CN113159088B - Fault monitoring and diagnosis method based on multi-feature fusion and width learning - Google Patents
Fault monitoring and diagnosis method based on multi-feature fusion and width learning Download PDFInfo
- Publication number
- CN113159088B CN113159088B CN202110020406.6A CN202110020406A CN113159088B CN 113159088 B CN113159088 B CN 113159088B CN 202110020406 A CN202110020406 A CN 202110020406A CN 113159088 B CN113159088 B CN 113159088B
- Authority
- CN
- China
- Prior art keywords
- signals
- signal
- fault
- historical process
- matrix
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 138
- 238000003745 diagnosis Methods 0.000 title claims abstract description 73
- 238000012544 monitoring process Methods 0.000 title claims abstract description 26
- 230000004927 fusion Effects 0.000 title claims abstract description 24
- 230000008569 process Effects 0.000 claims abstract description 107
- 238000004519 manufacturing process Methods 0.000 claims abstract description 23
- 239000013598 vector Substances 0.000 claims abstract description 18
- 238000012549 training Methods 0.000 claims abstract description 8
- 238000000605 extraction Methods 0.000 claims abstract description 7
- 238000011156 evaluation Methods 0.000 claims abstract description 6
- 239000011159 matrix material Substances 0.000 claims description 55
- 230000006870 function Effects 0.000 claims description 18
- 238000012360 testing method Methods 0.000 claims description 15
- 238000012545 processing Methods 0.000 claims description 14
- 230000004913 activation Effects 0.000 claims description 12
- 238000004458 analytical method Methods 0.000 claims description 12
- 238000005259 measurement Methods 0.000 claims description 12
- 238000010606 normalization Methods 0.000 claims description 7
- 238000009826 distribution Methods 0.000 claims description 6
- 238000013528 artificial neural network Methods 0.000 claims description 5
- 230000007246 mechanism Effects 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 3
- 230000009466 transformation Effects 0.000 claims description 3
- 238000007500 overflow downdraw method Methods 0.000 abstract description 3
- 238000004364 calculation method Methods 0.000 description 7
- 230000000694 effects Effects 0.000 description 6
- 238000000354 decomposition reaction Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000007274 generation of a signal involved in cell-cell signaling Effects 0.000 description 2
- 238000013139 quantization Methods 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 230000002087 whitening effect Effects 0.000 description 2
- 241000695274 Processa Species 0.000 description 1
- 230000005856 abnormality Effects 0.000 description 1
- 230000002411 adverse Effects 0.000 description 1
- 238000004422 calculation algorithm Methods 0.000 description 1
- 238000001311 chemical methods and process Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000009776 industrial production Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- 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)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (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)
- Testing And Monitoring For Control Systems (AREA)
Abstract
The invention provides a fault monitoring and diagnosing method based on multi-feature fusion and width learning, which comprises the following steps: acquiring historical process signals in an actual industrial process; performing multi-feature extraction based on historical process signals to construct multi-feature vectors; inputting the multi-feature vector into a width learning network for training to obtain a fault diagnosis model; applying the fault diagnosis model to an actual industrial process to carry out real-time fault diagnosis; evaluating the diagnosis performance of the fault diagnosis model by adopting an evaluation index F1-score, and if the Micro F1 is larger than a first set threshold and the Macro F1 is larger than a second set threshold, continuing to use the fault diagnosis model; otherwise, the training is resumed. The invention adopts the multi-feature fusion method, provides more dynamic information for fault diagnosis, improves the accuracy of fault diagnosis, solves the problem of reduced fault diagnosis efficiency caused by increased operation amount caused by multi-feature fusion, and can meet the requirements of real-time monitoring and diagnosis of the actual industry.
Description
Technical Field
The invention relates to the technical field of industrial process fault monitoring and diagnosis, in particular to a fault monitoring and diagnosis method based on multi-feature fusion and width learning.
Background
With the progress of science and technology and the development of modern production capacity, the automation degree of production equipment is greatly improved, the functions are more complete, and the structure is more complex. The performance of the equipment is reduced or the original function is lost due to a plurality of inevitable adverse factors in the industrial process, and even disastrous accidents can be caused if the equipment is not discovered and intervened in time. With the development of informatization and intelligent technology, the number of variables needing to be monitored in the industrial process is greatly increased, and the traditional diagnosis mode utilizing manual experience is difficult to deal with a large amount of industrial process data. Therefore, a rapid and accurate fault monitoring and diagnosis method is designed, and the method has important and practical significance for guaranteeing the safety of the modern industrial process.
The traditional multivariate statistical method generally considers that the industrial process is in a static state, namely, only the original process signal is judged and analyzed, and the dynamic information of each variable in the original process signal is not utilized, so that the capture of the dynamic change of the original process signal is insufficient. However, as faults occur, they are often accompanied by changes in characteristics such as trends, which can aid in the monitoring and diagnosis of faults. Therefore, modern industrial processes demand the extraction of dynamic features of the original process signal.
At present, a fault diagnosis method based on deep learning is a complex industrial process monitoring method which is widely concerned in recent years. The model established by the deep neural network is a black box system, so the complexity of the system structure does not need to be considered. In order to obtain better diagnosis effect, most of fault monitoring and diagnosis researches based on deep neural networks focus on stacking deeper structures or optimizing parameters of models, which results in occupying a large amount of computing resources in the process of adjusting the structures and optimizing the parameters. However, such expensive computational costs are often not affordable in actual industrial production, and the real-time, lightweight, and economical nature of the system is a great concern for the plant.
Disclosure of Invention
In order to solve the problems, the invention provides a fault monitoring and diagnosing method based on multi-feature fusion and width learning.
A fault monitoring and diagnosing method based on multi-feature fusion and width learning comprises the following steps:
s1, acquiring a historical process signal with a fault label in the actual industrial process; the historical process signals include fault signals and normal signals;
s2, performing multi-feature extraction based on the historical process signal to construct a multi-feature vector;
s3, inputting the multi-feature vectors into a width learning network for training to obtain a fault diagnosis model;
s4, applying the fault diagnosis model to an actual industrial process to carry out real-time fault diagnosis;
further, in step S2, based on the mechanism analysis, dividing the variables in the historical process signals into control variables, continuous measurement variables, and state variables, and extracting binary alarm signals in the control variables and slow characteristic signals, qualitative trend signals, and binary alarm signals in the continuous measurement variables, specifically:
s21, extracting the trend characteristics of the continuous measurement variables in the historical process signals by using a qualitative trend analysis method to obtain qualitative trend signals;
s22, calculating upper and lower alarm thresholds based on the normal signals in the historical process signals in the step S1, and acquiring binary alarm signals based on the upper and lower alarm thresholds;
s23, extracting slow features of continuous measurement variables in the historical process signals by using a slow feature analysis method to obtain slow feature signals;
the multi-feature vector is a vector formed by the historical process signal, the qualitative trend signal, the binary alarm signal and the slow feature signal;
further, in step S21, the specific process of acquiring the qualitative trend signal is as follows:
(1) preprocessing historical process signals by utilizing wavelet transformation, removing noise of local signals, and then performing Min-Max normalization processing to obtain preprocessed historical process signals;
(2) calculating a sample difference value of the preprocessed historical process signal;
(3) based on a qualitative trend analysis method, extracting the qualitative trend of the preprocessed historical process signal by using the sample difference value to obtain a qualitative trend signal;
further, the specific implementation process of step (3) is as follows:
when the sample difference is less than- δ, the qualitative trend is a decrease, represented by the qualitative trend signal value-1;
when the sample difference value is greater than or equal to-delta and less than or equal to delta, the qualitative trend is unchanged, and the qualitative trend signal is represented by a value of 0;
when the sample difference is greater than δ, the qualitative trend is increasing, the qualitative trend signal being represented by a value of 1; wherein, delta belongs to [0.01,0.05 ];
further, in step S22, if the normal signal in the historical process signal has an upper and lower alarm threshold set based on the priori knowledge, the upper and lower alarm threshold is used as the upper and lower alarm threshold; if the upper and lower alarm thresholds set based on the priori knowledge do not exist, processing according to the type of the normal signals of the historical process signals, specifically:
if the normal signals of the historical process signals accord with Gaussian distribution, calculating the mean value and the standard deviation of the normal signals in the historical process signals according to a 3 sigma principle, and adopting the standard deviation with adjustable mean value plus-minus multiple as the upper and lower alarm threshold values; if the normal signals of the historical process signals do not accord with Gaussian distribution, estimating a probability density function of the normal signals by using a kernel density estimation method, and calculating the upper and lower alarm thresholds based on the probability density and the significance level alpha; wherein α has a value of 0.005 or 0.01 or 0.05;
generating a binary alarm signal consisting of values 0 and 1 by judging whether the historical process signal exceeds the upper and lower alarm thresholds, specifically: when the historical process signal exceeds the upper alarm threshold and the lower alarm threshold, the binary alarm signal is 1 to indicate that abnormity occurs, otherwise, the binary alarm signal is 0 to indicate that no abnormity occurs;
further, the specific implementation process of step S3 is as follows:
s31, forming the multi-feature vector into a first input matrix XtrainAnd to said secondAn input matrix XtrainCarrying out Min-Max normalization processing to obtain a processed first input matrix Xtrain;
S32, based on the processed first input matrix X in the step S31trainGenerating an input node, the input node activation function being:
ZI=Φ[XtrainWei+βei]
where Φ is the linear activation function, βeiIs a bias matrix of input nodes, WeiIs XtrainWeight matrix of, and βeiAnd WeiIn (b) is [0,1 ]]Between randomly generated, input node ZI=[Z1,Z2,...Zi]I represents the total number of input nodes;
s33, for input node ZILinear combination is carried out to generate an enhanced node HJ=[H1,H2,…,Hj]The enhanced node activation function is:
HJ=ζ[ZIWhj+βhj]
where j represents the total number of enhanced nodes, ξ is the tanh activation function, WhjIs ZIWeight matrix of betahjIs HJAnd W is a bias matrix ofhjAnd betahjThe value of (1) is [0]Randomly generating;
s34, inputting the node Z in the step S32IAnd enhanced node H in step S33JForming a second input matrix A, which is an input of the width learning network, wherein A ═ ZI|HJ];
S35, calculating the pseudo inverse A of the second input matrix A according to the formula+The formula is as follows:
wherein λ is a constant, E is an identity matrix, ATRepresents the transpose of A;
s36 based onThe pseudo inverse A in step S35+Calculating a weight matrix, wherein a specific formula is as follows:
Wn=A+Yn
wherein, WnIs a weight matrix, i.e. the weight of the width learning network, YnA matrix of fault signatures representing historical process signals;
further, in step S4, the fault diagnosis model is applied to fault diagnosis in an actual industrial process, and the specific process is as follows:
s41, collecting process signals in an actual industrial process on line, and obtaining binary alarm signals of the process signals by judging whether the process signals exceed the upper alarm threshold and the lower alarm threshold;
s42, based on the binary alarm signal in step S41, determining whether an abnormality occurs, if so, extracting a qualitative trend signal and a slow characteristic signal of the process signal, and generating a multi-characteristic data matrix X composed of the process signal, the binary alarm signal, the qualitative trend signal and the slow characteristic signaltestStep S43 is executed; otherwise, judging the working condition to be a normal working condition state;
s43, converting the multi-feature data matrix XtestInputting the fault diagnosis model to obtain an output matrix Yn.testI.e. is the corresponding fault label, where Yn.test=XtestWn,WnIs a weight matrix;
s44, comparing the fault label in the step S43 with an actual fault label, evaluating the diagnosis performance of the fault diagnosis model by adopting an evaluation index F1-score, and if Micro F1 is larger than a first set threshold and Macro F1 is larger than a second set threshold, continuing to use the fault diagnosis model; otherwise, returning to the step S3, and retraining to obtain a new fault diagnosis model;
further, the first set threshold is 0.90, and the second set threshold is 0.85.
The invention has the beneficial effects that: by adopting the multi-feature fusion method, more dynamic information is provided for fault diagnosis, the accuracy of fault diagnosis is improved, the problem that the fault diagnosis efficiency is reduced due to the fact that the operation amount is increased caused by multi-feature fusion is solved, the requirements of real-time monitoring and diagnosis of the actual industry can be met, and the method specifically comprises the following steps:
extracting a trend signal, a binary alarm signal and a slow characteristic signal from the historical process signal, forming multi-characteristic fusion data with the historical process signal, and then adopting a width learning system as a neural network for processing the multi-characteristic fusion data; because the modeling of the width learning system does not depend on a deeper network structure, the calculation of the network weight is fast, the model building time is short, and the requirements of real-time monitoring and diagnosis of the actual industry can be met.
Drawings
The invention will be further described with reference to the accompanying drawings and examples, in which:
FIG. 1 is a flow chart of a fault monitoring and diagnosis method based on multi-feature fusion and breadth learning in an embodiment of the present invention;
FIG. 2 is a graph of historical fault signals in an embodiment of the invention;
FIG. 3 is a denoising effect map of a historical fault signal map in embodiment 2 of the present invention;
fig. 4 is a trend signal generation diagram of a historical failure signal in embodiment 2 of the present invention;
FIG. 5 is a binary alarm signal generation diagram of a historical fault signal in embodiment 2 of the present invention;
fig. 6 is a diagram showing generation of a binary alarm signal after a failure occurs in embodiment 2 of the present invention.
Detailed Description
For a more clear understanding of the technical features, objects and effects of the present invention, embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
The embodiment of the invention provides a fault monitoring and diagnosing method based on multi-feature fusion and width learning.
In the embodiment, the effectiveness of the fault diagnosis method based on multi-feature fusion is verified based on a simulation experiment of chemical process design of a public model TE (Tennessee Eastman). The data used were derived from TE simulation experiments with 52 observed variables for each sample in the TE set. The method comprises the steps of classifying 52 variables according to table 1, distinguishing original variables (such as table 1) needing feature extraction, dividing variables in historical process signals into control variables, continuous measurement variables and state variables based on mechanism analysis, extracting binary alarm signals in the control variables and slow feature signals, qualitative trend signals and binary alarm signals in the continuous measurement variables, and using the state variables as the historical process signals without extracting the features of the state variables.
Historical process signals generated by the actual running process are simulated by setting a fault in the TE. There are 21 faults, that is, 22 tags (21 fault tags +1 normal tags) are corresponded.
Table 1: information of 52 measured historical fault signals
Referring to fig. 1, fig. 1 is a flowchart of a fault monitoring and diagnosing method based on multi-feature fusion and width learning according to an embodiment of the present invention, including the following steps:
s1, acquiring historical process signals in the actual industrial process; the historical process signals include fault signals and normal signals;
s2, performing multi-feature extraction based on the historical process signal, and constructing a multi-feature vector, wherein the method specifically comprises the following steps:
s21, extracting the trend characteristics of the continuous measurement variables in the historical process signals by using a qualitative trend analysis method to obtain qualitative trend signals, wherein the method specifically comprises the following steps:
(1) preprocessing a historical process signal (such as a graph 2) by utilizing wavelet transformation, removing noise of a local signal (such as a graph 3), and then performing Min-Max normalization processing to obtain a preprocessed historical process signal;
the Min-Max normalization formula is:
wherein,is a normalized preprocessed historical process signal, HLIs the historical process signal after removal of local noise, HMaxIs HLMaximum value of (C), HMinIs HLMinimum value of (d);
the specific process of denoising by using wavelet transform comprises the following steps:
(1-1) selecting a fundamental wave type, determining the level of wavelet decomposition, and performing wavelet decomposition calculation on the historical fault signal, wherein a discrete wavelet function can be expressed as:
wherein psij,k(t) represents an orthogonal wavelet basis, a0Is the step size, b0J, k belongs to Z and t represents a time variable as a translation factor;
(1-2) performing threshold quantization processing on the wavelet coefficients decomposed in the step (1-1), specifically: selecting a threshold, removing general noise with wavelet coefficient lower than the threshold, and keeping useful historical process signal with wavelet coefficient higher than the threshold;
(1-3) performing wavelet reconstruction of the de-noised process based on the process signal subjected to the quantization processing in the step (1-2), specifically:
where f (t) represents the reconstructed denoised process signal, C is a constant independent of the historical fault signal, Cj,kIs a discrete wavelet variation coefficient;
(2) calculating a sample difference value of the preprocessed historical process signal;
(3) based on a qualitative trend analysis method, extracting a qualitative trend of the preprocessed historical process signal by using the sample difference value to obtain a qualitative trend signal (as shown in fig. 4), which specifically comprises the following steps:
when the sample difference is less than- δ, the qualitative trend is a decrease, represented by the qualitative trend signal value-1;
when the sample difference value is greater than or equal to-delta and less than or equal to delta, the qualitative trend is unchanged, and the qualitative trend signal is represented by a value of 0;
when the sample difference is greater than δ, the qualitative trend is increasing, and the qualitative trend signal is represented by a value of 1; wherein, delta belongs to [0.01,0.05 ].
S22, calculating upper and lower alarm thresholds based on the normal signals in the historical process signals in the step S1, and acquiring binary alarm signals based on the upper and lower alarm thresholds, specifically:
if the normal signals in the historical process signals have upper and lower alarm threshold values set based on priori knowledge, taking the upper and lower alarm threshold values as the upper and lower alarm threshold values; if the upper and lower alarm thresholds are not set based on the prior knowledge, processing is carried out according to the type of the normal signal of the historical process signal, specifically:
if the normal signals of the historical process signals accord with Gaussian distribution, calculating the mean value and the standard deviation of the normal signals in the historical process signals according to a 3 sigma principle, and adopting the standard deviation with adjustable mean value plus-minus multiple as the upper and lower alarm threshold values; if the normal signals of the historical process signals do not accord with Gaussian distribution, estimating a probability density function of the normal signals by using a kernel density estimation method, and calculating the upper and lower alarm thresholds based on the probability density and the significance level alpha; wherein α has a value of 0.005 or 0.01 or 0.05;
generating a binary alarm signal composed of values 0 and 1 by judging whether the historical process signal exceeds the upper and lower alarm thresholds (fig. 5 is a binary alarm signal in a normal state, and fig. 6 is a binary alarm signal after a fault occurs), specifically: when the historical process signal exceeds the upper alarm threshold and the lower alarm threshold, the binary alarm signal is 1 to indicate that abnormity occurs, otherwise, the binary alarm signal is 0 to indicate that no abnormity occurs;
s23, extracting slow characteristics of continuous measurement variables in the historical process signals by using a slow characteristic analysis method to obtain slow characteristic signals, wherein the slow characteristic signals are as follows:
linear features extracted using SFA algorithm: taking the historical process signal as an input signal vector x (t), firstly adopting whitening processing to eliminate the correlation before the variable, and completing whitening by using singular value decomposition. According to a formula s (t) (x (t)) g (x), searching a matrix g (x), making the slow characteristic variable s (t) slowly change along with time, writing the extracted slow characteristic into a linear combination of a historical process signal and a load matrix, and realizing the extraction of the slow characteristic signal;
the multi-feature vector is a vector formed by the historical process signal, the qualitative trend signal, the binary alarm signal and the slow feature signal;
s3, inputting the multi-feature vectors into a width learning network for training to obtain a fault diagnosis model, wherein the specific implementation process is as follows:
s31, forming the multi-feature vector into a first input matrix XtrainAnd for the first input matrix XtrainCarrying out Min-Max normalization processing to obtain a processed first input matrix Xtrain;
S32, based on the processed first input matrix X in the step S31trainGenerating an input node, the input node activation function being:
ZI=Φ[XtrainWei+βei]
where Φ is the linear activation function, βeiIs a bias matrix of input nodes, WeiIs XtrainAnd β iseiAnd WeiThe value of (1) is [0]Between randomly generated, input node ZI=[Z1,Z2,...Zi]I represents the total number of input nodes;
s33, input node ZIAre linearly combined to generateEnhanced node HJ=[H1,H2,…,Hj]The enhanced node activation function is:
HJ=ζ[ZIWhj+βhj]
where j represents the total number of booster nodes, ζ is the tanh activation function, WhjIs ZIWeight matrix of betahjIs HJA bias matrix of, and WhjAnd betahjIn (b) is [0,1 ]]Randomly generating;
s34, inputting the node Z in the step S32IAnd enhanced node H in step S33JForming a second input matrix A, which is an input to the width learning network, wherein A ═ ZI|HJ];
S35, calculating the pseudo inverse A of the second input matrix A according to the formula+The formula is as follows:
wherein λ is a constant, E is an identity matrix, ATRepresents the transpose of a;
s36, based on the pseudo-inverse A in step S35+Calculating a weight matrix, wherein the concrete formula is as follows:
Wn=A+Yn
wherein, WnIs a weight matrix, i.e. the weight of the width learning network, YnA matrix of fault signatures representing historical process signals;
s4, applying the fault diagnosis model to an actual industrial process to carry out real-time fault diagnosis, wherein the specific process is as follows:
s41, collecting process signals in an actual industrial process on line, and obtaining binary alarm signals of the process signals by judging whether the process signals exceed the upper alarm threshold and the lower alarm threshold;
s42, based on the binary alarm signal in step S41, judging whether a fault occurs, if so, extracting the processA qualitative trend signal and a slow signature signal of the signal, a multi-signature data matrix X consisting of the process signal, the binary alarm signal, the qualitative trend signal and the slow signature signal being generatedtestStep S43 is executed; otherwise, judging the data to be normal data, namely, no fault occurs;
s43, converting the multi-feature data matrix XtestInputting the fault diagnosis model to obtain an output matrix Yn.testI.e. the corresponding fault label, and further determine the type of the fault, wherein Yn.test=XtestWn, WnIs a weight matrix;
s44, comparing the fault label in the step S43 with an actual fault label, evaluating the diagnosis performance of the fault diagnosis model by adopting an evaluation index F1-score, and if the Micro F1 is more than 0.90 and the Macro F1 is more than 0.85, continuing to use the current fault diagnosis model; otherwise, the method returns to the step S3, and the training is repeated to obtain a new fault diagnosis model.
The evaluation index F1-score has two calculation modes, namely Micro F1 and Macro F1. If the historical fault signals are extremely unbalanced, the evaluation effect of the Micro F1 is poor; however, Macro F1 is susceptible to high recall rate and high accuracy.
Micro F1 was calculated as follows:
(1) calculating a Total Recall RecallmThe calculation formula is as follows:
(2) calculating the Total PrecisionmThe calculation formula is as follows:
(3) calculating the Micro F1, wherein the calculation formula is as follows:
wherein, TPnThe True Positive class of the nth class is determined as the Positive class, FPnThe False Positive negative class of the nth class is determined as a Positive class, TNnThe positive True class of the nth class is judged as a Negative class, FNnThe False Negative class of the nth class is judged as a Negative class, and n is 1, 2, … and 22;
macro F1 is calculated as follows:
1) calculate F1 for each category:
wherein, RecallnIndicating a recall rate for a single fault;
2) calculate Macro F1:
using different signals as data sources for fault diagnosis (Alarm represents a binary Alarm signal, Trend represents a qualitative Trend signal, slow represents a slow characteristic signal), F1-score for obtaining fault diagnosis of different signals is shown in table 2:
table 2: f1-score for fault diagnosis of different signals
When the Micro F1 is larger than 0.90 and the Macro F1 is larger than 0.85, the fault diagnosis model is proved to have good effect, and the current fault diagnosis model is continuously used; otherwise, the effect of the fault diagnosis model is proved to be poor, the step S3 is returned, and the training is carried out again to obtain a new fault diagnosis model;
meanwhile, the higher the scores of Micro F1 and Macro F1 are, the better the fault diagnosis model is.
The invention has the beneficial effects that: by adopting the multi-feature fusion method, more dynamic information is provided for fault diagnosis, the accuracy of fault diagnosis is improved, the problem that the fault diagnosis efficiency is reduced due to the fact that the operation amount is increased due to multi-feature fusion is solved, the requirements of real-time monitoring and diagnosis of the actual industry can be met, and the method specifically comprises the following steps:
extracting trend signals, binary alarm signals and slow characteristic signals from historical process signals, forming multi-characteristic fusion data with the historical process signals, and then adopting a width learning system as a neural network for processing the multi-characteristic fusion data; because the modeling of the width learning system does not depend on a deeper network structure, the network weight calculation is fast, the model building time is short, and the requirements of real-time monitoring and diagnosis of the actual industry can be met.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.
Claims (7)
1. A fault monitoring and diagnosing method based on multi-feature fusion and width learning is characterized in that: the method comprises the following steps:
s1, acquiring a historical process signal with a fault label in the actual industrial process; the historical process signals include fault signals and normal signals;
s2, performing multi-feature extraction based on the historical process signal to construct a multi-feature vector;
in step S2, based on the mechanism analysis, dividing the variables in the historical process signals into control variables, continuous measurement variables, and state variables, and extracting binary alarm signals in the control variables and slow characteristic signals, qualitative trend signals, and binary alarm signals in the continuous measurement variables, specifically:
s21, extracting the trend characteristics of the continuous measurement variables in the historical process signals by using a qualitative trend analysis method to obtain qualitative trend signals;
s22, calculating upper and lower alarm thresholds based on the normal signals in the historical process signals in the step S1, and acquiring binary alarm signals based on the upper and lower alarm thresholds;
s23, extracting slow characteristics of continuous measurement variables in the historical process signals by using a slow characteristic analysis method to obtain slow characteristic signals;
the multi-feature vector is a vector formed by the historical process signal, the qualitative trend signal, the binary alarm signal and the slow feature signal;
s3, inputting the multi-feature vectors into a width learning network for training to obtain a fault diagnosis model;
and S4, applying the fault diagnosis model to the fault diagnosis of the actual industrial process.
2. The fault monitoring and diagnosis method based on multi-feature fusion and width learning as claimed in claim 1, wherein: in step S21, the specific process of acquiring the qualitative trend signal is as follows:
(1) preprocessing historical process signals by utilizing wavelet transformation, removing noise of local signals, and then performing Min-Max normalization processing to obtain preprocessed historical process signals;
(2) calculating a sample difference value of the preprocessed historical process signal;
(3) and based on a qualitative trend analysis method, extracting the qualitative trend of the preprocessed historical process signal by using the sample difference value to obtain a qualitative trend signal.
3. The fault monitoring and diagnosis method based on multi-feature fusion and width learning as claimed in claim 2, characterized in that: the specific implementation process of the step (3) is as follows:
when the sample difference is less than- δ, the qualitative trend is a decrease, the qualitative trend signal being represented by a value of-1;
when the sample difference value is greater than or equal to-delta and less than or equal to delta, the qualitative trend is unchanged, and the qualitative trend signal is represented by a value of 0;
when the sample difference is greater than δ, the qualitative trend is increasing, and the qualitative trend signal is represented by a value of 1, where δ ∈ [0.01,0.05 ].
4. The fault monitoring and diagnosis method based on multi-feature fusion and width learning as claimed in claim 1, wherein: in step S22, if the normal signal in the historical process signal has an upper and lower alarm threshold set based on the priori knowledge, taking the upper and lower alarm threshold as the upper and lower alarm threshold; if the upper and lower alarm thresholds set based on the priori knowledge do not exist, processing according to the type of the normal signals of the historical process signals, specifically:
if the normal signals of the historical process signals conform to Gaussian distribution, calculating the mean value and the standard deviation of the normal signals in the historical process signals according to a 3 sigma principle, and adopting the standard deviation with adjustable mean value plus-minus multiple as the upper and lower alarm threshold values; if the normal signals of the historical process signals do not accord with Gaussian distribution, estimating a probability density function of the normal signals by using a kernel density estimation method, and calculating the upper and lower alarm thresholds based on the probability density and the significance level alpha; wherein α has a value of 0.005 or 0.01 or 0.05;
generating a binary alarm signal composed of values 0 and 1 by judging whether the historical process signal exceeds the upper and lower alarm thresholds, specifically: when the historical process signal exceeds the upper and lower alarm threshold values, the binary alarm signal is 1 to indicate that a fault occurs, otherwise, the binary alarm signal is 0 to indicate that no fault occurs.
5. The fault monitoring and diagnosis method based on multi-feature fusion and width learning as claimed in claim 1, characterized in that: the specific implementation process of step S3 is as follows:
s31, forming the multi-feature vector into a first input matrix XtrainAnd for the first input matrix XtrainCarrying out Min-Max normalization processing to obtain a processed first input matrix Xtrain;
S32, based on the processed first input matrix X in the step S31trainGenerating an input node, wherein the input node activation function is as follows:
ZI=[Φ[XtrainWei+βei]]
where Φ is the linear activation function, βeiIs a bias matrix of input nodes, WeiIs XtrainWeight matrix of, and βeiAnd WeiIn (b) is [0,1 ]]Is randomly generated, input node ZI=[Z1,Z2,...Zi]I represents the total number of input nodes;
s33, input node ZILinear combination is carried out to generate an enhanced node HJ=[H1,H2,…,Hj]The enhanced node activation function is:
HJ=[ζ[ZIWhj+βhj]]
where j represents the total number of booster nodes, ζ is the tanh activation function, WhjIs ZIWeight matrix of betahjIs HJAnd W is a bias matrix ofhjAnd betahjThe value of (1) is [0]Randomly generating;
s34, inputting the node Z in the step S32IAnd enhanced node H in step S33JForming a second input matrix A, which is an input of the neural network, wherein A ═ ZI|HJ];
S35, calculating the pseudo inverse A of the second input matrix A according to the formula+The formula is as follows:
wherein λ is a constant, E is an identity matrix, ATRepresents the transpose of a;
s36, based on the pseudo-inverse A in step S35+Calculating a weight matrix, wherein the concrete formula is as follows:
Wn=A+Yn
wherein, WnIs a weight matrix, i.e. the weight of the width learning network, YnA matrix of fault signatures representing historical process signals.
6. The fault monitoring and diagnosis method based on multi-feature fusion and width learning as claimed in claim 5, characterized in that: in step S4, the fault diagnosis model is applied to fault diagnosis in an actual industrial process, and the specific process is as follows:
s41, collecting process signals in an actual industrial process on line, and obtaining binary alarm signals of the process signals by judging whether the process signals exceed the upper alarm threshold and the lower alarm threshold;
s42, based on the binary alarm signal in step S41, determining whether a fault occurs, if so, extracting the qualitative trend signal and the slow characteristic signal of the process signal, and generating a multi-characteristic data matrix X composed of the process signal, the binary alarm signal, the qualitative trend signal and the slow characteristic signaltestStep S43 is executed; otherwise, judging the data to be normal data, namely, not generating faults;
s43, converting the multi-feature data matrix XtestInputting the fault diagnosis model to obtain an output matrix Yn,testI.e. is the corresponding fault label, where Yn,test=XtestWn,WnIs a weight matrix;
s44, comparing the fault label in the step S43 with the actual fault label, and adopting an evaluation index F1-score evaluating the diagnostic performance of said fault diagnosis model, if Micro F1 is greater than a first set threshold and Macro F1 is greater than a second set threshold, continuing to use said fault diagnosis model; otherwise, returning to step S3, performing training again to obtain a new fault diagnosis model.
