diagnosis in different scenes such as normal operation, emergency conditions, outdoor disasters, and the like, so as to complete an interpretable intelligent decision. Transaction processing between a meteorological system, a fresh air system and an air conditioning system: acquiring the air quality value, the indoor temperature value, the fresh air system parameter and the air conditioning system parameter in the meteorological system in real time; when the air quality value is detected to be not up to standard, starting a fresh air system, and automatically adjusting the running mode of the fresh air system according to different pollution programs; and judging whether to start the air conditioning system according to the monitored indoor temperature. And automatically stopping the fresh air system when the air quality reaches the standard, and stopping the air conditioning system after the temperature reaches the set temperature.
Another embodiment of the present invention provides a cross-scale time series data fusion system facing the internet of things, which includes:
the data acquisition module is used for acquiring cross-scale time sequence data from various subsystems and various types of sensors, denoising and filtering the acquired cross-scale time sequence data and then sending the acquired cross-scale time sequence data to the sensing data fusion module;
the sensing data fusion module is used for performing time window sliding treatment on the cross-scale time sequence data received from the data acquisition module, and the treated result comprises two types of data of equal interval multivariable time sequence and multielement association relation and is provided for the modeling feature fusion module through an internal interface;
the modeling feature fusion module is used for carrying out modeling feature fusion on the two types of data of the equal interval multivariable time sequence and the multielement association relationship processed by the sensing data fusion module, and synchronously, the modeling feature fusion module dynamically loads parameters along with the transaction processing period and operates, and provides fused implicit feature data for the transaction processing module through an internal interface;
the transaction processing module is used for processing the implicit characteristic data provided by the modeling characteristic fusion module according to specific steps and then applying the implicit characteristic data to safety monitoring, system situation assessment and inter-subsystem linkage operation decision.
Another embodiment of the invention provides a computer device (computer, server, smart phone, etc.) comprising a memory storing a computer program configured to be executed by the processor and a processor, the computer program comprising instructions for performing the steps of the method of the invention.
Another embodiment of the invention provides a computer readable storage medium (e.g., ROM/RAM, magnetic disk, optical disk) storing a computer program which, when executed by a computer, performs the steps of the method of the invention.
Claims (10)
1. The cross-scale time sequence data fusion method for the Internet of things is characterized by comprising the following steps of:
acquiring cross-scale time sequence data from multiple subsystems and multiple types of sensors;
performing time window sliding treatment on the cross-scale time sequence data, wherein the treated result comprises two types of data of equal interval multivariable time sequence and multielement association relation;
modeling feature fusion is carried out on two types of data of the multivariate time sequence and the multivariate association relation at equal intervals, and parameters are dynamically loaded and operated along with the transaction processing period synchronously, so that fused implicit feature data are obtained;
And processing the fused implicit characteristic data according to specific steps, and applying the processed implicit characteristic data to safety monitoring, system situation assessment and inter-subsystem linkage operation decision.
2. The method of claim 1, wherein the cross-scale timing data comprises one or more of the following: data with different time granularity, data with different communication modes, data with different communication protocols and data with different data types.
3. The method of claim 1, wherein the time window sliding process employs a time window to achieve sensory data fusion in combination with:
a) Converting the trans-scale time sequence data into an equidistant multivariable time sequence by algebraic interpolation, machine learning fitting, spline function or numerical integration equivalent method;
b) And dynamically extracting the multi-element association relation among the data in the set time window along with the running process of the system.
4. The method of claim 1, wherein the multivariate association relationship comprises a causal relationship between the cross-scale temporal data, and other relationships between the cross-scale temporal data established using a graph model, the other relationships comprising a spatial relationship, a device topology relationship.
5. The method of claim 1, wherein the modeling feature fusion comprises:
a) Taking historical data of the equal-interval multivariable time sequence as a reference sample, taking current and subsequent time sequence data as learning targets, and establishing a multivariable time sequence prediction model;
b) Taking the attribute of the measured entity in the Internet of things as a node, taking the multi-element association relationship as an edge, establishing a multi-element feature graph model, and mining potential association relationship or association relationship change;
c) Model parameters of the multivariate time sequence prediction model and the multivariate feature map model are saved as readable data files and updated regularly through a programmed automatic optimization solver.
