CN116975755A - Noise-tag-scene-oriented stepping robust semi-supervised gas turbine anomaly detection method and system - Google Patents

Noise-tag-scene-oriented stepping robust semi-supervised gas turbine anomaly detection method and system Download PDF

Info

Publication number
CN116975755A
CN116975755A CN202310656040.0A CN202310656040A CN116975755A CN 116975755 A CN116975755 A CN 116975755A CN 202310656040 A CN202310656040 A CN 202310656040A CN 116975755 A CN116975755 A CN 116975755A
Authority
CN
China
Prior art keywords
sample
classifier
complementary
noise
bangji
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
Application number
CN202310656040.0A
Other languages
Chinese (zh)
Inventor
叶建君
沈晓兵
车斌
滕昊真
孙健
张铁汉
赵春晖
杨佳阳
陈旭
宋鹏宇
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xiaoshan Power Plant Of Zhejiang Zhengneng Electric Power Co ltd
Zhejiang University ZJU
Original Assignee
Xiaoshan Power Plant Of Zhejiang Zhengneng Electric Power Co ltd
Zhejiang University ZJU
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
Application filed by Xiaoshan Power Plant Of Zhejiang Zhengneng Electric Power Co ltd, Zhejiang University ZJU filed Critical Xiaoshan Power Plant Of Zhejiang Zhengneng Electric Power Co ltd
Priority to CN202310656040.0A priority Critical patent/CN116975755A/en
Publication of CN116975755A publication Critical patent/CN116975755A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/2431—Multiple classes
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01D—MEASURING NOT SPECIALLY ADAPTED FOR A SPECIFIC VARIABLE; ARRANGEMENTS FOR MEASURING TWO OR MORE VARIABLES NOT COVERED IN A SINGLE OTHER SUBCLASS; TARIFF METERING APPARATUS; MEASURING OR TESTING NOT OTHERWISE PROVIDED FOR
    • G01D21/00—Measuring or testing not otherwise provided for
    • G01D21/02—Measuring two or more variables by means not covered by a single other subclass
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/10—Pre-processing; Data cleansing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • 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
    • 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
    • G06N3/0895—Weakly supervised learning, e.g. semi-supervised or self-supervised learning

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a noise label scene-oriented stepping robust semi-supervised gas turbine anomaly detection method and system. The invention designs an arctangent support vector machine model and an average absolute error neural network model based on a symmetry loss theory to construct a complementary Lu Bangji classifier, and has stronger robust characteristic on noise labels; secondly, a complementary noise filtering mechanism is provided, noise labels are filtered in a stepping mode, a complementary Lu Bangji classifier is updated, and the risk of overfitting on the noise labels is further reduced; in addition, the invention also designs a self-adaptive data supplementing strategy, and the problem of rare data labels of the gas turbine is overcome by utilizing a high-confidence pseudo label sample to realize the expansion of a labeled data set. The invention can effectively complete the task of detecting the abnormality of the operation of the gas turbine under the condition that the sample labels are rare and the label noise exists, and provides a new method for detecting the abnormality and safely operating and maintaining the gas turbine generator set.

Description

Noise-tag-scene-oriented stepping robust semi-supervised gas turbine anomaly detection method and system
Technical Field
The invention discloses a noise-tag-scene-oriented stepping robust semi-supervised gas turbine anomaly detection method and system. The invention belongs to the field of industrial anomaly detection, and particularly relates to anomaly detection of industrial process data with few labels and noise on labels.
Background
The gas turbine generator set is typical large-scale industrial equipment and has the characteristics of huge scale, difficult manual inspection, frequent change of operation load and the like. With the acceleration of modern industry to high-end, complex and intelligent, it becomes more difficult to realize efficient operation and maintenance of the gas turbine generator set. Anomaly detection is an important task in the operation and maintenance of a gas turbine, and specifically includes evaluating the state of the operation process of the gas turbine, and determining whether there is an anomaly in the equipment such as the gas turbine, steam turbine, and generator. Reliable anomaly detection is of great importance to ensure the safety of the power generation process of the combustion engine. Therefore, how to detect the abnormality of the gas turbine generator set timely and accurately in the running process of the gas turbine generator set is always a hot spot for technical staff to study.
With the advent of internet of things (Internet of Things, ioT) technology and artificial intelligence (Artificial Intelligence, AI), one can conveniently store large amounts of industrial process data, explore more efficient information from these data, and ultimately understand the underlying operating rules of an industrial process. Over the past three decades, many pattern recognition methods have been applied to anomaly detection tasks in industrial processes and have achieved satisfactory results. Most of the existing anomaly detection methods based on pattern recognition use some supervised learning algorithms to build classification models, and effective information indicating potential operation behaviors can be extracted from industrial process samples and labels by using the classification models, so that anomaly recognition is finally completed. However, the above method requires a large number of labeled samples as support, which has a significant limitation in practical applications. In the actual operation process of the gas engine, the unlabeled process data are relatively easy to obtain, the operation states of thousands of process data are judged one by one, and the time-consuming and expensive process of obtaining sufficient and clear indication of the equipment operation state data is realized. The lack of the label sample severely limits the application performance of the deep supervised learning model in the abnormal detection task of the gas turbine. In addition, some data may be marked incorrectly, i.e. carry noisy labels, during manual marking of the engine data due to lack of expert knowledge, occasional mistakes by the operator, etc. Such inaccurate label information may lead to erroneous guidance for training of the model, eventually resulting in reduced accuracy of anomaly detection. In the actual abnormal detection task of the operation of the combustion engine, the two problems often occur simultaneously, and a great challenge is brought to the traditional supervision abnormal detection.
Disclosure of Invention
The invention aims to design a stepping iteration complementary noise filtering strategy and further establish an anomaly detection method aiming at the problems that the data labels of a gas turbine are rare and noise exists. Firstly, an arctangent support vector machine model and an average absolute error neural network model are designed based on a symmetry loss theory to construct a complementary Lu Bangji classifier, and the complementary Lu Bangji classifier has strong robust characteristics on noise labels; secondly, a complementary noise filtering mechanism is provided, noise labels are filtered in a stepping mode, a complementary Lu Bangji classifier is updated, and the risk of overfitting of the system to the noise labels is further reduced; in addition, the invention designs a self-adaptive data supplementing strategy, and the problem of rare data labels of the fuel engine is overcome by utilizing a high-confidence pseudo label sample to realize the expansion of a labeled data set.
The aim of the invention is realized by the following technical scheme:
a step robust semi-supervised gas turbine anomaly detection method oriented to a noise label scene comprises the steps of obtaining gas turbine operation data, and inputting the gas turbine operation data into a trained complementary Lu Bangji classifier to obtain an anomaly detection result; wherein the complementary Lu Bangji classifier is obtained by training the following steps:
step 1: constructing a training set based on the collected combustion engine operation data, wherein the training set comprises a labeled training set and a non-labeled training set;
step 2: constructing a complementary Lu Bangji classifier, wherein the complementary Lu Bangji classifier model comprises a parallel arctangent support vector machine model and an average absolute error neural network model; the loss function adopted in the training of the arctangent support vector machine model is as follows:
wherein w is 1 And b 1 Representing parameters to be optimized of the arctangent support vector machine model, wherein xi represents a penalty coefficient, L represents the number of labeled samples, and L is L 2 Representing the square of the two norms,label for the ith labeled sample, +.>The fuel engine operation data of the ith tagged sample; the loss function adopted in the training of the average absolute error neural network model is the average absolute error; training the complementary Lu Bangji classifier by using the labeled training set to obtain a first complementary Lu Bangji classifier;
step 3: predicting each sample of the labeled training set by using a first complementary Lu Bangji classifier, judging whether the corresponding sample is a noise label or not according to a risk evaluation value of a prediction result, and filtering samples of the noise label; retraining the complementary Lu Bangji classifier with the noise-removed labeled training set to obtain a second complementary Lu Bangji classifier;
step 4: randomly selecting samples in the non-tag training set to form a non-tag subset, predicting each sample in the non-tag subset by using a second complementary Lu Bangji classifier to obtain corresponding pseudo tags, sorting according to the confidence of the pseudo tags, selecting one or more samples with the maximum confidence, deleting the samples from the non-tag training set, and forming a tagged sample with the corresponding pseudo tags, and adding the tagged sample into the tagged training set;
and (3) alternately performing complementary noise label filtering and self-adaptive data supplementing according to the steps 2-4 until the set number of rounds T is reached, and obtaining the final trained complementary Lu Bangji classifier.
Further, the collected fuel engine operation data is normalized fuel engine operation data.
Further, in the step 3, each sample in the labeled training set is predicted by using a first complementary Lu Bangji classifier, and whether the corresponding sample is a noise label or not is judged according to a risk evaluation value of a prediction result, and the sample of the noise label is filtered, specifically:
predicting each sample of the labeled training set by using a first complementary Lu Bangji classifier, and calculating a risk evaluation value of a prediction result; the risk evaluation value of the prediction result calculated based on the arctangent support vector machine model is a negative value of the product of the prediction label of each sample corresponding to the arctangent support vector machine model and the label carried by the sample; the risk evaluation value of the prediction result calculated based on the average absolute error neural network model is the absolute value of the label difference value carried by the sample after the average absolute error neural network model is in linear transformation corresponding to the prediction label of each sample and is mapped to the [ -1,1] interval;
according to the risk evaluation value of the prediction result, all the labeled samples are divided into 4 categories, including:
secure sample L c : the risk evaluation value of the prediction result calculated based on the arctangent support vector machine model does not exceed a first threshold mu 1 And the risk evaluation value of the prediction result calculated based on the mean absolute error neural network model does not exceed the second threshold mu 2 Is a sample of (2);
risk sample L d1 : risk evaluation value of prediction result calculated based on arctangent support vector machine model exceeds first threshold mu 1 While the risk evaluation value of the prediction result calculated based on the mean absolute error neural network model does not exceed the second threshold μ 2 Is a sample of (2);
second-class risk sample L d2 : the risk evaluation value of the prediction result calculated based on the mean absolute error neural network model exceeds a second threshold mu 2 While the risk evaluation value of the prediction result calculated based on the arctangent support vector machine model does not exceed the first threshold mu 1 Is a sample of (2);
dangerous sample L e : risk evaluation value of prediction result calculated based on arctangent support vector machine model exceeds first threshold mu 1 And the risk evaluation value of the prediction result calculated based on the mean absolute error neural network model exceeds a second threshold mu 2 Is a sample of (2);
the dangerous samples are judged to be samples containing noise labels, and the samples of the noise labels are filtered.
Further, the method further comprises the following steps:
retraining the complementary Lu Bangji classifier by using the labeled training set after filtering the dangerous samples to obtain a third complementary Lu Bangji classifier;
predicting each sample of the labeled training set after dangerous samples are filtered by using a third complementary Lu Bangji classifier, and calculating a risk evaluation value of a prediction resultAnd->Wherein (1)>The risk evaluation values of the prediction results calculated based on the average absolute error neural network model and the average absolute error neural network model are respectively marked with the index i as the serial number of the sample;
calculating a labeled training set after filtering dangerous samples and one type of dangerous samples simultaneouslyIs the mean difference delta of (1) 1 If delta 1 If the number is more than 0, judging the risk sample as a sample containing a noise label;
and/or calculating a labeled training set after filtering the dangerous samples and the second-class dangerous samples simultaneouslyIs the mean difference delta of (1) 2 If delta 2 And (3) judging the second-class risk samples as samples of the noisy labels.
Further, the confidence of the pseudo tag is expressed as:
c 1,j representing pseudo tagsConfidence of c 2,j Representing pseudo tag->Confidence of (2); />Representing the output of the arctangent support vector machine model and the average absolute error neural network model in the second complementary Lu Bangji classifier, respectively, corresponding to the ith labeled sample; h is the number of samples of the unlabeled subset.
Further, the method further comprises the following steps:
if the abnormal detection result of the arctangent support vector machine model or the average absolute error neural network model in the trained complementary Lu Bangji classifier is abnormal, triggering an abnormal alarm;
if the abnormal detection results of the arctangent support vector machine model or the average absolute error neural network model in the trained complementary Lu Bangji classifier are normal, an alarm is not triggered.
Further, the structure of the average absolute error neural network model is a full-connection layer neural network.
Further, the arctangent support vector machine model has a structure of a soft interval support vector machine.
The system comprises a memory, a processor and a computer program which is stored in the memory and can run on the processor, wherein the processor realizes the noise-tag-scene-oriented step-type robust semi-supervised gas engine abnormality detection method when executing the computer program.
The beneficial effects of the invention are as follows: aiming at the problems that the data labels of the gas turbine are rare and noise exists, the invention designs a stepping iteration complementary noise filtering strategy and further establishes an anomaly detection model. Firstly, an arctangent support vector machine model and an average absolute error neural network model are designed based on a symmetry loss theory to construct a classifier, and the classifier has strong robust characteristics on noise labels; secondly, a complementary noise filtering mechanism is provided, noise labels are filtered step by step, a classifier is updated, and the risk of overfitting of the system to the noise labels is further reduced; in addition, the invention designs a self-adaptive data supplementing strategy, and the problem of rare data labels of the fuel engine is overcome by utilizing a high-confidence pseudo label sample to realize the expansion of a labeled data set. Compared with the traditional supervised learning-based anomaly detection method, the method provided by the invention realizes reasonable utilization of insufficient labeled samples, cleans the labeled data set based on a stepping noise label filtering mechanism, builds a complementary Lu Bangji classifier as a reliable anomaly detection model, and improves the accuracy and F1 score of anomaly detection. Compared with the traditional supervised learning-based anomaly detection method, the method can effectively resist the influence of noise labels, and can establish an anomaly detection model with good performance under the condition of rare labels.
Drawings
FIG. 1 is a flow chart of a stepwise robust semi-supervised combustion engine anomaly detection method for a noise label scenario;
FIG. 2 is a schematic diagram of noise risk assessment;
FIG. 3 is a schematic diagram of the operation of the base classifier for complementary verification of the evaluation result;
Detailed Description
The present embodiment uses the actual engine operating data to verify the validity of the present method. The invention will now be described in further detail with reference to the drawings and to specific examples.
In order to prevent the heat channel parts of the gas turbine from being damaged due to over-temperature during the operation of the gas turbine generator set, the gas turbine is provided with a cooling air system. In the embodiment, 1000 samples before and after the occurrence of abnormal cooling air pressure of the compressor of the gas turbine of the unit in a certain period are collected as training sets to build a model, and 400 samples before and after the occurrence of the same abnormal state in another period are taken as two test sets to perform abnormality detection test; wherein 50 samples in the training samples are labeled samples, and the rest training samples are unlabeled samples; each sample contains 33 measuring points of process variables which are mainly collected by sensors positioned at the positions of a unit front-end module, a unit turbine, a combustion chamber and a unit compressor in a gas turbine generator set; the process variables include in particular: the natural gas volumetric flow rate of the front-end module, the natural gas mass flow rate of the front-end module, the gas inlet temperature (substituting ambient temperature) of the gas compressor, the gas compressor outlet pressure 1, the gas compressor outlet pressure 2, the gas compressor outlet temperature, the gas compressor bearing temperature 1, the gas compressor bearing temperature 2, the gas compressor bearing temperature 3, the gas compressor thrust tile bearing generator end temperature 1, the gas compressor thrust tile bearing generator end temperature 2, the gas compressor thrust tile bearing generator end temperature 3, the gas compressor thrust tile bearing gas engine end temperature 1, the gas compressor thrust tile bearing gas engine end temperature 2, the gas compressor thrust tile bearing gas engine end temperature 3 the method comprises the following steps of 1, 2 and 2 parts of compressor bearing vibration, 2 parts of compressor side large shaft vibration, 2 parts of compressor inlet channel pressure difference, 2 parts of turbine exhaust average temperature, 1 part of turbine side large shaft vibration, 2 parts of combustion chamber pressure difference, 1 part of combustion engine cooling air regulating valve position, 2 parts of combustion engine cooling air regulating valve position, 1 part of combustion engine buzzing, 2 parts of combustion engine rotating speed, 2 parts of combustion engine power, 1 part of combustion engine 2-stage stator blade ring holding cavity cooling air pressure, 2 parts of combustion engine 2-stage stator blade ring holding cavity cooling air pressure, 1 part of combustion engine 3-stage stator blade ring holding cavity cooling air pressure and 2 parts of combustion engine 3-stage stator blade ring holding cavity cooling air pressure.
According to the noise-tag-scene-oriented stepping robust semi-supervised gas turbine anomaly detection method, as shown in fig. 1, in an online stage, gas turbine operation data are acquired, and the gas turbine operation data are input into a trained complementary Lu Bangji classifier to obtain an anomaly detection result.
Specifically, the complementary Lu Bangji classifier was obtained by training the offline modeling phase as follows:
step 1: collecting the operation data of the gas engine as a training set, wherein the training set is divided into two parts, namely, the data which are collected during the operation of the gas engine and are used for clearly indicating the operation state, namely, a labeled sample; data which are collected during the running of the gas engine and do not explicitly indicate the running state, namely a label-free sample;
collecting the operation data of the combustion engine as a training set, and recording asWherein N is 1 And N 2 Respectively represent the number of labeled and unlabeled exemplars, wherein X train The data dimension of each sample is V, that is, each sample is the data of the running state corresponding to one time, including V process variables, in this embodiment v=33; x is X train Is of N 1 The samples correspond to the tag +.>The part of the sample and the label thereof form a label sample set which is marked asWith N 2 The samples are unlabeled samples, and form an unlabeled sample set, which is marked as +.>In the present embodiment, N 1 ＝50，N 2 =950; wherein the sample tag-> Indicating that the sample is marked as normal; />Indicating that the sample is marked as abnormal; sample tags have a certain proportion of noise conditions, i.e. for some samples, the tag marked is wrong;
in general, in order to ensure detection accuracy, data is subjected to normalization processing, specifically:
for a pair ofX train And (3) carrying out standardization processing, wherein the calculation formula is as follows:
wherein, the liquid crystal display device comprises a liquid crystal display device,x i,v,train representing the value of the v-th process variable in the i-th training sample collected;representing the mean of the v-th process variable in all training samples; s is(s) v,train Representing standard deviations of the v-th process variable in all training samples; x is x i,v,train Representing the standardized numerical values of the corresponding samples and the corresponding variables;
for training dataX train Data set obtained after normalizationX train With N 1 The samples correspond to the labels indicating the running state of the gas engine, and the samples and the labels form a label sample set which is recorded asWith N 2 The samples are unlabeled samples, and form an unlabeled sample set, which is marked as +.>
Step 2: constructing a complementary type Lu Bangji classifier, wherein the complementary type Lu Bangji classifier model comprises a parallel arctangent support vector machine model and an average absolute error neural network model; specifically:
(2.1) constructing an arctangent support vector machine model, which means a support vector machine model for constructing a loss function by adopting an arctangent function, wherein the support vector machine model has a structure of a conventional support vector machine, such as a soft interval support vector machine, and the like, and specifically, training an arctangent support vector machine classifier f 1 The loss function of (2) is:
wherein w is 1 And b 1 Representing parameters to be optimized of the arctangent support vector machine model, wherein xi represents a penalty coefficient, L represents the number of labeled samples, and L is L 2 Representing the square of the two norms, the penalty coefficient ζ in this embodiment is set to 0.078; the loss functionThe method has symmetrical property and strong robustness to tag noise; for a particular sample x, the output of the arctangent support vector machine model is noted as
o 1 ＝f 1 (x)＝w 1 T x+b 1
The arctangent support vector machine classifier f 1 The model parameters of (2) are updated in a gradient descent mode so as to complete training; the parameter updating mode is as follows:
wherein ε 1 Representing ginsengLearning rate of number update, epsilon in this embodiment 1 Set to 0.5; the value is an assignment operator, and the parameter is updated;
(2.2) constructing an average absolute error neural network model, wherein the average absolute error neural network model of the invention is a neural network model for constructing a loss function by adopting an average absolute error function, and the average absolute error neural network model can adopt a conventional neural network, such as a full-connection layer neural network and the like, and comprises an input layer and an output layer, and in the embodiment, the number of nodes of the input layer is 33, and the number of nodes of the output layer is 1; for a particular sample x, the output of the mean absolute error neural network model is recorded as
Wherein w is 2 And b 2 Model parameters representing an average absolute error neural network model; the mean absolute error neural network classifier f 2 Is updated by means of batch gradient descent to complete model training, wherein the parameter updating learning rate is epsilon 2 The number of samples in each training batch is M, ε in this example 2 Set to 0.001, m set to 10; training classifier f 2 Specifically the average absolute error function:
wherein, the liquid crystal display device comprises a liquid crystal display device,represent a certain labeled training sample->Corresponding tag o 2,train Representing that the mean absolute error neural network model corresponds to a training sample +.>An output of (2); the average absolute error loss also has symmetrical property, and has stronger robustness to the label noise;
(2.3) training the complementary Lu Bangji classifier by using the tag training set to obtain a first complementary Lu Bangji classifier; wherein the arctangent support vector machine model and the average absolute error neural network model can be trained simultaneously or sequentially, and the same training times or different training times can be adopted, in this embodiment, the arctangent support vector machine classifier f 1 Training run t 1 Set to 10; mean absolute error neural network classifier f 2 Training run t 2 Set to 20;
the trained arctangent support vector machine model and the average absolute error neural network model form a first complementary type Lu Bangji classifier, and the first complementary type Lu Bangji classifier has strong robustness to noise labels;
step 3: predicting each sample of the labeled training set by using a first complementary Lu Bangji classifier, judging whether the corresponding sample is a noise label or not according to a risk evaluation value of a prediction result, and filtering samples of the noise label; retraining the complementary Lu Bangji classifier with the noise-removed labeled training set to obtain a second complementary Lu Bangji classifier; specifically:
(3.1) noise risk assessment, wherein the method is schematically shown in fig. 2-3, specifically:
using two trained classifiers f 1 And f 2 Performing noise risk assessment; the two classifiers evaluate the label risks of the existing labeled data respectively, and each selects a sample exceeding a risk threshold; the threshold value corresponding to each classifier can be given in advance according to experience, f 1 The threshold value of (1) is expressed as mu 1 ，f 2 The threshold value of (1) is expressed as mu 2 Mu in this example 1 Sum mu 2 Respectively 0.2 and 0.9; model f based on arctangent support vector machine 1 The risk evaluation value of the calculated prediction result is the arctangent support directionThe negative value of the product of the predictive label corresponding to each sample and the label carried by that sample is expressed as:
wherein, the liquid crystal display device comprises a liquid crystal display device,classifier f representing arctangent support vector machine 1 The evaluation corresponds to the ith labeled training sampleRisk of (1)>Classifier f representing arctangent support vector machine 1 Corresponds to sample->An output of (2);
neural network model f based on average absolute error 2 The risk evaluation value of the calculated prediction result is that the average absolute error neural network model is linearly transformed to the prediction label corresponding to each sample and mapped to [ -1,1]The absolute value of the tag difference value carried by the sample after the interval is expressed as:
wherein, the liquid crystal display device comprises a liquid crystal display device,mean absolute error neural network model f 2 The evaluation corresponds to the ith labeled training sampleRisk of (1)>Mean absolute error neural network model f 2 Corresponds to sample->An output of (2);
according to the result of noise risk assessment, all labeled sample sets are divided into 4 categories, including:
secure sample L c : risk evaluation value of prediction result calculated based on arctangent support vector machine model, namely arctangent support vector machine classifier f 1 The noise risk imparted does not exceed the corresponding threshold mu 1 And the risk evaluation value of the prediction result calculated based on the mean absolute error neural network model is the mean absolute error neural network classifier f 2 The noise risk imparted does not exceed the corresponding threshold mu 2 Is a sample of (2);
risk sample L d1 : arctangent support vector machine classifier f 1 The noise risk conferred exceeds the corresponding threshold μ 1 And the mean absolute error neural network classifier f 2 The noise risk imparted does not exceed the corresponding threshold mu 2 Is a sample of (2);
second-class risk sample L d2 : mean absolute error neural network classifier f 2 The noise risk conferred exceeds the corresponding threshold μ 2 While arctangent support vector machine classifier f 1 The noise risk imparted does not exceed the corresponding threshold mu 1 Is a sample of (2);
dangerous sample L e : arctangent support vector machine classifier f 1 The noise risk conferred exceeds the corresponding threshold μ 1 And mean absolute error neural network classifier f 2 The noise risk conferred exceeds the corresponding threshold μ 2 Is a sample of (2);
(2.2) complementary verification of noise evaluation results
(2.2.1) wherein the dangerous sample is judged as a sample containing a noise tag, and the dangerous sample L e Performing direct filtering operation; i.e.
L′＝L-L e
Wherein L' represents a filtered dangerous sample L e A label sample set is obtained later;
(2.2.2) further comprising determining whether the first and second class risk samples are noise-containing label samples, specifically as follows:
(2.2.2.1) retraining the complementary Lu Bangji classifier with the remaining labeled samples L' to obtain a third complementary Lu Bangji classifier, the arctangent support vector machine classifier and the average absolute error neural network classifier of the third complementary Lu Bangji classifier being denoted as f, respectively 1 ' and f 2 ' re-calculating the risk of remaining label samples, respectively noted asAnd->
(2.2.2.2) calculating a labeled training set after filtering the dangerous sample and the risk sample simultaneouslyIs the mean difference delta of (1) 1 If delta 1 If the number is more than 0, judging the risk sample as a sample containing a noise label; the method comprises the following steps:
calculating a labeled training set L' after dangerous samples are filteredAverage risk, noted as
Calculating labeled training set L' -L after filtering dangerous samples and one type of dangerous samples simultaneously d1 A kind of electronic deviceAverage risk, noted as
Calculate the difference delta between the two 1 Is recorded as
(2.2.2.3) calculating a labeled training set after filtering the dangerous sample and the second-class dangerous sample simultaneouslyIs the mean difference delta of (1) 2 If delta 2 Judging the class II risk samples as samples containing noise labels if the number of the class II risk samples is more than 0; the method comprises the following steps:
calculating a labeled training set L' after dangerous samples are filteredAverage risk, noted as
Calculating labeled training set L' -L after filtering dangerous samples and class II dangerous samples simultaneously d2 A kind of electronic deviceAverage risk, noted as
Calculate the difference delta between the two 2 Is recorded as
(2.2.3) further cleaning the labeled data set by further judging whether the first-class risk sample and the second-class risk sample are samples of noisy labels, and obtaining a final remaining labeled sample set L ", which is specifically as follows:
if delta 1 > 0 and delta 2 > 0, then for a class of risk samples L d1 And L d2 The filtering operation is performed, namely the following operation is performed:
L″＝L′-L d1 -L d2
if delta 1 > 0 and delta 2 Less than or equal to 0, for a type of risk sample L d1 Executing filtering operation;
L″＝L′-L d1
if delta 1 Less than or equal to 0 and delta 2 > 0, then for the second class risk sample L d2 Executing filtering operation;
L″＝L′-L d2
if delta 1 Less than or equal to 0 and delta 2 Less than or equal to 0, for a type of risk sample L d1 And L d2 The reservation operation is performed, i.e. the following operations are performed:
L″＝L′
retraining the classifier by using the residual labeled sample L' after the filtering of the complementary noise label to obtain a second complementary Lu Bangji classifier, wherein the arctangent support vector machine classifier and the average absolute error neural network classifier are respectively marked as f 1 "sum f 2 ”；
Step 4: randomly selecting samples in the non-tag training set to form a non-tag subset, predicting each sample in the non-tag subset by using a second complementary Lu Bangji classifier to obtain corresponding pseudo tags, sorting according to confidence levels of the pseudo tags, selecting a plurality of samples with the largest confidence levels, deleting the samples from the non-tag training set, forming labeled samples with the corresponding pseudo tags, adding the labeled samples into the labeled training set, and performing self-adaptive data supplementation; the arctangent support vector machine model and the average absolute error neural network model can be respectively performed, and the method specifically comprises the following steps:
from no markRandomly selecting H samples from the tag data set U to form a non-tag subsetIn this embodiment, H is selected to be 400, and an arctangent support vector machine classifier f is used 1 U-feed 1 The label-free sample in (1) predicts the corresponding pseudo label, specifically:
wherein, the liquid crystal display device comprises a liquid crystal display device,representing the arctangent support vector machine classifier f 1 "give sample>Pseudo tag of->Representation classifier f 1 "corresponds to>Sign (·) represents a sign function, specifically:
wherein z represents an argument of the sign function;
calculating and evaluating the confidence of each pseudo tag:
wherein c 1,j The representation corresponds to a samplePseudo tag of->Confidence of (2);
all unlabeled exemplars are tested according to their corresponding confidence levels c 1,j Sorting from big to small, and selecting h with highest confidence coefficient of pseudo tag 1 The individual samples form a set of candidate samples, denoted asCollecting pseudo labels corresponding to the samples to form a pseudo label sample set, which is marked as +.>In this embodiment, h 1 1 is selected;
randomly picking H samples from the unlabeled dataset U to form an unlabeled subsetClassifier f using mean absolute error neural network 2 U-feed 2 The label-free sample in (1) predicts the corresponding pseudo label, specifically:
wherein, the liquid crystal display device comprises a liquid crystal display device,representing the mean absolute error neural network classifier f 2 "give sample>Is a pseudo tag of (a),representation classifier f 2 "corresponds to>And evaluate the confidence of each pseudo tag:
wherein c 2,j The representation corresponds to a samplePseudo tag of->Confidence of (2);
all unlabeled exemplars are tested according to their corresponding confidence levels c 2,j Sorting from big to small, and selecting h with highest confidence coefficient of pseudo tag 2 The samples, forming a set of candidate samples, are denoted asCollecting pseudo labels corresponding to the samples to form a pseudo label sample set, which is marked as +.>In this embodiment, h 2 1 is selected;
sample set Q 1 And Q 2 Adding the samples and the pseudo tags thereof in the tagged data set, namely executing the following operation;
L←L∪Q 1 ∪Q 2
wherein, U represents the complement operation of the set;
sample set P 1 And P 2 The samples in (a) are removed from the unlabeled dataset, i.e., the following operations are performed:
U←U-(P 1 ∪P 2 ).
repeating the operations of noise risk assessment, complementation verification on the noise assessment result and self-adaptive data supplementation according to the steps 2-4 until the set number of rounds T is reached to obtain a final trained complementary Lu Bangji classifier; t in this example is selected to be 20; record final trainingThe finished arctangent support vector machine classifier and the average absolute error neural network classifier are f 1 final And f 2 final 。
Abnormality detection can be performed by using a trained complementary Lu Bangji classifier, specifically as follows:
inputting a real-time collected fuel engine operation data sample into a trained complementary Lu Bangji classifier, and fusing a discrimination result of the base classifier on an online sample operation state so as to detect abnormality of the fuel engine;
wherein, the operation data of the combustion engine collected in real time is standardized, taking a test sample as an example,x test is V, i.e., the sample corresponds to V process variables; the specific standardized mode is as follows:
wherein, the liquid crystal display device comprises a liquid crystal display device,x v,test and x v,test Respectively representing values before and after the standardization of the v-th process variable in the collected test sample; the standardized test data x test Input to a trained classifier f 1 final Giving a prediction result; the method comprises the following steps:
result 1 ＝sign(o 1,test )
wherein o is 1,test Representation classifier f 1 final Corresponds to x test Is specifically expressed as
o 1,test ＝f 1 final (x test )
result 1 =1 indicates that the discrimination result is normal, result 1 = -1 indicates that the discrimination result is abnormal;
the standardized test data x test Input to a trained classifierIn the following, the prediction result is given:
result 2 ＝sign(o 2,test -0.5)
wherein o is 2,test Representation classifierCorresponds to x test Is specifically expressed as
result 2 =1 indicates that the discrimination result is normal, result 2 = -1 indicates that the discrimination result is abnormal;
further, if the anomaly detection result of the arctangent support vector machine model or the mean absolute error neural network model is anomaly, i.e., the classifier f 1 final Or classifierJudging the online sample as an abnormal sample, and triggering an abnormal alarm;
if the abnormal detection results of the arctangent support vector machine model or the average absolute error neural network model are normal, namely the classifier f 1 final ClassifierThe online sample is determined to be a normal sample, and an alarm is not triggered.
The effect of the invention is explained by using the sample of the gas turbine compressor in abnormal cooling air pressure as the test set to perform fault detection analysis. In the aspect of noise setting, the invention respectively examines the situations that the initial tag noise proportion is 20% and 30%; each set of experiments was repeated 5 times at the same parameter settings.
In order to more clearly embody the superiority of the complementary noise filtering robust semi-supervised combustion engine anomaly detection task, two existing noise label learning methods are selected, and mutual teaching strategies co-learning (Han B, yao Q, yu X, et al Co-learning: robust training of deep neural networks with extremely noisy labels [ J ]. Advances in neural information processing systems,2018,31 ]), and importance weighting importance reweighting (Liu T, tao D.classification with noisy labels by importance reweighting [ J ]. IEEE Transactions on pattern analysis and machine intelligence,2015,38 (3): 447-461.) are compared. Adding a common co-tracking strategy on the basis of a co-tracking method in an experiment to complete self-adaptive data supplementation; the self-adaptive data supplementation is completed by adding a common self-tracking strategy on the basis of the method of importance reweighting.
For more visual comparison, tables 1 and 2 list the mean and variance of the anomaly detection Accuracy (Accuracy) and the F1 score (F1-score) over the test set for 5 replicates, respectively, when the initial tag noise ratios were 20% and 30%, respectively. The computational expressions of Accuracy and F1-score are:
it can be found that when the noise proportion in the label sample is low (20%), good abnormality detection effect can be obtained by all three methods, wherein the accuracy of the method and the accuracy of the F1 fraction in five experiments are higher than those of other two methods, and the variance is smaller; when the noise proportion in the label sample is higher (30%), the abnormality detection performance of the comparison method is obviously reduced, but the precision and the F1 fraction of the method provided by the invention can still be kept at a high level, and the precision and the F1 fraction of the method can obtain the optimal results in the three methods on the indexes of mean and variance. The method provided by the invention has obvious robustness on the noise label and strong stability on model training, and further proves the feasibility and effectiveness of the method.
Table 1 comparison of anomaly detection performance at 20% tag noise ratio for different methods
Table 2 comparison of anomaly detection performance at a tag noise ratio of 30% for different methods
Corresponding to the embodiment of the noise-tag-scene-oriented step-by-step robust semi-supervised combustion engine anomaly detection method, the invention also provides the embodiment of the noise-tag-scene-oriented step-by-step robust semi-supervised combustion engine anomaly detection system.
The invention discloses a noise-tag-scene-oriented stepping type robust semi-supervised gas turbine anomaly detection system, which comprises a memory, a processor and a computer program stored on the memory and capable of running on the processor, wherein the processor realizes the noise-tag-scene-oriented stepping type robust semi-supervised gas turbine anomaly detection method when executing the computer program.
For system embodiments, reference is made to the description of method embodiments for the relevant points, since they essentially correspond to the method embodiments. The system embodiments described above are merely illustrative, wherein the elements illustrated as separate elements may or may not be physically separate, and the elements shown as elements may or may not be physical elements, may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the purposes of the present invention. Those of ordinary skill in the art will understand and implement the present invention without undue burden.
It is apparent that the above examples are given by way of illustration only and are not limiting of the embodiments. Other variations or modifications of the above teachings will be apparent to those of ordinary skill in the art. It is not necessary or exhaustive of all embodiments. And obvious variations or modifications thereof are contemplated as falling within the scope of the present invention.

Claims (9)

1. The step-by-step robust semi-supervised gas turbine anomaly detection method for the noise label scene is characterized by comprising the steps of acquiring gas turbine operation data, and inputting the gas turbine operation data into a trained complementary Lu Bangji classifier to obtain an anomaly detection result; wherein the complementary Lu Bangji classifier is obtained by training the following steps:
step 1: constructing a training set based on the collected combustion engine operation data, wherein the training set comprises a labeled training set and a non-labeled training set;
step 2: constructing a complementary Lu Bangji classifier, wherein the complementary Lu Bangji classifier model comprises a parallel arctangent support vector machine model and an average absolute error neural network model; the loss function adopted in the training of the arctangent support vector machine model is as follows:
wherein w is 1 And b 1 Representing parameters to be optimized of the arctangent support vector machine model, wherein xi represents a penalty coefficient, L represents the number of labeled samples, and L is L 2 Representing the square of the two norms,label for the ith labeled sample, +.>The fuel engine operation data of the ith tagged sample; the loss function adopted in the training of the average absolute error neural network model is the average absolute error; complementary pairs using labeled training setsTraining the classifier of the formula Lu Bangji to obtain a first complementary Lu Bangji classifier;
step 3: predicting each sample of the labeled training set by using a first complementary Lu Bangji classifier, judging whether the corresponding sample is a noise label or not according to a risk evaluation value of a prediction result, and filtering samples of the noise label; retraining the complementary Lu Bangji classifier with the noise-removed labeled training set to obtain a second complementary Lu Bangji classifier;
step 4: randomly selecting samples in the non-tag training set to form a non-tag subset, predicting each sample in the non-tag subset by using a second complementary Lu Bangji classifier to obtain corresponding pseudo tags, sorting according to the confidence of the pseudo tags, selecting one or more samples with the maximum confidence, deleting the samples from the non-tag training set, and forming a tagged sample with the corresponding pseudo tags, and adding the tagged sample into the tagged training set;
and (3) alternately performing complementary noise label filtering and self-adaptive data supplementing according to the steps 2-4 until the set number of rounds T is reached, and obtaining the final trained complementary Lu Bangji classifier.
2. The method of claim 1, wherein the collected engine operation data is normalized engine operation data.
3. The method of claim 1, wherein in the step 3, each sample in the labeled training set is predicted by using a first complementary Lu Bangji classifier, and whether the corresponding sample is a noise label is determined according to a risk evaluation value of a prediction result, and the sample of the noise label is filtered, specifically:
predicting each sample of the labeled training set by using a first complementary Lu Bangji classifier, and calculating a risk evaluation value of a prediction result; the risk evaluation value of the prediction result calculated based on the arctangent support vector machine model is a negative value of the product of the prediction label of each sample corresponding to the arctangent support vector machine model and the label carried by the sample; the risk evaluation value of the prediction result calculated based on the average absolute error neural network model is the absolute value of the label difference value carried by the sample after the average absolute error neural network model is in linear transformation corresponding to the prediction label of each sample and is mapped to the [ -1,1] interval;
according to the risk evaluation value of the prediction result, all the labeled samples are divided into 4 categories, including:
secure sample L c : the risk evaluation value of the prediction result calculated based on the arctangent support vector machine model does not exceed a first threshold mu 1 And the risk evaluation value of the prediction result calculated based on the mean absolute error neural network model does not exceed the second threshold mu 2 Is a sample of (2);
risk sample L d1 : risk evaluation value of prediction result calculated based on arctangent support vector machine model exceeds first threshold mu 1 While the risk evaluation value of the prediction result calculated based on the mean absolute error neural network model does not exceed the second threshold μ 2 Is a sample of (2);
second-class risk sample L d2 : the risk evaluation value of the prediction result calculated based on the mean absolute error neural network model exceeds a second threshold mu 2 While the risk evaluation value of the prediction result calculated based on the arctangent support vector machine model does not exceed the first threshold mu 1 Is a sample of (2);
dangerous sample L e : risk evaluation value of prediction result calculated based on arctangent support vector machine model exceeds first threshold mu 1 And the risk evaluation value of the prediction result calculated based on the mean absolute error neural network model exceeds a second threshold mu 2 Is a sample of (2);
the dangerous samples are judged to be samples containing noise labels, and the samples of the noise labels are filtered.
4. A method according to claim 3, further comprising:
retraining the complementary Lu Bangji classifier by using the labeled training set after filtering the dangerous samples to obtain a third complementary Lu Bangji classifier;
using third complementary robustnessThe base classifier predicts each sample of the labeled training set after dangerous samples are filtered, and calculates a risk evaluation value of a predicted resultAnd->Wherein (1)>The risk evaluation values of the prediction results calculated based on the average absolute error neural network model and the average absolute error neural network model are respectively marked with the index i as the serial number of the sample;
calculating a labeled training set after filtering dangerous samples and one type of dangerous samples simultaneouslyIs the mean difference delta of (1) 1 If delta 1 If the number is more than 0, judging the risk sample as a sample containing a noise label;
and/or calculating a labeled training set after filtering the dangerous samples and the second-class dangerous samples simultaneouslyIs the mean difference delta of (1) 2 If delta 2 And (3) judging the second-class risk samples as samples of the noisy labels.
5. The method of claim 1, wherein the confidence level of the pseudo tag is expressed as:
c 1,j representing pseudo tagsConfidence of c 2,j Representing pseudo tag->Confidence of (2); />Respectively representing the output of the arctangent support vector machine model and the average absolute error neural network model in the second complementary Lu Bangji classifier corresponding to the ith unlabeled exemplar; h is the number of samples of the unlabeled subset.
6. The method as recited in claim 1, further comprising:
if the abnormal detection result of the arctangent support vector machine model or the average absolute error neural network model in the trained complementary Lu Bangji classifier is abnormal, triggering an abnormal alarm;
if the abnormal detection results of the arctangent support vector machine model or the average absolute error neural network model in the trained complementary Lu Bangji classifier are normal, an alarm is not triggered.
7. The method of claim 1, wherein the mean absolute error neural network model is structured as a full-connected layer neural network.
8. The method of claim 1, wherein the arctangent support vector machine model is structured as a soft-spaced support vector machine.
9. A noise-tag-scene-oriented step-by-step robust semi-supervised combustion engine anomaly detection system, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements a noise-tag-scene-oriented step-by-step robust semi-supervised combustion engine anomaly detection method as recited in any one of claims 1-8 when executing the computer program.
CN202310656040.0A 2023-06-05 2023-06-05 Noise-tag-scene-oriented stepping robust semi-supervised gas turbine anomaly detection method and system Pending CN116975755A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310656040.0A CN116975755A (en) 2023-06-05 2023-06-05 Noise-tag-scene-oriented stepping robust semi-supervised gas turbine anomaly detection method and system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310656040.0A CN116975755A (en) 2023-06-05 2023-06-05 Noise-tag-scene-oriented stepping robust semi-supervised gas turbine anomaly detection method and system

