Binary fault feature adaptive mining and diagnosis method based on binomial adversarial representation learning

Abstract

The application discloses a binarization fault feature adaptive mining and diagnosis method based on binomial confrontation representation learning, and the method comprises the following steps: obtaining standardized vibration signals and corresponding true health state labels under various health states according to original vibration signals of a machine under various health states; performing unsupervised confrontation training on a binomial representation encoder by using binomial confrontation samples, the standardized vibration signals and the corresponding true health state labels, to obtain a trained binomial representation encoder; forming an initial mechanical fault diagnosis classifier by connecting the trained binomial representation encoder and a full-connection classification network in series, performing supervised fine-tuning on the initial mechanical fault diagnosis classifier, and obtaining a practical mechanical fault diagnosis classifier, so that the practical mechanical fault diagnosis classifier is used for fault diagnosis of a machine to be diagnosed.

Description

TECHNICAL FIELD

[0001] The application relates to a binarization fault feature adaptive mining and diagnosis method based on binomial confrontation representation learning. BACKGROUND

[0002] With the progress of science and technology, mechanical equipment becomes more and more complex, and is usually operated under complex or harsh conditions. Once a failure occurs, the equipment may not be able to complete the prescribed task, and even may cause a catastrophic accident, causing economic losses and casualties.

[0003] Mechanical fault diagnosis technology can utilize measurement data to extract key fault features, so as to identify the health condition in the operation process of the machinery, and is an effective means to ensure the reliability and safety of industrial systems. However, in the big data era, how to mine useful information for fault diagnosis from a large amount of monitoring data has become a problem to be urgently studied. The data-driven fault diagnosis method establishes a mapping relationship between monitoring data and mechanical health state, and then realizes the discrimination of different fault modes. Therefore, the data-driven mechanical fault diagnosis method has become a popular research direction.

[0004] The general process of the data-driven fault diagnosis method includes data preprocessing, feature extraction and health state classification. Among them, the quality of feature extraction is a key factor affecting the performance of fault diagnosis. The feature extraction method includes manual feature extraction and adaptive feature extraction. Compared with the manual feature extraction method, the adaptive feature extraction method can automatically extract a feature vector that can represent the health state information from the original signal, and does not depend on expert knowledge, so it has been more widely used in mechanical fault diagnosis.

[0005] In recent years, the deep learning (DL) method has been widely used in fault feature extraction and fault diagnosis tasks due to its strong representation learning and nonlinear mapping ability. The DL-based fault diagnosis method alleviates the dependence of traditional manual feature extraction on expert knowledge, and can realize adaptive representation learning. However, the existing DL-based fault diagnosis representation learning method has the following two problems.

[0006] 1) Most of the DL-based representation learning methods can be regarded as a black box, and the learning process is uncontrollable. Therefore, the learned representation may have low quality and the training data may be over-fitted.

[0007] 2) The good performance of the DL model depends on sufficient training data. Therefore, when the training data is insufficient, overfitting phenomenon occurs, and the diagnosis performance inevitably decreases significantly. SUMMARY

[0008] To overcome the above problems, the application provides a binary fault feature adaptive mining and diagnosis method based on binomial adversarial representation learning.

[0009] The application provides a binary fault feature adaptive mining and diagnosis method based on binomial adversarial representation learning, which comprises the following steps: obtaining standardized vibration signals of a machine in various health states and corresponding true health state labels according to original vibration signals of the machine in the various health states; using binomial adversarial samples sampled from a binomial distribution, the standardized vibration signals of the machine and the corresponding true health state labels to perform unsupervised adversarial training on a binomial representation encoder for binomial adversarial representation learning in an auto-encoder through a generation-discrimination learning mechanism, so as to obtain the binomial representation encoder capable of learning a representation vector subject to a binomial distribution from original vibration signals of the machine, as a binary fault feature extractor for mechanical fault diagnosis; connecting the binomial representation encoder and a fully connected classification network in series to form an initial mechanical fault diagnosis classifier; performing supervised fine-tuning on the initial mechanical fault diagnosis classifier by using the standardized vibration signals of the machine in the various health states and the corresponding true health state labels, so as to obtain a practical mechanical fault diagnosis classifier; and using the practical mechanical fault diagnosis classifier to perform fault diagnosis on a machine to be diagnosed.

[0010] Preferably, the obtaining of the standardized vibration signals of the machine in various health states and the corresponding true health state labels comprises: collecting original vibration signals of the machine in various health states, wherein the various health states include a normal state and various fault states; dividing the original vibration signals into a plurality of non-overlapping signal segments according to a fixed window length; performing standardization processing on each signal segment to obtain each standardized vibration signal; and adding the true health state labels of the machine to each standardized vibration signal according to the health state of the original vibration signal corresponding to each standardized vibration signal.

[0011] Preferably, the binomial adversarial samples sampled from the binomial distribution, the standardized vibration signals of the machine and the corresponding true health status labels are used to generate a discriminative learning mechanism to perform unsupervised adversarial training on a binomial representation encoder in an autoencoder for binomial adversarial representation learning, so as to obtain a binomial representation encoder capable of learning a representation vector subject to a binomial distribution from the original vibration signals of the machine, as a binary fault feature extractor for mechanical fault diagnosis, comprising: encoding the standardized vibration signals of the machine by using the binomial representation encoder in the autoencoder to obtain a representation vector, and decoding the representation vector by using a decoder in the autoencoder to obtain a reconstructed signal; using a binomial discriminator for identifying a binomial distribution to discriminate whether the representation vector and the binomial adversarial sample are subject to a binomial distribution to obtain a binomial discrimination result; determining a reconstruction loss according to the standardized vibration signals of the machine and the reconstructed signal; determining a binomial loss according to the representation vector and the binomial discrimination result; continuously adjusting the parameters of the binomial representation encoder, the parameters of the decoder and the parameters of the binomial discriminator until the reconstruction loss and the binomial loss are minimized or stabilized, and taking the binomial representation encoder when the reconstruction loss and the binomial loss are minimized or stabilized as the binary fault feature extractor for mechanical fault diagnosis.

[0012] Preferably, the using a binomial discriminator for identifying a binomial distribution to discriminate whether the representation vector and the binomial adversarial sample are subject to a binomial distribution to obtain a binomial discrimination result comprises: inputting the representation vector into the binomial discriminator to make the binomial discriminator discriminate whether the representation vector is subject to a binomial distribution to obtain a binomial discrimination result of the representation vector; and inputting the binomial adversarial sample into the binomial discriminator to make the binomial discriminator discriminate whether the binomial adversarial sample is subject to a binomial distribution to obtain a binomial discrimination result of the binomial adversarial sample.

[0013] Preferably, the determining a binomial loss according to the representation vector and the binomial discrimination result comprises: determining a binomial loss of the binomial discriminator according to the binomial discrimination results of the representation vector and the binomial adversarial sample; and determining a binomial loss of the autoencoder according to the binomial discrimination result of the representation vector.

[0014] Preferably, the continuously adjusting the parameters of the binomial representation encoder, the parameters of the decoder and the parameters of the binomial discriminator until the reconstruction loss and the binomial loss are minimum or stable comprises: adjusting the parameters of the binomial discriminator according to the binomial loss of the binomial discriminator and the binomial loss of the autoencoder; adjusting the parameters of the autoencoder according to the reconstruction loss, the binomial loss of the binomial discriminator and the binomial loss of the autoencoder; alternately performing the above two steps until the reconstruction loss, the binomial loss of the binomial discriminator and the binomial loss of the autoencoder are minimum or stable.

[0015] Preferably, the fully connected classification network comprises a fully connected layer and a Softmax classification layer.

[0016] Preferably, the supervised fine-tuning of the initial mechanical fault diagnosis classifier by using the standardized vibration signals of the mechanical in various health states and the corresponding true health state labels to obtain a practical mechanical fault diagnosis classifier comprises: inputting the standardized vibration signals of the mechanical in various health states into the initial mechanical fault diagnosis classifier to obtain the classification results output by the initial mechanical fault diagnosis classifier; continuously adjusting the parameters of the initial mechanical fault diagnosis classifier to minimize the cross-entropy loss between the output classification results and the true health state labels corresponding to the input standardized vibration signals; and taking the mechanical fault diagnosis classifier at the time of minimum cross-entropy loss as the practical mechanical fault diagnosis classifier.

[0017] Preferably, the fault diagnosis of the mechanical to be diagnosed by using the practical mechanical fault diagnosis classifier comprises: acquiring the original vibration signals of the mechanical to be diagnosed and processing them into standardized vibration signals; inputting the standardized vibration signals into the practical mechanical fault diagnosis classifier to perform feature extraction and fault diagnosis classification on the standardized vibration signals by the practical mechanical fault diagnosis classifier, and obtaining the health state of the mechanical to be diagnosed.

[0018] The application provides a binarized fault feature adaptive mining and diagnosis method based on binomial adversarial representation learning. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is a flowchart of the binarized fault feature adaptive mining and diagnosis method based on binomial adversarial representation learning provided by the application;

[0020] Figure 2 is a mechanical fault diagnosis model training process diagram based on BARL provided by the application.

[0021] Figure 3 This is a structural diagram of the mechanical fault diagnosis model based on BARL provided by the present invention;

[0022] Figure 4 is an NC-index diagram of the four methods on the CWRU dataset provided by this invention. Detailed Implementation

[0023] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In the following description, suffixes such as "module," "part," or "unit" used to denote elements are used only for the purpose of illustrative purposes and have no inherent meaning. Therefore, "module," "part," or "unit" may be used interchangeably.

[0024] To address the problems mentioned in the background section, this invention proposes a Binomial Adversarial Representation Learning (BARL) method for mechanical fault diagnosis. This method combines an autoencoder (AE) and adversarial learning to learn representation vectors following a binomial distribution from the raw vibration signals of machinery. On one hand, the adversarial learning mechanism forces the model to learn a representation space that falls on the prior distribution, thus enabling the learning of better fault representations in a controllable manner. On the other hand, the binomial distribution makes it easier for representation vectors to have small intra-class distances (ICD) and large inter-class distances (CCD), thereby improving the robustness and generalization ability of downstream mechanical fault diagnosis tasks, especially under conditions of insufficient training data.

[0025] See Figure 1 This invention provides an adaptive mining and diagnosis method for binary fault features based on binomial adversarial representation learning. The method may include the following steps:

[0026] Step S101: Based on the original vibration signals of the machine under various health conditions, obtain the standardized vibration signals of the machine under various health conditions and the corresponding real health status labels.

[0027] The raw vibration signals of the machine are collected under various health conditions, including normal and various fault conditions. Then, the raw vibration signals are divided into multiple non-overlapping signal segments according to a fixed window length. Each signal segment is standardized to obtain a standardized vibration signal. Based on the health status of the raw vibration signal corresponding to each standardized vibration signal, the actual health status label of the machine is added to each standardized vibration signal.

[0028] Step S102: Using binomial adversarial samples sampled from the binomial distribution, the standardized vibration signal of the machine, and the corresponding real health status label, an unsupervised adversarial training is performed on the binomial representation encoder in the autoencoder for binomial adversarial representation learning using a generative-discriminative learning mechanism. This results in a binomial representation encoder that can learn representation vectors that follow a binomial distribution from the original vibration signal of the machine, serving as a binarized fault feature extractor for mechanical fault diagnosis.

[0029] Specifically, the standardized vibration signal of the machine is encoded using a binomial representation encoder in the automatic encoder to obtain a representation vector, and the representation vector is decoded using a decoder in the automatic encoder to obtain a reconstructed signal. A binomial discriminator, used to identify binomial distributions, is used to determine whether the representation vector and the binomial adversarial sample follow a binomial distribution, obtaining a binomial discrimination result. Based on the standardized vibration signal of the machine and the reconstructed signal, a reconstruction loss is determined. Based on the representation vector and the binomial discrimination result, a binomial loss is determined. The parameters of the binomial representation encoder, the decoder, and the discriminator are continuously adjusted until the reconstruction loss and the binomial loss are minimized or stabilized. The binomial representation encoder at which the reconstruction loss and the binomial loss are minimized or stabilized is used as a binary fault feature extractor for mechanical fault diagnosis.

[0030] The step of using a binomial discriminator to identify binomial distributions to determine whether the representation vector and the binomial adversarial sample follow a binomial distribution, and obtaining a binomial discrimination result, includes: inputting the representation vector into the binomial discriminator so that the binomial discriminator determines whether the representation vector follows a binomial distribution, thereby obtaining a binomial discrimination result for the representation vector; and inputting the binomial adversarial sample into the binomial discriminator so that the binomial discriminator determines whether the binomial adversarial sample follows a binomial distribution, thereby obtaining a binomial discrimination result for the binomial adversarial sample.

[0031] The step of determining the binomial loss based on the representation vector and the binomial discrimination result includes: determining the binomial loss of the binomial discriminator based on the representation vector and the binomial discrimination result of the binomial adversarial example; and determining the binomial loss of the autoencoder based on the binomial discrimination result of the representation vector.

[0032] The step of continuously adjusting the parameters of the binomial characterizing encoder, the decoder, and the binomial discriminator until the reconstruction loss and the binomial loss are minimized or stabilized includes: adjusting the parameters of the binomial discriminator based on the binomial loss of the binomial discriminator and the binomial loss of the autoencoder; adjusting the parameters of the autoencoder based on the reconstruction loss, the binomial loss of the binomial discriminator, and the binomial loss of the autoencoder; and alternately performing the above two steps until the reconstruction loss, the binomial loss of the binomial discriminator, and the binomial loss of the autoencoder are minimized or stabilized.

[0033] Step S103: By connecting the binomial characterization encoder in series with the fully connected classification network, an initial mechanical fault diagnosis classifier is formed.

[0034] The fully connected classification network includes a fully connected layer and a Softmax classification layer.

[0035] Step S104: Using the standardized vibration signals of the machine under various health conditions and the corresponding real health status labels, the initial mechanical fault diagnosis classifier is fine-tuned in a supervised manner to obtain a practical mechanical fault diagnosis classifier.

[0036] Specifically, the standardized vibration signals of the machine under various health conditions are input into the initial mechanical fault diagnosis classifier to obtain the classification result output by the initial mechanical fault diagnosis classifier; the parameters of the initial mechanical fault diagnosis classifier are continuously adjusted to minimize the cross-entropy loss between the output classification result and the true health state label corresponding to the input standardized vibration signal; the mechanical fault diagnosis classifier with the minimum cross-entropy loss is used as the practical mechanical fault diagnosis classifier.

[0037] Step S105: Use the practical mechanical fault diagnosis classifier to diagnose the faults in the machine to be diagnosed.

[0038] Specifically, the original vibration signal of the machine to be diagnosed is acquired and processed into a standardized vibration signal; the standardized vibration signal is input into the practical mechanical fault diagnosis classifier, which performs feature extraction and fault diagnosis classification on the standardized vibration signal to obtain the health status of the machine to be diagnosed.

[0039] This invention proposes a controllable adaptive fault representation learning mechanism that induces the representation space to fall on a favorable prior distribution, thereby improving the robustness and generalization ability of fault representations even with insufficient training data. Adversarial representation (AE) is a good deep learning-based representation learning architecture; adversarial learning can force the network to learn representations that approximate a certain prior distribution through a generative-discriminative learning mechanism. Therefore, the AE model can be combined with an adversarial learning mechanism to control the representation learning process of AE. The binomial distribution can only have values ​​of 0 or 1, possessing the advantage of sparsity. Compared to continuous distributions such as the Gaussian distribution, the binomial distribution's binary nature (either 0 or 1) makes it easier for representations of different health states to be separated further in the representation space, thus making it difficult for features of different health states to overlap. Therefore, the binomial distribution is a suitable distribution for robust fault representation learning when training data is insufficient.

[0040] The following is combined with Figure 2 and Figure 3 Please provide a detailed explanation.

[0041] The training process of the BARL-based mechanical fault diagnosis model is as follows: Figure 2 As shown, the process includes three stages: data preprocessing, unsupervised adversarial training, and supervised fine-tuning. The specific execution process of these three stages is as follows.

[0042] Phase 1: Data Preprocessing

[0043] First, vibration signals were collected for each health state (including healthy and all failure modes). To reduce sample complexity and increase the number of samples input to the network, the raw vibration signals were divided into multiple segments with a fixed window length. Then, all segments were normalized to -1 to 1. Normalized samples and their labels are then presented. This is the input data for BARL, where N is... T This represents the number of samples in the training set. Where x... i It is a standardized sample, y i It is a label indicating your true health status.

[0044] Phase Two: Unsupervised Combat Training

[0045] During the unsupervised adversarial training phase, two parallel training branches are executed alternately: one for the autoencoder and the other for the binomial discriminator (or discriminator). The autoencoder consists of a binomial representation encoder and a decoder, while the discriminator distinguishes whether the input representation vector follows a binomial distribution. The model structure is as follows: Figure 3 As shown.

[0046] First, the forward propagation process is executed. Input signal x iAfter being input into the binomial representation encoder, the data is processed by convolutional layers (Conv), batch normalization layers (BN), flattening layers, and fully connected layers (FC) to be encoded into a representation vector h. i =BiEnc(x i ), where BiEnc(·) is a binomial mapping representing the nonlinear function expressed by the encoder. The decoder uses a transposed convolutional layer (TransConv) to transform h i Decode the signal to obtain the reconstructed signal. Where Dec(·) is the nonlinear function mapping expressed by the decoder. Simultaneously, binomial adversarial samples b are sampled from the binomial distribution B~b(n,p). i , where n is a binomial vector b i In the dimension b, p is the probability that each dimension has a value of 1. That is, b i The values ​​of each dimension follow a Bernoulli distribution with probability p. The representation vector h output by the encoder... i And binomial adversarial sample vector b i The input binomial discriminator, after processing through an FC layer and a Sigmoid output layer, yields the discriminant result D(h). i ) and D(b i ).

[0047] Based on the learning objectives of each part of the model, the loss function can be written in the following form.

[0048] The binomial discriminator attempts to distinguish whether an input vector is a true vector sampled from a binomial distribution. Therefore, the binomial loss function of the discriminator is:

[0049]

[0050] Where N b and N h These are the number of binomial vectors and representation vectors in the same batch of samples, respectively.

[0051] The reconstruction loss of the autoencoder is:

[0052]

[0053] Where N s N represents the number of input vibration signal samples. s =N h .

[0054] The binomial representation encoder and the binomial discriminator play a game with each other, attempting to learn a representation vector following a binomial distribution from the input signal. Therefore, the binomial loss of the encoder is:

[0055]

[0056] The trainable parameters in the binomial discriminator, binomial representation encoder, and decoder are respectively represented as Θ. D Θ BiEnc and Θ Dec The optimization objectives for the three parts of the model are as follows:

[0057]

[0058]

[0059]

[0060] Based on the above optimization objective, the parameters of the binomial discriminator, binomial representation encoder, and decoder are updated alternately until the loss value stabilizes and reaches Nash equilibrium.

[0061] Phase Three: Oversighted Fine-tuning

[0062] After the unsupervised adversarial training in Phase Two, the trained binomial representation encoder can learn key features from the input vibration signal and convert the input signal into a representation vector following a binomial distribution. Therefore, the binomial representation encoder from the autoencoder is extracted as the feature extractor for the mechanical fault diagnosis model; and a fully connected layer and a Softmax classification layer are then connected to it to form a complete fault diagnosis classifier. The nonlinear function mapping expressed by this complete fault diagnosis classifier is denoted as C(·). In the supervised fine-tuning phase, the fault diagnosis classifier parameters Θ are updated during training. c This makes the model applicable to sample x. i The given classification result C(x) i ) and True Health Status Labels i Minimize the cross-entropy loss between them.

[0063] After the fault diagnosis classifier parameters converge to the optimal value, the model can diagnose the health status of each input vibration signal sample.

[0064] Application Case

[0065] 1. Data Description

[0066] In this case study, the rolling bearing dataset provided by Case Western Reserve University (CWRU) was used to validate the effectiveness of the proposed method. The test bench consisted of a 2-horsepower motor, a torque sensor / encoder, a force gauge, and control electronics. The motor load was 1 horsepower. Single-point faults were introduced into the inner ring, rolling elements, and outer ring of different bearings, with the diameter of the injected fault at each location being 0.007 inches, 0.014 inches, and 0.021 inches, respectively. Therefore, there were a total of 10 different health states, including the bearing being in a fault-free (normal) state, as shown in Table 1.

[0067] Table 1. Description of the CWRU dataset

[0068]

[0069] Under each health condition, vibration signals were acquired using an accelerometer at a sampling frequency of 48 kHz for 10 seconds. The vibration signals were then segmented using a fixed-width window of 1200 data points, with each segment not overlapping. Therefore, the vibration signals under each health condition were divided into 4000 samples of 1200 data points each. All samples were then divided into training and test sets according to a certain ratio, and the values ​​were standardized to the interval [-1, 1] based on the maximum and minimum values ​​of the training samples.

[0070] 2. Parameter Settings

[0071] (1) Model parameters

[0072] In the model established according to this method, the hyperparameters representing the encoder and decoder are shown in Tables 2 and 3.

[0073] Table 2. Hyperparameters of the binomial encoder in the model

[0074] Layer Number Hyper-parameters 1 1d convolution, number of filters = 8, kernel size = 64 2 1d convolution, number of filters = 8, kernel size = 32 3 1d convolution, number of filters = 16, kernel size = 8 4 1d convolution, number of filters = 32, kernel size = 3 5 Flattening layer 6 Fully connected layer, number of units = 256 7 Fully connected layer, number of units = 512

[0075] Table 3. Hyperparameters of the decoder in the model

[0076]

[0077]

[0078] The binomial discriminator is a fully connected neural network with 512-256-128-64-1 neurons per layer. The first four layers are activated by the ReLU function, and the output layer is activated by the Sigmoid function. Before supervised fine-tuning, a four-layer fully connected classification network with 256-128-64-10 neurons per layer is connected after the adversarially trained binomial representation encoder. The first three layers of the classification network are activated by the ReLU function, and the output layer is activated by the Softmax function. After supervised fine-tuning, the model formed by concatenating the binomial representation encoder and the fully connected classification network constitutes a complete fault diagnosis classification model.

[0079] (2) Training parameters

[0080] In the unsupervised adversarial training phase, the batch size is 100 and the number of training rounds is 2000. In the supervised fine-tuning phase, the batch size is 100 and the number of training rounds is 200.

[0081] To test the model's performance under different numbers of available training samples, the proportion of training samples to all samples was set to 0.2, 0.4, 0.6, and 0.8, respectively, for the experiment.

[0082] 3. Comparison Methods

[0083] To illustrate the superiority of the proposed method, and specifically to examine the effectiveness of the binomial adversarial representation learning mechanism, three deep learning-based representation learning and fault diagnosis methods were compared with the proposed method. The hyperparameter settings of the comparison methods are shown in Table 4.

[0084] Table 4. Hyperparameter settings for the comparison method

[0085]

[0086] 4. Diagnostic Results and Discussion

[0087] (1) In most cases, especially under conditions of insufficient training data, BARL has higher diagnostic accuracy compared to the comparative method.

[0088] Table 5. Diagnostic results of BARL and contrastive methods on the CWRU dataset.

[0089]

[0090] The diagnostic results of BARL and the comparative methods proposed in this invention are shown in Table 5. Five independent trials were conducted under each training sample ratio condition to eliminate the influence of randomness in the results. The figures in the table consist of (mean accuracy ± standard deviation of 5 trials). According to the diagnostic results, BARL outperforms other comparative methods in all cases. Especially when faced with insufficient training samples (only 20% of the samples are available for model training), the advantage of BARL over other comparative methods is even more significant.

[0091] (2) Compared with other methods, the representation learned by BARL has a larger CCD value and a smaller ICD value.

[0092] As mentioned above, the purpose of learning representations that follow a binomial distribution is to obtain a larger CCD and a smaller ICD, i.e., better natural clustering characteristics. To verify the conclusion, four different indices (NC-index) of the representations learned by the model were calculated. These indices reflect the natural clustering characteristics of the representations, including: 1) the ratio of CCD to ICD under Euclidean distance, 2) the ratio of CCD to ICD under cosine distance, 3) the silhouette coefficient (Sil-Coef), and 4) the Calinski-Harabasz Index (CHI).

[0093] To demonstrate the generalization ability of the method, the training sample ratio was set to 0.6, and representations were extracted from the test samples. The representation vector of BARL / GARL refers to the output vector of their binomial / Gaussian encoder. The representation vector of CNN refers to the output vector of its last convolutional layer. The representation vector of SSAE refers to the encoding vector of its deepest hidden layer.

[0094] For NC-index calculations using Euclidean and cosine distance metrics, the ICD value is the average distance between representation vectors of samples with the same label, while the CCD value is the average distance between representation vectors of samples with different labels. Sil-Coef and CHI measure natural clustering properties by training a K-means clustering model using the learned representations, and then calculate the distance between samples in a cluster and the cluster center, as well as the distance between different cluster centers. Higher Sil-Coef and CHI values ​​indicate that the representations have better natural clustering properties.

[0095] The NC-index calculation results of each model are shown in Figure 4. BARL achieved the best result among all NC-index methods. The CCD value obtained by learning the representation through BARL is much larger than that of ICD, and its natural clustering characteristics are far better than other comparative methods.

[0096] The beneficial effects of this invention are as follows:

[0097] (1) By combining adversarial learning mechanism and representation learning, this invention proposes a binomial adversarial representation learning method for mechanical fault diagnosis. It can force the model to learn representation vectors that follow a specified prior distribution from the original signal, thereby solving the uncontrollability problem of traditional deep learning representation learning methods and improving the quality of fault feature extraction.

[0098] (2) The BARL proposed in this invention can embed the health status information contained in the mechanical monitoring signal into the learned binomial representation vector. The binarization characteristic of the binomial distribution makes the fault representation have good natural clustering characteristics of small intra-class distance and large inter-class distance. This improves the robustness and accuracy of this method in fault diagnosis tasks, especially under the condition of insufficient training data.

[0099] The preferred embodiments of the present invention have been described above with reference to the accompanying drawings, but this does not limit the scope of the invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and spirit of the present invention should be within the scope of the present invention.

Claims

1. A binarized fault feature adaptive mining and diagnosis method based on binomial adversarial representation learning, characterized in that, The method includes the following steps: Based on the original vibration signals of the machine under various health conditions, the standardized vibration signals of the machine under various health conditions and the corresponding real health status labels are obtained. Using binomial adversarial samples sampled from a binomial distribution, the standardized vibration signal of the machine, and the corresponding real health status labels, an unsupervised adversarial training is performed on the binomial representation encoder in the autoencoder for binomial adversarial representation learning using a generative-discriminative learning mechanism. This results in a binomial representation encoder that can learn representation vectors that follow a binomial distribution from the original vibration signal of the machine, which serves as a binarized fault feature extractor for mechanical fault diagnosis. By concatenating the binomial characterization encoder with a fully connected classification network, an initial mechanical fault diagnosis classifier is formed; By using the standardized vibration signals of the machine under various health conditions and the corresponding real health condition labels, the initial mechanical fault diagnosis classifier is fine-tuned in a supervised manner to obtain a practical mechanical fault diagnosis classifier. The practical mechanical fault diagnosis classifier described above is used to diagnose faults in the machine to be diagnosed.

2. The method according to claim 1, characterized in that, The process of obtaining standardized vibration signals and corresponding real health status labels for the machine under various health conditions based on the original vibration signals of the machine under various health conditions includes: Collect raw vibration signals of the machine under various health conditions, including normal conditions and various fault conditions; The original vibration signal is divided into multiple non-overlapping signal segments according to a fixed window length; Each signal segment is standardized to obtain a standardized vibration signal. Based on the health status of the original vibration signal corresponding to each standardized vibration signal, a label representing the actual health status of the machine is added to each standardized vibration signal.

3. The method according to claim 1, characterized in that, The method involves using binomial adversarial samples sampled from a binomial distribution, the standardized vibration signal of the machine, and the corresponding real health status labels to perform unsupervised adversarial training on the binomial representation encoder in the autoencoder for binomial adversarial representation learning via a generative-discriminative learning mechanism. This results in a binomial representation encoder capable of learning representation vectors following a binomial distribution from the original vibration signal of the machine. This encoder serves as a binarized fault feature extractor for mechanical fault diagnosis. The standardized vibration signal of the machine is encoded using the binomial characterization encoder in the automatic encoder to obtain a characterization vector, and the characterization vector is decoded using the decoder in the automatic encoder to obtain a reconstructed signal. Using a binomial discriminator to identify binomial distributions, the representation vector and the binomial adversarial sample are judged to determine whether they follow a binomial distribution, and the binomial discrimination result is obtained. The reconstruction loss is determined based on the standardized vibration signal of the machine and the reconstruction signal; Based on the representation vector and the binomial discrimination result, determine the binomial loss; The parameters of the binomial characterization encoder, the decoder, and the binomial discriminator are continuously adjusted until the reconstruction loss and the binomial loss are minimized or stabilized. The binomial characterization encoder at which the reconstruction loss and the binomial loss are minimized or stabilized is used as the binarized fault feature extractor for mechanical fault diagnosis.

4. The method according to claim 3, characterized in that, The step of using a binomial discriminant to identify binomial distributions to determine whether the representation vector and the binomial adversarial sample follow a binomial distribution, and obtaining the binomial discrimination result includes: The representation vector is input into the binomial discriminator so that the binomial discriminator can determine whether the representation vector follows a binomial distribution, thereby obtaining the binomial discrimination result of the representation vector; The binomial adversarial sample is input into the binomial discriminator so that the binomial discriminator can determine whether the binomial adversarial sample follows a binomial distribution, thereby obtaining the binomial discrimination result of the binomial adversarial sample.

5. The method according to claim 4, characterized in that, The step of determining the binomial loss based on the representation vector and the binomial discrimination result includes: Based on the representation vector and the binomial discrimination result of the binomial adversarial example, the binomial loss of the binomial discriminator is determined; The binomial loss of the autoencoder is determined based on the binomial discrimination result of the representation vector.

6. The method according to claim 5, characterized in that, The step of continuously adjusting the parameters of the binomial encoder, the decoder, and the binomial discriminator until the reconstruction loss and the binomial loss are minimized or stabilized includes: The parameters of the binomial discriminator are adjusted based on the binomial loss of the binomial discriminator and the binomial loss of the automatic encoder; The parameters of the autoencoder are adjusted based on the reconstruction loss, the binomial loss of the binomial discriminator, and the binomial loss of the autoencoder. The two steps described above are performed alternately until the reconstruction loss, the binomial loss of the binomial discriminator, and the binomial loss of the autoencoder are minimized or stabilized.

7. The method according to claim 1, characterized in that, The fully connected classification network includes a fully connected layer and a Softmax classification layer.

8. The method according to claim 1, characterized in that, The step of using standardized vibration signals of the machinery under various health conditions and corresponding real health condition labels to perform supervised fine-tuning on the initial mechanical fault diagnosis classifier to obtain a practical mechanical fault diagnosis classifier includes: The standardized vibration signals of the machine under various health conditions are input into the initial mechanical fault diagnosis classifier to obtain the classification results output by the initial mechanical fault diagnosis classifier. The parameters of the initial mechanical fault diagnosis classifier are continuously adjusted to minimize the cross-entropy loss between the output classification result and the true health status label corresponding to the input standardized vibration signal. The mechanical fault diagnosis classifier with the minimum cross-entropy loss is used as the practical mechanical fault diagnosis classifier.

9. The method according to claim 1, characterized in that, The process of using the practical mechanical fault diagnosis classifier to diagnose faults in the machine to be diagnosed includes: The original vibration signal of the machine to be diagnosed is acquired and processed into a standardized vibration signal; The standardized vibration signal is input into the practical mechanical fault diagnosis classifier, which performs feature extraction and fault diagnosis classification on the standardized vibration signal to obtain the health status of the machine to be diagnosed.

Images