6. The method of claim 5, wherein the multivariate timing prediction model is an integrated model formed from a single-step prediction target fitting model, a cross-cycle prediction target fitting model, and a sequence-to-sequence model, and learning the prediction tasks of the system situation following the steps of:
1) And (3) data acquisition:
receiving a multi-element time sequence formed by the collection of multiple types of internet of things data along with time, wherein the multi-element time sequence is represented by the following two-dimensional data:
wherein X is i, Sample data representing the ith sensor at the jth time; the length of a processing time window of the sampling data is t, the total number of sensors of the Internet of things incorporated into the model is M, and the cross-scale property of the data X is reflected in different sensors i and X i, The dimensions and ranges of (2) may be different; the timing of the data X is embodied for different sampling instants j, X i, May be different, exhibit dynamic characteristics;
2) Data dimension reduction:
and ranking importance by the two modules of the single-step prediction target fitting model and the cross-period prediction target fitting model according to the feature importance coefficients, wherein the steps are as follows:
the single-step prediction target fitting model receives multiple time sequence data X as input, and k prediction targets in the next step, namely 't+1 moment', are used as model learning tasks:
wherein the single target prediction task is learned when k=1; learning a multi-objective prediction task when k > 1;
and outputting importance coefficients of corresponding features of each dimension sensor after the single-step prediction target fitting model is trained:
the cross-period prediction target fitting model focuses on a system signal with periodic fluctuation, receives multi-element time sequence data X as input, and takes the following k prediction targets of 't+s moment' after multiple periods as model learning tasks:
wherein a single-target prediction task is learned when k=1, and a multi-target prediction task is learned when k > 1;
after the cross-period prediction target fitting model is trained, the importance coefficients of the corresponding characteristics of each dimension of sensor are output:
And integrating the evaluation results of the single-step prediction target fitting model and the cross-period prediction target fitting model, and converting the feature importance corresponding to the 1 st to M th sensors into the feature importance corresponding to the 1 st to M th sensors:
normalizing to obtain importance coefficients of final features in all dimensions:
will v (1) ～v (M) In descending order, directly selecting top M dimension features according to the system characteristics and actual demands or screening features according to a threshold value, and recording the finally reserved feature dimension as M, namely reserving data records of M sensors from M sensors, wherein M is less than M;
3) Sequence-to-sequence modeling:
m×t two-dimensional data which is reserved after the data is subjected to dimension reduction:
seen as a sequence of column vectors
Wherein the method comprises the steps of
Will be serialized
As t inputs of the sequence to the sequence model, from historical data samples
R sequential output targets as the sequence-to-sequence model according to +.>
Selecting different construction large numbers of samples at starting time points, and learning model parameters in a neural network structure of the sequence-to-sequence model through end-to-end training;
4) Predicting system situation:
two situations are distinguished: (A) If the predicted target is completely contained in the sequence-to-sequence model
In the output, the predicted value is directly from +. >
Acquiring a target value of a sensor corresponding to the corresponding moment in a formed matrix; (B) Otherwise, the predicted target comes from->
Indirect integration of known information, in which case the hidden vectors of the sequence into the output layer within the sequence model structure are taken out, denoted +.>
A decoding model with a transducer mechanism is added downstream, and the decoding model is specifically expressed as follows:
the three key parameters Q, K, V of the transducer are calculated by the following formula:
wherein the method comprises the steps of
For model parameters d k The dimension of the vector after conversion for the multi-head attention mechanism of the transducer is as follows: hd k =m; multi-head attention calculation mode:
selecting a nonlinear unit RELU or GELU according to the predicted target property, and finally predicting and outputting the system situation:
or (b)
Wherein W is Y For the model parameters of the output layer,
optimization target: model output for each sample
True "predicted" values from samples derived from historical data +.>
The construction of the objective function is as follows:
7. the method of claim 5, wherein the multivariate feature map model is used for fusing the correlation knowledge among the measurement points to implement more accurate anomaly detection or automatically identifying the data linkage relation which is not marked manually, taking time window time sequence data acquired by each sensor as multivariate input, performing relation learning by using the correlation map formed by the existing measurement points, and constructing a feature representation model of each node in the correlation map; the identification task of the multi-element characteristic diagram model is completed by following steps:
1) Entity association stability determination:
first check the sampling time window t 0 ,t 1 ,...,t N ) Whether entity links or deletes are arranged in the Internet of things or not is checked by checking whether the key value pairs/the relational database/the graph database record are in the record according to the actual storage formThe links are added and deleted;
A. such as time window t 0 ,t 1 ,...,t N ) The link-free addition and deletion in the Internet of things show that the linkage in the Internet of things is stable, and the steps are directly skipped
2)；
B. Such as time window t 0 ,t 1 ,...,t N ) With link additions or deletions therein, e.g. links varying ij For example, when e is within window period ij The occupied time of=0 (1) is e ij When k times the time taken by=1 (0) and the k value is large, it is considered that e ij =0 (1) is [ t ] 0 ,t 1 ,...,t N ) Normal association within, jump to step 2), and e ij A sample corresponding to=1 (0) is taken as an abnormal sample identifiable by the model;
C. otherwise, the two conditions are not satisfied, and the description is not in accordance with the characteristics of the internet of things facing the patent, and the data sampling period t needs to be reselected i -t i-1 Or give up a time window [ t ] 0 ,t 1 ,...,t N ) Training samples by all models in the model;
2) Node feature coding of fusion stability association information:
considering that entities contained in the Internet of things also have heterogeneity, splicing the one-hot coding of the characterization type and the acquired time sequence data to serve as original multi-element coding of the node i:
Wherein the method comprises the steps of
Indicating that node i belongs to the kth type of device, +.>
Indicating that node i is in window period t 0 ,t 1 ,...,t N ) Collected time sequence data; taking the node i as the original code, and obtaining the node i feature code after the association information is fused by using the TransH or the TransR:
wherein the method comprises the steps of
Representing the characteristic codes of the fused nodes i, and sampling the characteristic codes in a time window [ t ] 0 ,t 1 ,...,t N ) The inner effect is achieved;
the algorithm with smaller actual measurement error is identified according to the known links in the heterogeneous network by adopting the TransH or the TransR;
3) Multi-element feature map model structure:
to be used for
Representing the feature vector of the ith entity obtained by projection/rotation of the TransH/TransR algorithm on the corresponding semantic space; in order to expand relation judgment in a TransH or TransR algorithm, model input is defined as follows for anomaly detection of heterogeneous association between an entity i and an entity j in the Internet of things:
the design model structure is a multi-layer perceptron for outputting K dimensions:
wherein the method comprises the steps of
K represents the total number of association types; outputting the corresponding r-th dimension element Y ij (r) →1 indicates that there is a high probability that the time window [ t ] is between entity i and entity j 0 ,t 1 ,...,t N ) The interior has the r-type linkage relation, otherwiseY ij (r) →0 then no such linkage relationship is considered;
for AND in input dimension
The regular term is additionally added to the directly-related model parameters, so that overfitting caused by overutilization of multi-dimensional time sequence information in a window period is prevented;
4) Training and application:
the training samples of the model are from the relation samples which are not sampled by the TransH or TransR algorithm in the relation which is stably existing in the A of the step 1) and the samples which are not involved in the operation of the step 2) due to the relation change in the B of the step 1) and are not involved in the operation of the step 2) because of the abnormal relation, and are regarded as negative sampling in the model training process when being used as abnormal relation change which can be recognized by the model,
training samples are in the form of
Each dimension element Y in (3) ij (r) E {0,1}, training the neural network of the multi-layer perceptron based on the samples, and optimizing model parameters by adopting a random gradient descent method and a batch gradient descent method in the actual training process; generating new samples as the system is running, so the model is typically regularly updated by training;
after the model is put into use, the acquired multi-metadata is acquired
Processing the K-dimensional relationship into an input format required by the model, and generating the K-dimensional relationship by the model to determine +.>
For r=1, 2, …, K, when Y ij (r) Window period [ t ] with the actual System 0 ,t 1 ,...,t N ) When the inter-set linkage rule marks are inconsistent, that is, the time is consideredInter-system operation occurs an association anomaly between entities i and j.
8. The method of claim 1, wherein the specific step comprises:
a) Carrying out different relation association on the data after the modeling feature fusion or the data after the time window sliding treatment, and deducing the result of another variable by inputting different values of single or multiple variables to form a fuzzy rule;
b) Establishing a data classifier for the implicit features after the modeling features are fused; the acquired data is finally transmitted to a link formed by the output of a data classifier through the mapping of the two layers of fusion modules, and the link is used as a rule formed by data mining;
c) Based on the mining rules of running data dynamic verification and expert approval, creating the transaction-based ECA rules.
9. The method of claim 8, wherein a transaction relationship between subsystems is established using the transaction-based ECA rules to form a coordinated control logic relationship between subsystems for security monitoring, system situational assessment, and inter-subsystem coordinated operation decisions.
10. The trans-scale time sequence data fusion system for the Internet of things is characterized by comprising:
the data acquisition module is used for acquiring cross-scale time sequence data from various subsystems and various types of sensors, denoising and filtering the acquired cross-scale time sequence data and then sending the acquired cross-scale time sequence data to the sensing data fusion module;
The sensing data fusion module is used for performing time window sliding treatment on the cross-scale time sequence data received from the data acquisition module, and the treated result comprises two types of data of equal interval multivariable time sequence and multielement association relation and is provided for the modeling feature fusion module through an internal interface;
the modeling feature fusion module is used for carrying out modeling feature fusion on the two types of data of the equal interval multivariable time sequence and the multielement association relationship processed by the sensing data fusion module, and synchronously, the modeling feature fusion module dynamically loads parameters along with the transaction processing period and operates, and provides fused implicit feature data for the transaction processing module through an internal interface;
the transaction processing module is used for processing the implicit characteristic data provided by the modeling characteristic fusion module according to specific steps and then applying the implicit characteristic data to safety monitoring, system situation assessment and inter-subsystem linkage operation decision.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310077575.2A CN116049764A (en) | 2023-01-16 | 2023-01-16 | Cross-scale time sequence data fusion method and system for Internet of things |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310077575.2A CN116049764A (en) | 2023-01-16 | 2023-01-16 | Cross-scale time sequence data fusion method and system for Internet of things |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116049764A true CN116049764A (en) | 2023-05-02 |
Family
ID=86131215
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310077575.2A Pending CN116049764A (en) | 2023-01-16 | 2023-01-16 | Cross-scale time sequence data fusion method and system for Internet of things |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116049764A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117011092A (en) * | 2023-09-28 | 2023-11-07 | 武昌理工学院 | Intelligent building equipment management monitoring system and method |
| CN117251736A (en) * | 2023-11-13 | 2023-12-19 | 杭州海康威视数字技术股份有限公司 | Internet of things data aggregation method and device based on neural network and space-time correlation degree |
| CN118195361A (en) * | 2024-05-17 | 2024-06-14 | 国网吉林省电力有限公司经济技术研究院 | Big data-based energy management method and system |
-
2023
- 2023-01-16 CN CN202310077575.2A patent/CN116049764A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117011092A (en) * | 2023-09-28 | 2023-11-07 | 武昌理工学院 | Intelligent building equipment management monitoring system and method |
| CN117011092B (en) * | 2023-09-28 | 2023-12-19 | 武昌理工学院 | Intelligent building equipment management monitoring system and method |
| CN117251736A (en) * | 2023-11-13 | 2023-12-19 | 杭州海康威视数字技术股份有限公司 | Internet of things data aggregation method and device based on neural network and space-time correlation degree |
| CN118195361A (en) * | 2024-05-17 | 2024-06-14 | 国网吉林省电力有限公司经济技术研究院 | Big data-based energy management method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN116049764A (en) | Cross-scale time sequence data fusion method and system for Internet of things | |
| CN110647900B (en) | Intelligent safety situation prediction method, device and system based on deep neural network | |
| CN112101480B (en) | Multivariate clustering and fused time sequence combined prediction method | |
| CN116757534B (en) | Intelligent refrigerator reliability analysis method based on neural training network | |
| CN114282443B (en) | Residual service life prediction method based on MLP-LSTM supervised joint model | |
| CN116562514B (en) | Method and system for immediately analyzing production conditions of enterprises based on neural network | |
| CN116739562A (en) | LSTM-based power distribution network stability operation and maintenance method | |
| CN116557787B (en) | Intelligent evaluation system and method for pipe network state | |
| CN115953900A (en) | Traffic flow prediction method based on multi-dimensional time and space dependence mining | |
| Liu et al. | Model fusion and multiscale feature learning for fault diagnosis of industrial processes | |
| CN113036913A (en) | Method and device for monitoring state of comprehensive energy equipment | |
| CN115561416A (en) | Method and device for detecting inlet water quality of sewage treatment plant in real time | |
| CN117435997B (en) | Elevator early warning and on-demand maintenance method based on transducer and time sequence compression mechanism | |
| Deng et al. | Toward smart multizone HVAC control by combining context-aware system and deep reinforcement learning | |
| CN115128978A (en) | Internet of things environment big data detection and intelligent monitoring system | |
| CN114444811A (en) | Aluminum electrolysis mixing data superheat degree prediction method based on attention mechanism | |
| CN117116045A (en) | Traffic flow prediction method and device based on space-time sequence deep learning | |
| CN114912169B (en) | Industrial building heat supply autonomous optimization regulation and control method based on multisource information fusion | |
| CN114399024B (en) | Oil gas concentration big data intelligent detection system | |
| CN114386672B (en) | Environment big data Internet of things intelligent detection system | |
| Boulmaiz et al. | Optimizing occupant actions to enhance his comfort while reducing energy demand in buildings | |
| CN110009132A (en) | A kind of short-term electric load fining prediction technique based on LSTM deep neural network | |
| CN112862213B (en) | Heat supply demand prediction method, system and equipment based on periodic feedback LSTM | |
| CN115062764A (en) | Big data system of illuminance intelligent regulation and environmental parameter thing networking | |
| CN112906856A (en) | Livestock and poultry house ambient temperature detection system based on cloud platform |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |