A cross-domain fault sample translation generation method

Abstract

The application discloses a cross-domain fault sample translation generation method, comprising the following steps: obtaining a majority normal sample set and a minority fault sample set of a device; constructing a normal-fault sample translation model NFST composed of a normal sample autoencoder, a fault sample autoencoder, a domain adversarial discriminator and a generative adversarial discriminator; training the NFST by using the majority normal sample set and the minority fault sample set, obtaining a trained NFST, and extracting a trained normal sample autoencoder and a trained fault sample autoencoder from the trained NFST; obtaining a plurality of normal samples of a target device, and performing cross-domain sample translation conversion by using each normal sample, the trained normal sample autoencoder and the trained fault sample autoencoder to obtain converted generated fault samples.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of prognostics and health management (PHM), in particular to a cross-domain fault sample translation generation method. BACKGROUND

[0002] With the continuous progress of science and technology, the equipment in the fields of aerospace, shipbuilding, industry, etc. gradually develops in the direction of complexity, automation and intelligence, and the reliability and safety of various complex equipment are increasingly valued. Once the large and complex equipment fails, it may lead to task failure, or even cause major losses of personnel safety and economic property, and even catastrophic consequences. The PHM technology can determine the current and future state of the system according to the changes of sensor parameters and usage, and then determine the best way to perform tasks and strengthen and support maintenance and support. As a key technology in the field of PHM, fault diagnosis has always been the focus of research and attention by domestic and foreign scholars. Among them, the data-driven fault diagnosis method does not depend on analytical models or expert knowledge, and can analyze, mine and extract fault-related information from data, with higher usability and wider applicability.

[0003] The general process of the data-driven fault diagnosis method includes two stages of model training and model application, and the core is data. Therefore, good basic data conditions are an important prerequisite for the successful application of data-driven fault diagnosis methods. However, in actual engineering, due to the improvement of equipment reliability, most of the time it runs in a healthy state, and the frequency of failure is low, resulting in that most of the actual available data are samples in a healthy state, and the number of fault samples is small, and the number of healthy and fault samples is extremely unbalanced. Therefore, there is an urgent need for a low-cost way to achieve a large expansion of fault samples under the condition of extreme imbalance between healthy and fault samples. SUMMARY

[0004] The present application provides a cross-domain fault sample translation generation method to solve the technical problem of how to achieve a large expansion of fault samples under the condition of extreme imbalance between healthy and fault samples.

[0005] The embodiment of the present application provides a cross-domain fault sample translation generation method, comprising: obtaining a majority class normal sample set and a minority class fault sample set of a device; constructing a normal-fault sample translation model NFST composed of a normal sample autoencoder, a fault sample autoencoder, a domain adversarial discriminator and a generative adversarial discriminator; training the NFST by using the majority class normal sample set and the minority class fault sample set, obtaining a trained NFST, and extracting a trained normal sample autoencoder and a trained fault sample autoencoder from the trained NFST; obtaining a plurality of normal samples of a target device, and performing cross-domain sample translation conversion by using each normal sample, the trained normal sample autoencoder and the trained fault sample autoencoder, to obtain converted generated fault samples.

[0006] Preferably, the training of the NFST by using the majority class normal sample set and the minority class fault sample set to obtain the trained NFST comprises: in each round of training, performing first stage training on the normal sample autoencoder and the fault sample autoencoder in the NFST by using the majority class normal sample set and the minority class fault sample set, to obtain a first stage trained normal sample autoencoder and a first stage trained fault sample autoencoder; introducing a domain adversarial network in the NFST on the basis of the first stage trained normal sample autoencoder and the first stage trained fault sample autoencoder, and performing second stage training on the first stage trained normal sample autoencoder, the first stage trained fault sample autoencoder and the domain adversarial network by using the majority class normal sample set and the minority class fault sample set, to obtain a second stage trained normal sample autoencoder, a second stage trained fault sample autoencoder and a second stage trained domain adversarial network; introducing a generative adversarial discriminator in the NFST on the basis of the second stage trained normal sample autoencoder and the second stage trained fault sample autoencoder, and performing third stage training on the second stage trained normal sample autoencoder, the second stage trained fault sample autoencoder and the generative adversarial discriminator by using the majority class normal sample set and the minority class fault sample set, to obtain a third stage trained normal sample autoencoder, a third stage trained fault sample autoencoder and a third stage trained generative adversarial discriminator; and repeating the above steps until maximum round of training, to obtain the trained NFST.

[0007] Preferably, the extracting the trained normal sample autoencoder and the trained fault sample autoencoder from the trained NFST comprises: extracting a third stage trained normal sample autoencoder and a third stage trained fault sample autoencoder from the trained NFST, and taking the third stage trained normal sample autoencoder as the trained normal sample autoencoder, and taking the third stage trained fault sample autoencoder as the trained fault sample autoencoder.

[0008] Preferably, the first stage training comprises in-domain training and cross-domain training, wherein the first stage training the normal sample autoencoder and the fault sample autoencoder in the NFST by using the majority class normal sample set and the minority class fault sample set comprises: training the normal sample autoencoder in the NFST by using the majority class normal sample set to obtain a trained normal sample autoencoder, and training the fault sample autoencoder in the NFST by using the minority class fault sample set to obtain a trained fault sample autoencoder; and training the trained normal sample autoencoder and the trained fault sample autoencoder by using the majority class normal sample set and the minority class fault sample set to obtain the first stage trained normal sample autoencoder and the first stage trained fault sample autoencoder.

[0009] Preferably, the training the normal sample autoencoder in the NFST by using the majority class normal sample set to obtain a trained normal sample autoencoder, and training the fault sample autoencoder in the NFST by using the minority class fault sample set to obtain a trained fault sample autoencoder comprises: compressing and reconstructing each normal sample in the majority class normal sample set by using the normal sample autoencoder in the NFST to obtain a reconstructed normal sample, and compressing and reconstructing each fault sample in the minority class fault sample set by using the fault sample autoencoder in the NFST to obtain a reconstructed fault sample; calculating a reconstruction loss of each normal sample according to each normal sample and the corresponding reconstructed normal sample, and calculating a reconstruction loss of each fault sample according to each fault sample and the corresponding reconstructed fault sample; and training the normal sample autoencoder in the NFST according to the reconstruction loss of each normal sample to obtain a trained normal sample autoencoder, and training the fault sample autoencoder in the NFST according to the reconstruction loss of each fault sample to obtain a trained fault sample autoencoder.

[0010] Preferably, the cross-domain training of the trained normal sample autoencoder and the trained failure sample autoencoder by using the majority class normal sample set and the minority class failure sample set to obtain the first stage trained normal sample autoencoder and the first stage trained failure sample autoencoder comprises: calculating a first semantic consistency loss of a conversion process from a normal sample to a failure sample of each normal sample by using the encoder in the trained normal sample autoencoder and the decoder and the encoder in the trained failure sample autoencoder; calculating a second semantic consistency loss of a conversion process from a failure sample to a normal sample of each failure sample by using the encoder in the trained failure sample autoencoder and the decoder and the encoder in the trained normal sample autoencoder; and performing cross-domain training of the trained normal sample autoencoder and the trained failure sample autoencoder according to the first semantic consistency loss and the second semantic consistency loss to obtain the first stage trained normal sample autoencoder and the first stage trained failure sample autoencoder.

[0011] Preferably, the second stage training of the first stage trained normal sample autoencoder, the first stage trained failure sample autoencoder and the domain adversarial network by using the majority class normal sample set and the minority class failure sample set to obtain the second stage trained normal sample autoencoder, the second stage trained failure sample autoencoder and the second stage trained domain adversarial network comprises: inputting each normal sample in the majority class normal sample set into the encoder in the first stage trained normal sample autoencoder to obtain a feature vector of each normal sample, and simultaneously inputting each failure sample in the minority class failure sample set into the encoder in the first stage trained failure sample autoencoder to obtain a feature vector of each failure sample; assigning a label 0 to the feature vector of each normal sample and a label 1 to the feature vector of each failure sample to form a domain adversarial training data set, and performing second stage training of the first stage trained normal sample autoencoder, the first stage trained failure sample autoencoder and the domain adversarial network by using the domain adversarial training data set to obtain the second stage trained normal sample autoencoder, the second stage trained failure sample autoencoder and the second stage trained domain adversarial network.

[0012] Preferably, the third stage training of the second stage trained normal sample autoencoder, the second stage trained fault sample autoencoder and the generative adversarial discriminator by using the majority class normal sample set and the minority class fault sample set comprises: calculating a converted generated fault sample of each normal sample by using an encoder in the second stage trained normal sample autoencoder and a decoder in the second stage trained fault sample autoencoder, inputting a fault sample in the fault sample set and the generated fault sample into the generative adversarial discriminator to obtain a first discrimination result, and calculating a first generative adversarial loss of a translation generation process from a normal sample to a fault sample by using the first discrimination result; calculating a converted generated normal sample of each fault sample by using an encoder in the second stage trained fault sample autoencoder and a decoder in the second stage trained normal sample autoencoder, inputting a normal sample in the normal sample set and the generated normal sample into the generative adversarial discriminator to obtain a second discrimination result, and calculating a second generative adversarial loss of a translation generation process from a fault sample to a normal sample by using the second discrimination result; and third stage training the second stage trained normal sample autoencoder, the second stage trained fault sample autoencoder and the generative adversarial discriminator according to the first generative adversarial loss and the second generative adversarial loss to obtain a third stage trained normal sample autoencoder, a third stage trained fault sample autoencoder and a third stage trained generative adversarial discriminator.

[0013] Preferably, the cross-domain sample translation conversion by using the each normal sample, the trained normal sample autoencoder and the trained fault sample autoencoder to obtain the converted generated fault sample comprises: obtaining a feature vector of each normal sample by inputting the each normal sample into an encoder in the trained normal sample autoencoder; and obtaining the converted generated fault sample by inputting the feature vector of the each normal sample into a decoder in the trained fault sample autoencoder for cross-domain sample translation conversion.

[0014] Preferably, the cross-domain sample translation conversion by using the each normal sample, the trained normal sample autoencoder and the trained fault sample autoencoder to obtain the converted generated fault sample comprises: obtaining a feature vector of each normal sample by inputting the each normal sample into an encoder in the trained normal sample autoencoder; and obtaining the converted generated fault sample by inputting the feature vector of the each normal sample into a decoder in the trained fault sample autoencoder for cross-domain sample translation conversion.

[0015] The beneficial effects of the present application are as follows:

[0016] (1) By imposing semantic consistency loss and domain adversarial loss in the encoding-decoding model, the common features contained in health and failure samples are adaptively extracted and aligned;

[0017] (2) The effective diagnostic information contained in a large number of health samples is transferred to the failure sample generation process, breaking through the information bottleneck usually faced by existing technologies, and realizing large-scale and high-quality failure sample translation generation;

[0018] (3) The case verification is carried out by using the real data of rotating machinery object, under the condition that the amount of failure sample data is small, the failure sample generation task is carried out based on the method, and the failure diagnosis is carried out based on the generated sample, which has effectiveness in sample generation. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is a flowchart of a cross-domain failure sample translation generation method based on common feature space provided by the application;

[0020] Figure 2 is a same-domain sample reconstruction schematic diagram provided by the application;

[0021] Figure 3 is a cross-domain sample conversion schematic diagram provided by the application;

[0022] Figure 4 is a common feature space construction schematic diagram provided by the application;

[0023] Figure 5 is a sample quality improvement schematic diagram provided by the application;

[0024] Figure 6 is a bearing failure sample translation generation case A group test generation result schematic diagram provided by the application;

[0025] Figure 7 is a bearing failure sample translation generation case B group test generation result schematic diagram provided by the application;

[0026] Figure 8 is a bearing failure sample translation generation case C group test generation result schematic diagram provided by the application;

[0027] Figure 9 is a schematic diagram of a cross-domain failure sample translation generation technology based on common feature space provided by the application;

[0028] Figure 10 is a detailed flowchart of a cross-domain failure sample translation generation method provided by the application. DETAILED DESCRIPTION

[0029] It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application. In the following description, the suffixes "module", "part" or "unit" used for components are merely intended for facilitating explanation of the present application, and are by no means intended to limit the present application. Therefore, "module", "part" or "unit" can be mixedly used.

[0030] The embodiment of the present application provides a flow chart of a cross-domain fault sample translation generation method, as shown in the figure, including the following steps. Figure 10 S101: acquiring a majority class normal sample set and a minority class fault sample set of a device; S102: constructing a normal-fault sample translation model NFST composed of a normal sample auto-encoder, a fault sample auto-encoder, a domain adversarial discriminator and a generative adversarial discriminator; S103: training the NFST by using the majority class normal sample set and the minority class fault sample set, obtaining a trained NFST, and extracting a trained normal sample auto-encoder and a trained fault sample auto-encoder from the trained NFST; S104: acquiring a plurality of normal samples of a target device, and performing cross-domain sample translation conversion by using each normal sample, the trained normal sample auto-encoder and the trained fault sample auto-encoder, to obtain converted generated fault samples.

[0031] Further, the training of the NFST by using the majority-class normal sample set and the minority-class fault sample set comprises: in each training round, performing first-stage training of the normal sample auto-encoder and the fault sample auto-encoder in the NFST by using the majority-class normal sample set and the minority-class fault sample set, to obtain a first-stage trained normal sample auto-encoder and a first-stage trained fault sample auto-encoder; introducing the domain adversarial network in the NFST based on the first-stage trained normal sample auto-encoder and the first-stage trained fault sample auto-encoder, and performing second-stage training of the first-stage trained normal sample auto-encoder, the first-stage trained fault sample auto-encoder and the domain adversarial network by using the majority-class normal sample set and the minority-class fault sample set, to obtain a second-stage trained normal sample auto-encoder, a second-stage trained fault sample auto-encoder and a second-stage trained domain adversarial network; introducing the generative adversarial discriminator in the NFST based on the second-stage trained normal sample auto-encoder and the second-stage trained fault sample auto-encoder, and performing third-stage training of the second-stage trained normal sample auto-encoder, the second-stage trained fault sample auto-encoder and the generative adversarial discriminator by using the majority-class normal sample set and the minority-class fault sample set, to obtain a third-stage trained normal sample auto-encoder, a third-stage trained fault sample auto-encoder and a third-stage trained generative adversarial discriminator; and repeating the above steps until maximum training rounds, to obtain the trained NFST.

[0032] Specifically, the extracting of the trained normal sample auto-encoder and the trained fault sample auto-encoder from the trained NFST comprises: extracting the third-stage trained normal sample auto-encoder and the third-stage trained fault sample auto-encoder from the trained NFST, and taking the third-stage trained normal sample auto-encoder as the trained normal sample auto-encoder, and taking the third-stage trained fault sample auto-encoder as the trained fault sample auto-encoder.

[0033] Further, the first stage training comprises in-domain training and cross-domain training, wherein the first stage training of the normal sample autoencoder and the fault sample autoencoder in the NFST by using the majority class normal sample set and the minority class fault sample set comprises: training the normal sample autoencoder in the NFST by using the majority class normal sample set to obtain a trained normal sample autoencoder, and simultaneously training the fault sample autoencoder in the NFST by using the minority class fault sample set to obtain a trained fault sample autoencoder; and training the trained normal sample autoencoder and the trained fault sample autoencoder by using the majority class normal sample set and the minority class fault sample set to obtain the first stage trained normal sample autoencoder and the first stage trained fault sample autoencoder.

[0034] Further, the training of the normal sample autoencoder in the NFST by using the majority class normal sample set to obtain a trained normal sample autoencoder, and simultaneously training the fault sample autoencoder in the NFST by using the minority class fault sample set to obtain a trained fault sample autoencoder comprises: compressing and reconstructing each normal sample in the majority class normal sample set by using the normal sample autoencoder in the NFST to obtain a reconstructed normal sample, and simultaneously compressing and reconstructing each fault sample in the minority class fault sample set by using the fault sample autoencoder in the NFST to obtain a reconstructed fault sample; calculating a reconstruction loss of each normal sample according to each normal sample and the corresponding reconstructed normal sample, and simultaneously calculating a reconstruction loss of each fault sample according to each fault sample and the corresponding reconstructed fault sample; and training the normal sample autoencoder in the NFST according to the reconstruction loss of each normal sample to obtain a trained normal sample autoencoder, and simultaneously training the fault sample autoencoder in the NFST according to the reconstruction loss of each fault sample to obtain a trained fault sample autoencoder.

[0035] Further, the cross-domain training of the trained normal sample autoencoder and the trained failure sample autoencoder by using the majority class normal sample set and the minority class failure sample set to obtain the first stage trained normal sample autoencoder and the first stage trained failure sample autoencoder comprises: calculating a first semantic consistency loss of a conversion process from a normal sample to a failure sample of each normal sample by using an encoder in the trained normal sample autoencoder and a decoder and an encoder in the trained failure sample autoencoder; calculating a second semantic consistency loss of a conversion process from a failure sample to a normal sample of each failure sample by using an encoder in the trained failure sample autoencoder and a decoder and an encoder in the trained normal sample autoencoder; and performing cross-domain training of the trained normal sample autoencoder and the trained failure sample autoencoder according to the first semantic consistency loss and the second semantic consistency loss to obtain the first stage trained normal sample autoencoder and the first stage trained failure sample autoencoder.

[0036] Further, the second stage training of the first stage trained normal sample autoencoder, the first stage trained failure sample autoencoder and the domain adversarial network by using the majority class normal sample set and the minority class failure sample set to obtain the second stage trained normal sample autoencoder, the second stage trained failure sample autoencoder and the second stage trained domain adversarial network comprises: inputting each normal sample in the majority class normal sample set into an encoder in the first stage trained normal sample autoencoder to obtain a feature vector of each normal sample, and simultaneously inputting each failure sample in the minority class failure sample set into an encoder in the first stage trained failure sample autoencoder to obtain a feature vector of each failure sample; assigning a label 0 to the feature vector of each normal sample and a label 1 to the feature vector of each failure sample to form a domain adversarial training data set, and performing second stage training of the first stage trained normal sample autoencoder, the first stage trained failure sample autoencoder and the domain adversarial network by using the domain adversarial training data set to obtain the second stage trained normal sample autoencoder, the second stage trained failure sample autoencoder and the second stage trained domain adversarial network.

[0037] Specifically, the third stage training of the second stage trained normal sample autoencoder, the second stage trained fault sample autoencoder and the generative adversarial discriminator by using the majority class normal sample set and the minority class fault sample set comprises: calculating a converted generated fault sample of each normal sample by using an encoder in the second stage trained normal sample autoencoder and a decoder in the second stage trained fault sample autoencoder, inputting a fault sample in the fault sample set and the generated fault sample into the generative adversarial discriminator to obtain a first discrimination result, and calculating a first generative adversarial loss of a translation generation process from a normal sample to a fault sample by using the first discrimination result; calculating a converted generated normal sample of each fault sample by using an encoder in the second stage trained fault sample autoencoder and a decoder in the second stage trained normal sample autoencoder, inputting a normal sample in the normal sample set and the generated normal sample into the generative adversarial discriminator to obtain a second discrimination result, and calculating a second generative adversarial loss of a translation generation process from a fault sample to a normal sample by using the second discrimination result; and third stage training the second stage trained normal sample autoencoder, the second stage trained fault sample autoencoder and the generative adversarial discriminator according to the first generative adversarial loss and the second generative adversarial loss to obtain a third stage trained normal sample autoencoder, a third stage trained fault sample autoencoder and a third stage trained generative adversarial discriminator.

[0038] Specifically, the cross-domain sample translation conversion by using each normal sample, the trained normal sample autoencoder and the trained fault sample autoencoder to obtain a converted generated fault sample comprises: obtaining a feature vector of each normal sample by inputting the each normal sample into an encoder in the trained normal sample autoencoder; and obtaining a converted generated fault sample by inputting the feature vector of the each normal sample into a decoder in the trained fault sample autoencoder for cross-domain sample translation conversion.

[0039] The application further comprises: obtaining a feature vector of each fault sample by inputting the each fault sample into an encoder in the trained fault sample autoencoder; and obtaining a converted generated normal sample by inputting the feature vector of the each fault sample into a decoder in the trained normal sample autoencoder for cross-domain sample translation conversion.

[0040] The application provides a cross-domain fault sample translation generation technology based on a common feature space.

[0041] The cross-domain fault sample translation generation method provided by the application can be applied under the precondition that there are common semantic components between the translation domains, that is, the source domain sample contains information that can be transferred to the target domain sample generation task and is effective for the fault diagnosis task. If there is no common semantic between the two domains (for example, data from different objects without any similarity), the fault sample translation generation cannot be realized by using the method.

[0042] The cross-domain fault sample translation generation technology based on the common feature space provided by the application, as shown in Figure 1 and Figure 9 , specifically comprises the following steps.

[0043] Step 1: same-domain sample reconstruction based on an automatic encoder

[0044] The automatic encoder usually comprises two parts: an encoder and a decoder, wherein the encoder and the decoder are realized by a neural network. The encoder and the decoder can be expressed as function mappings and , f The sample x is mapped from the original data space to the feature space, and g the reconstruction of the sample z is realized by mapping x from the feature space to the original data space. For a given training set , , the automatic encoder can be expressed as follows:

[0045]

[0046] , wherein W and b are weight and bias parameters of the encoder, , These are the weights and bias parameters of the decoder.

[0047] The training process of an autoencoder is the process of optimizing the following loss function:

[0048]

[0049] in, This refers to the set of encoder and decoder parameters. Optimization problems are typically solved using gradient descent, a method commonly used in neural network training.

[0050] After training, the autoencoder can sequentially compress and reconstruct the original samples, and the error between the reconstructed samples and the original samples is very small. Therefore, the representations learned and mined from the original samples by the autoencoder contain the key information of the data. The autoencoder model is a powerful tool for unsupervised feature mining of samples.

[0051] An autoencoder is constructed, and through reconstruction using samples from the same neighborhood, NFST acquires the most basic ability to extract key representations from the original samples. A schematic diagram of the same-neighborhood sample reconstruction is shown below. Figure 2 As shown.

[0052] For normal sample set and fault sample set ,in , These represent the number of normal and faulty samples, respectively. , They represent the first One normal sample and one faulty sample. Construct a pair of dual autoencoders for each. , ,in , These represent the encoder and decoder used for normal samples, respectively. , These represent the encoder and decoder used for faulty samples, respectively. For the original normal samples... and fault samples Compression-reconstruction is performed to obtain the reconstructed normal sample. and fault samples :

[0053]

[0054] Calculate the reconstruction loss for normal samples and faulty samples respectively. and :

[0055]

[0056] Use respectively and Normal sample autoencoder and failure sample autoencoder Training optimization:

[0057]

[0058] where , are the parameter sets of normal sample autoencoder and failure sample autoencoder respectively; , is the expression after optimization of the pair of autoencoders. and

[0059] Step 2: Cross-domain sample conversion based on semantic consistency

[0060] By constraining the consistency degree of the original sample features and the output sample features after cross-domain conversion, it is required to maintain the semantic consistency of the information contained in the samples before and after translation. Taking the conversion process from normal to failure samples as an example, first, the normal sample is input into the normal sample encoder, at this time the output feature vector should contain most of the key information contained in the input sample; the feature vector of the normal sample is input into the failure sample decoder, and the failure sample obtained by translation conversion is output; then the failure sample is input into the failure sample encoder, and the feature vector of the translated sample is obtained; if the original normal sample and the translated failure sample contain common information, then the feature vector of the original sample and the feature vector of the translated sample should be very close. Based on the above deduction, the cross-domain sample conversion process based on semantic consistency is designed, and the cross-domain sample conversion schematic diagram is shown in Figure 3

[0061] For the normal sample set , first input it into the normal sample encoder to obtain the feature vector of the normal sample :

[0062]

[0063] After obtaining , unlike the same domain sample reconstruction stage, at this time is input into the failure sample decoder to output the failure sample obtained by translation conversion :

[0064]

[0065] Then, input to the failure sample encoder to obtain the feature of the translated failure sample:

[0066] ​​

[0067] According to and The semantic consistency loss of the conversion process from normal samples to fault samples is calculated :

[0068]

[0069] Similarly, the semantic consistency loss of the conversion process from fault samples to normal samples is calculated According to the following formula:

[0070]

[0071] The pair of dual autoencoders is trained and optimized according to the two-part semantic consistency loss:

[0072]

[0073] wherein, , are the parameter sets of the normal sample and fault sample autoencoders, respectively; + is the expression of the pair of dual autoencoders optimized according to the two-part semantic consistency loss.

[0074] Step 3: Common feature space construction based on domain adversarial network

[0075] Domain adversarial adaptation is a method in domain adaptation, which refers to mapping data from source and target domains into the same feature space through adversarial means to solve the problem of inconsistent data distribution between source and target domains and its adverse impact on machine learning models. Domain adversarial neural network (DANN) is the most widely used method in domain adversarial adaptation, and its core idea is to learn a common feature space for two domains using deep neural networks, thereby reducing the difference between the marginal distributions of the two domains at the representation level and .

[0076] GAN introduces a discriminator to describe the difference between the real data distribution and the generated data distribution. Inspired by this, DANN introduces a domain discriminator to distinguish the features extracted by the feature extractor in different domain data. The discrimination accuracy of the domain discriminator describes the difference between the marginal distributions of the two domain data, and the purpose of the feature generator is to deceive the domain discriminator, thereby reducing the difference in feature marginal distribution. Through domain adversarial adaptation, the features of samples from the two domains can be mapped into the same space.

[0077] On the basis of the dual autoencoder, a domain adversarial network is introduced to constrain the common features learned by the normal sample encoder and the fault sample encoder to fall within the same subspace through domain adaptation, and this subspace is defined as the common feature space. Only when the common features fall within the same subspace can the cross-domain encoding and decoding of the features have practical significance, and the common feature space construction diagram is as shown in Figure 4 .

[0078] The binary classifier is used as the domain adversarial network. First, the normal sample set and the fault sample set are sent into the respective encoders to obtain the feature vectors:

[0079]

[0080] The normal sample feature vectors are assigned a label of 0, and the fault sample feature vectors are assigned a label of 1 to form a domain adversarial training data set ; the autoencoder and are subjected to adversarial training using the data set, and the optimization is performed according to the following formula:

[0081]

[0082] wherein , , , are the parameter sets of the normal sample encoder, the fault sample encoder, and the domain adversarial network , and are the discrimination results given by the to the normal sample feature vectors and the fault sample feature vectors, i.e. , ; is a set of expressions of the dual autoencoder optimized using the domain adversarial training data set.

[0083] Step 4: Sample quality improvement based on generative adversarial

[0084] In the same field sample reconstruction stage, by applying the L2 norm loss between the reconstructed sample and the original sample, the similarity between the reconstructed sample and the real sample can be controlled. While in the cross-domain sample conversion stage, only the L2 norm loss between the original sample features and the translated sample features is applied, so only the maximum common information between the original sample and the translated sample can be guaranteed, but the similarity between the translated sample and the real sample cannot be guaranteed. Inspired by the idea of generative adversarial network, a generative adversarial discriminator is introduced into the model, the model used in the cross-domain sample conversion process is used as the generator, and the two together form a generative adversarial network, and a generative adversarial loss is applied to continuously improve the similarity between the translated sample and the real sample. The sample quality improvement schematic diagram is shown in Figure 5 .

[0085] Taking the quality improvement of the normal-fault sample translation conversion process as an example, the sample quality improvement process based on the generative adversarial network is described. For the normal sample set , first input the normal sample encoder to obtain the feature vector of the normal sample:

[0086]

[0087] The normal sample feature vector is input into the fault sample decoder to perform cross-domain sample translation conversion to obtain the converted generated fault sample:

[0088]

[0089] At this time, the real fault sample is sampled from the real fault sample set , and the generated fault sample is input into the normal-fault sample translation generation discriminator to obtain the discrimination result . The generative adversarial loss of the normal-to-fault sample translation generation process is calculated using , and the model parameters involved in the forward propagation of the process are trained and optimized:

[0090]

[0091] Among them, , , are the model parameters of the normal sample encoder, the fault sample decoder, and the normal-fault sample translation generation discriminator, is the generative adversarial loss of the normal-to-fault sample translation generation process, and the calculation method is as follows:

[0092]

[0093] For the fault-to-normal sample translation conversion process, the adversarial loss is generated, i.e., the training method is similar to the normal-to-fault process, only the positions of the normal and fault samples, the encoder and the decoder, and the discriminator are interchanged in the calculation process.

[0094] After the above four-step training is completed, the health sample is used as the input of the normal sample encoder to obtain the feature vector of the normal sample:

[0095]

[0096] Further, the normal sample feature vector is sent to the fault sample decoder for cross-domain sample translation conversion to obtain the converted generated fault sample:

[0097]

[0098] Embodiment:

[0099] First, the experimental data used in the case is introduced:

[0100] The proposed method is verified using a university CWRU bearing dataset. The test bench consists of a 2-horsepower motor, a torque sensor / decoder, a power tester, and an electronic controller. Single-point faults of different diameters are introduced on the test bearing using electrical discharge machining. An acceleration sensor placed on the bearing seat is used to collect the vibration acceleration signals of the faulty bearing at a frequency of 48 kHz. In the original dataset, vibration acceleration signal data recorded under 0, 1, 2, and 3 horsepower load conditions are included. In this case, normal (no fault) data and inner ring, rolling element, and outer ring fault data containing faults of 0.007-inch diameter under 1-horsepower load conditions are selected to constitute the dataset used for method verification.

[0101] The dataset used in this case is shown in the following table.

[0102] Table 1. University bearing dataset information

[0103]

[0104] In the above four health states, each health state contains a raw vibration signal with a length of 480000 points, and in this case the length of a single sample is set to 1200. The raw vibration signal is divided into 400 different samples for each health state. In each health state, 200 samples are randomly selected as the training set, and the remaining 200 samples are used as the test set. In order to simulate the condition of rare fault data with unbalanced normal and fault samples, in each test, all normal samples in the training set are used for training, and a number of fault samples from the training set are randomly selected, which is 0.1 times the number of normal samples. Then this part of the sample is repeated 10 times and used for training. Therefore, in this case, the extreme case of rare fault samples with a normal to fault sample ratio of 10:1 is simulated.

[0105] In order to verify the translation generation capability of the model for samples under different fault modes, three groups of tests are set up, which are translation generation between normal samples and inner ring fault samples (N-IR, group A test), translation generation between normal samples and rolling body fault samples (N-B, group B test), and translation generation between normal samples and outer ring fault samples (N-OR, group C test).

[0106] (2) Test parameter setting

[0107] In order to verify the robustness of the method proposed in this chapter, the specific deep network hyperparameters used in the case are not optimized on a large scale, but a relatively simple network model is designed according to previous experience and field conventions, so as to better illustrate that the method does not rely on a specific network architecture and does not require complex parameter tuning skills, and can still achieve good fault sample generation results.

[0108] In this case, the encoder and decoder use a fully connected network architecture that is symmetrical to each other. In order to make the model better learn the shared semantic representation between normal samples and fault samples, the encoder and decoder of normal and fault samples respectively use the weight sharing strategy in the last few layers of the output layer and the last few layers of the input layer, that is, a set of network parameters are shared at the output end of the encoder and the input end of the decoder. For the encoder and decoder network, before training begins, all layers of weights are initialized using the Kaiming normal method. The encoder and decoder network hyperparameters are shown in the following table.

[0109] Table 2 Encoder hyperparameters

[0110]

[0111] Table 3 Decoder hyperparameters

[0112]

[0113] The field adversarial discriminator for realizing the common semantic space construction of normal and fault samples is implemented by using a full connection structure, and before the start of training, the weights of all layers are initialized by using the Xavier normal method. The field adversarial discriminator hyperparameters are set as shown in the following table.

[0114] Table 4 Field adversarial discriminator hyperparameters

[0115]

[0116] Since the present research mainly focuses on realizing the translation process from normal samples to fault samples, in the link of improving sample quality based on generative adversarial, only the discriminator network for fault samples is set, and the generator composed of the encoder and the decoder in series constitutes a generative adversarial network. For the above generative adversarial discriminator network, before the start of training, the weights of all layers are initialized by using the Xavier normal method. The discriminator network is implemented by using a full connection structure, and the hyperparameters are set as shown in the following table.

[0117] Table 5 Generative adversarial network discriminator hyperparameters

[0118]

[0119] In the training process, the training iteration number epoch is set to 2000, and the sample batch size for performing gradient descent is 20. In the loss function, the weights of the reconstruction loss, the semantic consistency loss, the field adversarial loss and the generative adversarial loss are set to 1, 0.5, 1 and 1 respectively. The optimizers used by each part of the network are all Adam, wherein the initial learning rate used by the encoder and the decoder part is 1e-4, the initial learning rate used by the field discriminator part is 2e-4, and the initial learning rate used by the generative adversarial discriminator part is 1e-3. The remaining hyperparameters all use the default values of the Adam optimizer.

[0120] (3) Sample translation generation result

[0121] For the three groups of experiments, a real fault sample and two translation generated fault samples are randomly selected, and the original time domain waveform and the spectrum after Fast Fourier Transform (FFT) processing are visualized, as shown in Figures 6-8 , to analyze the fault sample translation generation effect.

[0122] From the results of the samples of the three groups of tests, it can be seen that the NFST model proposed in the paper can use normal samples to generate realistic fault samples, and the generated fault samples have a high degree of consistency in both time domain waveform and spectrum with the real samples, proving the fault sample translation generation ability of the method under the condition of small amount of data.

[0123] (4) Comparison of the improvement effect of the generated samples on fault diagnosis accuracy

[0124] The essence of fault sample translation generation under the condition of small amount of data is to expand the training data amount of the diagnosis model and improve the performance of the fault diagnosis model under the condition of small amount of samples. Therefore, the fault samples generated by the model are added to the training set of the fault diagnosis model as supplementary samples, and the improvement effect of the generated fault samples on the performance of the diagnosis model under the condition of small amount of samples is observed. In this case, in order to reduce the negative impact of different label samples imbalance on fault diagnosis, SVM which is not sensitive to small samples and imbalanced samples is selected as the diagnosis model for analyzing the improvement effect of the generated samples on fault diagnosis accuracy.

[0125] 200 normal samples and 20 fault samples of each fault mode are randomly selected from the real samples to simulate the unbalanced training set of the diagnosis model under the condition of small amount of fault data. In addition, 200 normal samples and 200 fault samples of each fault mode are selected as the fault diagnosis test set, which do not overlap with the training set samples. Different numbers of generated fault samples are supplemented in the unbalanced training set, and then the diagnosis model is trained, and the diagnosis accuracy results obtained on the test set are as follows. The results show that the fault samples generated by NFST have a positive effect on improving the accuracy of the diagnosis model.

[0126] Table 6 Accuracy of fault diagnosis using different numbers of generated fault samples

[0127] The above describes only some specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for translating and generating cross-domain fault samples, characterized in that, include: Obtain the majority class of normal samples and the minority class of faulty samples for the device; Construct a normal-fault sample translation model NFST consisting of a normal sample autoencoder, a fault sample autoencoder, a domain adversarial discriminator, and a generative adversarial discriminator; The NFST is trained using the majority class normal sample set and the minority class fault sample set to obtain a trained NFST, and a trained normal sample autoencoder and a trained fault sample autoencoder are extracted from the trained NFST. Multiple normal samples of the target device are acquired, and cross-domain sample translation is performed using each normal sample, a trained normal sample autoencoder, and a trained fault sample autoencoder to obtain the converted generated fault samples. The step of training the NFST using the majority class normal sample set and the minority class fault sample set to obtain a trained NFST includes: In each training round, the normal sample autoencoder and the faulty sample autoencoder in the NFST are trained using the majority class normal sample set and the minority class faulty sample set, including first-stage training and cross-domain training, to obtain the first-stage trained normal sample autoencoder and the first-stage trained faulty sample autoencoder, which include: The normal sample autoencoder in the NFST is trained in the same domain using the majority class normal sample set to obtain the trained normal sample autoencoder. At the same time, the fault sample autoencoder in the NFST is trained in the same domain using the minority class fault sample set to obtain the trained fault sample autoencoder. The trained normal sample autoencoder and the trained fault sample autoencoder are trained across domains using the majority class normal sample set and the minority class fault sample set to obtain a first-stage trained normal sample autoencoder and a first-stage trained fault sample autoencoder. This includes: calculating a first semantic consistency loss for the transformation process from a normal sample to a fault sample for each normal sample using the encoder in the trained normal sample autoencoder and the decoder and encoder in the trained fault sample autoencoder; calculating a second semantic consistency loss for the transformation process from a fault sample to a normal sample for each fault sample using the encoder in the trained fault sample autoencoder and the decoder and encoder in the trained normal sample autoencoder; and performing cross-domain training on the trained normal sample autoencoder and the trained fault sample autoencoder based on the first semantic consistency loss and the second semantic consistency loss to obtain a first-stage trained normal sample autoencoder and a first-stage trained fault sample autoencoder. Based on the normal sample autoencoder and the fault sample autoencoder trained in the first stage, the domain adversarial network in the NFST is introduced, and the normal sample autoencoder, the fault sample autoencoder trained in the first stage, and the domain adversarial network are trained in the second stage using the majority class normal sample set and the minority class fault sample set, to obtain the normal sample autoencoder, the fault sample autoencoder trained in the second stage, and the domain adversarial network trained in the second stage. Based on the normal sample autoencoder trained in the second stage and the fault sample autoencoder trained in the second stage, the generative adversarial discriminator in the NFST is introduced, and the normal sample autoencoder trained in the second stage, the fault sample autoencoder trained in the second stage, and the generative adversarial discriminator are trained in the third stage using the majority class normal sample set and the minority class fault sample set, to obtain the normal sample autoencoder trained in the third stage, the fault sample autoencoder trained in the third stage, and the generative adversarial discriminator trained in the third stage. Repeat the above steps until the maximum number of training rounds is reached, and a trained NFST is obtained.

2. The method according to claim 1, characterized in that, The step of extracting the trained normal sample autoencoder and the trained fault sample autoencoder from the trained NFST includes: The normal sample autoencoder and the fault sample autoencoder trained in the third stage are extracted from the trained NFST. The normal sample autoencoder trained in the third stage is used as the trained normal sample autoencoder, and the fault sample autoencoder trained in the third stage is used as the trained fault sample autoencoder.

3. The method according to claim 2, characterized in that, The step of training the normal sample autoencoder in the NFST using the majority class normal sample set to obtain a trained normal sample autoencoder, and simultaneously training the fault sample autoencoder in the NFST using the minority class fault sample set to obtain a trained fault sample autoencoder, includes: The normal sample autoencoder in the NFST is used to compress and reconstruct each normal sample in the majority class normal sample set to obtain the reconstructed normal sample. At the same time, the fault sample autoencoder in the NFST is used to compress and reconstruct each fault sample in the minority class fault sample set to obtain the reconstructed fault sample. Based on each normal sample and its corresponding reconstructed normal sample, calculate the reconstruction loss for each normal sample. At the same time, based on each faulty sample and its corresponding reconstructed faulty sample, calculate the reconstruction loss for each faulty sample. The normal sample autoencoder in the NFST is trained in the same neighborhood based on the reconstruction loss of each normal sample to obtain the trained normal sample autoencoder. At the same time, the fault sample autoencoder in the NFST is trained in the same neighborhood based on the reconstruction loss of each fault sample to obtain the trained fault sample autoencoder.

4. The method according to claim 2, characterized in that, The step of using the majority class normal sample set and the minority class faulty sample set to perform a second-stage training on the normal sample autoencoder trained in the first stage, the faulty sample autoencoder trained in the first stage, and the domain adversarial network to obtain the normal sample autoencoder trained in the second stage, the faulty sample autoencoder trained in the second stage, and the domain adversarial network trained in the second stage includes: By inputting each normal sample in the majority class normal sample set into the encoder of the normal sample autoencoder trained in the first stage, the feature vector of each normal sample is obtained. At the same time, by inputting each fault sample in the minority class fault sample set into the encoder of the fault sample autoencoder trained in the first stage, the feature vector of each fault sample is obtained. By assigning a label of 0 to the feature vector of each normal sample and a label of 1 to the feature vector of each faulty sample, a domain adversarial training dataset is constructed. The domain adversarial training dataset is then used to perform a second-stage training on the normal sample autoencoder trained in the first stage, the faulty sample autoencoder trained in the first stage, and the domain adversarial network, resulting in the normal sample autoencoder trained in the second stage, the faulty sample autoencoder trained in the second stage, and the domain adversarial network trained in the second stage.

5. The method according to claim 2, characterized in that, The step of using the majority class normal sample set and the minority class faulty sample set to perform a third-stage training on the normal sample autoencoder trained in the second stage, the faulty sample autoencoder trained in the second stage, and the generative adversarial discriminator trained in the third stage, to obtain the normal sample autoencoder trained in the third stage, the faulty sample autoencoder trained in the third stage, and the generative adversarial discriminator trained in the third stage, includes: Using the encoder in the normal sample autoencoder trained in the second stage and the decoder in the fault sample autoencoder trained in the second stage, the generated fault sample after the transformation of each normal sample is calculated, and the fault samples in the fault sample set and the generated fault samples are input into the generative adversarial discriminator to obtain the first discrimination result, and the first generative adversarial loss of the translation generation process from normal sample to fault sample is calculated using the first discrimination result. Using the encoder in the fault sample autoencoder trained in the second stage and the decoder in the normal sample autoencoder trained in the second stage, the generated normal sample after the transformation of each fault sample is calculated, and the normal samples in the normal sample set and the generated normal samples are input into the generative adversarial discriminator to obtain the second discrimination result, and the second generative adversarial loss of the translation generation process from fault sample to normal sample is calculated using the second discrimination result. The normal sample autoencoder trained in the second stage, the faulty sample autoencoder trained in the second stage, and the generative adversarial discriminator trained in the second stage are trained in the third stage based on the first generative adversarial loss and the second generative adversarial loss, to obtain the normal sample autoencoder trained in the third stage, the faulty sample autoencoder trained in the third stage, and the generative adversarial discriminator trained in the third stage.

6. The method according to claim 1, characterized in that, The method of using each normal sample, the trained normal sample autoencoder, and the trained fault sample autoencoder to perform cross-domain sample translation conversion, resulting in the generated fault samples, includes: By inputting each normal sample into the encoder of the trained normal sample autoencoder, the feature vector of each normal sample is obtained; By inputting the feature vector of each normal sample into the decoder of the trained fault sample autoencoder, cross-domain sample translation is performed to obtain the converted generated fault sample.

7. The method according to claim 1, characterized in that, Also includes: By inputting each fault sample into the encoder of the trained fault sample autoencoder, the feature vector of each fault sample is obtained; By inputting the feature vector of each faulty sample into the decoder of the trained normal sample autoencoder, cross-domain sample translation is performed to obtain the converted normal sample.

Images