Bearing fault classification method and device based on tensor embedding

Abstract

The invention discloses a bearing fault classification method and device based on tensor embedding, and the method comprises the steps: collecting the vibration acceleration time series data generated by a bearing, decomposing the multi-channel time series data through employing a multi-element variational mode decomposition process for a multi-channel signal processing technology, and obtaining a classification result; obtaining a decomposed high-dimensional multi-channel limited bandwidth intrinsic mode component; carrying out time domain and frequency domain feature calculation on the decomposed multi-channel limited bandwidth intrinsic mode component to obtain time-frequency domain feature signals of a plurality of decomposed signals, forming a sensing node tensor by using a high-dimensional tensor signal of each sensing node to form a tensor graph structure, and processing the tensor graph structure by adopting a TGE algorithm to obtain a low-dimensional feature tensor graph; and importing the low-dimensional feature tensor diagram into an affine propagation clustering algorithm to realize the distinguishing of different fault types. According to the method, the multi-sensing signal can be processed under the condition of keeping the inherent structure information of the multi-sensing signal, and the fault classification precision is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment monitoring, and in particular to a bearing fault classification method and device based on tensor embedding. Background Art

[0002] Bearings are important supporting parts in most electromechanical equipment and mechanical equipment. Generally, bearings need to continuously support the equipment load since installation. Therefore, bearings can easily fail due to long-term or overload operation, causing accidents such as unexpected downtime, which may cause economic losses and casualties in serious cases. When a bearing failure occurs, timely and accurate judgment of the bearing failure type is one of the keys to ensure the safe operation of electromechanical systems, mechanical equipment, etc.

[0003] In recent years, with the rapid development of big data technology, many data-driven fault identification methods have emerged, which can effectively mine and operate big data and reduce the dependence of operators on expert knowledge in the fault identification process. However, data-driven methods require a large amount of data. Since multi-sensors can reflect fault status from multiple dimensions and can obtain more data than a single sensor, multi-sensors are gradually being used more and more in bearing fault monitoring.

[0004] However, as the sensor structure becomes more and more complex, for example, when multiple sensor nodes are high-dimensional data, traditional technologies generally stack similar high-dimensional data into low-dimensional calculations, or directly calculate them one by one. This approach has two disadvantages: first, stacking into low-dimensional calculations will destroy the inherent information such as the position and structure of multiple sensors, reducing the accuracy of fault identification; second, calculating them one by one does not consider the sensor structure and the spatial distribution of high-dimensional data, which greatly increases the calculation time. The above practices will affect the recognition accuracy and speed.

[0005] Based on this, a bearing fault classification method and device based on tensor graph structure and tensor graph embedding are designed. The main purpose is to retain the inherent structural information of multi-sensor signals, carry out parallel computing and synchronous analysis, and effectively improve the fault classification accuracy while reasonably utilizing high-dimensional multi-point data. Summary of the invention

[0006] The purpose of the present invention is to provide a bearing fault classification method and device based on tensor embedding. The main purpose of the present invention is to process multi-sensor signals while retaining their inherent structural information, thereby effectively improving the fault classification accuracy.

[0007] To achieve the above object, the present invention provides a bearing fault classification method based on tensor embedding, comprising the following steps:

[0008] Step 1: Arrange multiple vibration acceleration sensors at key positions, each vibration acceleration sensor is a sensing node, and use the vibration acceleration sensor to collect and obtain the vibration acceleration time series data generated by the bearing during the operation of the equipment; Arrange multiple vibration acceleration sensors at key positions of the equipment under test, each vibration acceleration sensor is a sensing node, and each vibration acceleration sensor is multi-channel, and use the above method to collect and obtain the vibration acceleration time series data during the operation of the equipment;

[0009] Step 2, performing multivariate variational mode decomposition on the multi-channel time series data obtained by each sensor node; using the multivariate variational mode decomposition process to decompose the multi-channel time series data using multi-channel signal processing technology to obtain the decomposed high-dimensional multi-channel limited bandwidth intrinsic mode components;

[0010] The multi-channel time series data obtained from each sensor node are decomposed by using the multivariate variational mode decomposition (MVMD) signal processing technology to obtain the decomposed high-dimensional multi-component node signal; the multi-channel vibration acceleration time series signal of each node is subjected to multivariate variational mode decomposition processing, and the multi-channel signal of each node is decomposed in parallel to obtain multiple components; acceleration sensors are arranged at key measuring points (such as motor output shaft box, transmission gear box, etc.), and the acceleration sensor of each measuring point is multi-channel, so as to obtain the vibration acceleration time series signal during the operation of the equipment;

[0011] Step 3, perform time domain and frequency domain feature calculations on the decomposed multi-channel limited bandwidth intrinsic mode components to obtain time and frequency domain feature signals of multiple decomposed signals, and then construct the time and frequency domain feature signals into a high-dimensional tensor signal; perform time domain and frequency domain numerical feature calculations on the multi-group components to obtain time domain and frequency domain statistical features, and use the number of sensor channels, the number of multi-components after decomposition, and the number of time and frequency domain statistical features to construct a third-order node tensor, which becomes a node tensor; after the tensor node representation is completed, use the calculation method of the fully connected graph to establish all sensor node tensors into a tensor graph structure, in which the calculation of the fully connected weight is implemented using the tensor distance;

[0012] Step 4, the high-dimensional tensor signal of each sensor node is combined into a sensor node tensor to form a tensor graph structure, and the tensor graph structure is processed by the TGE algorithm to obtain a low-dimensional feature tensor graph; the present invention uses a fully connected graph to establish all the obtained sensor node tensors into a tensor graph; the tensor graph embedding (TGE) method is used to reduce the dimension of the tensor graph to obtain low-dimensional features;

[0013] Step 5: Import the low-dimensional feature tensor graph into the Affinity Propagation (AP) clustering algorithm to distinguish different fault types. The tensor graph is input into the tensor graph embedding algorithm to obtain a low-dimensional embedding sequence of the tensor graph; the low-dimensional embedding sequence is input into the affine propagation clustering algorithm to complete the data classification, that is, to distinguish different fault categories.

[0014] The above scheme is further preferred, in step 4, using the TGE algorithm to process the tensor graph structure comprises the following steps:

[0015] Step 40, forming a third-order node tensor using the number of channels × the number of multi-channel finite bandwidth intrinsic mode components × the number of characteristics;

[0016] Step 41, connect the three node tensors with edge weights, so that each node tensor is connected to other node tensors to form a tensor graph structure, and use the weight matrix W to represent the tensor structure graph, where the edge weight w ij satisfy:

[0017]

[0018] In the formula, represents the distance between tensor i and tensor j; Satisfy, forming a tensor graph structure, where the edge weight satisfy:

[0019]

[0020] Parameter g lm Represents the position information of the internal elements of the tensor, parameter g lm satisfy:

[0021]

[0022] ||p l -p m || represents the positional relationship between high-dimensional sequences within the tensor, that is, the position distance;

[0023] Step 42, through the edge weight w ij Calculate the weight matrix W of the tensor graph structure,

[0024] W←w ij ;

[0025] For any sensor node tensor, the degree de represents the sum of the weights of all connections to the current sensor node tensor, and the degree matrix D is filled with the degree de, which satisfies: D←de i , the degree satisfies:

[0026]

[0027] Step 43, using the weight matrix W and the degree matrix D to calculate the graph Laplacian matrix L, and solve the solution vector, that is, satisfying: L = DW;

[0028] Step 43, select the eigenvectors corresponding to the smallest u non-zero eigenvalues ​​as the embedding vectors to obtain the features after dimensionality reduction; that is, the dimensionality reduction result is obtained.

[0029] The above scheme is further preferred in that each of the vibration acceleration sensors is multi-channel.

[0030] In the above scheme, it is further preferred that the sensed node tensor is a second-order or third-order tensor.

[0031] The above scheme is further preferred that the tensor graph structure is established in a fully connected manner, connecting each node tensor with other node tensors.

[0032] According to another aspect of the present invention, the present invention provides a bearing fault classification device based on tensor graph structure and tensor graph embedding, characterized in that the bearing fault classification device includes a data acquisition module, a signal processing module, a feature extraction module, a tensor processing module and an intelligent classification module;

[0033] The data acquisition module uses a multi-channel vibration acceleration sensor to collect vibration acceleration time series data generated during the operation of the bearing to obtain a multi-channel signal;

[0034] The signal processing module uses a multivariate variational mode decomposition process to decompose the multichannel time series data using a multichannel signal processing technology to obtain a decomposed high-dimensional multichannel limited bandwidth intrinsic mode component;

[0035] The feature extraction module is used to perform time domain and frequency domain feature calculations on the decomposed multi-channel limited bandwidth intrinsic mode components to obtain time-frequency domain feature signals of multiple decomposed signals;

[0036] The tensor processing module is used to combine the high-dimensional tensor signal of each sensor node into a sensor node tensor to form a tensor graph structure, and use a tensor graph embedding algorithm to process the tensor graph structure to obtain a low-dimensional feature tensor graph;

[0037] The intelligent classification module imports the low-dimensional feature tensor graph into the affine propagation clustering algorithm for calculation to achieve the distinction between different fault types.

[0038] In summary, since the present invention adopts the above-mentioned technical scheme, the present invention has the following beneficial technical effects: the present invention provides a feasible way for the efficient use of multi-sensor data in the fault monitoring and classification device, and can make full use of the spatial structure information of multi-sensor and multi-channel data, and at the same time integrate the above-mentioned information into the final classification result, thereby achieving good fault classification effect, and effectively improving the fault classification accuracy under the condition of reasonable use of high-dimensional multi-measurement point data. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0040] Figure 1 It is a classification flow chart of the bearing fault classification device of the present invention.

[0041] Figure 2 It is a flow chart of the multivariate variational mode decomposition algorithm of the present invention.

[0042] Figure 3 It is a tensor diagram of sensor nodes established under different fault types and normal conditions in the present invention.

[0043] Figure 4 It is a tensor graph structure established under the multi-sensing nodes of the outer ring fault in the present invention.

[0044] Figure 5 It is a flow chart of the tensor graph embedding algorithm of the present invention.

[0045] Figure 6 This is a comparison effect diagram of the results of bearing fault classification using different methods in the present invention;

[0046] Figure 7 It is a flow chart of a bearing fault classification method based on tensor embedding of the present invention. DETAILED DESCRIPTION

[0047] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0048] For demonstration purposes, see Figure 1The present invention provides a bearing fault classification device based on tensor graph structure and tensor graph embedding, the bearing fault classification device comprises a data acquisition module, a signal processing module, a feature extraction module, a tensor processing module and an intelligent classification module; the data acquisition module uses a multi-channel vibration acceleration sensor to collect the vibration acceleration time series data generated during the operation of the bearing to obtain a multi-channel signal; the signal processing module uses a multivariate variational mode decomposition process to decompose the multi-channel time series data using a multi-channel signal processing technology to obtain a decomposed high-dimensional multi-channel finite bandwidth intrinsic mode component; the feature extraction module is used to extract the decomposed multi-channel finite bandwidth intrinsic mode component. The time domain and frequency domain feature calculations are performed on the state components to obtain the time and frequency domain feature signals of multiple decomposed signals; the tensor processing module is used to combine the high-dimensional tensor signals of each sensor node into a sensor node tensor to form a tensor graph structure, and the tensor graph embedding algorithm is used to process the tensor graph structure to obtain a low-dimensional feature tensor graph; the intelligent classification module imports the low-dimensional feature tensor graph into the affine propagation clustering algorithm for calculation to achieve the distinction of different fault types; this embodiment uses the subway train drive motor bearing as an example, and pre-sets 3 types of bearing faults. The faults are all processed by electrical spark machining, and the preset fault values ​​are respectively the inner ring fault of the bearing, the outer ring fault of the bearing, and the rolling element fault of the bearing. The vibration acceleration sensors are arranged at the output end of the drive motor, the non-output end, the middle of the motor housing, and the output shaft box, that is, there are 4 sensor nodes in total.

[0049] The vibration sensor in this embodiment is a four-directional vibration sensor, that is, each sensor node has four channels, and the sampling frequency of the vibration acceleration sensor used is 25600 Hz.

[0050] See also Figures 1 to 7 The present invention provides a bearing fault classification method based on tensor graph structure and tensor graph embedding. The overall process is as follows: Figure 7 As shown; the bearing fault classification method includes the following steps:

[0051] Step 1, based on the arranged vibration acceleration sensors, obtain the high-dimensional vibration acceleration sequence data generated when the train drive motor is running under multi-sensor nodes and multi-channels;

[0052] Specifically, the vibration acceleration sequence data obtained is X C ={x 1 (t),x 2 (t),...,x i (t)}; where X C Represents a high-dimensional vibration acceleration sequence data set of C sensor nodes; C represents the number of sensor nodes (the number is C sensor nodes), and the distance between the node tensors will be calculated in pairs in the subsequent steps;

[0053] Specifically, this embodiment has 4 sensor nodes, so there are 4 high-dimensional vibration acceleration sequence data sets, namely X 1 , X 2 , X 3 , X 4 ; High-dimensional vibration acceleration sequence data set X C The sub-dataset x in i (t) represents the t-th sampling point in the i-th sensing channel. Obviously, in this embodiment, i={1,2,3,4}. In addition, the number of sampling points t is set to 4096 in this embodiment;

[0054] Step 2, performing multivariate variational mode decomposition on the multi-channel data in each sensor node;

[0055] Specifically, for the multi-channel time series data multi-channel signal X C ={x 1 (t),x 2 (t),...,x i (t)}, using Figure 2 The multi-channel band-limited intrinsic mode function (BLIMF) signal is calculated by following the multivariate variational mode decomposition process shown in The kth layer of the multivariate variational mode decomposition layer, where k = 1, 2, ..., K;

[0056] Specifically, in this embodiment, the number of multivariate variational mode decomposition layers K is set to 4;

[0057] Step 3, perform feature calculation on the BLIMF obtained from the multi-component signal of each sensor node;

[0058] Specifically, this embodiment uses 10 frequency domain features and 10 time domain features, totaling 20, as shown in formula (1) to formula (20);

[0059]

[0060] F 5 =max|x(n)| (5)

[0061]

[0062]

[0063] Specifically, F 1 to F 10 represents the statistical characteristics in the time domain, where x(n) represents the vibration point in the time series, and n = 1, 2, ..., N represents the number of vibration data points;

[0064] Specifically, F 11 to F 20 represents the frequency domain statistical characteristics, where s(k) is the spectrum of the corresponding time series x(n), k = 1, 2, ..., K, represents the number of spectral lines, and f k is the frequency value of the kth spectral line;

[0065] Step 4: The high-dimensional signal of each sensor node is combined into a node tensor to form a tensor graph structure; the tensor graph structure is processed by the TGE algorithm to obtain a low-dimensional feature tensor graph. Low-dimensional features refer to features with only two or three dimensions. If feature extraction and calculation are performed according to multi-channel data, hundreds or thousands of dimensional features will be generated. Therefore, such a large number of features is very unfavorable for classification. Therefore, these features are first represented by a tensor graph structure, and then the tensor graph structure is processed to reduce the feature dimension to 2 or 3 dimensions. The low-dimensional features are then introduced into the clustering algorithm to divide different signals and distinguish faults.

[0066] S41 uses the number of channels × the number of BLIMFs × the number of features to form a third-order node tensor;

[0067] Specifically, the third-order node tensor of this embodiment is 4×4×20;

[0068] Specifically, the above node tensor can be expressed as Where c represents the cth sensor node;

[0069] Specifically, after constructing the tensors of all fault types and normal types, the tensor of each sensor node 1 is displayed in Figure 3 In the example, the number of channels × the number of BLIMFs × the number of features form a third-order node tensor. A tensor is stacked into a three-dimensional array. This three-dimensional array is composed of the number of channels, the number of BLIMFs, and the number of features (the number of features). Figure 3 We can observe that the maximum dimension of the number of channels is 4, the maximum dimension of the number of BLIMFs is also 4, and the maximum dimension of the number of features is 20.

[0070] S42 connects the node tensors in S41 with edge weights to form a tensor graph structure; specifically, the tensor graph of this embodiment adopts a fully connected mode, that is, each node tensor is connected to other node tensors;

[0071] Specifically, the number of sensor nodes in this embodiment is 4, so each node tensor is connected to the other 3 sensor nodes; wherein the tensor graph structure formed by the 4 sensor nodes of the outer ring fault is as follows: Figure 4 As shown;

[0072] Specifically, when each node tensor is connected to other node tensors, the edge connection is made through weight. Each edge is actually a numerical value, which is called the weight, that is, the edge weight. The edge weight w ij Calculate according to formula (21), where δ is a hyperparameter. The larger the δ value, the greater the impact of the edge weight. In this embodiment, the δ value is set to 1;

[0073]

[0074] Specifically, the parameters represents the distance between tensor i and tensor j, Calculate according to formula (22);

[0075]

[0076] Specifically, the parameter g lm Used to express the position information of the internal elements of the tensor, g lm Calculated according to formula (23), where γ is a regularization parameter, which is set to 1 in this embodiment;

[0077]

[0078] Specifically, ||p l -p m || represents the positional relationship between high-dimensional sequences within a tensor, which is called position distance;

[0079] Specifically, for the third-order tensor in this embodiment, where l=(l 1 ,l 2 ,l 3 ) and m=(m 1 ,m 2 ,m 3 ) represents an element x in the node tensor l and x m The position coordinates, ||p l -p m ||Calculate according to formula (24);

[0080]

[0081] S43 uses the tensor graph structure formed in S42 Figure 5 The TGE algorithm process shown in the figure is used to calculate and obtain the features after dimensionality reduction. It should be noted that Figure 5 The TGE algorithm part shown has been completed in step S42, for example, the weight w ij Calculation of g lm The calculation and The specific calculation process is as follows: First, use formula (21) to calculate the elements in the weight matrix W in the tensor graph structure (the elements in W are determined by the edge weights w ij Composition), the weight matrix filled according to formula (25);

[0082] W←w ij ; (25)

[0083] (Secondly, for any node tensor X in the tensor graph C , where C represents the number of sensor nodes, such as Figure 4 There are 4 sensor nodes shown, so C = 4. The degree de represents the sum of the weights of all the tensor connections to this node. De can be calculated according to formula (26), and then filled in according to formula (27) to obtain the degree matrix D, which represents the sum of the weights of the node tensor connections;

[0084]

[0085] D←de i (27)

[0086] Again, the graph Laplacian matrix L is calculated using the weight matrix and the degree matrix according to formula (28);

[0087] L=DW (28)

[0088] Finally, the Lagrangian function matrix is ​​constructed using the graph Laplace matrix L and the degree matrix D according to formula (29);

[0089] Ly i =λ i Dy i (29)

[0090] Formula (29) can be regarded as a generalized eigenvalue solution problem, which can be solved by eigendecomposition, which is a common calculation method. After calculation, the solution vector y of formula (28) is obtained: i Then, the eigenvectors corresponding to the smallest u non-zero eigenvalues ​​are selected as the embedding vectors, that is, the dimensionality reduction result obtained, where u is the dimension reduced to. By selecting the first u smallest eigenvectors, u is a constant value set by the user. From the calculated eigenvalues ​​and eigenvectors, only the first u smallest eigenvectors are selected as the result output. In the process of solving the eigenvalues, the number of eigenvectors is the same as the dimension of the matrix D, so u cannot be set to a dimension exceeding D. For visualization, the u value in this embodiment is set to 2. The specific u value setting can be set according to user needs. It should be noted that in the process of solving the eigenvalues, the number of eigenvectors is the same as the dimension of the matrix D, so u cannot be set to a dimension exceeding D.

[0091] Step 5, input the u-dimensional features obtained in S43 into the AP clustering algorithm to perform AP clustering, complete the classification of different fault types, and obtain the fault classification results. After using the AP clustering algorithm, the collected fault signals can be divided into different categories, that is, the fault classification is completed; and the establishment of the u-dimensional features is based on the calculation of formula 21 to formula 29. The original features have been converted into, for example, weights w, distance values ​​d, etc., and then Laplace calculations are performed to obtain the degree matrix and solve the Lagrangian function to obtain the eigenvalues ​​and their corresponding eigenvectors. These eigenvectors are used as dimensionality reduction results; such as Figure 6 As shown; among them, Figure 6 (e) is the result obtained by the method of the present invention. According to the above u is set to 2, that is, one of the eigenvectors is used as the X axis and the other eigenvector is used as the Y axis. The dot graph formed by X and Y is plotted on Figure 6 (e) in.

[0092] In order to better understand the technical solution of the present invention, the embodiment uses another 5 methods to compare with the present invention to reflect the superiority of the present invention. The final fault classification results are shown in Figure 6 In the following, we will further explain the various methods in the figure:

[0093] The other five comparison methods used are: Principal Components Analysis (PCA), Locality Preserving Projections (LPP), Locally Linear Embedding (LLE), t-distributed Stochastic Neighbor Embedding (t-SNE), and Isometric Feature Mapping (ISOMAP). For the rationality of the comparison, the comparison method sets the dimension reduction dimension to 2;

[0094] Specifically, Figure 6 In (a), the PCA algorithm has difficulty distinguishing inner circle faults from outer circle faults. Figure 6 The LPP algorithm in (b) also confuses the inner and outer race faults. Figure 6 The LLE algorithm in (c) is very loose, and the distinction between the inner ring, outer ring and cage is not high. Figure 6 (d) The T-SNE still has difficulty separating the inner and outer ring faults. Figure 6 The method of the present invention in (e) divides the four types of faults well. Figure 6(f) The ISOMAP classification effect is still poor. The main reason is that PCA, LPP, LLE, t-SNE, and ISOMAP cannot process tensor data and can only convert tensor data into vectors for calculation, thus destroying the spatial structure of the data itself and achieving poor results. The method of the present invention can make full use of the spatial structure information of multi-sensor and multi-channel data, and integrate the above information into the final classification results, achieving good fault classification results.

[0095] The above is only a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that for those skilled in the art, without departing from the principle of the present invention, several improvements, equivalent substitutions and modifications made should also be regarded as the protection scope of the present invention.

Claims

1. A bearing fault classification method based on tensor embedding, characterized in that: The following steps are involved: Step 1: Arrange multiple vibration acceleration sensors at key locations, each vibration acceleration sensor being a sensing node, and use the vibration acceleration sensors to collect and obtain the vibration acceleration time series data generated by the bearing during the operation of the equipment; Step 2, multivariate variational mode decomposition is performed on the multi-channel time series data obtained by each sensor node, and the multi-channel signal processing technology is used to decompose the multi-channel time series data to obtain the decomposed high-dimensional multi-channel limited bandwidth intrinsic mode components; Step 3, performing time domain and frequency domain feature calculations on the decomposed multi-channel limited bandwidth intrinsic mode components to obtain time-frequency domain feature signals of multiple decomposed signals, and then constructing the time-frequency domain feature signals into a high-dimensional tensor signal; Step 4, the high-dimensional tensor signal of each sensor node is combined into a sensor node tensor to form a tensor graph structure, and the tensor graph structure is processed by the TGE algorithm to obtain a low-dimensional feature tensor graph; Step 5: Import the low-dimensional feature tensor graph into the affine propagation clustering algorithm to distinguish different fault types.

2. A bearing fault classification method based on tensor embedding according to claim 1, characterized in that: In step 4, using the TGE algorithm to process the tensor graph structure includes the following steps: Step 40, forming a third-order node tensor using the number of channels × the number of multi-channel finite bandwidth intrinsic mode components × the number of characteristics; Step 41, connect the three node tensors with edge weights, so that each node tensor is connected to other node tensors to form a tensor graph structure, and use the weight matrix W to represent the tensor structure graph, where the edge weight w ij satisfy: In the formula, represents the distance between tensor i and tensor j; where Satisfy, forming a tensor graph structure, where the edge weight satisfy: Parameter g lm Represents the position information of the internal elements of the tensor, parameter g lm satisfy: ||p l -p m || represents the positional relationship between high-dimensional sequences within the tensor, that is, the position distance; Step 42, through the edge weight w ij Calculate the weight matrix W of the tensor graph structure, W←w ij ; For any sensor node tensor, the degree de represents the sum of the weights of all connections to the current sensor node tensor, and the degree matrix D is filled with the degree de, which satisfies: D←de i , the degree satisfies: Step 43, using the weight matrix W and the degree matrix D to calculate the graph Laplacian matrix L, and solve the solution vector, that is, satisfying: L = DW; Step 43, select the eigenvectors corresponding to the smallest u non-zero eigenvalues ​​as the embedding vectors to obtain the features after dimensionality reduction; that is, the dimensionality reduction result is obtained.

3. A bearing fault classification method based on tensor embedding according to claim 1, characterized in that: Each of the vibration acceleration sensors is multi-channel.

4. A bearing fault classification method based on tensor graph structure and tensor graph embedding as claimed in claim 1, characterized in that: The sensed node tensor is a second-order or third-order tensor.

5. A bearing fault classification method based on tensor embedding according to claim 1 or 2, characterized in that: The tensor graph structure is established in a fully connected manner, connecting each node tensor to other node tensors.

6. A bearing fault classification device based on tensor graph structure and tensor graph embedding, characterized in that: The bearing fault classification device includes a data acquisition module, a signal processing module, a feature extraction module, a tensor processing module and an intelligent classification module; The data acquisition module uses a multi-channel vibration acceleration sensor to collect vibration acceleration time series data generated during the operation of the bearing to obtain a multi-channel signal; The signal processing module uses a multivariate variational mode decomposition process to decompose the multichannel time series data using a multichannel signal processing technology to obtain a decomposed high-dimensional multichannel limited bandwidth intrinsic mode component; The feature extraction module is used to perform time domain and frequency domain feature calculations on the decomposed multi-channel limited bandwidth intrinsic mode components to obtain time-frequency domain feature signals of multiple decomposed signals; The tensor processing module is used to combine the high-dimensional tensor signal of each sensor node into a sensor node tensor to form a tensor graph structure, and use a tensor graph embedding algorithm to process the tensor graph structure to obtain a low-dimensional feature tensor graph; The intelligent classification module imports the low-dimensional feature tensor graph into the affine propagation clustering algorithm for calculation to achieve the distinction between different fault types.

Images