A Bearing Abnormality Detection Method and System Based on Temporal Association Difference

Abstract

The present application discloses a bearing anomaly detection method and system based on temporal correlation difference, including: constructing a convolutional neural network composed of multiple CNN modules, and using each CNN module to extract multi-dimensional features of the bearing in the time domain and frequency domain from different scales; extracting adjacent temporal features and non-adjacent temporal features within the bearing sliding window through an adjacent attention module and a non-adjacent attention module respectively; calculating the correlation difference between the adjacent temporal features and the non-adjacent temporal features by using a bearing temporal correlation difference mechanism based on symmetric KL divergence; inputting the correlation difference into the convolutional neural network for training to increase the correlation difference between the normal state and the abnormal state of the bearing and enhance the discrimination of feature representation; using the trained convolutional neural network to map the multi-dimensional features of the bearing into a probability distribution to realize the detection of the abnormal state of the bearing. The present application improves the accuracy of the bearing anomaly detection task through the above method.

Description

Technical Field

[0001] The present invention relates to the technical field of fault detection in the intelligent manufacturing industry, and specifically relates to a bearing anomaly detection method and system based on temporal correlation difference. Background Art

[0002] As a core component of rotating machinery, the operating state of a bearing directly affects the reliability and safety of equipment. Traditional bearing anomaly detection methods include vibration analysis-based methods and sound analysis-based methods. Among them, the vibration analysis-based method diagnoses the bearing state by monitoring the vibration signal of the bearing during operation, uses an acceleration sensor to capture the vibration data of the bearing, and uses signal processing technology to identify the characteristic frequency of bearing faults. However, during the process of capturing data, it is difficult to capture the temporal correlation between multiple physical fields, resulting in a relatively high missed detection rate of early weak anomalies. The sound analysis-based method detects anomalies by analyzing the sound signal generated during the operation of the bearing. This method can use a sound sensor such as a microphone to capture data, and uses sound processing technology such as Mel Frequency Cepstral Coefficients to analyze the anomaly characteristics of the signal. However, it is affected by environmental noise interference and has low accuracy.

[0003] In recent years, deep learning-based bearing anomaly detection methods have become a research hotspot, and a deep neural network model is trained to identify anomaly patterns in bearing data. The deep learning-based bearing anomaly detection method can process a large amount of data and learn complex feature representations from it, thereby improving the accuracy and reliability of bearing fault detection. However, in the existing methods during the bearing detection process, only the temporal features of the bearing are learned, and the adjacent temporal features and non-adjacent temporal features of the bearing cannot be accurately obtained, resulting in low accuracy of bearing anomaly detection.

[0004] Therefore, how to improve the accuracy of bearing anomaly detection is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] In order to solve the above technical problems, the present application proposes the following technical solutions:

[0006] In a first aspect, an embodiment of the present application provides a bearing anomaly detection method based on temporal correlation difference, including:

[0007] Construct a convolutional neural network composed of multiple CNN modules, and use each CNN module to extract multi-dimensional features of the bearing in the time domain and frequency domain from different scales;

[0008] Extract the adjacent temporal features and non-adjacent temporal features within the bearing sliding window through an adjacent attention module and a non-adjacent attention module respectively;

[0009] Calculate the correlation difference between the adjacent time series features and the non - adjacent time series features by using the bearing time series correlation difference mechanism based on symmetric KL divergence;

[0010] Input the correlation difference into a convolutional neural network for training to increase the correlation difference between the normal state and the abnormal state of the bearing and enhance the discrimination of feature representation;

[0011] Use the trained convolutional neural network to map the multi - dimensional features of the bearing into a probability distribution to realize the detection of the abnormal state of the bearing.

[0012] In a possible implementation manner, the extraction of adjacent time series features and non - adjacent time series features within the bearing sliding window by the adjacent attention module and the non - adjacent attention module respectively includes:

[0013] Uniformly divide the input bearing features into a first sub - feature and a second sub - feature according to the channel dimension;

[0014] When the first sub - feature and the second sub - feature pass through the adjacent attention module and the non - adjacent attention module respectively, extract the adjacent time series features and non - adjacent time series features of the bearing.

[0015] In a possible implementation manner, when the first sub - feature passes through the adjacent attention module, the extraction of the adjacent time series features of the bearing includes:

[0016] Input the first sub - feature into the adjacent attention module, and by defining the query - key - value relationship between the current moment and multiple historical moments, use the cross - attention module to calculate the correlation features between the current moment and the adjacent moments step by step; the calculation formula of the cross - attention module is:

[0017]

[0018] In the formula, 、 、 represent the query vector, the key vector, and the value vector respectively, represents 、 the dimension sizes of;

[0019] Fuse the correlation features of the multiple historical moments to form a comprehensive representation of adjacent time series features.

[0020] In a possible implementation manner, the calculation of the correlation features between the current moment and the adjacent moments step by step by defining the query - key - value relationship between the current moment and multiple historical moments and using the cross - attention module includes:

[0021] Define \(T_n\) as the current moment, \(T_{n - 1}\) as the previous moment, \(T_{n - 2}\) as the moment two steps ago, and so on until \(T_{n - t}\) as the moment \(t\) steps ago;

[0022] Take the current moment \(T_n\) as the query vector, the previous moment \(T_{n - 1}\) as the key vector and value vector, and input them into the cross-attention module to obtain the correlation features of the moment \(T_n\) relative to the previous moment;

[0023] Take the current moment \(T_n\) as the query vector, the previous two moments \(T_{n - 1}\) and \(T_{n - 2}\) as the key vector and value vector, and input them into the cross-attention module to obtain the correlation features of the moment \(T_n\) relative to the previous two moments;

[0024] Until the current moment \(T_n\) is used as the query vector, and the previous \(t\) moments \(T_{n - 1}, T_{n - 2}, \ldots, T_{n - t}\) are used as the key vector and value vector, and input them into the cross-attention module to obtain the correlation features of the moment \(T_n\) relative to the previous \(t\) moments.

[0025] In a possible implementation, when the second sub-feature passes through the non-local attention module, the non-local time-series features of the bearing are extracted, including:

[0026] The non-local attention module performs sliding window sampling on the second sub-feature;

[0027] Use the sparse sampling method to obtain the sparse representation of the input bearing features;

[0028] Convert the sparse features of the bearing into query vectors, key vectors, and value vectors through the first linear layer, and use matrix multiplication to calculate the correlation degree between the query vector and the key vector;

[0029] Apply the Softmax activation function to convert the correlation degree into a probability distribution vector;

[0030] Perform matrix multiplication on the probability distribution vector and the value vector, and convert through the second linear layer to obtain the non-local time-series features of the current moment of the bearing and the entire bearing sequence.

[0031] In a possible implementation, calculating the correlation difference between the local time-series features and the non-local time-series features by using the bearing time-series correlation difference mechanism based on symmetric KL divergence includes:

[0032] Use KL divergence to calculate the difference between the local time-series features and the non-local time-series features, and the calculation formula is:

[0033]

[0034] where and represent the probability distributions of the local time-series features and the non-local time-series features respectively;

[0035] The KL divergence is used to calculate the difference between non - adjacent time - series features and adjacent time - series features. The calculation formula is as follows:

[0036]

[0037] The symmetric KL divergence is used to calculate the point - wise correlation difference between adjacent time - series features and non - adjacent time - series features. The calculation formula is as follows:

[0038]

[0039] Among them, is the prior correlation and the sequence correlation The KL divergence calculated between two discrete distributions in each row, is the point - wise correlation difference relative to the multi - layer prior correlation and the sequence correlation .

[0040] In a possible implementation, the calculation formula for increasing the correlation difference between the normal state and the abnormal state of the bearing is as follows:

[0041]

[0042] Among them, represents the correlation difference between the prior correlation P and the sequence correlation S within the same sliding window; is a hyper - parameter; represents the value of the loss function.

[0043] In a second aspect, an embodiment of the present application provides a bearing anomaly detection system based on time - series correlation difference, including:

[0044] A feature extraction module, a convolutional neural network composed of multiple CNN modules, which uses each CNN module to extract multi - dimensional features of the bearing in the time domain and frequency domain from different scales;

[0045] A time - series feature learning module, which is used to extract adjacent time - series features and non - adjacent time - series features within the bearing sliding window through an adjacent attention module and a non - adjacent attention module respectively;

[0046] A time - series correlation difference calculation module, which is used to calculate the correlation difference between the adjacent time - series features and the non - adjacent time - series features based on the bearing time - series correlation difference mechanism of the symmetric KL divergence;

[0047] A network training module, which is used to input the correlation difference into the convolutional neural network for training, increase the correlation difference between the normal state and the abnormal state of the bearing, and enhance the discrimination of the feature representation;

[0048] Anomaly detection module, which uses the trained convolutional neural network to map the multi-dimensional features of the bearing into a probability distribution to achieve the detection of the abnormal state of the bearing.

[0049] Compared with the prior art, the beneficial effects of this application are as follows:

[0050] In this application, a proximity attention module and a non-proximity attention module are designed at the level of bearing adjacent time-series feature extraction and bearing non-adjacent time-series feature extraction. Among them, the proximity attention module focuses on the features of adjacent time points, enabling the network to better identify potential fault patterns, effectively capture the dynamic changes of the bearing within a local range, and enhance the sensitivity and accuracy of the model to the bearing state changes. The non-proximity attention module can efficiently capture the dynamic changes of the bearing on a larger time scale, which is particularly important for fault features that are not obvious in the short term but gradually appear during long-term operation.

[0051] In this application, at the level of bearing time-series correlation difference processing, a bearing time-series correlation difference mechanism based on symmetric KL divergence is designed, and the adjacent time-series features and non-adjacent time-series features within the same sliding window are calculated based on the bearing time-series correlation difference mechanism based on symmetric KL divergence. During training, by increasing the correlation difference between the normal state and the abnormal state, the discrimination ability of the feature representation is improved to effectively distinguish the normal points and abnormal points of the bearing. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a schematic flowchart of a bearing anomaly detection method based on time-series correlation difference provided by an embodiment of this application;

[0053] Figure 2 It is the overall block diagram of the convolutional neural network provided by an embodiment of this application;

[0054] Figure 3 It is the schematic diagram of the bearing Transformer structure provided by an embodiment of this application;

[0055] Figure 4 It is the schematic diagram of the bearing attention structure provided by an embodiment of this application;

[0056] Figure 5 It is the schematic diagram of the structure of the proximity attention module provided by an embodiment of this application;

[0057] Figure 6 It is the schematic diagram of the structure of the non-proximity attention module provided by an embodiment of this application;

[0058] Figure 7 It is the schematic diagram of a bearing anomaly detection system based on time-series correlation difference provided by an embodiment of this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] The present solution will be described below in conjunction with the accompanying drawings and specific embodiments.

[0060] Figure 1 It is a schematic flowchart of a bearing anomaly detection method based on temporal correlation differences provided by an embodiment of the present application. Refer to Figure 1 , a bearing anomaly detection method based on temporal correlation differences in this embodiment includes:

[0061] S101, construct a convolutional neural network composed of multiple CNN modules, and use each CNN module to extract multi-dimensional features of the bearing in the time domain and frequency domain from different scales.

[0062] In the bearing anomaly detection model, the extraction of time-domain features is crucial because the richness of these features directly affects the accuracy of anomaly detection. As Figure 2 shown, this embodiment constructs a convolutional neural network composed of five CNN blocks. Each CNN module consists of 1D convolution, bearing Transformer, ReLU, batch normalization, and max pooling, which are used to capture the characteristics of the bearing at different frequencies and time domains, realize the extraction of multi-dimensional features of the bearing at different scales, obtain a multi-scale representation of the bearing features, and enhance the sensitivity of the model to changes in the bearing state. Among them, the core of the CNN module is the bearing Transformer, and its architecture is as Figure 3 shown. The design of the bearing Transformer includes two key parts: the proximity attention module and the non-proximity attention module. These two modules are responsible for extracting the proximity temporal features and non-proximity temporal features in the bearing data respectively. The proximity attention module focuses on capturing the relationship between time points and local patterns and short-term dependencies in the bearing signal. These features are usually closely related to the immediate state of the bearing, and its architecture is as Figure 5 shown. The non-proximity attention module, on the other hand, aims to identify the relationship between time points and long-term dependencies and global patterns in the bearing signal, which is crucial for understanding the long-term operating state and potential degradation trends of the bearing, and its architecture is as Figure 6 shown. By combining these two types of attention modules, the bearing Transformer module can comprehensively extract the time-domain features of the bearing. Use 1D convolution to extract bearing features, introduce non-linearity through the ReLU activation function to help the model learn complex patterns, use batch normalization to stabilize the learning process, accelerate the convergence speed, and help reduce internal covariate shift and improve the generalization ability of the model. Use the max pooling layer to reduce the spatial dimension of the features while retaining the most important information. This pooling strategy not only reduces the number of parameters but also improves the robustness of the model to minor changes and perturbations.

[0063] S102. Extract the adjacent time series features and non - adjacent time series features within the bearing sliding window through the adjacent attention module and the non - adjacent attention module respectively.

[0064] In this embodiment, in the CNN block, a bearing Transformer is designed to obtain the adjacent time series features and non - adjacent time series features of the bearing. First, the input bearing features are evenly divided into two sub - features along the channel dimension. The first sub - feature passes through the adjacent attention module to extract the adjacent time series features of the bearing, and the second sub - feature passes through the non - adjacent attention module to extract the non - adjacent time series features of the bearing. When fusing the adjacent time series features and non - adjacent time series features, the adjacent time series features obtained through the adjacent attention module and the non - adjacent time series features obtained through the non - adjacent attention module are concatenated along the channel dimension. The concatenated features are restored to the input video frame feature dimension using a linear layer. A feed - forward network is used to extract useful feature representations from the concatenated features to obtain the fusion result of the adjacent time series features and non - adjacent time series features, which is used for the subsequent network to learn the bearing features.

[0065] Among them, when the first sub - feature passes through the adjacent attention module to extract the adjacent time series features of the bearing, a method of defining time relationships, feature extraction, and feature fusion is adopted for extracting the adjacent time series features of the bearing. When defining time relationships, let Tn be the current time, Tn - 1 be the previous time, Tn - 2 be the time two steps ago, and so on until Tn - t is the time t steps ago. When extracting features, the current time Tn is used as the query vector Q, the previous time Tn - 1 is used as the key vector K and the value vector V, and they are input into the cross - attention module to obtain the correlation features of the Tn time relative to its previous time. This process is extended. Tn is used as the query vector Q, the previous two times Tn - 1 and Tn - 2 are used as the key vector K and the value vector V, and they are input into the cross - attention module to obtain the correlation features of the Tn time relative to its previous two times. This process continues until Tn is used as the query vector Q, the previous t times Tn - 1, Tn - 2,..., Tn - t are used as the key vector K and the value vector V, and they are input into the cross - attention module to obtain the correlation features of the Tn time relative to its previous t times. When fusing features, all the features obtained above are input into the cross - attention module, and this module automatically fuses the above - mentioned correlation features to form a comprehensive representation of adjacent time series features.

[0066] When the second sub - feature extracts the non - adjacent time - series features of the bearing through the non - adjacent attention module, first, the input features are sampled by a sliding window, and the sparse sampling method is used to achieve the sparse representation of the input bearing features. The linear layer converts the sparse features of the bearing into query vector (Q), key vector (K), and value vector (V). Then, matrix multiplication is used to calculate the correlation degree between the query vector features and the key vector features. To adaptively adjust these correlation degrees, the attention weight scaling method is used. By introducing a learnable scaling factor, the model is allowed to dynamically adjust the correlation strength between features according to the specific requirements of the task. Subsequently, the Softmax activation function is applied to convert the correlation degree into a probability distribution vector. The obtained probability distribution vector is multiplied by the value vector V through matrix multiplication, and then converted through another linear layer to obtain the non - adjacent time - series features of the bearing at time Tn and the bearing sequence within the sliding window.

[0067] All the features obtained above are input into the cross - attention module, which will automatically fuse the correlation features at time Tn and its adjacent times to form a comprehensive representation of adjacent time - series features. In this embodiment, the calculation formula of the cross - attention module is:

[0068]

[0069] In the formula, 、 、 represent the query vector, key vector, and value vector respectively, represents 、 the dimension sizes of.

[0070] S103. Calculate the correlation difference between adjacent time - series features and non - adjacent time - series features by using the bearing time - series correlation difference mechanism based on symmetric KL - divergence.

[0071] See Figure 4 In this embodiment, the bearing time - series correlation difference mechanism based on symmetric KL - divergence is used to calculate the correlation difference between adjacent time - series features and non - adjacent time - series features of the bearing. It not only considers the forward correlation difference but also the reverse correlation difference, thus providing a comprehensive measure of correlation difference. First, the KL - divergence is used to calculate the difference between adjacent time - series features and non - adjacent time - series features. The calculation formula is:

[0072]

[0073] Among them, and represent the probability distributions of adjacent time - series features and non - adjacent time - series features respectively;

[0074] The KL divergence is used to calculate the difference between non - adjacent time - series features and adjacent time - series features. The calculation formula is as follows:

[0075]

[0076] To solve the problem of the asymmetry of the KL divergence, the symmetric KL divergence is used to calculate the point - by - point correlation difference between adjacent time - series features and non - adjacent time - series features, ensuring that the feature differences at each time point can be accurately measured, thus providing rich time - series information for the anomaly detection of bearings. The calculation formula is as follows:

[0077]

[0078] Among them, is the prior correlation and the sequence correlation the KL divergence calculated between two discrete distributions of each row, is the point - by - point correlation difference relative to the multi - layer prior correlation and the sequence correlation Anomaly points will show a larger AssDis(P, S; X) than normal time points, which makes AssDis essentially distinguishable.

[0079] S104, input the correlation difference into a convolutional neural network for training to increase the correlation difference between the normal state and the abnormal state of the bearing and enhance the distinguishability of feature representation.

[0080] In this embodiment, by increasing the correlation difference between the normal state and the abnormal state, the distinguishability of feature representation is improved to effectively distinguish the normal points and abnormal points of the bearing. Specifically, when increasing the correlation difference, the correlation difference between the normal time points of the bearing and the adjacent time points and the correlation difference between the normal time points of the bearing and the non - adjacent time points change little, that is, the correlation difference is small, because the normal points of the bearing usually follow the general mode of bearing operation. In the time series, the changes between normal points and adjacent points are small, and the changes with non - adjacent points are not significant either, which indicates that the normal points show high consistency in the sequence. While the correlation difference between the abnormal time points of the bearing and the adjacent time points and the correlation difference between the abnormal time points of the bearing and the non - adjacent time points change greatly, that is, the correlation difference is large, which reflects the isolation and inconsistency of the abnormal points in the time series. Therefore, increasing the correlation difference can better achieve the detection of abnormal points, enabling the effective distinction between the normal points and abnormal points of the bearing. During the training process, the calculation formula for increasing the correlation difference between the normal state and the abnormal state of the bearing is as follows:

[0081]

[0082] Among them, Represents the association difference between the prior association P and the sequence association S within the same sliding window; Is a hyperparameter used to adjust the weight of the association difference; Represents the value of the loss function. The network adjusts Can improve the accuracy of bearing anomaly detection.

[0083] S105, using the trained convolutional neural network to map the multi-dimensional features of the bearing into a probability distribution to achieve the detection of the abnormal state of the bearing.

[0084] Corresponding to the method for bearing anomaly detection based on temporal sequence association difference provided in the above embodiment, the present application also provides an embodiment of a bearing anomaly detection system based on temporal sequence association difference.

[0085] See Figure 7 As shown in, an embodiment of a bearing anomaly detection system 20 based on temporal sequence association difference provided by the embodiment of the present application includes:

[0086] Feature extraction module 201, a convolutional neural network composed of multiple CNN modules, using each CNN module to extract multi-dimensional features of the bearing in the time domain and frequency domain from different scales.

[0087] Temporal feature learning module 202, used to extract adjacent temporal features and non-adjacent temporal features within the bearing sliding window through an adjacent attention module and a non-adjacent attention module respectively.

[0088] Temporal sequence association difference calculation module 203, used to calculate the association difference between the adjacent temporal features and the non-adjacent temporal features based on the bearing temporal sequence association difference mechanism of symmetric KL divergence.

[0089] Network training module 204, used to input the association difference into the convolutional neural network for training, increase the association difference between the normal state and the abnormal state of the bearing, and enhance the discrimination of the feature representation.

[0090] Anomaly detection module 205, using the trained convolutional neural network to map the multi-dimensional features of the bearing into a probability distribution to achieve the detection of the abnormal state of the bearing.

[0091] In the embodiment of the present application, "a plurality of" means two or more. "And / or" describes the association relationship of the associated objects, indicating that three relationships can exist. For example, A and / or B can represent the situation of A existing alone, A and B existing simultaneously, and B existing alone. Where A and B can be singular or plural. The character " / " generally represents an "or" relationship between the preceding and following associated objects.

[0092] It should be noted that in this text, the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article or device including the said element.

[0093] As described above, the above is only the specific implementation manner of this application. Any person skilled in the art within the technical scope disclosed by this application can easily think of changes or substitutions, which should be covered within the protection scope of this application. The protection scope of this application shall be subject to the protection scope of the said claims.

Claims

1. A bearing anomaly detection method based on temporal correlation differences, characterized in that, Including: Construct a convolutional neural network composed of multiple CNN modules, and use each CNN module to extract multi-dimensional features of the bearing in the time domain and frequency domain from different scales; Extract the adjacent time series features and non-adjacent time series features within the bearing sliding window through the adjacent attention module and the non-adjacent attention module respectively, including: Evenly divide the input bearing features into a first sub-feature and a second sub-feature according to the channel dimension; When the first sub-feature passes through the adjacent attention module, extract the adjacent time series features of the bearing, including: Input the first sub-feature into the adjacent attention module, and by defining the query-key value relationship between the current moment and multiple historical moments, use the cross-attention module to calculate the correlation features between the current moment and the adjacent moments step by step; the calculation formula of the cross-attention module is: where Q, K, and V represent the query vector, key vector, and value vector respectively, and d k represents the dimensionality of Q and K; Fuse the correlation features of multiple historical moments to form a comprehensive representation of adjacent time series features; When the second sub-feature passes through the non-adjacent attention module, extract the non-adjacent time series features of the bearing, including: The non-adjacent attention module performs sliding window sampling on the second sub-feature; Use the sparse sampling method to obtain the sparse representation of the input bearing features; Convert the sparse features of the bearing into a query quantity, a key vector, and a value vector through the first linear layer, and use matrix multiplication to calculate the correlation degree between the query quantity and the key vector; Apply the Softmax activation function to convert the correlation degree into a probability distribution vector; Perform matrix multiplication on the probability distribution vector and the value vector, and convert through the second linear layer to obtain the non-adjacent time series features of the current moment of the bearing and the entire bearing sequence; Adopt a bearing time series correlation difference mechanism based on symmetric KL divergence to calculate the correlation difference between the adjacent time series features and the non-adjacent time series features; Input the correlation difference into the convolutional neural network for training to increase the correlation difference between the normal state and the abnormal state of the bearing and enhance the discrimination of the feature representation; Use the trained convolutional neural network to map the multi-dimensional features of the bearing into a probability distribution to realize the detection of the abnormal state of the bearing.

2. The bearing anomaly detection method based on timing correlation difference according to claim 1, wherein The step of defining the query-key value relationship between the current moment and multiple historical moments and using the cross-attention module to calculate the correlation features between the current moment and the adjacent moments step by step includes: Define Tn as the current moment, Tn-1 as the previous moment, Tn-2 as the previous two moments, and so on to Tn-t as the previous t moments; Take the current moment Tn as the query vector, the previous moment Tn-1 as the key vector and the value vector, and input them into the cross-attention module to obtain the correlation feature of the Tn moment relative to the previous moment; Take the current moment Tn as the query vector, the previous two moments Tn-1 and Tn-2 as the key vector and the value vector, and input them into the cross-attention module to obtain the correlation feature of the Tn moment relative to the previous two moments; Until the current moment Tn is used as the query vector, the previous t moments Tn-1, Tn-2,..., Tn-t are used as the key vector and the value vector, and input into the cross-attention module to obtain the correlation feature of the Tn moment relative to the previous t moments.

3. The bearing anomaly detection method based on temporal correlation difference according to claim 1, characterized in that Calculating the correlation difference between the adjacent time-series features and the non-adjacent time-series features by using the bearing time-series correlation difference mechanism based on symmetric KL divergence, including: Calculating the difference between the adjacent time-series features and the non-adjacent time-series features by using KL divergence, and the calculation formula is: KL(P||S) = 0.5 * KL(P||(P + S) / 2) + 0.5 * KL(S||((P + S) / 2)) Wherein, P and S respectively represent the probability distributions of the adjacent time-series features and the non-adjacent time-series features; Calculating the difference between the non-adjacent time-series features and the adjacent time-series features by using KL divergence, and the calculation formula is: KL(S||P) = 0.5 * KL(S||(P + S) / 2) + 0.5 * KL(P||((P + S) / 2)) Calculating the pointwise correlation difference between the adjacent time-series features and the non-adjacent time-series features by using symmetric KL divergence, and the calculation formula is: Among them, is the KL divergence calculated between two discrete distributions for each row of the prior association P and the sequence association S, and AssDis(P, S; X) is the pointwise association difference of X with respect to the multi-layer prior association P and the sequence association S.

4. The bearing anomaly detection method based on temporal correlation difference according to claim 1, characterized in that, The calculation formula for increasing the correlation difference between the normal state and the abnormal state of the bearing is: L Total (P, S, λ; X) = -× ||AssDis(P, S; X)||1 Among them, AssDis(P, S; X) represents the association difference between the prior association P and the sequence association S within the same sliding window; λ is a hyperparameter; L Total (P, S, λ; X) represents the value of the loss function.

5. A bearing anomaly detection system based on temporal correlation differences, characterized in that Including: A feature extraction module, a convolutional neural network composed of multiple CNN modules, and each CNN module is used to extract multi-dimensional features of the bearing in the time domain and the frequency domain from different scales; A time-series feature learning module, which is used to extract the adjacent time-series features and the non-adjacent time-series features in the bearing sliding window through an adjacent attention module and a non-adjacent attention module respectively, including: Uniformly dividing the input bearing features into a first sub-feature and a second sub-feature according to the channel dimension; When the first sub-feature passes through the adjacent attention module, extracting the adjacent time-series features of the bearing, including: Inputting the first sub-feature into the adjacent attention module, defining the query-key value relationship between the current moment and multiple historical moments, and using the cross-attention module to calculate the correlation features between the current moment and the adjacent moments step by step; the calculation formula of the cross-attention module is: Wherein, Q, K, and V respectively represent the query vector, key vector, and value vector, and d k represents the dimensionality of Q and K; Fusing the correlation features of multiple historical moments to form a comprehensive representation of adjacent time-series features; When the second sub-feature passes through the non-adjacent attention module, extracting the non-adjacent time-series features of the bearing, including: The non-adjacent attention module performs sliding window sampling on the second sub-feature; Using a sparse sampling method to obtain a sparse representation of the input bearing features; Converting the sparse features of the bearing into a query quantity, a key vector and a value vector through a first linear layer, and calculating the correlation degree between the query quantity and the key vector by using matrix multiplication; Applying a Softmax activation function to convert the correlation degree into a probability distribution vector; Performing matrix multiplication on the probability distribution vector and the value vector, and converting through a second linear layer to obtain the non-adjacent time-series features of the current moment of the bearing and the entire bearing sequence; A time-series correlation difference calculation module, which is used to calculate the correlation difference between the adjacent time-series features and the non-adjacent time-series features by using the bearing time-series correlation difference mechanism based on symmetric KL divergence; A network training module, which is used to input the correlation difference into a convolutional neural network for training, increase the correlation difference between the normal state and the abnormal state of the bearing, and enhance the discrimination of the feature representation; An anomaly detection module, which uses the trained convolutional neural network to map the multi-dimensional features of the bearing into a probability distribution to realize the detection of the abnormal state of the bearing.

Images