CN114091504A - Rotary machine small sample fault diagnosis method based on generation countermeasure network - Google Patents

Rotary machine small sample fault diagnosis method based on generation countermeasure network Download PDF

Info

Publication number
CN114091504A
CN114091504A CN202110813313.9A CN202110813313A CN114091504A CN 114091504 A CN114091504 A CN 114091504A CN 202110813313 A CN202110813313 A CN 202110813313A CN 114091504 A CN114091504 A CN 114091504A
Authority
CN
China
Prior art keywords
sample
data
training
discriminator
generator
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202110813313.9A
Other languages
Chinese (zh)
Inventor
曹智
伏洪勇
王珂
李振祥
张俊华
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Chinese Academy of Sciences
Technology and Engineering Center for Space Utilization of CAS
Original Assignee
University of Chinese Academy of Sciences
Technology and Engineering Center for Space Utilization of CAS
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by University of Chinese Academy of Sciences, Technology and Engineering Center for Space Utilization of CAS filed Critical University of Chinese Academy of Sciences
Priority to CN202110813313.9A priority Critical patent/CN114091504A/en
Publication of CN114091504A publication Critical patent/CN114091504A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01H—MEASUREMENT OF MECHANICAL VIBRATIONS OR ULTRASONIC, SONIC OR INFRASONIC WAVES
    • G01H17/00—Measuring mechanical vibrations or ultrasonic, sonic or infrasonic waves, not provided for in the preceding groups
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M13/00—Testing of machine parts
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/086—Learning methods using evolutionary algorithms, e.g. genetic algorithms or genetic programming
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Mathematical Physics (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Signal Processing (AREA)
  • Physiology (AREA)
  • Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)

Abstract

The invention discloses a rotary machine small sample fault diagnosis method based on a generation countermeasure network, which comprises the steps of firstly converting collected rotary machine time domain signals into frequency spectrum signals with characteristics easy to observe through Fast Fourier Transform (FFT); and then, on the basis of ACWGAN-GP, an ACWGAN-GP-ARGMAX diagnosis model is obtained by introducing an ARGMAX multi-classification idea. The diagnosis model endows the discriminator with classification and identification capabilities, enhances the classification and identification capabilities of the discriminator, and can effectively improve the accuracy and efficiency of fault diagnosis of small samples of rotary machines.

Description

Rotary machine small sample fault diagnosis method based on generation countermeasure network
Technical Field
The invention relates to the technical field of fault diagnosis of rotary machines, in particular to a fault diagnosis method for a small sample of a rotary machine based on a generation countermeasure network.
Background
The rotating machinery is very commonly applied in the fields of aerospace, ships, wind power, petrochemicals and the like, is one of key components in an industrial scene, and is equipment which is easy to break down. Most rotary machines are of a circularly symmetric structure and are stable under normal operating conditions. However, when the rotary machine fails, its symmetry is broken, resulting in unstable performance and easily causing a serious accident. In order to avoid catastrophic failures of the rotating machine, it is of great significance to continuously monitor and diagnose the operating state of the rotating machine.
The operating state of a rotating machine can be detected and diagnosed by measuring signals such as vibration, noise, and oil. At present, the collection and analysis technology of vibration signals is relatively mature, and fault diagnosis of rotary machines based on vibration signals is one of the most common and widely applied methods. Vibration signal characteristics of rotary machines are complex, and non-professional people generally do not know and cannot identify fault characteristics of vibration signals, so that intelligent fault diagnosis research based on multidisciplinary intersection is the main development direction of fault diagnosis of rotary machines at present. However, in the training process of the intelligent fault diagnosis method, corresponding data distribution needs to be learned from a large amount of historical data, so that accurate fault diagnosis is realized. However, in a real industrial scenario, due to: 1) the failure occurrence rate is low, and the comprehensive real operation data corresponding to various health states are difficult to acquire; 2) the fault injection experiment of the rotating machinery has high cost and has great risk, and the personnel safety and the property safety can be endangered; 3) even if a large amount of operation monitoring data exist, a large amount of manpower and material resources are consumed for effectively labeling the data, the cost is extremely high, and therefore, various types of data which are enough to be effectively labeled are generally difficult to obtain for model training. Therefore, it is necessary to develop a diagnostic method that can achieve a high-precision fault diagnosis result by training with only a small sample.
Disclosure of Invention
The invention aims to provide a rotary machine small sample fault diagnosis method based on a generation countermeasure network, so as to solve the problems in the prior art.
In order to achieve the purpose of the invention, the technical scheme adopted by the invention is as follows:
a rotary machine small sample fault diagnosis method based on a generation countermeasure network comprises the following steps:
s1, collecting vibration signals of the rotating machine;
s2, preprocessing the rotary mechanical vibration signal, and dividing the signal into a training set and a testing set;
s3, establishing an ACWGAN-GP-ARGMAX network model;
s4, training the ACWGAN-GP-ARGMAX network model;
s5, respectively carrying out countermeasure training on the data in each health state in the training set, and fusing the trained network models into a diagnosis model;
s6: and inputting the test set into the diagnosis model for diagnosis test, and outputting a diagnosis result.
Further, a specific implementation manner of step S1 includes arranging a vibration acceleration sensor on the rotary machine, and acquiring data of one-dimensional time series vibration acceleration of the rotary machine in various health states by using the vibration acceleration sensor to form a raw data set, where the raw data set includes 1 normal state and N-1 fault states.
Further, the specific implementation manner of step S2 includes performing feature extraction on the vibration signals in various health states in the original data set by a fast fourier transform and standard normalization processing manner, and constructing a new data set from the processed data;
the standard normalization method is as follows: x ═ X (X-X)min)/(xmax-xmin) Wherein x ismaxIs the maximum value, x, of the data sampleminIs the minimum value of the data sample;
according to algorithm requirements, dividing the new data set into a training set and a testing set, wherein the training set is expressed as
Figure BDA0003169013860000021
Where k represents the k-th health state,
Figure BDA0003169013860000022
the ith training sample in the healthy state.
Further, the ACWGAN-GP-ARGMAX network model in the step S3 includes a generator and a discriminator;
the generator is represented by a network model structure comprising 5 deconvolution layers, and input data is 4096-dimensional random noise which follows normal distribution; in the generator, the random noise is firstly input into a full-connection layer consisting of 4096 neurons, and then passes through the 5-layer deconvolution layer in sequence; the number of the filters in the first four layers of deconvolution layers is 256, 128, 64 and 32 in turn, and the last layer of deconvolution layer only has 1 filter; the size of the first layer of deconvolution kernel is 4 x 4 and the step size is 1, and the sizes of the last four layers of deconvolution kernels are all 3 x 3 and the step size is 2; through the five deconvolution layers, input data is converted into a 64 x 64 two-dimensional matrix;
the discriminator is expressed as a network model structure comprising 5 layers of convolution layers, and input data is a 64 x 64 two-dimensional matrix generated by the generator and a 64 x 64 two-dimensional matrix obtained by shaping real data; in the discriminator, the number of the filters in the first four convolutional layers is 32, 64, 128 and 256 in turn, and the last convolutional layer only has 1 filter; the sizes of the first four convolutional layer convolutional cores are 3 multiplied by 3 and the step length is 2, and the sizes of the last convolutional layer convolutional cores are 4 multiplied by 4 and the step length is 1; through the five convolutional layers, the input data is converted into a tag value.
Further, the specific implementation manner of step S4 includesSelecting Gaussian distribution as noise distribution PnoiseAnd assigning a corresponding class label from which a subset of size n is sampled, denoted as
Figure RE-GDA0003393588660000031
Random noise subset with class label
Figure RE-GDA0003393588660000032
Input to a generator G corresponding to the health state kkForward propagation results in generating a subset of samples
Figure RE-GDA0003393588660000033
Training set P from real data of said health state kk～dataTaking the data sample as a positive sample, and recording as
Figure RE-GDA0003393588660000034
Then training set P from real data outside the health state k(C≠k)～dataTaking the data sample as a negative sample, and recording as
Figure RE-GDA0003393588660000035
By
Figure RE-GDA0003393588660000036
And
Figure RE-GDA0003393588660000037
composing the true sample of the health state k, and generating the sample by
Figure RE-GDA0003393588660000038
Combined to form a training set subset { Xk}；
Setting the discriminator DkAnd the generator GkThe ratio of training times is m, first using the training set subset { XkThe algorithm pairs the discriminator D with AdamkPerforming m times of training; then using said random noise subset with class labels
Figure RE-GDA0003393588660000039
And Adam's algorithm to the generator GkPerforming 1 training; selecting a cross entropy loss function as an objective function, and specifically defining the cross entropy loss function as follows:
Figure RE-GDA00033935886600000310
Figure RE-GDA00033935886600000311
where x represents the true sample or samples and where,
Figure RE-GDA00033935886600000312
it is indicated that the generation of the sample,
Figure RE-GDA00033935886600000313
by adding at x and
Figure RE-GDA00033935886600000314
random samples obtained by linear interpolation, lambda being a gradient penalty factor,
Figure RE-GDA00033935886600000315
the expectation of the distribution of the real data is represented,
Figure RE-GDA00033935886600000316
indicating the desire to generate a distribution of data,
Figure RE-GDA00033935886600000317
denotes a gradient penalty term, P (C ═ C | Xreal) Representing the conditional probability distribution of the real data on class labels,
Figure RE-GDA00033935886600000318
representing a conditional probability distribution of the generated data on the class label,
according to the above-mentioned ratio m of training times of each round, for the said discriminator DkAnd the generator GkPerforming iterative training until the discriminator DkAnd the generator GkAchieve nash equilibrium between them.
Further, the specific implementation manner of step S5 includes,
firstly, marking data samples except the kth health state as negative samples and marking the data samples as class 0; correspondingly, marking the data sample in the kth health state as a positive sample and marking the data sample as a class 1;
second, the labeled data is inputted as a real sample to the discriminator DkPerforming countermeasure training until the model loss function converges to the minimum value, thereby obtaining a generator and a discriminator of a kth sample;
thirdly, performing the training on each type of sample in turn to obtain n trained ACWGAN-GP-ARGMAX models; separating n discriminators from the data to be used as a trained diagnosis model;
and fourthly, model fusion is carried out on the n discriminators, the probability values of the output results are combined into n-dimensional vectors, and the n-dimensional vectors are accessed into a Softmax activation layer for activation, so that the probability that each value output by the output layer corresponds to each test sample and belongs to each health state is obtained.
Further, between the step S4 and the step S5, there is further included a step S500: and performing quality evaluation on the network model trained in the step S4, and if the network model cannot pass the evaluation, repeating the step S4.
Further, the quality evaluation in step S500 includes:
first, evaluate the generator GkThe method comprises the steps of comparing the frequency spectrums of a generated sample and a real sample qualitatively, and then measuring the similarity of the generated sample and the real sample quantitatively by calculating the Pearson Correlation Coefficient (PCC) and the Cosine Similarity (CS) of the generated sample and the real sample;
second, the discriminator D is evaluatedkThe classification and identification capability of (2) is specifically realized by using the trained discriminator DkGenerating a data sample if said discriminator DkCapable of generating sufficiently realistic generationSamples, these generated samples are further used to continuously align the discriminator DkTraining is performed so that the discriminator DkThe hidden distribution in the current sample is learned.
The invention has the beneficial effects that:
1) on the basis of the ACWGAN-GP model, the invention introduces the multi-classification idea of 'ARGMAX', so that the ACWGAN-GP-ARGMAX model has strong generating capability and strong classification and identification capability.
2) The invention integrates the generation of the countermeasure network and the fault diagnosis model, and overcomes the defect that the generation of the countermeasure network cannot be classified under the unsupervised condition, so that the fault diagnosis can be realized by directly utilizing the data set without additionally introducing a discrimination model.
3) The rotary machine small sample fault diagnosis model based on the generated countermeasure network has the advantages of simple structure, few parameters, low requirement on hardware resources and high economy, and has certain generalization capability, and the gear box data set and the bearing data set experiments prove that the rotary machine small sample fault diagnosis accuracy can be effectively improved.
Drawings
FIG. 1 is a basic flow diagram of the rotary machine small sample fault diagnosis method based on the generation countermeasure network of the present invention;
FIG. 2 is a flow chart of an embodiment implementation of the present invention;
FIG. 3 is a diagram of the ACWGAN-GP-ARGMAX network model architecture according to the present invention;
FIG. 4 is a time domain diagram of the raw vibration signal of the class 8 gearbox according to the embodiment of the invention;
FIG. 5 is a graph comparing various types of generated data with the spectrum of the original data in the embodiment of the present invention;
FIG. 6 is a graph of the diagnostic results of 10 independent experiments at different sample ratios in the examples described herein;
FIG. 7 is a graph of the mean values of the diagnostic results of 10 independent experiments at different sample ratios in the example of the present invention;
FIG. 8 is a confusion matrix of the mean values of the diagnostic results of 10 independent experiments under different sample ratios in the example described in the present invention;
FIG. 9 is the average of the diagnostic results of 10 independent experiments under different sample ratios for different diagnostic methods in the examples described herein.
Detailed Description
The invention relates to a rotary machine small sample fault diagnosis method based on a generated countermeasure network, which comprises the following steps:
and S1, acquiring vibration signals of the rotating machine. The method comprises the steps of arranging a vibration acceleration sensor on a rotary machine, collecting data of one-dimensional time series vibration acceleration of the rotary machine in various health states by using the vibration acceleration sensor, and obtaining an original data set, wherein the data set comprises 1 normal state and N-1 fault states.
And S2, preprocessing the rotating mechanical vibration signal. And performing feature extraction on the vibration signals in various health states in the data set by processing modes such as Fast Fourier Transform (FFT), standard normalization and the like, and constructing the processed data into a new data set. The standard normalization method comprises the following steps: x ═ X (X-X)min)/(xmax-xmin) Wherein x ismaxIs the maximum value of the data sample, xminIs the minimum value of the data sample. And according to algorithm requirements, dividing the new data set into a training set and a testing set. The training set may be represented as
Figure BDA0003169013860000061
Where k represents the k-th health state,
Figure BDA0003169013860000062
the ith training sample in the healthy state.
And S3, establishing an ACWGAN-GP-ARGMAX network model corresponding to the health state. In view of the excellent performance of CNNs in feature extraction and classification, the present invention builds generators and discriminators on the basis of two-dimensional CNNs.
The network model structure of the generator consists of 5 deconvolution layers, and input data is 4096-dimensional random noise which follows normal distribution. In the generator, random noise is first input into a fully connected layer of 4096 neurons, then sequentially through 5 deconvolution layers. The number of the filters in the first four deconvolution layers is 256, 128, 64 and 32 in sequence, and the last deconvolution layer only has 1 filter; the size of the first deconvolution layer convolution kernel is 4 x 4 and the step size is 1, and the sizes of the last four layers of deconvolution layer convolution kernels are all 3 x 3 and the step size is 2; through the five deconvolution layers, the input data is converted into a 64 × 64 two-dimensional matrix.
The network structure of the arbiter is symmetric to the generator. The network model structure of the discriminator is composed of 5 layers of convolution layers, and input data is a 64 x 64 two-dimensional matrix generated by a generator and a 64 x 64 two-dimensional matrix obtained by shaping real data. In the discriminator, the number of filters in the first four convolutional layers is 32, 64, 128 and 256 in turn, and the last convolutional layer only has 1 filter; the sizes of the convolution kernels of the first four layers are 3 multiplied by 3 and the step length is 2, and the sizes of the convolution kernels of the last layer are 4 multiplied by 4 and the step length is 1; through the five convolutional layers, the input data is converted to a tag value.
And S4, training the network model. Selecting Gaussian distribution as noise distribution PnoiseAnd assigning a corresponding class label, from which a subset of size n is sampled, denoted as
Figure RE-GDA0003393588660000063
Random noise subset with class label
Figure RE-GDA0003393588660000064
Input to a generator G corresponding to the health state kkForward propagation results in generating a subset of samples
Figure RE-GDA0003393588660000065
Training set P of real data from health state kk～dataTaking the data sample as a positive sample, and recording as
Figure RE-GDA0003393588660000066
Then training set P from real data outside the health state k(C≠k)～dataTaking the data sample as a negative sample, and recording as
Figure RE-GDA0003393588660000067
By
Figure RE-GDA0003393588660000068
And
Figure RE-GDA0003393588660000069
forming a real sample under the condition of the health state k, and generating the sample by
Figure RE-GDA00033935886600000610
Combined to form a training set subset { Xk}. Setting discriminator DkAnd generator GkThe ratio of training times is m, first using a subset of the training set { X }kDiscriminator D of the pair of Adam and AlgorithmkTraining for m times; then with random noise subsets of class labels