Publications (1)

Publication Number Publication Date
CN116975755A true CN116975755A (en) 2023-10-31

Family

ID=88478579

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310656040.0A Pending CN116975755A (en) 2023-06-05 2023-06-05 Noise-tag-scene-oriented stepping robust semi-supervised gas turbine anomaly detection method and system

Country Status (1)

Country Link
CN (1) CN116975755A (en)

Similar Documents

Publication Publication Date Title
JP7069269B2 (en) Semi-supervised methods and systems for deep anomaly detection for large industrial surveillance systems based on time series data using digital twin simulation data
Jang et al. Adversarial autoencoder based feature learning for fault detection in industrial processes
JP2022037241A (en) Abnormality detection system, abnormality detection method, abnormality detection program, and method for generating learned model
US7251582B2 (en) Fault diagnosis
CN113065581B (en) Vibration fault migration diagnosis method for reactance domain self-adaptive network based on parameter sharing
CN110757510B (en) Method and system for predicting remaining life of robot
CN116448419A (en) Zero sample bearing fault diagnosis method based on depth model high-dimensional parameter multi-target efficient optimization
CN114282443B (en) Residual service life prediction method based on MLP-LSTM supervised joint model
CN112418306B (en) Gas turbine compressor fault early warning method based on LSTM-SVM
CN112926642B (en) Multi-source domain self-adaptive intelligent mechanical fault diagnosis method and system
CN109298633A (en) Chemical production process fault monitoring method based on adaptive piecemeal Non-negative Matrix Factorization
CN114297918A (en) Aero-engine residual life prediction method based on full-attention depth network and dynamic ensemble learning
Farahani et al. On the application of domain adversarial neural network to fault detection and isolation in power plants
CN116929815A (en) Equipment working state monitoring system and method based on Internet of things
CN116336400A (en) Baseline detection method for oil and gas gathering and transportation pipeline
CN113239610A (en) Domain self-adaptive rolling bearing fault diagnosis method based on Wasserstein distance
Huang et al. Static and dynamic joint analysis for operation condition division of industrial process with incremental learning
CN116842323A (en) Abnormal detection method for operation data of water supply pipeline
Alinezhad et al. A modified bag-of-words representation for industrial alarm floods
CN116975755A (en) Noise-tag-scene-oriented stepping robust semi-supervised gas turbine anomaly detection method and system
CN116070134A (en) Intelligent equipment fault diagnosis method and system based on prototype learning
CN115618506A (en) Method for predicting power of single-shaft combined cycle gas turbine
CN115017978A (en) Fault classification method based on weighted probability neural network
Lu et al. Visual monitoring of industrial operation states based on kernel fisher vector and self-organizing map networks
Lööv Comparison of undersampling methods for prediction of casting defects based on process parameters

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