CN113485863B - Method for generating heterogeneous imbalance fault samples based on improved generation of countermeasure network - Google Patents
Method for generating heterogeneous imbalance fault samples based on improved generation of countermeasure network Download PDFInfo
- Publication number
- CN113485863B CN113485863B CN202110796968.XA CN202110796968A CN113485863B CN 113485863 B CN113485863 B CN 113485863B CN 202110796968 A CN202110796968 A CN 202110796968A CN 113485863 B CN113485863 B CN 113485863B
- Authority
- CN
- China
- Prior art keywords
- fault
- data
- heterogeneous
- sample
- continuous
- 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
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F11/00—Error detection; Error correction; Monitoring
- G06F11/07—Responding to the occurrence of a fault, e.g. fault tolerance
- G06F11/0703—Error or fault processing not based on redundancy, i.e. by taking additional measures to deal with the error or fault not making use of redundancy in operation, in hardware, or in data representation
- G06F11/079—Root cause analysis, i.e. error or fault diagnosis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F11/00—Error detection; Error correction; Monitoring
- G06F11/07—Responding to the occurrence of a fault, e.g. fault tolerance
- G06F11/0703—Error or fault processing not based on redundancy, i.e. by taking additional measures to deal with the error or fault not making use of redundancy in operation, in hardware, or in data representation
- G06F11/0706—Error or fault processing not based on redundancy, i.e. by taking additional measures to deal with the error or fault not making use of redundancy in operation, in hardware, or in data representation the processing taking place on a specific hardware platform or in a specific software environment
- G06F11/0736—Error or fault processing not based on redundancy, i.e. by taking additional measures to deal with the error or fault not making use of redundancy in operation, in hardware, or in data representation the processing taking place on a specific hardware platform or in a specific software environment in functional embedded systems, i.e. in a data processing system designed as a combination of hardware and software dedicated to performing a certain function
- G06F11/0739—Error or fault processing not based on redundancy, i.e. by taking additional measures to deal with the error or fault not making use of redundancy in operation, in hardware, or in data representation the processing taking place on a specific hardware platform or in a specific software environment in functional embedded systems, i.e. in a data processing system designed as a combination of hardware and software dedicated to performing a certain function in a data processing system embedded in automotive or aircraft systems
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F11/00—Error detection; Error correction; Monitoring
- G06F11/22—Detection or location of defective computer hardware by testing during standby operation or during idle time, e.g. start-up testing
- G06F11/26—Functional testing
- G06F11/263—Generation of test inputs, e.g. test vectors, patterns or sequences ; with adaptation of the tested hardware for testability with external testers
-
- 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
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- Quality & Reliability (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Probability & Statistics with Applications (AREA)
- Computer Hardware Design (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Test And Diagnosis Of Digital Computers (AREA)
- Testing And Monitoring For Control Systems (AREA)
Abstract
The invention provides a method for generating heterogeneous imbalance fault samples based on improved generation of an countermeasure network. The hybrid automatic encoder and the double discriminators are used for generating an countermeasure network MAE-D2GAN, aiming at the serious unbalance problem of monitoring data, the discriminators obtained by training normal state samples are added into the countermeasure process, and an improved generation countermeasure network with the double discriminators is established; designing an automatic encoder comprising an encoder and a decoder, encoding heterogeneous monitoring data into continuous potential features through the encoder, and inputting the continuous potential features into a double discriminator to generate an countermeasure network model; the decoder output layer uses a micro Gumbel-softmax to process the double discriminators corresponding to the discrete variables to generate continuous features against network generation, and heterogeneous fault samples are obtained. The fault sample generated by the hybrid automatic encoder and the double discriminators is closer to the real fault sample, so that the influence of the class overlap problem can be effectively reduced, and the prediction accuracy of the fault diagnosis model is better improved.
Description
Technical Field
The present application relates to the field of fault diagnostics, and in particular to the handling of class imbalance problems in complex system or equipment fault diagnostics, employing improved generation countermeasure networks to enable data expansion of heterogeneous fault samples.
Background
With the rapid development of information technology, the intelligent and comprehensive degree of the system is gradually improved, and the structure and the function of the system are more and more complex. The cross influence of each component in the complex system can cause a chain reaction due to a small fault, so that the whole system is damaged. This not only brings great economic loss, but also endangers the life safety of the relevant personnel. Therefore, the state monitoring and fault diagnosis technology is becoming more and more important as a predictive maintenance means in system health and safety management. Currently, there are many methods for fault diagnosis, such as expert system models, physical models, data-driven models, and the like. The data driving method is widely applied to fault diagnosis of a complex system. By adopting the data driving model, the abnormality inspection model can be obtained quickly and cheaply only by enough related monitoring data and maintenance data, and the dependence on the priori knowledge of the equipment can be avoided. Although the method needs massive data to obtain a high-precision model, the defect of the data volume can be overcome by an optimization algorithm, simulation data and reinforcement learning.
Most data-driven fault diagnosis methods assume that the data sets are evenly distributed, i.e. the number of samples of different categories is close. However, the data in practical use is often unbalanced, and particularly for a complex system or apparatus with high reliability, the faulty sample is inevitably far less than the normal sample. When these data-driven classification algorithms are used directly for fault diagnosis, it is difficult to obtain satisfactory results. The prediction results tend to be biased toward most categories, so that the accuracy of fault diagnosis is very low. However, in practical applications, fault class data is clearly more important. Thus, the deviations caused by unbalanced data must be overcome in the face of this.
Currently, there are three main types of methods for processing unbalanced data:
(1) The resampling method comprises the following steps: such as majority undersampling, minority oversampling, and comprehensive sampling;
(2) The algorithm level method comprises the following steps: such as modifying a loss function, modifying a classification threshold, cost sensitive learning, etc.;
(3) And (3) ensemble learning: such as iterative training using enhanced combinations, etc.
Most of the methods can improve the classification accuracy to a certain extent, and can be verified and popularized in various fields. Among them, the synthetic minority sample technique (Synthetic Minority Oversampling Technique, SMOTE) is a common method, and the data distribution is adjusted by adding the synthetic minority sample, so as to improve the classification performance. Recently proposed generation antagonism networks (Generative Adversarial Network, GAN) are also used for the generation of synthetic samples due to their efficiency and flexibility. Unlike SMOTE and its variants, which rely primarily on expert knowledge to design rules of generation that synthesize minority groups, GAN methods can automatically learn their inherent distribution and generate minority samples that are similar to real samples. A GAN includes two variable networks: a generator and a discriminator, denoted G and D, respectively, are trained in GAN to play with each other. The sample generated by the generator G is judged and evaluated by the discriminator D, and then the generator G is optimized according to the evaluation result, so that the efficiency and quality of the sample generation process can be greatly improved. Currently, GAN and its variants have been successfully applied to various fields such as image restoration, scene synthesis, face recognition, etc., and GAN-based studies have been gradually developed in recent years for unbalanced data in fault diagnosis.
Previous studies have focused mainly on continuous and high sampling frequency data, whereas actual data often contain both numerical and classification variables of low sampling frequency, so-called heterogeneous data. For example, factors affecting a brake system of a high-speed train include numerical variables such as voltage and current, and the like, as well as variables such as operation mode and brake state.
In existing GANs, the type variable, known as the discrete variable, is often encoded as a numerical value using one-time thermal encoding, but this leaves the resulting sample without any engineering interpretation, the value no longer discrete and may exceed the value range of the original variable. In addition to the heterogeneity of data, previous work mostly assumed that there were enough failure samples available for training the GAN model, while failure data were limited for a complex system. Training the resulting GAN model on limited failure samples often results in serious overfitting problems.
Disclosure of Invention
To overcome the deficiencies of the prior art, it is an object of the present invention to propose a method of generating heterogeneous fault samples with an improved generation countermeasure network with dual discriminators in combination with an improved Auto Encoder (AE). In this method, discrete and continuous variables can be automatically identified and processed separately. The improved AE includes an encoder for extracting the valid latent variables and feeding them to the GAN with double discriminators and a decoder for converting the generated continuous data into heterogeneous data.
To achieve the above object, in one aspect, the present invention provides a method for generating heterogeneous imbalance fault samples against network generation based on improvement, which includes the steps of:
step 1: collecting fault states and monitoring data of an electromechanical product system or device, and obtaining historical monitoring data;
step 2: performing data preprocessing on the historical monitoring data obtained in the step 1 to obtain a preprocessed data set [ X, Y ], wherein X is a related variable of a fault state of an electromechanical product system or equipment, Y is a fault state of the electromechanical product system or equipment, X is heterogeneous data, and the X comprises a continuous variable and a discrete variable;
step 3: setting an automatic encoder to realize mutual conversion of heterogeneous data and continuous characteristics, adding a normal sample discriminator into an existing classical generation countermeasure network, establishing an improved generation countermeasure network with double discriminators, inputting the preprocessed data set [ X, Y ] obtained in the step 2 into the automatic encoder and the improved generation countermeasure network with double discriminators to obtain heterogeneous fault samples U, and specifically comprising the following steps of:
step 31: the automatic encoder includes an encoder and a decoder; the function of the encoder is to encode heterogeneous monitoring data into successive latent features; the decoder output layer uses a micro Gumbel-softmax to realize discretization of continuous features, the function of the decoder is to decode continuous fault feature data into heterogeneous fault samples, and the preprocessed data set [ X, Y ] obtained in the step 2 is input into the encoder to obtain fault sample feature data X and normal sample feature data;
step 32: said step 3 improved generation countermeasure network with double discriminators comprises a generator G and two discriminators D and F; training the discriminator F based on the normal sample feature data obtained in the step 31; inputting white noise z into a generator G of the improved generation countermeasure network with double discriminators, and obtaining a characteristic data set G (z) corresponding to a potential space fault sample;
step 33: training the generator G and the discriminator D according to the fault sample characteristic data x and the characteristic data set G (z) corresponding to the fault samples in potential space, wherein the improved generation of the loss function V (D, G) with the double discriminators in the training process is as follows:
wherein: d (x) is the probability that the generated samples are from the fault sample feature data x;
solving expected values for all x; d (G (z)) is a probability that the discriminator D discriminates whether or not the continuous variable set G (z) is true; f (G (z)) is a probability that the discriminator F discriminates whether or not the continuous variable set G (z) is normal; />
Calculating expected values for all z set by the taking pass;
the objective function of the improved generation countermeasure network with double discriminators is
Based on the objective function->
Iteratively training the generator G and discriminator D, the iterative training comprising training the generator G to minimize the loss function V (D, G) and training the discriminator D to maximize the loss function V (D, G); when the difference between the maximum value and the minimum value of the loss function V (D, G) in 100 continuous iterative training of the model is smaller than 0.01, the iterative training is stopped; obtaining a trained generator G after the iterative training is stopped; inputting a group of white noise z into the trained generator G to obtain a characteristic data set G (z) corresponding to a potential space fault sample;
step 34: inputting the characteristic data set G (z) corresponding to the potential space fault sample obtained in the step 33 into a decoder in the automatic encoder to obtain a heterogeneous fault sample U;
step 4: and adding a column of fault state I to the heterogeneous fault sample U obtained in the step 34 to obtain a heterogeneous fault sample set [ U, I ], wherein the fault state I is a column vector with elements of 1, and the dimension of the fault state I is consistent with the dimension of the heterogeneous fault sample U.
Further, the specific steps of obtaining the heterogeneous fault sample U in the step 34 are: inputting the characteristic data set G (z) corresponding to the potentially spatially faulty sample obtained in the step 33 into the decoder to obtain a decoded data set, where the decoded data set includes a column corresponding to a discrete variable and a column corresponding to a continuous variable, and performing discretization on the column corresponding to the discrete variable by using a reducible gummel-softmax to obtain a discretized column, and obtaining a heterogeneous faulty sample U according to the column corresponding to the continuous variable and the discretized column.
Further, the encoding process by the encoder in the step 31 specifically includes: and (3) performing single-heat coding on discrete variables in the preprocessed data set [ X, Y ] obtained in the step (2) to obtain single-heat coded variables, and uniformly coding the single-heat coded variables and continuous variables in the preprocessed data set [ X, Y ] obtained in the step (2) to obtain fault sample characteristic data X and normal sample characteristic data.
Preferably, the fault state in step 1 includes faulty and non-faulty.
Preferably, the monitoring data in step 1 is heterogeneous data, and the heterogeneous data includes continuous features, discrete features and signal features.
Further, the data preprocessing in the step 2 comprises filling missing values, replacing abnormal values, processing dimension differences and digitizing; the missing value filling adopts a local mean filling method; the process dimension gap is data normalization, and the data normalization adopts a z-score method.
The invention also provides a method for diagnosing real-time faults according to the method for generating heterogeneous imbalance fault samples of the countermeasure network based on improvement, which specifically comprises the following steps:
step 1: collecting fault states and monitoring data of an electromechanical product system or device, and obtaining historical monitoring data;
step 2: performing data preprocessing on the historical monitoring data obtained in the step 1 to obtain a preprocessed data set [ X, Y ], wherein X is a related variable of a fault state of an electromechanical product system or equipment, Y is a fault state of the electromechanical product system or equipment, X is heterogeneous data, and the X comprises a continuous variable and a discrete variable;
step 3: setting an automatic encoder to realize the mutual conversion of heterogeneous data and continuous characteristics, adding a normal sample discriminator into the existing classical generation countermeasure network, establishing an improved generation countermeasure network with double discriminators, inputting the preprocessed data set [ X, Y ] obtained in the step 2 into the automatic encoder and the improved generation countermeasure network with double discriminators to obtain heterogeneous fault samples U;
step 4: adding a column of fault state I to the heterogeneous fault sample U obtained in the step 3 to obtain a heterogeneous fault sample set [ U, I ], wherein the fault state I is a column vector with elements of 1, and the dimension of the fault state I is consistent with the dimension of the heterogeneous fault sample U;
step 5: the heterogeneous fault sample set [ U, I ] obtained in the step 4 is processed]Incorporating the preprocessed data set [ X, Y ] obtained in said step 2]In obtaining a new data set
Based on the new dataset +.>
Training an artificial neural network ANN algorithm to obtain a fault diagnosis model;
step 6: and (3) collecting real-time monitoring data, preprocessing the real-time monitoring data with the same data in the step (2) to obtain preprocessed real-time monitoring data, inputting the preprocessed real-time monitoring data into the fault diagnosis model obtained in the step (5) to obtain a real-time fault state, and completing fault diagnosis.
Compared with the prior art, the invention provides a method for generating heterogeneous imbalance fault samples against network generation based on improvement, which has the beneficial effects that:
1. the AE provided by the invention can automatically identify and simultaneously process continuous variables and discrete variables, and the discrete variables are reduced by using a micro Gumbel-softmax, so that an optimal model is easier to obtain;
2. the GAN with the double discriminators is used for data expansion, so that the influence of the class overlapping problem can be effectively reduced;
3. the hybrid automatic encoder and the double discriminator established by the method generate a countermeasure network (Mixed AE-2Discriminators GAN,MAE-D2 GAN) for fault diagnosis, and have the beneficial effects that: the generated heterogeneous fault sample is more real and effective, accords with engineering practice, and can better improve the fault diagnosis precision;
4. meanwhile, the method not only can be used for fault diagnosis, but also can be used for other operation and maintenance scenes, and has a good use prospect.
Drawings
FIG. 1 is a block diagram of an improved generation countermeasure network with dual discriminators for generating heterogeneous fault samples in accordance with an embodiment of the present invention;
FIG. 2 is a block diagram of an AE according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a fault diagnosis model based on an artificial neural network ANN algorithm in an embodiment of the present invention;
FIG. 4 is a flow chart of a 10-fold cross-validation experiment in this embodiment of the invention;
FIG. 5 is a two-dimensional projection of the fault status monitoring data of a high-speed rail brake system according to the present embodiment of the invention.
Detailed Description
Hereinafter, embodiments of the present invention will be described with reference to the drawings.
The embodiment of the invention provides a real-time fault diagnosis method for an electromechanical product system or equipment based on width learning, which comprises the following specific steps:
step 1: collecting fault states of electromechanical product systems or equipment within a period of time and relevant monitoring data as historical monitoring data;
typically, the collected system or equipment fault condition monitoring data is heterogeneous data, including continuous features, discrete features, and signal features.
Step 2: and carrying out data preprocessing on the historical monitoring data, including missing values, abnormal values, dimension differences, digital processing and the like, so as to obtain a preprocessed data set [ X, Y ]. Where X is an electromechanical product system or equipment fault state related variable, X is heterogeneous data, including continuous and discrete variables; y is the fault state of the electromechanical product system or equipment, and the monitoring data can be converted into numerical data by extracting key indexes of signal characteristics and digitizing discrete characteristics. The present application is based on the converted numeric data for research to include numeric data of continuous and discrete features as a preprocessed data set.
If the original fault signature data set contains fault signatures of only one state, then it is not valuable to diagnose whether the system is faulty from the data, so that such fault signatures need to be deleted and subsequent research is performed based on the remaining discrete data fault signatures and continuous data fault signatures.
Firstly, data cleaning and conversion are needed to be carried out on historical monitoring data, such as filling missing values, replacing abnormal values and processing dimension gaps. The filling of the missing values adopts a local mean filling method, namely, the current missing values are filled by the mean value of the data of the first 5 rows and the last 5 rows, which are adjacent to the current monitoring data; the process dimension gap, i.e., data normalization, takes the z-score method, i.e., subtracting the mean value and dividing by the standard deviation.
Various faults can occur in the system or the equipment, and only faults and no faults are distinguished in the invention, and no fault type distinction is performed.
Classical generation of the countermeasure network GAN can well process continuous data, and an optimal model is obtained through first-order differentiation, but parameters are difficult to update for discrete variables, and creating a GAN model for heterogeneous class unbalanced data faces more challenges, such as difficulty in obtaining an optimal model, non-compliance of a generated sample, and the like. In view of this, the present invention proposes a method to combine the improved GAN and AE with dual discriminators and the differentiable gummel-softmax function to generate high quality heterogeneous fault samples. The improved GAN with double discriminators is obtained by adding a discriminator F on the basis of classical GAN, wherein the discriminator F is used for discriminating whether the generated sample is normal or not, the sample is obtained only by training based on normal sample characteristic data, only loss values are provided when the improved generation countermeasure network with double discriminators is trained, and the discriminator F does not perform iterative optimization.
Step 3: the improvement of training with double discriminators after encoding using AE generates double discriminators D and F in the countermeasure network from the preprocessed data set [ X, Y ], wherein F is optimized based on only normal sample feature data, which is feature data obtained after AE encoding of normal samples in the preprocessed data set [ X, Y ], and the D and G combinations are iteratively optimized based on the generated continuous data set and the faulty sample feature data. The fault sample characteristic data X is the characteristic data of the fault samples in the preprocessed data set [ X, Y ] after AE coding. The input of white noise z into a generator G with dual discriminators to generate a continuous data set G (z) of potential feature space against the network. Converting the generated continuous data set G (z) into a heterogeneous data set U using a decoder in AE;
under normal conditions, the d-dimensional vector h can be normalized by using a softmax function to obtain each classification in the vector, and the sampling result is an element h 1 ,h 2 ,...,h i ,...,h d To convert h to a probability represented at h 1 ,h 2 ,...,h i ,...,h d Discrete variables of the samples.
h＝[h 1 ,h 2 ,...,h i ,...,h d ] (1)
y＝softmax(h)＝[p 1 ,p 2 ,...,p i ,...,p d ] (2)
Wherein: y is a vector of d-dimensional vector h normalized by softmax function; p is p i Is the i-th element in the normalized vector.
softmax tends to make the probability of the largest element in h significantly greater than the other elements, but the vector p representing the probability has virtually no probabilistic significance and is not differentiable based on softmax output during model training.
The micro-function Gumbel-softmax is proposed based on the softmax transformation, and parameter updating can be enabled to be towards an optimal solution, so that a generated sample is more real and effective, and a vector h is obtained by normalizing the Gumbel-softmax function:
wherein: y ' is a vector of d-dimensional vector h normalized by Gumbel-softmax function; g= [ g ] 1 ,g 2 ,...,g i ,...,g d ]，g i (i=1, 2, …, d) independent of each other and obeys gummel distribution; τ is the control parameter for softness in the Gumbel-softmax function, y' →y when τ→0. When τ→infinity, y' approximately satisfies a uniform distribution.
The invention adopts an encoder and a decoder to form AE to reconstruct the characteristics of the collected historical monitoring data, namely, the heterogeneous monitoring data is encoded into continuous variables and extracted, and the dimension of the monitoring data is reduced by eliminating redundant information. In order to encode and decode both continuous and discrete variables in an automatic encoder, the present invention employs a hybrid model as shown in fig. 2. Also, to prevent information loss, the discrete variable is converted to a multi-dimensional vector based on one-hot encoding, and then to generate discrete data, the decoder output needs to be one-hot vector, which is typically implemented by a softmax function in the output layer. The present invention uses a micronized Gumbel-softmax function instead of the classical softmax function.
In order to make the generated fault sample similar to the actual fault data and different from the actual normal data, the countermeasure for generating the countermeasure network adopts a two-dimensional design, namely, one countermeasure judges whether the generated sample is true or not, the other countermeasure judges whether the generated sample is normal or not, as shown in fig. 1, an improved generation countermeasure network structure diagram for generating heterogeneous fault samples is shown, and an additional item is added in a classical loss function so as to prevent the generated data from aggravating the class overlapping problem.
The improvement with double discriminators generates the loss function of the antagonism network as:
wherein: x is fault sample characteristic data, z is Gaussian white noise input into a generator G, G (z) is a generated continuous data set, namely a characteristic data set corresponding to a fault sample in a potential space, D (x) is probability of generating the sample from the fault sample characteristic data x, D (G (z)) is probability of distinguishing whether G (z) is true or not by D, and F (G (z)) is probability of distinguishing whether G (z) is normal or not by F.
To take all x's, we find the expected value +.>
All z values set are referred to as expected values.
Establishing an improved generation countermeasure network with dual discriminators as an objective function
Training generator G minimizes the trend of the loss function, training discriminator D maximizes the trend of the loss function until the loss function V (D, G) stops training when the difference between the maximum and minimum values in 100 consecutive iterations is less than 0.01, and the resulting generator G can be used to generate a feature data set G (z) corresponding to a potentially spatially faulty sample.
Step 4: and (3) decoding G (z) by using AE to obtain a heterogeneous fault feature set U, adding a column of fault labels to the U to obtain a generated heterogeneous fault sample set [ U, I ], wherein I is a fault state, I is a column vector with elements of 1, and the dimension of the fault state I is consistent with the number of the fault samples U.
The hybrid automatic encoder and the double discriminator created by the method of steps 1-4 generate a countermeasure network (Mixed AE-2Discriminators GAN,MAE-D2 GAN) for fault diagnosis, and the specific steps are as follows:
heterogeneous fault sample set [ U, I ] generated in step 4]Incorporating pre-processed datasets [ X, Y ]]In, a new data set is obtained
Based on the new data set training artificial neural network ANN model, a fault diagnosis model is obtained, and the method can effectively solve the problem of unbalanced category and improve the fault diagnosis precision.
The artificial neural network ANN model structure is shown in fig. 3. The number of nodes of the input layer is the same as the number of columns of X, only one node exists in the output layer in the two-class fault diagnosis scene, the number of layers and the number of nodes of the hidden layer are approximately set to be complex according to the sample size and the feature number, and then the structure is manually adjusted to obtain an approximate local optimal solution.
And (3) collecting real-time monitoring data related to the fault state of the system or the equipment, preprocessing the real-time monitoring data in the same way as the step (2) to obtain preprocessed real-time monitoring data, inputting the preprocessed real-time monitoring data into a fault diagnosis model to obtain the real-time fault state of the product system or the equipment, and completing fault diagnosis.
In order to verify the effectiveness of the proposed method, based in part on a monitoring dataset of a high speed train brake system during one year of operation, a comparative experiment was designed to verify that the method was more effective than the usual oversampling method. The comparison method comprises the following steps: random oversampling, integrated oversampling technique SMOTE, marginal oversampling technique SMOTE-borperline 1, marginal oversampling technique SMOTE-borperline 2, ADASYN, and addition of a single thermally encoded classical GAN. The experimental flow is shown in fig. 4, and in 10-fold cross validation, the actual training data of each fault diagnosis model is a data set composed of original training data and generated data.
The brake system is a device for ensuring that the high-speed train can effectively decelerate, and is one of the most important components of the high-speed train. Data relating to the state of health of the brake system may be collected by locating sensors, capturing state information. The monitoring dataset contains 43 variables that may lead to failure of the braking system, including 21 continuous variables and 22 discrete variables such as speed, voltage, current, temperature, mode of operation, braking status, etc.
Data cleaning and conversion of raw data collected by the sensor, such as filling in missing values, data normalization, are required before a new faulty sample is generated using a different method. Data normalization is to balance the dimensional differences between numerical variables. Furthermore, some variables in the raw data need to be converted to numerical values according to the data requirements of the model. After data processing, 43 variables in the input data all become numerical variables between [0,1 ].
To avoid being affected by random factors, the experimental observation shows the average precision and the running time of the ten-fold cross validation. In order to obtain the approximate local optimal solution, all models are trained and optimized through gradually increasing structural complexity and iteration times until the generalization precision basically keeps stable or decreases. Under the same operating environment, the model outputs the average F1-score, the average G-mean and the average AUC of the fault classes in the test results through 10 times of cross validation.
FIG. 5 is a two-dimensional projection of a processed dataset embedded in t-SNE based on a t-distribution random neighborhood, 1.0 representing a faulty sample and 0.0 representing a normal sample. the t-SNE algorithm is one of the most common and efficient methods of analyzing high-dimensional data visualization. This method converts the similarity between data points into probabilities, and by projecting the high-dimensional data into a two-dimensional or three-dimensional space, visualization of the high-dimensional data can be achieved. Since the faulty sample will be completely covered with all data, only 10% of the normal sample is shown in the figure. By comparison, the 10% normal samples in the graph are substantially consistent with the t-SNE projection distribution of the entire dataset.
Based on the improved generation of heterogeneous fault samples of the countermeasure network, whether the generated samples are added with training data can be compared, and the fault diagnosis accuracy can be effectively improved. It can also be seen from fig. 5 that the data is seriously unbalanced, and the classification result of the fault diagnosis model is severely biased to the normal state, so that the prediction accuracy of the fault class is very low. Obviously, in the actual scene, we only pay attention to the prediction accuracy of the fault state, so that the overall generalization accuracy cannot be used as a comparison index, and the accuracy and recall rate of the fault state are used. However, it can be seen from fig. 5 that the intersection of the two types of data is strong, which makes it impossible to achieve very high levels of accuracy and recall of the fault state at the same time. Thus, the combination metrics F1-score, G-means and AUC are used in the present invention as evaluation metrics for different oversampling methods.
Under the same operation environment, a classical artificial neural network model is adopted as a fault diagnosis model, the influence of different oversampling modes on a fault diagnosis result is compared according to a comparison experiment flow, and the experiment result is shown as follows:
table 1 comparison of different methods of generation
| F1-score | G-means | AUC | |
| Untreated | 0.38095 | 0.81494 | 0.63299 |
| RandomOversampling | 0.58065 | 0.74922 | 0.79879 |
| SMOTE-regular | 0.54054 | 0.67361 | 0.83125 |
| SMOTE-borderline1 | 0.63636 | 0.69469 | 0.96407 |
| SMOTE-borderline2 | 0.52632 | 0.65881 | 0.83108 |
| ADASYN | 0.62222 | 0.68301 | 0.96389 |
| GAN | 0.84404 | 0.85449 | 0.99705 |
| MAE-D2GAN | 0.875 | 0.88192 | 0.99723 |
It can be seen that the class imbalance can seriously affect the result of the fault diagnosis, in which case the conventional over-sampling method can improve the accuracy of the fault diagnosis. Wherein, the MAE-D2GAN combined with the improved generation countermeasure network with the double discriminators and the improved automatic encoder can further improve the fault diagnosis result, which is superior to the traditional GAN.
Compared with a typical oversampling method, in a high-speed train braking system, the MAE-D2GAN can generate a fault sample with higher quality, so that the prediction precision of a data-driven fault diagnosis model is greatly improved. In addition, MAE-D2GAN has potential value for failure prediction and regression scenarios for other PHM analyses of high reliability systems.
The above examples are only illustrative of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention, and various modifications and improvements made by those skilled in the art to the technical solution of the present invention should fall within the scope of protection defined by the claims of the present invention without departing from the spirit of the present invention.
Claims (5)
1. A method for real-time fault diagnosis based on improved generation of heterogeneous imbalance fault samples against network generation, comprising the steps of:
step 1: collecting fault states and monitoring data of an electromechanical product system or device, and obtaining historical monitoring data;
step 2: performing data preprocessing on the historical monitoring data obtained in the step 1 to obtain a preprocessed data set [ X, Y ], wherein X is a variable related to a fault state of an electromechanical product system or equipment and is heterogeneous data, and the variables comprise continuous variables and discrete variables; the Y is an electromechanical product system or equipment fault state;
step 3: setting an automatic encoder to realize mutual conversion of heterogeneous data and continuous characteristics, adding a normal sample discriminator in a generation countermeasure network, establishing an improved generation countermeasure network with a double discriminator, inputting the preprocessed data set [ X, Y ] obtained in the step 2 into the automatic encoder and the improved generation countermeasure network with the double discriminator to obtain heterogeneous fault samples U, wherein the step 3 specifically comprises the following substeps:
step 31: the automatic encoder includes an encoder and a decoder; the function of the encoder is to encode heterogeneous monitoring data into successive latent features; the decoder output layer uses a micro Gumbel-softmax to realize discretization of continuous features, the function of the decoder is to decode continuous fault feature data into heterogeneous fault samples, and the preprocessed data set [ X, Y ] obtained in the step 2 is input into the encoder to obtain fault sample feature data X and normal sample feature data;
step 32: said step 3 improved generation countermeasure network with double discriminators comprises a generator G and two discriminators D and F; the discriminator F is used for discriminating whether the generated sample is normal or not, is obtained only based on the normal sample characteristic data, only provides a loss value when the improved generation countermeasure network with the double discriminators is trained, and does not perform iterative optimization; training the discriminator F based on the normal sample feature data obtained in the step 31; inputting white noise z into a generator G of the improved generation countermeasure network with double discriminators, and obtaining a characteristic data set G (z) corresponding to a potential space fault sample;
step 33: training the generator G and the discriminator D according to the fault sample characteristic data x and the characteristic data set G (z) corresponding to the fault samples in potential space, wherein the improved generation of the loss function V (D, G) with the double discriminators in the training process is as follows:
wherein: d (x) is the probability that the generated samples are from the fault sample feature data x;
solving expected values for all x; d (G (z)) is a probability that the discriminator D discriminates whether or not the continuous variable set G (z) is true; f (G (z)) is a probability that the discriminator F discriminates whether or not the continuous variable set G (z) is normal; />
For taking all z set-ups, find the expected value of log (F (G (z)),)>
The expected value of log (1-D (G (z))) is calculated for all z's set up for the pass;
the objective function of the improved generation countermeasure network with double discriminators is
Based on the objective function->
Iteratively training the generator G and discriminator D, the iterative training comprising training the generator G to minimize the loss function V (D, G) and training the discriminator D to maximize the loss function V (D, G); when the difference between the maximum value and the minimum value of the loss function V (D, G) in 100 continuous iterative training of the model is smaller than 0.01, the iterative training is stopped; obtaining a trained generator G after the iterative training is stopped; inputting a group of white noise z into the trained generator G to obtain a characteristic data set G (z) corresponding to a potential space fault sample;
step 34: inputting the characteristic data set G (z) corresponding to the potential space fault sample obtained in the step 33 into a decoder in the automatic encoder to obtain a heterogeneous fault sample U; inputting the feature data set G (z) corresponding to the potentially spatially faulty sample obtained in the step 33 into the decoder to obtain a decoded data set, where the decoded data set includes a column corresponding to a discrete variable and a column corresponding to a continuous variable, and performing discretization on the column corresponding to the discrete variable by using a microable gummel-softmax to obtain a discretized column, and obtaining a heterogeneous faulty sample U according to the column corresponding to the continuous variable and the discretized column;
step 4: adding a column of fault state I to the heterogeneous fault sample U obtained in the step 34 to obtain a heterogeneous fault sample set [ U, I ], wherein the fault state I is a column vector with elements of 1, and the dimension of the fault state I is consistent with the dimension of the heterogeneous fault sample U;
step 5: the heterogeneous fault sample set [ U, I ] obtained in the step 4 is processed]Incorporating the preprocessed data set [ X, Y ] obtained in said step 2]In obtaining a new data set
Based on the new dataset +.>
Training an artificial neural network ANN algorithm to obtain a fault diagnosis model;
step 6: and (3) collecting real-time monitoring data, preprocessing the real-time monitoring data with the same data in the step (2) to obtain preprocessed real-time monitoring data, inputting the preprocessed real-time monitoring data into the fault diagnosis model obtained in the step (5) to obtain a real-time fault state, and completing fault diagnosis.
2. The method for real-time fault diagnosis based on improved generation of anti-network generation heterogeneous imbalance fault samples according to claim 1, wherein the encoding process using the encoder in step 31 is specifically: and (3) performing single-heat coding on discrete variables in the preprocessed data set [ X, Y ] obtained in the step (2) to obtain single-heat coded variables, and uniformly coding the single-heat coded variables and continuous variables in the preprocessed data set [ X, Y ] obtained in the step (2) to obtain fault sample characteristic data and normal sample characteristic data.
3. The method for real-time fault diagnosis based on improved generation of heterogeneous imbalance fault samples for network generation according to claim 1, wherein the fault conditions in step 1 include faulty and non-faulty.
4. The method for real-time fault diagnosis based on improved generation of anti-network generated heterogeneous imbalance fault samples according to claim 1, wherein the monitored data in step 1 is heterogeneous data, and the heterogeneous data includes continuous features, discrete features and signal features.
5. The method for real-time fault diagnosis based on improved generation of anti-network generation heterogeneous imbalance fault samples according to claim 1, wherein the data preprocessing in step 2 includes filling missing values, replacing outliers, processing dimension differences and digitizing; the missing value filling adopts a local mean filling method; the process dimension gap is data normalization, and the data normalization adopts a z-score method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110796968.XA CN113485863B (en) | 2021-07-14 | 2021-07-14 | Method for generating heterogeneous imbalance fault samples based on improved generation of countermeasure network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110796968.XA CN113485863B (en) | 2021-07-14 | 2021-07-14 | Method for generating heterogeneous imbalance fault samples based on improved generation of countermeasure network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113485863A CN113485863A (en) | 2021-10-08 |
| CN113485863B true CN113485863B (en) | 2023-05-05 |
Family
ID=77938762
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110796968.XA Active CN113485863B (en) | 2021-07-14 | 2021-07-14 | Method for generating heterogeneous imbalance fault samples based on improved generation of countermeasure network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113485863B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116821697B (en) * | 2023-08-30 | 2024-05-28 | 聊城莱柯智能机器人有限公司 | Mechanical equipment fault diagnosis method based on small sample learning |
| CN117854716B (en) * | 2024-03-08 | 2024-06-21 | 长春师凯科技产业有限责任公司 | Method and system for filling heart disease diagnosis missing data based on HF-GAN |
| CN118013443B (en) * | 2024-04-08 | 2024-07-23 | 华侨大学 | Online real-time vacuum dry pump abnormality detection method based on generation model algorithm |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110428004A (en) * | 2019-07-31 | 2019-11-08 | 中南大学 | Component of machine method for diagnosing faults under data are unbalance based on deep learning |
| CN110567720A (en) * | 2019-08-07 | 2019-12-13 | 东北电力大学 | method for diagnosing depth confrontation of fault of fan bearing under unbalanced small sample scene |
| CN110647923A (en) * | 2019-09-04 | 2020-01-03 | 西安交通大学 | Variable working condition mechanical fault intelligent diagnosis method based on self-learning under small sample |
| CN112039687A (en) * | 2020-07-14 | 2020-12-04 | 南京邮电大学 | Small sample feature-oriented fault diagnosis method based on improved generation countermeasure network |
| CN113032917A (en) * | 2021-03-03 | 2021-06-25 | 安徽大学 | Electromechanical bearing fault detection method based on generation countermeasure and convolution cyclic neural network and application system |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109948117B (en) * | 2019-03-13 | 2023-04-07 | 南京航空航天大学 | Satellite anomaly detection method for network self-encoder |
| CN111006865A (en) * | 2019-11-15 | 2020-04-14 | 上海电机学院 | Motor bearing fault diagnosis method |
| CN111767861B (en) * | 2020-06-30 | 2024-03-12 | 苏州兴钊防务研究院有限公司 | SAR image target recognition method based on multi-discriminant generation countermeasure network |
| CN113095402B (en) * | 2021-04-12 | 2023-06-13 | 江南大学 | Code input-based generation countermeasure network fault detection method and system |
-
2021
- 2021-07-14 CN CN202110796968.XA patent/CN113485863B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110428004A (en) * | 2019-07-31 | 2019-11-08 | 中南大学 | Component of machine method for diagnosing faults under data are unbalance based on deep learning |
| CN110567720A (en) * | 2019-08-07 | 2019-12-13 | 东北电力大学 | method for diagnosing depth confrontation of fault of fan bearing under unbalanced small sample scene |
| CN110647923A (en) * | 2019-09-04 | 2020-01-03 | 西安交通大学 | Variable working condition mechanical fault intelligent diagnosis method based on self-learning under small sample |
| CN112039687A (en) * | 2020-07-14 | 2020-12-04 | 南京邮电大学 | Small sample feature-oriented fault diagnosis method based on improved generation countermeasure network |
| CN113032917A (en) * | 2021-03-03 | 2021-06-25 | 安徽大学 | Electromechanical bearing fault detection method based on generation countermeasure and convolution cyclic neural network and application system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113485863A (en) | 2021-10-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113485863B (en) | Method for generating heterogeneous imbalance fault samples based on improved generation of countermeasure network | |
| CN111914873B (en) | Two-stage cloud server unsupervised anomaly prediction method | |
| CN107561997B (en) | A kind of power equipment state monitoring method based on big data decision tree | |
| CN107657250B (en) | Bearing fault detection and positioning method and detection and positioning model implementation system and method | |
| CN108459955B (en) | Software defect prediction method based on deep self-coding network | |
| CN116300691B (en) | State monitoring method and system for multi-axis linkage numerical control machining | |
| CN111027615B (en) | Middleware fault early warning method and system based on machine learning | |
| Yacout | Fault detection and diagnosis for condition based maintenance using the logical analysis of data | |
| CN110851654A (en) | Industrial equipment fault detection and classification method based on tensor data dimension reduction | |
| CN117421582A (en) | Equipment health analysis method based on multi-source data driving | |
| CN117131110B (en) | Method and system for monitoring dielectric loss of capacitive equipment based on correlation analysis | |
| CN113064873B (en) | Log anomaly detection method with high recall rate | |
| Amruthnath et al. | Factor analysis in fault diagnostics using random forest | |
| CN114487673A (en) | Power equipment fault detection model based on Transformer and electronic equipment | |
| CN115310837A (en) | Complex electromechanical system fault detection method based on causal graph attention neural network | |
| CN110779988A (en) | Bolt life prediction method based on deep learning | |
| CN116467652A (en) | Bearing fault prediction method and system based on convolution noise reduction self-encoder | |
| CN115564075B (en) | Main and auxiliary integrated fault collaborative diagnosis method and system for urban power grid | |
| CN116910559A (en) | Index anomaly detection method for intelligent operation and maintenance application of power grid supercomputer center | |
| CN114781450B (en) | Train rolling bearing state identification method based on parameter optimization MOMEDA-MIA-CNN | |
| CN115730255A (en) | Motor fault diagnosis and analysis method based on transfer learning and multi-source information fusion | |
| CN115392710A (en) | Wind turbine generator operation decision method and system based on data filtering | |
| CN116007937A (en) | Intelligent fault diagnosis method and device for mechanical equipment transmission part | |
| CN115659135A (en) | Anomaly detection method for multi-source heterogeneous industrial sensor data | |
| CN115470854A (en) | Information system fault classification method and classification system |
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 |