Figure RE-GDA00033935886600000611
And Adam Algorithm Pair Generator GkPerforming 1 training; selecting a cross entropy loss function as an objective function, which is specifically defined as follows:
Figure RE-GDA0003393588660000071
Figure RE-GDA0003393588660000072
where x represents the true sample or samples and where,
Figure RE-GDA0003393588660000073
it is indicated that the generation of the sample,
Figure RE-GDA0003393588660000074
by adding at x and
Figure RE-GDA0003393588660000075
random samples obtained by linear interpolation, lambda being a gradient penalty factor,
Figure RE-GDA0003393588660000076
the expectation of the distribution of the real data is represented,
Figure RE-GDA0003393588660000077
indicating the desire to generate a distribution of data,
Figure RE-GDA0003393588660000078
denotes a gradient penalty term, P (C ═ C | Xreal) Representing the conditional probability distribution of the real data on class labels,
Figure RE-GDA0003393588660000079
representing a conditional probability distribution of the generated data on the class label,
according to the above-mentioned ratio m of training times of each round, for the said discriminator DkAnd the generator GkPerforming iterative training until the discriminator DkAnd the generator GkAchieve nash equilibrium between them.
And S5, performing quality evaluation on the network model trained in the S4. The trained model is generally evaluated for quality in two ways. The similarity of generated samples is qualitatively and quantitatively evaluated, namely, the generating capacity of a model generator is evaluated. The method comprises the steps of comparing frequency spectrums of a generated sample and a real sample qualitatively, and then measuring similarity of the generated sample and the real sample quantitatively by calculating a Pearson Correlation Coefficient (PCC) and a Cosine Similarity (CS) of the generated sample and the real sample; and secondly, evaluating the influence of the generated sample on fault diagnosis under the condition of a small sample by sample expansion, namely evaluating the classification and identification capability of the model discriminator. Generating data samples by using the trained generator, and if the generator can generate sufficiently vivid generated samples, further continuously training the discriminator by using the generated samples to ensure that the discriminator DkThe hidden distribution in the current sample is learned.
S6, respectively carrying out countermeasure training on the data in each health state in the data set, firstly, marking the data samples except the kth health state as negative samples and marking the negative samples as class 0; accordingly, the data sample in the k-th health state is labeled as a positive sample, labeled as class 1. Secondly, inputting the marked data as a true sample into a discriminator for countermeasure training until the model loss function converges to the minimum value, thereby obtaining a generator and a discriminator of the kth sample. Thirdly, the training is carried out on each type of sample in turn to obtain n trained ACWGAN-GP-ARGMAX models. From which n discriminators D are connectediAnd separating the model to be used as a well-trained diagnosis model. Fourthly, D isiModel fusion is carried out, and n D are obtained through model fusioniAnd combining probability values of the output results into an n-dimensional vector, and accessing a Softmax activation layer for activation, so that each value output by the output layer corresponds to the probability that the test sample belongs to each health state.
Thus, a complete fault diagnosis process is achieved.
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail below with reference to the accompanying drawings. It should be understood that the detailed description and examples, while indicating the preferred embodiment of the invention, are given by way of illustration only, not limitation.
Example (b):
FIG. 2 is a flow chart of an embodiment of the present invention based on a rotating machinery small sample fault diagnosis method for generating a countermeasure network. As shown in fig. 2, the method for diagnosing a fault of a small sample of a rotating machine based on a generated countermeasure network of the present invention specifically includes the steps of:
and S1, collecting vibration signals of the rotary machine by using the vibration acceleration sensor.
And S2, preprocessing the acquired vibration signals of the rotating machine to construct an original data set, and dividing the original data set into a training set and a testing set.
S3, constructing a network model based on ACWGAN-GP-ARGMAX.
And S4, performing model training on the network model based on ACWGAN-GP-ARGMAX, and fusing the trained network models in each health state to construct a diagnosis model.
And S5, inputting the test set into a diagnosis model to perform diagnosis test.
FIG. 3 is a diagram of the ACWGAN-GP-ARGMAX network model architecture according to the present invention. As shown in fig. 3, the ACWGAN-GP-ARGMAX network model comprises the following modules: the device comprises a generator module, a discriminator module, a data input module and a diagnosis module.
The present invention and its effects are specifically described below by way of an example.
The experimental data for this example is from the gearbox dataset provided by the IEEE PHM challenge in 2009, which is representative of general industrial gearbox data. The physical diagram of the gearbox is shown in figure 3. The data set contains a total of 8 time-domain vibration signals of different health states. All data in the data set were collected from the input shaft under high load conditions, with a shaft speed of 30Hz, a sampling frequency of 66.7kHz, and a sampling time of 4 s. On the premise of ensuring that the information contained in the diagnosis signal of the gearbox is not lost, 400 health state data samples of each type are selected. The sample length is 8192, and each sample is a time domain signal, so after the FFT, each type of data sample becomes a spectrum vector with a sample length of 4096. Details of the samples used are shown in table 1, with the time domain waveforms shown in fig. 4.
TABLE 1 gearbox data set operating mode details
Figure BDA0003169013860000091
IS:Input Shaft；ID:Idler Shaft；OS:Output Slide.
G:Good；C:Chipped；E:Eccentric；Br:Broken；B:Ball；I:Inner race；O:Outer race；Im:Imbalance；K:Keyway sheared.
Taking 200 samples of each type of healthy state in the data set as training sets respectivelyAnd used for testing. A part of samples are taken from the training set to be used as a training subset, and the ratio of the number of training subset samples to the number of test set samples is called a sample ratio. In the experiment, sample comparison is divided according to 1:20, 1:10, 1:5, 1:4, 1:2 and 1:1 respectively, and the number of various training samples in each sample proportion is shown in table 2. During the course of the confrontation training, the size of minipatch is set to 8 at a sample ratio of 1:20, and the size of minipatch is set to 16 in all the other cases. In the anti-training process of each wheel, the ratio of the update iteration times of the discriminator to the generator is set to be 5:1, namely, the generator is updated for 1 time when the discriminator is updated for 5 times. The learning rate is set to 0.0001, and the number of training iterations is 10000; arbiter for learning kth health status
Figure BDA0003169013860000092
For example, all training samples in the kth health state are selected as positive samples, the label type is 1, and part of the training samples in the remaining 7 health states are selected as negative samples, and the label type is 0, where the number of the positive samples and the number of the negative samples are approximately equal.
TABLE 2 number of training samples in various health states under different sample ratios
Figure BDA0003169013860000093
Figure BDA0003169013860000101
After the network model is pre-trained, 200 new samples are generated for each type of health state using each generator model. In order to evaluate a new sample, firstly, the frequency spectrums of the real sample and the generated sample are visually displayed in the same graph, as shown in fig. 5; next, the generated samples were evaluated by the values of the similarity indices PCC and CS, which are shown in table 3.
TABLE 3 PCC and CS between generated data and raw data under each type of health status
Figure BDA0003169013860000102
As can be seen from fig. 5, the overlap ratio of the generated sample and the true sample of each health state is high; as shown in table 3, the PCC and CS were higher than 0.5 for all healthy state samples, indicating that the generated samples have a higher similarity to the real samples. In conclusion, the generator subjected to the countermeasure training can accurately learn the implicit data distribution in the real sample and generate the sample with high similarity to the real sample.
In order to test the performance of the fault diagnosis method based on ACWGAN-GP-ARGMAX, the comparison experiment based on the ACWGAN-GP diagnosis model is arranged. For each sample ratio, 10 independent experiments were performed, and the experimental results are shown in fig. 6 and 7. Wherein data1 represents ACWGAN-GP diagnostic model and data2 represents ACWGAN-GP-ARGMAX diagnostic model. As can be seen from FIG. 6, the ACWGAN-GP-ARGMAX diagnostic model has significantly improved diagnostic accuracy and stability compared to the ACWGAN-GP diagnostic model. As can be seen from FIG. 7, with the increase of the proportion of the training samples, the diagnosis accuracy of the ACWGAN-GP-ARGMAX diagnosis model and the ACWGAN-GP diagnosis model is obviously improved, but the diagnosis effect of the ACWGAN-GP-ARGMAX diagnosis model is better. In order to more intuitively explain the diagnosis result of the ACWGAN-GP-ARGMAX diagnosis model, the diagnosis process is visualized by using a confusion matrix by taking each sample ratio as an example, as shown in FIG. 8. As can be seen from FIG. 8, when the sample ratio is 1:20, some of the data in Class2,4,6, and 8 are classified as erroneous. With the increasing sample proportion, the sample diagnosis accuracy rate is increased gradually, and when the sample proportion reaches 1:1, only few samples are classified wrongly.
In order to prove the superiority of the method, several common methods are adopted to be compared with ACWGAN-GP-ARGMAX in the aspect of solving the problem of a small sample for fault diagnosis, and the common methods comprise the following steps: CNN, DNN, SVM, SAE, GAN. The fault diagnosis accuracy of all the above methods was obtained by training and testing each method under the same conditions, as shown in fig. 9.
As can be seen from fig. 9, the failure accuracy of all methods increases with the sample scale, indicating that these methods are effective in failure diagnosis. Meanwhile, as is also apparent from fig. 9, the method provided by the invention has higher fault diagnosis accuracy and better effect under the condition of small samples compared with other methods.
In summary, according to the rotary machine small sample fault diagnosis method based on the generation countermeasure network, the generation countermeasure network is combined with the ARGMAX multi-classification idea to construct the diagnosis model based on the ACWGAN-GP-ARGMAX, and therefore the accuracy of rotary machine small sample fault diagnosis can be effectively improved.
The foregoing is only a preferred embodiment of the present invention, and it should be noted that it is obvious to those skilled in the art that various modifications and improvements can be made without departing from the principle of the present invention, and such modifications and improvements should be considered within the scope of the present invention.

Claims (8)

1. A rotary machine small sample fault diagnosis method based on a generation countermeasure network is characterized by comprising the following steps:
s1, collecting vibration signals of the rotating machine;
s2, preprocessing the rotary mechanical vibration signal, and dividing the signal into a training set and a testing set;
s3, establishing an ACWGAN-GP-ARGMAX network model;
s4, training the ACWGAN-GP-ARGMAX network model;
s5, respectively carrying out countermeasure training on the data in each health state in the training set, and fusing the trained network models into a diagnosis model;
s6: and inputting the test set into the diagnosis model for diagnosis test, and outputting a diagnosis result.
2. The method for diagnosing the fault of the rotating machine based on the small sample generated the countermeasure network as claimed in claim 1, wherein the step S1 is implemented by disposing a vibration acceleration sensor on the rotating machine, and acquiring data of one-dimensional time series vibration acceleration of the rotating machine in various health states by using the vibration acceleration sensor to form a raw data set, wherein the raw data set includes 1 normal state and N-1 fault states.
3. The method for diagnosing the fault of the rotating machinery small sample based on the generation countermeasure network as claimed in claim 1, wherein the step S2 is implemented by performing feature extraction on the vibration signals under various health states in the original data set through fast fourier transform and standard normalization processing, and constructing the processed data into a new data set;
the standard normalization method is as follows: x ═ X (X-X)min)/(xmax-xmin) Wherein x ismaxIs the maximum value of the data sample, xminIs the minimum value of the data sample;
according to algorithm requirements, dividing the new data set into a training set and a testing set, wherein the training set is expressed as
Figure FDA0003169013850000011
Where k represents the k-th health state,
Figure FDA0003169013850000012
the ith training sample in the healthy state.
4. The rotating machine small sample fault diagnosis method based on generation countermeasure network as claimed in claim 3, wherein said ACWGAN-GP-ARGMAX network model in step S3 comprises a generator and a discriminator;
the generator is represented by a network model structure comprising 5 deconvolution layers, and input data is 4096-dimensional random noise which follows normal distribution; in the generator, the random noise is firstly input into a full-connection layer consisting of 4096 neurons, and then passes through the 5-layer deconvolution layer in sequence; the number of the filters in the first four layers of deconvolution layers is 256, 128, 64 and 32 in turn, and the last layer of deconvolution layer only has 1 filter; the size of the first layer of deconvolution kernel is 4 x 4 and the step size is 1, and the sizes of the last four layers of deconvolution kernels are all 3 x 3 and the step size is 2; through the five deconvolution layers, the input data is converted into a 64 x 64 two-dimensional matrix;
the discriminator is expressed as a network model structure comprising 5 layers of convolution layers, and input data is a 64 x 64 two-dimensional matrix generated by the generator and a 64 x 64 two-dimensional matrix obtained by shaping real data; in the discriminator, the number of the filters in the first four convolutional layers is 32, 64, 128 and 256 in turn, and the last convolutional layer only has 1 filter; the sizes of the first four convolutional layer convolutional cores are 3 multiplied by 3 and the step length is 2, and the sizes of the last convolutional layer convolutional cores are 4 multiplied by 4 and the step length is 1; through the five convolutional layers, the input data is converted into a tag value.
5. The method for fault diagnosis of rotating machinery small sample based on generation countermeasure network as claimed in claim 4, wherein the step S4 is implemented by selecting Gaussian distribution as noise distribution PnoiseAnd assigning a corresponding class label from which a subset of size n is sampled, denoted as
Figure RE-FDA0003393588650000021
Random noise subset with class label
Figure RE-FDA0003393588650000022
Input to a generator G corresponding to the health state kkForward propagation results in generating a subset of samples
Figure RE-FDA0003393588650000023
Training set P from real data of said health state kk～dataIs taking a data sample as positiveSample, is marked as
Figure RE-FDA0003393588650000024
Then training set P from real data outside the health state k(C≠k)～dataTaking the data sample as a negative sample, and recording as
Figure RE-FDA0003393588650000025
By
Figure RE-FDA0003393588650000026
And
Figure RE-FDA0003393588650000027
composing the true sample of the health state k, and generating the sample by
Figure RE-FDA0003393588650000028
Combined to form a training set subset { Xk}；
Setting the discriminator DkAnd the generator GkThe ratio of training times is m, first using the training set subset { XkThe algorithm pairs the discriminator D with AdamkPerforming m times of training; then using said random noise subset with class labels
Figure RE-FDA0003393588650000029
And Adam's algorithm to the generator GkPerforming 1 training; selecting a cross entropy loss function as an objective function, which is specifically defined as follows:
Figure RE-FDA00033935886500000210
Figure RE-FDA00033935886500000211
wherein x represents a true sampleThe utility model relates to a novel water-saving device,
Figure RE-FDA00033935886500000212
it is indicated that the generation of the sample,
Figure RE-FDA00033935886500000213
by adding at x and
Figure RE-FDA00033935886500000218
random samples obtained by linear interpolation, lambda being a gradient penalty factor,
Figure RE-FDA00033935886500000215
the expectation of the distribution of the real data is represented,
Figure RE-FDA00033935886500000216
indicating the desire to generate a distribution of data,
Figure RE-FDA00033935886500000217
denotes a gradient penalty term, P (C ═ C | Xreal) Representing the conditional probability distribution of the real data on the class label,
Figure RE-FDA0003393588650000031
representing a conditional probability distribution of the generated data on the class label,
according to the above-mentioned ratio m of training times of each round, for the said discriminator DkAnd the generator GkPerforming iterative training until the discriminator DkAnd the generator GkAchieve nash equilibrium between them.
6. The rotating machinery small sample fault diagnosis method based on generation countermeasure network as claimed in claim 5, wherein the step S5 includes,
firstly, marking data samples except the kth health state as negative samples and marking the data samples as class 0; correspondingly, marking the data sample in the kth health state as a positive sample and marking the data sample as a type 1;
second, the labeled data is inputted as a real sample to the discriminator DkPerforming countermeasure training until the model loss function converges to the minimum value, thereby obtaining a generator and a discriminator of a kth sample;
thirdly, performing the training on each type of sample in turn to obtain n trained ACWGAN-GP-ARGMAX models; separating n discriminators from the data to be used as a trained diagnosis model;
and fourthly, model fusion is carried out on the n discriminators, the probability value of the output result is combined into an n-dimensional vector, and the n-dimensional vector is accessed into a Softmax activation layer for activation, so that the probability that each value output by the output layer corresponds to each test sample and belongs to each health state is obtained.
7. The rotating machinery small sample fault diagnosis method based on generation of countermeasure network as claimed in claim 5, wherein between said step S4 and said step S5 further comprising the step S500: and performing quality evaluation on the network model trained in the step S4, and if the network model cannot pass the evaluation, repeating the step S4.
8. The method for diagnosing faults of small samples of rotating machinery based on generation of countermeasure network as claimed in claim 5, wherein the quality evaluation in the step S500 comprises:
first, evaluate the generator GkThe method comprises the steps of comparing the frequency spectrums of a generated sample and a real sample qualitatively, and then measuring the similarity of the generated sample and the real sample quantitatively by calculating the Pearson Correlation Coefficient (PCC) and the Cosine Similarity (CS) of the generated sample and the real sample;
second, the discriminator D is evaluatedkThe classification and identification capability of (2) is specifically realized by using the trained discriminator DkGenerating a data sample if said discriminator DkCan generate sufficiently realistic generated samples, the generated samples are further used to continuously perform a comparison with the discriminator DkTraining is performed so that the discriminator DkThe hidden distribution in the current sample is learned.
CN202110813313.9A 2021-07-19 2021-07-19 Rotary machine small sample fault diagnosis method based on generation countermeasure network Pending CN114091504A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110813313.9A CN114091504A (en) 2021-07-19 2021-07-19 Rotary machine small sample fault diagnosis method based on generation countermeasure network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110813313.9A CN114091504A (en) 2021-07-19 2021-07-19 Rotary machine small sample fault diagnosis method based on generation countermeasure network

Publications (1)

Publication Number Publication Date
CN114091504A true CN114091504A (en) 2022-02-25

Family

ID=80296059

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110813313.9A Pending CN114091504A (en) 2021-07-19 2021-07-19 Rotary machine small sample fault diagnosis method based on generation countermeasure network

Country Status (1)

Country Link
CN (1) CN114091504A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115796238A (en) * 2022-12-02 2023-03-14 中国地质大学（武汉） Implicit data enhancement based mechanical fault diagnosis method and device for small sample
CN116204786A (en) * 2023-01-18 2023-06-02 北京控制工程研究所 Method and device for generating designated fault trend data
CN117250970A (en) * 2023-11-13 2023-12-19 青岛澎湃海洋探索技术有限公司 Method for realizing AUV fault detection based on model embedding generation countermeasure network
CN117407784A (en) * 2023-12-13 2024-01-16 北京理工大学 Sensor data abnormality-oriented intelligent fault diagnosis method and system for rotary machine

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115796238A (en) * 2022-12-02 2023-03-14 中国地质大学（武汉） Implicit data enhancement based mechanical fault diagnosis method and device for small sample
CN116204786A (en) * 2023-01-18 2023-06-02 北京控制工程研究所 Method and device for generating designated fault trend data
CN116204786B (en) * 2023-01-18 2023-09-15 北京控制工程研究所 Method and device for generating designated fault trend data
CN117250970A (en) * 2023-11-13 2023-12-19 青岛澎湃海洋探索技术有限公司 Method for realizing AUV fault detection based on model embedding generation countermeasure network
CN117250970B (en) * 2023-11-13 2024-02-02 青岛澎湃海洋探索技术有限公司 Method for realizing AUV fault detection based on model embedding generation countermeasure network
CN117407784A (en) * 2023-12-13 2024-01-16 北京理工大学 Sensor data abnormality-oriented intelligent fault diagnosis method and system for rotary machine
CN117407784B (en) * 2023-12-13 2024-03-12 北京理工大学 Sensor data abnormality-oriented intelligent fault diagnosis method and system for rotary machine

Similar Documents

Publication Publication Date Title
Li et al. Fault diagnosis of rotating machinery based on combination of deep belief network and one-dimensional convolutional neural network
Wu et al. A hybrid classification autoencoder for semi-supervised fault diagnosis in rotating machinery
Han et al. An adaptive spatiotemporal feature learning approach for fault diagnosis in complex systems
Li et al. A novel deep autoencoder and hyperparametric adaptive learning for imbalance intelligent fault diagnosis of rotating machinery
CN114091504A (en) Rotary machine small sample fault diagnosis method based on generation countermeasure network
Li et al. Unsupervised rotating machinery fault diagnosis method based on integrated SAE–DBN and a binary processor
Li et al. WavCapsNet: An interpretable intelligent compound fault diagnosis method by backward tracking
CN112257530B (en) Rolling bearing fault diagnosis method based on blind signal separation and support vector machine
Zhang et al. Triplet metric driven multi-head GNN augmented with decoupling adversarial learning for intelligent fault diagnosis of machines under varying working condition
Yang et al. Sparse autoencoder-based multi-head deep neural networks for machinery fault diagnostics with detection of novelties
CN112147432A (en) BiLSTM module based on attention mechanism, transformer state diagnosis method and system
CN111397902A (en) Rolling bearing fault diagnosis method based on feature alignment convolutional neural network
CN112633098A (en) Fault diagnosis method and system for rotary machine and storage medium
CN115358259A (en) Self-learning-based unsupervised cross-working-condition bearing fault diagnosis method
Zhao et al. A novel deep fuzzy clustering neural network model and its application in rolling bearing fault recognition
CN114429152A (en) Rolling bearing fault diagnosis method based on dynamic index antagonism self-adaption
Yang et al. Few-shot learning for rolling bearing fault diagnosis via siamese two-dimensional convolutional neural network
Chou et al. SHM data anomaly classification using machine learning strategies: A comparative study
Hu et al. Fault diagnosis based on multi-scale redefined dimensionless indicators and density peak clustering with geodesic distances
Guan et al. Fault diagnosis of rolling bearing with imbalanced small sample scenarios
Zhang et al. A Flexible Monitoring Framework via Dynamic-Multilayer Graph Convolution Network
CN112541524B (en) BP-Adaboost multisource information motor fault diagnosis method based on attention mechanism improvement
Zhang et al. Multi‐attribute quantitative bearing fault diagnosis based on convolutional neural network
Xing et al. Intelligent fault diagnosis of rotating machinery using locally connected restricted boltzmann machine in big data era
Xu et al. A Graph Convolutional Shrinkage Network-based Fault Diagnosis Method for Industrial Process

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