Bearing unknown fault detection method based on steady-state and transient-state characteristic similarity mining

Abstract

The application belongs to the technical field of fault detection, and discloses a bearing unknown fault detection method based on steady-state and non-steady-state feature similarity mining, which comprises phase one, a pre-training phase, which is used for training similar labels or dissimilar labels of paired sample data on a labeled set, and completing training of a similarity prediction network; phase two, a new class discovery phase, which is used for identifying and discovering new classes; the application proposes a similarity feature extraction framework based on differential diagnosis, provides a fault diagnosis method based on similarity measurement and differentiation, and adds a time-frequency attention mechanism to integrate information in two dimensions of time and frequency, improve the accuracy of feature representation, and weaken the influence of noise; a salient feature deep fusion module is used to maximize the use of information, deeply mine information correlation between features, and improve the accuracy of detection results.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of fault detection, and particularly relates to a bearing unknown fault detection method based on steady-state and non-steady-state feature similarity mining. BACKGROUND

[0002] Fault diagnosis technology realizes early fault detection and reduces errors in the running process by mastering the running state of a machine, and has important significance for mechanical, nuclear power, chemical and other fields. Unknown fault detection is a process of learning knowledge representation on known class fault data and applying it to new class fault data discovery. At present, many scholars have proposed various new methods for fault diagnosis problems. With the development of deep learning technology, deep learning-based fault diagnosis methods have attracted widespread attention. There are fault diagnosis technologies based on autoencoders, fault diagnosis technologies based on restricted Boltzmann machines, fault diagnosis technologies based on convolutional neural networks, and fault diagnosis technologies based on transfer learning.

[0003] The existing methods have the following problems: first, they cannot provide differentiated diagnosis for different fault types. The main manifestation is that different types of fault signals have different frequency, amplitude and other characteristics (steady-state corresponds to signals with stable frequency and uniform changes; non-steady-state corresponds to signals with abrupt changes or severe jitter). The same feature extraction method is used for each fault type, which ignores the differential component analysis of the signal and cannot capture all the key information, thereby limiting the accuracy of fault diagnosis. Second, they cannot use time-frequency features for collaborative training. The time domain features and frequency domain features of the signal can provide reliable information for fault diagnosis. Time features and frequency features complement each other in fault diagnosis and provide more comprehensive and multi-angle signal information. The existing methods ignore the feature extraction of the time domain and frequency domain dimensions of the signal and do not fully mine all the information in the signal, only focusing on a single perspective from the time domain or frequency domain, which cannot fully coordinate the time domain effective information to cope with the diversity of fault signals. Third, the salient features are not fully perceived, i.e., the salient feature mining when steady-state and non-steady-state are fused is not focused on. Not all extracted fault features contain effective information, and a single fusion strategy cannot mine the deep semantics between data, which also includes some redundant features in the fusion process, thereby limiting the representation ability of the fused features. SUMMARY

[0004] In view of the deficiencies of the prior art, the application provides a bearing unknown fault detection method based on steady-state and non-steady-state feature similarity mining. The existing fault class data and new fault class data are used as inputs, and the knowledge representation is trained on the existing fault data to train new classes on the new fault data.

[0005] To solve the above technical problems, the technical scheme adopted by the application is as follows:

[0006] The bearing unknown fault detection method based on steady-state and non-steady-state feature similarity mining, characterized in that, comprising a pre-training stage and a new class discovery stage two stages:

[0007] Stage one, the pre-training stage, is used for training the similarity label or dissimilarity label of the paired sample data on the labeled set, and the training of the similarity prediction network is completed:

[0008] The similarity prediction network adopts a twin network with shared weights as a feature extractor to learn the steady-state similarity and non-steady-state similarity of the sample. First, the twin network takes the similar sample or dissimilar sample as input, divides each sample data into steady-state signal and non-steady-state signal, and extracts features respectively. In the feature extraction, the feature extractors of the steady-state signal share weights, and the feature extractors of the non-steady-state signal also share weights. In the feature extraction process of the non-steady-state signal, time-frequency attention is introduced to guide the feature extractor to focus on the signal fluctuations in time and frequency. Finally, the total feature distance of the paired sample is obtained by adding the steady-state feature distance and the non-steady-state feature distance of the paired sample, and the relationship label of the paired sample is predicted according to the total feature distance, and then whether the paired sample belongs to the similar class is predicted.

[0009] Stage two, the new class discovery stage, is used for identifying and discovering new classes:

[0010] The input data of the new class discovery stage is paired unlabeled data of different categories from the training data of the pre-training stage. In this stage, the training process is divided into the following steps: first, the paired unlabeled data is input into the similarity prediction network trained in stage one to obtain the pseudo label S UL, if the pair of unlabeled data is predicted to be the same class, the pseudo label is 1, otherwise 0; second, the pair of unlabeled data is sent into the feature extraction module in batches to obtain the steady-state feature and the non-steady-state feature of each sample, and the operation of the feature extraction module in this step is specifically to decouple the sample into a steady-state signal and a non-steady-state signal, the feature extractors of the steady-state signal share weights among each other, the feature extractors of the non-steady-state signal also share weights among each other, and a time-frequency attention is introduced in the feature extraction process of the non-steady-state signal; but the weights of the steady-state signal and the non-steady-state signal feature extractors in this step are independent of the steady-state signal and the non-steady-state signal feature extractors of the similarity prediction network in the first step of stage two, that is, the two are not the same weight; third, the steady-state feature and the non-steady-state feature of each sample are fused to obtain the total feature of the sample; fourth, the total feature obtained from the input pair of unlabeled data is sent into a softmax layer to obtain clustering assignments P and Q; fifth, the KL divergence of the clustering assignments P and Q is calculated as the basis for judging whether the two samples are the same class, if the two samples are the same class, the KL divergence is used as a loss function combined with the predicted relationship label, if the two samples are different classes, the hinge loss is used as a loss function combined with the predicted relationship label to train the feature extraction module.

[0011] Further, stage one is to extract similarity features based on differential diagnosis to obtain the relationship prediction value of the pair of labeled data Specifically, for the input pair of samples, after each signal is decoupled into a steady-state signal and a non-steady-state signal, it is sent to a steady-state branch and a non-steady-state branch respectively to obtain its representation in an embedding space; the relationship label of the pair of samples is predicted according to the distance between the representations to train a similarity prediction network;

[0012] Among them, the steady-state branch is aimed at the stationarity and periodicity of the steady-state signal, and is used to mine the gradual failure and stability failure in the signal; the steady-state branch includes two parts of feature extraction and distance measurement: the pair of steady-state signals obtains a pair of steady-state features after feature extraction, the feature extraction networks of the two steady-state signals adopt the same network structure and share parameters to obtain the similarity of the samples in the steady-state component; finally, the distance between the pair of steady-state features is calculated;

[0013] The non-steady branch is aimed at the instantaneous characteristics of the non-stationary signal, and is used to mine the instantaneous faults in the signal. The non-steady branch includes feature extraction, discriminative feature time-frequency focusing module and distance measurement. First, the one-dimensional non-steady signal is converted into a two-dimensional time-frequency graph, and then is sent to the discriminative feature time-frequency focusing module to obtain the attention score of the two-dimensional time-frequency graph from the time dimension and the frequency dimension, so as to obtain the self-attention matrix, and the original two-dimensional time-frequency graph is adjusted to obtain the adjusted non-steady signal time-frequency graph. Then, the adjusted non-steady signal time-frequency graph is sent to the non-steady feature extraction network to obtain the non-steady feature. The feature extraction networks of the two non-steady signals also adopt the same network structure and share parameters, so as to align the non-steady signals to the same space for comparison and similarity mining, and capture the similarity between the non-steady signals. After obtaining the representation of the embedding space, the distance between the pair of non-steady features is calculated. The feature extraction networks of the two non-steady signals share weights, but the parameters in the discriminative feature time-frequency focusing module do not share weights.

[0014] Finally, the pair of steady-state feature distances and the pair of non-steady feature distances are added to obtain the total feature distance of the input signal, and then the total feature distance is sent to the fully connected layer to obtain the final output, that is, the relationship prediction value of the pair of labeled data

[0015] Further, the discriminative feature time-frequency focusing module takes a two-dimensional time-frequency graph X freq×time as input, and obtains an enhanced time-frequency graph under the adjustment of the time-frequency attention mechanism Wherein, the time-frequency attention is realized by parallel time attention and frequency attention mechanism, the time attention mechanism obtains the attention score T A ∈R 1×time of the time dimension, and the frequency attention obtains the attention score F A ∈R freq×1 of the frequency dimension; the one-dimensional attention scores obtained by calculating the two dimensions are multiplied by a matrix to obtain the total time-frequency attention matrix TF A ∈R freq×time ; the time-frequency attention matrix and the original two-dimensional time-frequency graph are point multiplied to obtain the attention data as output, that is, the adjusted non-steady signal time-frequency graph, wherein freq represents frequency, and time represents time.

[0016] Further, the adjustment of the time-frequency attention is divided into two parts: information aggregation and attention matrix generation. The information aggregation part is realized by maximum pooling and average pooling on the time dimension and the frequency dimension. On the frequency dimension, since the maximum pooling can obtain the significant time signal T Max under the frequency, the surge of the fault signal at a certain moment under the frequency can be captured, so the maximum pooling is set to focus on the time sequence mutation type fault, as shown in the following formula:

[0017] T Max (t) = max X(t) (5)

[0018] where X(t) represents the frequency data at the t-th time, and T Max ∈ R 1×time is obtained by taking the maximum value of the frequency data at each time. Max (t) represents the value of T Max at the t-th time; similarly, in the time dimension, since the max pooling can obtain the significant frequency signal at the time, it can capture the sudden increase of the fault signal at a certain frequency at the time, so the max pooling is set to focus on the frequency domain mutation type fault, as shown in the following formula:

[0019] F Max (f) = max X(f) (6)

[0020] where X(f) represents the time data at the f-th frequency component, and F Max ∈ R freq×1 is obtained by taking the maximum value of the time data at each frequency component. Max (f) represents the value of F Max at the f-th frequency component.

[0021] In the frequency dimension, the average pooling is set to obtain the fluctuation of the average value at the frequency, which supplements the capture of the time sequence progressive type fault, as shown in the following formula:

[0022]

[0023] where X represents the input two-dimensional time-frequency graph, and T avg (t) represents the t-th element in the obtained average pooling vector; similarly, in the time dimension, the average pooling is set to obtain the fluctuation of the average frequency at the time, which supplements the capture of the frequency domain progressive type fault, as shown in the following formula:

[0024]

[0025] where F avg (f) represents the f-th element in the obtained average pooling vector.

[0026] Subsequently, the vectors generated by the two-dimensional max pooling and average pooling are mapped to generate attention scores T A and F A by convolution operation; and the obtained one-dimensional attention scores are matrix multiplied according to the following formula to obtain a two-dimensional attention matrix TF A :

[0027]

[0028] wherein, denotes a matrix multiplication operation, and an attention matrix TF is obtained A After point multiplication operation with the original two-dimensional time-frequency diagram, the output of the discriminative feature time-frequency focusing module is obtained

[0029] Further, in the second phase of new class discovery, the steady-state feature and the non-steady-state feature are fused through the saliency feature deep fusion module, wherein, s denotes the steady-state feature input into the saliency feature deep fusion module, n denotes the non-steady-state feature input into the saliency feature deep fusion module, and f denotes the total feature after saliency deep fusion.

[0030] The saliency feature deep fusion module adds a saliency feature extractor and a double-layer fusion mechanism on the basis of the traditional gating mechanism. The saliency feature extractor is realized by double-layer convolution operation, removes redundant interference, and acquires key features. The steady-state feature f s and the non-steady-state feature f n are obtained through the saliency feature extractor. s n ;

[0031] The double-layer fusion mechanism sets up a nested fusion structure, and on the basis of the shallow information obtained by the initial fusion, the features after the cross are fused by gating, so that the fusion of the steady-state feature and the non-steady-state feature is promoted to the semantic level. In the first layer fusion, the steady-state and non-steady-state features are spliced first, and the weight vector W1 with the same dimension as f s and f n is output through convolution and activation operation, which is expressed by the formula as follows:

[0032] W1 = σ(Conv(Concat[f′ s ,f′ n ])) (12)

[0033] The weight vector W1 and f s are multiplied to obtain the steady-state feature f″ s after adaptive adjustment.

[0034] f″ s = W1 ⊙ f′ s (13)

[0035] Similarly, the non-steady-state feature f″ n after the initial fusion is obtained according to the formula as follows,

[0036] f″ n = (1-W1) ⊙ f′​n (14)

[0037] According to the adjusted steady-state feature f" s and the non-steady-state feature f" n , obtain the secondary fused weight W2,

[0038] W2 = sigma (Conv (Concat [f" s , f" n ])) (15)

[0039] Finally, the fused feature f is obtained,

[0040] f = W2 o f' s + (1-W2) o f' n (16)

[0041] Wherein, sigma represents a Sigmoid activation function, Conv represents a convolution operation, and Concat represents feature splicing.

[0042] Compared with the prior art, the present application has the following advantages:

[0043] (1) The present application proposes a new similarity feature extraction framework based on differential diagnosis, and provides a fault diagnosis strategy based on similarity measurement and differentiation. On the basis of using a twin network to mine the similarity of input signals, the signals are decoupled into steady-state components and non-steady-state components, further improving the feature discrimination. And when the similarity result is generated, the similarity of the steady-state signal and the non-steady-state signal is calculated respectively, further strengthening the extraction of the similarity information of the input signal.

[0044] (2) The present application proposes a new discriminative feature time-frequency focusing module. The time-frequency attention mechanism is added to integrate the information in the time and frequency dimensions, so as to capture time-frequency mutations, which can further improve the richness and accuracy of feature representation. In addition, the time-frequency attention can effectively distinguish background noise and fault signals, weaken the influence of noise, and guide the network to more effectively extract high-quality features, thereby promoting the integration of the new class discovery stage in clustering.

[0045] (3) The present application proposes a new saliency feature deep fusion module, which maximizes the use of information and deeply mines the information correlation between features. On the basis of the original gated fusion, saliency dimensionality reduction is used on the input features to capture key information in the data and filter out features that are irrelevant or redundant to the task. Secondly, the nested fusion design can further extract and integrate the relevant information between the steady-state and non-steady-state features, helping to correct the possible bias of single fusion. By multiple fusion, more data details and context information are captured, improving the accuracy and robustness of the overall fusion result. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without any creative effort based on these drawings.

[0047] Figure 1 Flow chart for the pre-training stage of the present application;

[0048] Figure 2 Flow chart for the new class discovery stage of the present application;

[0049] Figure 3 Framework diagram for the similarity feature extraction based on differential diagnosis of the present application;

[0050] Figure 4 Schematic diagram of the discriminative feature time-frequency focusing module of the present application;

[0051] Figure 5 Schematic diagram of the discriminative feature time-frequency focusing in the frequency dimension of the embodiment of the present application;

[0052] Figure 6 Schematic diagram of the discriminative feature time-frequency focusing in the time dimension of the embodiment of the present application;

[0053] Figure 7 Schematic diagram of the salient feature deep fusion module of the present application. DETAILED DESCRIPTION

[0054] The present application will be further described below in combination with the drawings and specific embodiments.

[0055] In combination with Figure 1 and Figure 2 shown, the present embodiment provides a bearing unknown fault detection method based on steady-state and non-steady-state feature similarity mining, which includes two stages of pre-training stage and new class discovery stage. The design idea is to use a twin network to learn the steady-state similarity and non-steady-state similarity of samples on the basis of signal decoupling, and to increase a discriminative feature time-frequency focusing module to guide the feature extractor to focus on the time-frequency fluctuations in the non-steady state. In the figure, S L represents the actual relationship label of the pair of labeled data, represents the relationship prediction value of the pair of labeled data, S UL represents the similarity relationship label value (i.e. pseudo label) obtained by the similarity prediction network for the pair of unlabeled data, and P and Q represent the clustering assignments obtained for the pair of unlabeled data.

[0056] The present application will be further described below in combination with the drawings and specific embodiments. Figure 1 and Figure 2The method is introduced in detail.

[0057] Phase one, pre-training phase, is used for training similar labels or dissimilar labels of paired sample data on a label set, and completing the training of the similarity prediction network.

[0058] The similarity prediction network uses a twin network sharing weights as a feature extractor to learn the steady-state similarity and non-steady-state similarity of samples. Firstly, the twin network takes similar samples or dissimilar samples as input, and considering the different characteristics of steady-state signals and non-steady-state signals reflecting the running state of the equipment, each sample data is divided into steady-state signals and non-steady-state signals for feature extraction. During feature extraction, the feature extractors of the steady-state signals share weights, and the feature extractors of the non-steady-state signals also share weights. Notably, the time-frequency attention is introduced in the feature extraction process of the non-steady-state signals to guide the feature extractor to focus on the signal fluctuations in time and frequency in the non-steady-state. Finally, the total feature distance of the paired samples is obtained by adding the steady-state feature distance and the non-steady-state feature distance of the paired samples, and the relationship label of the paired samples is predicted according to the total feature distance, and then whether the paired samples belong to similar categories is predicted.

[0059] More specifically, different types of faults show different signal fluctuations. The steady-state signal focuses more on expressing the basic characteristics and long-term behavior of the system, and the observation of the steady-state signal can detect persistent or progressive faults. The non-steady-state signal focuses more on expressing the dynamic changes and mutations of the system, and the observation of the non-steady-state signal can detect instantaneous or sudden faults. Therefore, in order to accurately diagnose different types of faults, the similarity feature extraction based on differential diagnosis is designed in phase one to obtain the relationship prediction value of the paired labeled data That is, used to obtain the relationship label of the paired samples. It includes a steady-state processing branch and a non-steady-state processing branch. The steady-state branch is used to mine progressive faults and stability faults in the signal according to the stationarity and periodicity of the steady-state signal. The non-steady-state branch is used to mine instantaneous faults in the signal according to the instantaneous characteristics of the non-stationary signal.

[0060] As shown in Figure 3 , specifically, for the input paired samples, after each signal is decoupled into steady-state signals and non-steady-state signals, they are sent to the steady-state branch and the non-steady-state branch respectively to obtain their representations in the embedding space; the relationship label of the paired samples is predicted according to the distance between the representations to train the similarity prediction network. Figure 3 In the formula, X L represents the labeled data of known categories, Y L represents the label of the labeled data, (x i ,x j ) represents the feature of the labeled data X LThe two samples taken out are combined into paired data, S L actual relationship label representing the labeled data, relationship prediction value representing the labeled data.

[0061] The design of the two branches will be described below respectively.

[0062] The steady-state branch includes two parts: feature extraction and distance measurement. The paired steady-state signals pass through feature extraction to obtain paired steady-state features. The feature extraction networks of the two steady-state signals adopt the same network structure and share parameters to obtain the similarity of the samples in the steady-state component. Finally, the distance between the paired steady-state features is calculated.

[0063] As a preferred embodiment, first, each of the input paired data (x i ,x j ) of known categories is decoupled into a steady-state component and a non-steady-state component to obtain and The steady-state signals and are sent to the steady-state branch. In feature extraction, first, passes through a 5-layer convolutional feature extraction network WDCNN to obtain steady-state features Similarly, also passes through a 5-layer convolutional feature extraction network WDCNN to obtain steady-state features It is worth noting that in the present application, the two feature extraction networks are set to the same network structure and parameters, i.e., the twin network for the steady-state signal is set to obtain the similarity of the samples in the steady-state component. Then, the distance between the paired steady-state features is calculated according to the following formula:

[0064]

[0065] The non-steady-state branch includes three parts: feature extraction, discriminative feature time-frequency focusing module, and distance measurement. First, the one-dimensional non-steady-state signal is converted into a two-dimensional time-frequency graph, which is then sent to the discriminative feature time-frequency focusing module to obtain the attention score of the two-dimensional time-frequency graph from the time dimension and the frequency dimension, thereby obtaining the self-attention matrix to adjust the original two-dimensional time-frequency graph and obtain the adjusted non-steady-state signal time-frequency graph. Then, the adjusted non-steady-state signal time-frequency graph is sent to the non-steady-state feature extraction network to obtain the non-steady-state features. Here, the feature extraction networks of the two non-steady-state signals also adopt the same network structure and share parameters to align the non-steady-state signals to the same space for comparison and similarity mining, so that they can better capture the similarity relationship between the non-steady-state signals. After obtaining the representation of the embedding space, the distance between the paired non-steady-state features is calculated.

[0066] Because the non-steady-state signal has the characteristic of changing with time, it can reflect the mutation or abnormal event in the system. When the discriminative feature time-frequency focusing module is added in the non-steady-state branch, the difference of the non-steady-state signal is further highlighted from the two dimensions of time and frequency, and the fault contained in the non-steady-state signal is further dug. The details of this module are described in the next part.

[0067] As a preferred embodiment, for the non-steady-state signal and it is sent into the non-steady-state branch. First, the one-dimensional non-steady-state signal is converted into a two-dimensional time-frequency graph through short-time Fourier transform, and then it is sent into the discriminative feature time-frequency focusing module to obtain the adjusted non-steady-state signal time-frequency information. Then the adjusted non-steady-state signal time-frequency graph is sent into the non-steady-state feature extraction network to obtain the non-steady-state feature Similarly, the non-steady-state feature of another non-steady-state signal is obtained. After obtaining the representation of the embedding space, the distance between the pair of non-steady-state features is calculated according to the following formula:

[0068]

[0069] It should be noted here that the feature extraction networks of the two non-steady-state signals share weights, but the parameters in the discriminative feature time-frequency focusing module do not share weights.

[0070] Finally, the pair of steady-state feature distances and the pair of non-steady-state feature distances are added to obtain the total feature distance of the input signal, and then in order to obtain the possibility that the pair of input signals belong to similar samples according to the feature distance, it is sent into the fully connected layer to obtain the final output, that is, the relationship prediction value of the pair of label data as shown in the following formula.

[0071]

[0072] Where FC represents the fully connected layer, and sig represents the sigmoid activation function.

[0073] In the training process, since the categories of the label data are known, the actual relationship label S L of the pair of samples can be obtained according to the corresponding category of each sample. If x i and x j belong to the same category, the label of this pair of samples is 1, otherwise it is 0. Therefore, according to the following loss function, the feature extractor can be trained to optimize the feature extraction process, and the loss L formula is as follows:

[0074]

[0075] The discriminative feature time-frequency focusing module is introduced below.

[0076] The time domain feature can help locate the node and duration of signal occurrence, and the frequency domain feature can analyze the existence and intensity of specific frequency domain in the signal. Therefore, the time feature and the frequency feature can complement each other and discover faults from different angles. Therefore, in order to further highlight the difference of non-stationary signals in time-frequency sensitivity, a discriminative feature time-frequency focusing module is designed to help the network focus on the discriminative part of the signal and improve the network's perception ability of key time points in the signal from two dimensions.

[0077] As shown in Figure 4 , the discriminative feature time-frequency focusing module obtains an enhanced time-frequency graph X' under the adjustment of the time-frequency attention mechanism, where X' ∈ R freq×time is taken as input (it should be noted that in the two-dimensional time-frequency graph, the horizontal axis represents time and the vertical axis represents frequency. Taking a row of data F s ∈ R 1×time , it can be understood as the information of all time points at this frequency, that is, the frequency dimension information. Similarly, taking a column of data T s ∈ R freq×1 , it can be understood as the information of all frequencies at this time, that is, the time dimension information). The time-frequency attention is realized by parallel time attention and frequency attention mechanisms. The time attention mechanism obtains the time dimension attention score T A ∈ R 1×time , and the frequency attention obtains the frequency dimension attention score F A ∈ R freq×1 ; the one-dimensional attention scores calculated in the two dimensions are multiplied by a matrix to obtain the total time-frequency attention matrix TF A ∈ R freq×time ; the point multiplication operation is performed on the time-frequency attention matrix and the original two-dimensional time-frequency graph to obtain the attention data as output, that is, the adjusted non-stationary signal time-frequency graph, where freq represents frequency and time represents time.

[0078] The specific details are described below in conjunction with Figure 5 and Figure 6 .

[0079] The adjustment of the time-frequency attention includes two parts: information aggregation and attention matrix generation. The information aggregation part is realized by maximum pooling and average pooling in the time dimension and the frequency dimension. In the frequency dimension, the maximum pooling can obtain the significant time signal T Max at this frequency, which can capture the surge of the fault signal at a time point at this frequency. Therefore, the maximum pooling is set to focus on the time sequence mutation type fault, as shown in the following formula:

[0080] T Max(t) = max X(t) (5)

[0081] where X(t) represents the frequency data at the t-th time, and T Max ∈ R 1×time is obtained by obtaining the maximum value of the frequency data at each time. Max (t) represents the value of T Max at the t-th time. Similarly, in the time dimension, since the max-pooling can obtain the significant frequency signal at the time, it can capture the sudden increase of the fault signal at a certain frequency at the time, so the max-pooling is set to focus on the frequency domain mutation type fault, as shown in the following formula:

[0082] F Max (f) = max X(f) (6)

[0083] where X(f) represents the time data at the f-th frequency component, and F Max ∈ R freq×1 is obtained by obtaining the maximum value of the time data at each frequency component. Max (f) represents the value of F Max at the f-th frequency component.

[0084] In the frequency dimension, the average pooling is set to obtain the fluctuation of the average value at the frequency, which supplements the capture of the time sequence progressive type fault, as shown in the following formula:

[0085]

[0086] where X represents the input two-dimensional time-frequency graph, and T avg (t) represents the t-th element in the obtained average pooling vector; similarly, in the time dimension, the average pooling is set to obtain the fluctuation of the average frequency at the time, which supplements the capture of the frequency domain progressive type fault, as shown in the following formula:

[0087]

[0088] where F avg (f) represents the f-th element in the obtained average pooling vector.

[0089] Subsequently, the vectors generated by the two-dimensional max-pooling and average pooling are mapped to generate attention scores T A and F A by convolution operation; and the obtained one-dimensional attention scores are matrix multiplied according to the following formula to obtain a two-dimensional attention matrix TF A :

[0090]

[0091] where, denotes a matrix multiplication operation to obtain an attention matrix TF A After point multiplication operation with the original two-dimensional time-frequency graph, the output of the discriminative feature time-frequency focusing module is obtained

[0092] Phase two, new class discovery phase, is used to identify and discover new classes.

[0093] The input data of the new class discovery phase is a pair of unlabeled data of different classes from the training data of the pre-training phase. In this phase, the training process is divided into the following steps: first, input the pair of unlabeled data (x p ,x q ) to the similarity prediction network trained in phase one to predict the pseudo label S UL . If the pair of unlabeled data is predicted to be the same class, the pseudo label is 1, otherwise it is 0; second, input the pair of unlabeled data (x p ,x q ) into the feature extraction module in batches to obtain the steady-state feature and the non-steady-state feature of each sample. The operation of the feature extraction module here is to decouple the sample into a steady-state signal and a non-steady-state signal. The feature extractors of the steady-state signal share weights among them, and the feature extractors of the non-steady-state signal also share weights among them, and time-frequency attention is introduced in the feature extraction process of the non-steady-state signal. However, the weights of the steady-state signal and non-steady-state signal feature extractors here are independent of the weights of the steady-state signal and non-steady-state signal feature extractors of the similarity prediction network in the first step of phase two, that is, the two are not the same weights. Third, the steady-state feature and the non-steady-state feature of each sample are fused to obtain the total feature of the sample. Fourth, the total feature (f p ,f q ) obtained by the input pair of unlabeled data (x p ,x q ) is sent to the softmax layer to obtain the clustering assignment P and Q. Fifth, the KL divergence of the clustering assignment P and Q is calculated as the basis for judging whether the two samples are the same class. If the two samples are the same class, use the predicted relationship label and the KL divergence as the loss function, if the two samples are different classes, use the predicted relationship label and the hinge loss as the loss function to train the feature extraction module.

[0094] In the new class discovery phase, the feature vector of each input signal needs to be calculated to obtain the corresponding cluster distribution to calculate the loss, train the network, so the steady-state and non-steady-state features need to be fused to obtain the total features of the signal. However, considering that there may be noise or other interference in the extracted signal, there may be useless information in the features, thereby weakening the distance measurement between different distributions. Moreover, single feature fusion may only stay in the weight adjustment of the data, and cannot obtain global perception of the information. Therefore, in the present application, a saliency feature deep fusion module is proposed to replace the direct element addition fusion method.

[0095] In combination Figure 7 As shown in FIG. 2, in the second phase of the new class discovery phase, the steady-state feature and the non-steady-state feature are fused by the saliency feature deep fusion module, wherein f s represents the steady-state feature input into the saliency feature deep fusion module, f n represents the non-steady-state feature input into the saliency feature deep fusion module, and f represents the total feature after saliency deep fusion.

[0096] The saliency feature deep fusion module adds a saliency feature extractor and a double-layer fusion mechanism on the basis of the traditional gating mechanism. The saliency feature extractor is realized by double-layer convolution operation, removes redundant interference, and obtains key features with rich information. The steady-state feature f s and the non-steady-state feature f n are respectively subjected to the saliency feature extractor to obtain the steady-state saliency feature f' s and the non-steady-state saliency feature f' n . Taking the steady-state feature f s ∈R l×1 as an example, the process is represented as:

[0097]

[0098] wherein f' s ∈R l×1 , represents the first convolution operation of the steady-state feature, and δ and σ respectively represent the Relu activation function and the Sigmoid activation function. Similarly, the non-steady-state saliency feature f' n ∈R l×1 is obtained by the following formula:

[0099]

[0100] wherein represents the first convolution operation of the non-steady-state feature.

[0101] The double-layer fusion mechanism sets a nested fusion structure, and on the basis of obtaining shallow information in the primary fusion, the cross features are gate fused to promote the fusion of steady-state features and non-steady-state features to the semantic level and obtain more rich features. In the first layer fusion, the steady-state and non-steady-state features are spliced first, and the weight vector W1 with the same dimension as f' s and f' n is output through convolution and activation operation, which is expressed by formula as follows:

[0102] W1=σ(Conv(Concat[f′ s ,f′ n ])) (12)

[0103] The weight vector W1 and f' s are multiplied to obtain the steady-state feature f″ s after adaptive adjustment;

[0104] f″ s =W1⊙f′ s (13)

[0105] Similarly, the non-steady-state feature f″ n after the primary fusion is obtained according to the following formula,

[0106] f″ n =(1-W1)⊙f′ n (14)

[0107] According to the adjusted steady-state feature f″ s and the non-steady-state feature f″ n , the weight W2 of the secondary fusion is obtained,

[0108] W2=σ(Conv(Concat[f″ s ,f″ n ])) (15)

[0109] Finally, the fused feature f is obtained,

[0110] f=W2⊙f′ s +(1-W2)⊙f′ n (16)

[0111] Wherein, σ represents the Sigmoid activation function, Conv represents the convolution operation, and Concat represents the feature splicing.

[0112] In summary, the present application provides a more refined similarity measure and a differentiated fault diagnosis strategy. By mining the similarities between steady-state signals and non-steady-state signals, compared to the traditional method of directly comparing sample similarities, a more accurate and detailed measurement method can be provided. This detailed measurement method can more accurately identify and compare the key features of each component in the signal, and better improve the discrimination of the features. And since both steady-state signals and non-steady-state signals may contain fault-related information, similarity measurement for different types of signals can obtain a more comprehensive set of difference features, optimize the feature selection process, and thus improve the generalization ability and diagnostic effect of the model.

[0113] The present application captures discriminative changes and quickly focuses on the training of main features. The time-frequency attention mechanism designed in the present application guides the feature extractor to focus on signals at different times and frequencies, effectively avoiding the network's attention to noise or other mixed parts. Time-frequency attention can focus on both time and frequency domain features of the signal. First, in fault diagnosis, many fault signals often exist in the form of time-frequency domain, such as resonance frequency or energy change in a specific frequency band in vibration signals. Using time-frequency attention can effectively capture these key features. Second, in fault diagnosis, ordinary attention mechanisms may not be able to effectively distinguish fault signals from background noise or interference signals, resulting in key fault signals that may be submerged in strong noise or other interference. Time-frequency attention can be weighted for specific time-frequency regions, making it easier to detect and identify fault signals in complex environments and improving the ability to perceive weak signals.

[0114] The present application maximizes the use of information and deeply mines the information correlation between features. The salient feature deep fusion module can effectively filter out features that are not related or redundant to the task. This helps to reduce noise and interference in the data and extract more representative and useful features, thereby improving the robustness and generalization ability of the model. Second, multiple fusions can further extract and integrate relevant information between steady-state and non-steady-state features, enhancing the expression ability of the features. Each fusion can capture different aspects of the data through different ways and weight adjustments, thereby providing richer and more representative feature representations.

[0115] Of course, the above description is not a limitation on the present application, and the present application is not limited to the above examples. Those skilled in the art can make changes, modifications, additions or substitutions within the scope of the present application, and such changes, modifications, additions or substitutions should be within the scope of the present application.

Claims

1. A bearing unknown fault detection method based on steady-state and non-steady-state feature similarity mining, characterized in that, The method comprises a pre-training stage and a new class discovery stage: In the first stage, the pre-training stage, the similarity labels or dissimilarity labels of paired sample data are trained on a labeled set, and the training of the similarity prediction network is completed: The similarity prediction network adopts a twin network with shared weights as a feature extractor to learn the steady-state similarity and non-steady-state similarity of samples. First, the twin network takes similar samples or dissimilar samples as input, and divides each sample data into steady-state signals and non-steady-state signals for feature extraction. During feature extraction, the feature extractors of the steady-state signals share weights, and the feature extractors of the non-steady-state signals also share weights. Time-frequency attention is introduced in the feature extraction process of the non-steady-state signals to guide the feature extractor to focus on the signal fluctuations in time and frequency. Finally, the total feature distance of the paired samples is obtained by adding the steady-state feature distance and the non-steady-state feature distance of the paired samples, and the relationship label of the paired samples is predicted according to the total feature distance, and then whether the paired samples belong to the similar category is predicted. In the second stage, the new class discovery stage, new classes are identified and discovered: The input data of the new class discovery stage is paired unlabeled data of different categories from the training data of the pre-training stage. In this stage, the training process is divided into the following steps: first, the paired unlabeled data is input into the similarity prediction network trained in the first stage to obtain pseudo labels S UL . If the paired unlabeled data is predicted to be of the same category, the pseudo label is 1, otherwise it is 0; second, the paired unlabeled data is sent into the feature extraction module in batches to obtain the steady-state feature and the non-steady-state feature of each sample. The operation of the feature extraction module in this step is specifically to decouple the sample into a steady-state signal and a non-steady-state signal. The feature extractors of the steady-state signal share weights among each other, and the feature extractors of the non-steady-state signal also share weights among each other. In addition, time-frequency attention is introduced in the feature extraction process of the non-steady-state signal. However, the weights of the steady-state signal and the non-steady-state signal feature extractors in this step are independent of the weights of the steady-state signal and the non-steady-state signal feature extractors of the similarity prediction network in the first step of the second stage, that is, they are not the same weights. Third, the steady-state features and non-steady-state features of each sample are fused to obtain the total features of the sample. Fourth, the total features obtained from the input paired unlabeled data are sent to the softmax layer to obtain the clustering assignments P and Q. Fifth, the KL divergence of the clustering assignments P and Q is calculated as the basis for judging whether the samples belong to the same class. If the two samples belong to the same class, the KL divergence is used as the loss function combined with the predicted relationship label. If the two samples belong to different classes, the hinge loss is used as the loss function combined with the predicted relationship label to train the feature extraction module.

2. The bearing unknown fault detection method based on steady-state non-steady-state feature similarity mining according to claim 1, characterized in that, The first stage is to extract similarity features based on differential diagnosis to obtain relationship prediction values of the pair of labeled data Specifically, for the input pair of samples, each signal is decoupled into a steady-state signal and a non-steady-state signal, and then sent to a steady-state branch and a non-steady-state branch respectively to obtain its representation in an embedding space; the similarity prediction network is trained according to the distance between the representations to predict the relationship label of the pair of samples; The steady-state branch is used to mine the gradual faults and stability faults in the signals according to the stationarity and periodicity of the steady-state signals. The steady-state branch includes feature extraction and distance measurement: paired steady-state signals are extracted to obtain paired steady-state features, and the feature extraction networks of the two steady-state signals use the same network structure and share parameters to obtain the similarity of the samples in the steady-state component. Finally, the distance between the paired steady-state features is calculated. The non-stationary branch is used for mining the instantaneous faults in the signal according to the instantaneous characteristics of the non-stationary signal, and includes three parts of feature extraction, discriminative feature time-frequency focusing module and distance measurement. First, the one-dimensional non-stationary signal is converted into a two-dimensional time-frequency graph, and then is sent into the discriminative feature time-frequency focusing module to obtain the attention score of the two-dimensional time-frequency graph from the time dimension and the frequency dimension, thereby obtaining a self-attention matrix, and the original two-dimensional time-frequency graph is adjusted to obtain an adjusted non-stationary signal time-frequency graph. Then, the adjusted non-stationary signal time-frequency graph is sent into a non-stationary feature extraction network to obtain non-stationary features. The feature extraction networks of two non-stationary signals also adopt the same network structure and share parameters, so that the non-stationary signals are aligned to the same space for comparison and similarity mining, and the similarity between the non-stationary signals is captured. After obtaining the representation of the embedding space, the distance between the pair of non-stationary features is calculated. The feature extraction networks of the two non-stationary signals share weights, but the parameters in the discriminative feature time-frequency focusing module do not share weights. Finally, the pair-wise steady-state feature distance and the pair-wise non-steady-state feature distance are added to obtain the total feature distance of the input signal, which is then sent to the fully connected layer to obtain the final output, i.e. the relationship prediction value of the pair label data 3. The bearing unknown fault detection method based on steady-state non-steady-state feature similarity mining according to claim 2, characterized in that, The discriminative feature time-frequency focusing module obtains a two-dimensional time-frequency graph X ∈ R freq×time As input, an enhanced time-frequency graph is obtained under the regulation of the time-frequency attention mechanism Wherein, the time-frequency attention is realized by parallel time attention and frequency attention mechanisms, the time attention mechanism obtains time dimension attention score T A ∈R 1×time , and the frequency attention obtains frequency dimension attention score F A ∈R freq×1 ; the one-dimensional attention scores obtained by calculating the two dimensions are multiplied by a matrix to obtain a total time-frequency attention matrix TF A ∈R freq×time ; the time-frequency attention matrix and the original two-dimensional time-frequency graph are point multiplied to obtain attention data as output, that is, the regulated non-stationary signal time-frequency graph, wherein freq represents frequency, and time represents time.

4. The bearing unknown fault detection method based on steady-state non-steady-state feature similarity mining according to claim 3, characterized in that, The adjustment of time-frequency attention is divided into two parts: information aggregation and attention matrix generation. The information aggregation part is realized by maximum pooling and average pooling in the time dimension and the frequency dimension. In the frequency dimension, the maximum pooling can obtain the significant time signal T Max at this frequency, so the maximum pooling is set to focus on the time sequence mutation fault, as shown in the following formula: T Max (t) = max X(t) (5) Wherein, X(t) represents the frequency data at the t time, and the maximum value of the frequency data at each time is obtained to obtain T Max ∈R 1 ×time , T Max (t) represents the value of T Max at the t time; similarly, in the time dimension, since the maximum pooling can obtain the significant frequency signal at the time, the sudden increase of the fault signal at the time can be captured, so that the maximum pooling is set to focus on the frequency domain mutation type fault, as shown in the following formula: F Max (f) = max X(f) (6) where X(f) represents the time data at the fth frequency component, F is obtained by taking the maximum of the time data at each frequency component Max ∈ R freq×1 , F Max (f) represents the value of F at the fth frequency component Max ; On the frequency dimension, average pooling is set to obtain the fluctuation of the average value at the frequency, and to supplement the capture of the time sequence progressive fault, as represented by the following formula: wherein X represents the input two-dimensional time-frequency graph, T avg (t) represents the tth element in the obtained average pooling vector; similarly, in the time dimension, the average pooling is set to obtain the average frequency fluctuation at this time, which supplements the capture of the frequency domain progressive fault, as shown in the following formula: where F avg (f) denotes the f-th element in the obtained average-pooled vector. Subsequently, the vectors generated after two-dimensional maximum pooling and average pooling are mapped by convolution operation to generate attention score T A and F A ; and the obtained one-dimensional attention score is matrix multiplied according to the following formula to obtain a two-dimensional attention matrix TF A : wherein, denotes a matrix multiplication operation, resulting in an attention matrix TF A The output of the discriminative feature time-frequency focusing module is obtained after point multiplication operation between the original two-dimensional time-frequency graph and the output of the attention module 5. The method for bearing unknown fault detection based on steady-state and non-steady-state feature similarity mining according to claim 1, characterized in that, In the stage two new class discovery stage, the steady-state feature and the non-steady-state feature are fused through a saliency feature deep fusion module, wherein, f s represents the steady-state feature input into the saliency feature deep fusion module, f n represents the non-steady-state feature input into the saliency feature deep fusion module, and f represents the total feature after saliency deep fusion; The saliency feature deep fusion module adds a saliency feature extractor and a double-layer fusion mechanism on the basis of a traditional gating mechanism, the saliency feature extractor is realized by double-layer convolution operation, removes redundant interference, and obtains key features; the steady-state feature f s and the non-steady-state feature f n After the saliency feature extractor, the steady-state saliency feature f' s and the non-steady-state saliency feature f' n are obtained. The double-layer fusion mechanism sets a nested fusion structure, on the basis of obtaining shallow information in the primary fusion, performs gated fusion on the cross features, and promotes the fusion of steady-state features and non-steady-state features to the semantic level; in the first layer fusion, the steady-state and non-steady-state features are spliced first, and the weight vector W1 with the same dimension as f s and f' n is output through convolution and activation operation, which is expressed by the formula as follows: W1 = σ(Conv(Concat[f s ,f n ])) (12) The weight vector W1 and f s Point multiplication obtains the steady-state feature f s ; By the same token, the non-steady-state characteristic f" after the initial fusion n is obtained according to the following equation, According to the adjusted steady-state characteristic f s and the non-steady-state characteristic f n , a weight W2 of the secondary fusion is obtained, W2 = σ(Conv(Concat[f s ,f n ])) (15) Finally, the fused features f are obtained, wherein σ represents a Sigmoid activation function, Conv represents a convolution operation, and Concat represents feature splicing.

Images