7. The fault monitoring and diagnosis method based on multi-feature fusion and width learning as claimed in claim 6, characterized in that: the first set threshold is 0.90, and the second set threshold is 0.85.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110020406.6A CN113159088B (en) | 2021-01-07 | 2021-01-07 | Fault monitoring and diagnosis method based on multi-feature fusion and width learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110020406.6A CN113159088B (en) | 2021-01-07 | 2021-01-07 | Fault monitoring and diagnosis method based on multi-feature fusion and width learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113159088A CN113159088A (en) | 2021-07-23 |
CN113159088B true CN113159088B (en) | 2022-07-15 |
Family
ID=76878369
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110020406.6A Active CN113159088B (en) | 2021-01-07 | 2021-01-07 | Fault monitoring and diagnosis method based on multi-feature fusion and width learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113159088B (en) |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113722140B (en) * | 2021-08-30 | 2023-08-04 | 中国地质大学(武汉) | Industrial alarm flooding source diagnosis method based on small sample learning and storage medium |
CN115891741B (en) * | 2022-09-30 | 2023-09-22 | 南京邮电大学 | Remote fault early warning method and device suitable for electric automobile charging process |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102520341A (en) * | 2011-12-05 | 2012-06-27 | 南京航空航天大学 | Analog circuit fault diagnosis method based on Bayes-KFCM (Kernelized Fuzzy C-Means) algorithm |
CN106647718A (en) * | 2017-01-20 | 2017-05-10 | 中国石油大学(华东) | Non-linear industrial process fault detection method based on Bayes kernel slow feature analysis |
CN107341504A (en) * | 2017-06-07 | 2017-11-10 | 同济大学 | A kind of Trouble Diagnostic Method of Machinery Equipment based on the popular study of time series data |
CN109828552A (en) * | 2019-02-22 | 2019-05-31 | 北京工业大学 | A kind of batch process Fault monitoring and diagnosis method based on width learning system |
CN111255674A (en) * | 2020-01-21 | 2020-06-09 | 武汉瑞莱保能源技术有限公司 | System and method for detecting state of rotating mechanical equipment |
-
2021
- 2021-01-07 CN CN202110020406.6A patent/CN113159088B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102520341A (en) * | 2011-12-05 | 2012-06-27 | 南京航空航天大学 | Analog circuit fault diagnosis method based on Bayes-KFCM (Kernelized Fuzzy C-Means) algorithm |
CN106647718A (en) * | 2017-01-20 | 2017-05-10 | 中国石油大学(华东) | Non-linear industrial process fault detection method based on Bayes kernel slow feature analysis |
CN107341504A (en) * | 2017-06-07 | 2017-11-10 | 同济大学 | A kind of Trouble Diagnostic Method of Machinery Equipment based on the popular study of time series data |
CN109828552A (en) * | 2019-02-22 | 2019-05-31 | 北京工业大学 | A kind of batch process Fault monitoring and diagnosis method based on width learning system |
CN111255674A (en) * | 2020-01-21 | 2020-06-09 | 武汉瑞莱保能源技术有限公司 | System and method for detecting state of rotating mechanical equipment |
Non-Patent Citations (3)
Title |
---|
Deep learning technique for process fault detection and diagnosis in the presence of incomplete data;Cen Guo et al.;《Chinese Journal of Chemical Engineering》;20200620;第2358–2367页 * |
Quantitative trend fault diagnosis of a rolling bearing based on Sparsogram and Lempel-Ziv;Lingli Cui et al.;《Measurement》;20180625;第410–418页 * |
多传感器信息融合的燃气轮机状态监测***;尚文 等;《机械设计与制造》;20140331(第3期);第171-174页 * |
Also Published As
Publication number | Publication date |
---|---|
CN113159088A (en) | 2021-07-23 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Wang et al. | CHMM for tool condition monitoring and remaining useful life prediction | |
CN105738109B (en) | Bearing fault classification diagnosis method based on rarefaction representation and integrated study | |
Zou et al. | Bearing fault diagnosis based on combined multi-scale weighted entropy morphological filtering and bi-LSTM | |
CN110309603B (en) | Short-term wind speed prediction method and system based on wind speed characteristics | |
CN108445752B (en) | Random weight neural network integrated modeling method for self-adaptively selecting depth features | |
CN113642754B (en) | Complex industrial process fault prediction method based on RF noise reduction self-coding information reconstruction and time convolution network | |
CN113159088B (en) | Fault monitoring and diagnosis method based on multi-feature fusion and width learning | |
CN112116198B (en) | Data-driven process industrial state perception network key node screening method | |
Tran et al. | Algorithm for estimating online bearing fault upon the ability to extract meaningful information from big data of intelligent structures | |
CN112904810B (en) | Process industry nonlinear process monitoring method based on effective feature selection | |
CN117290800B (en) | Timing sequence anomaly detection method and system based on hypergraph attention network | |
Li et al. | Gear pitting fault diagnosis using raw acoustic emission signal based on deep learning | |
CN112504682A (en) | Chassis engine fault diagnosis method and system based on particle swarm optimization algorithm | |
CN111709577B (en) | RUL prediction method based on long-range correlation GAN-LSTM | |
CN110766060A (en) | Time series similarity calculation method, system and medium based on deep learning | |
CN111222689A (en) | LSTM load prediction method, medium, and electronic device based on multi-scale temporal features | |
CN114897138A (en) | System fault diagnosis method based on attention mechanism and depth residual error network | |
Zhang et al. | MS-TCN: A multiscale temporal convolutional network for fault diagnosis in industrial processes | |
CN114861349A (en) | Rolling bearing RUL prediction method based on model migration and wiener process | |
CN113762967A (en) | Risk information determination method, model training method, device, and program product | |
CN116680639A (en) | Deep-learning-based anomaly detection method for sensor data of deep-sea submersible | |
Liu et al. | Prognostics for rotating machinery using variational mode decomposition and long short-term memory network | |
Wang et al. | Explainable machine learning for motor fault diagnosis | |
CN114298413A (en) | Hydroelectric generating set runout trend prediction method | |
Liu et al. | Online anomaly detection with streaming data based on fine-grained feature forecasting |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |