ICBL-based intelligent fault diagnosis method for gear box of cutting unit of coal mining machine

Abstract

The invention discloses an ICBL-based intelligent fault diagnosis method for a gear box of a cutting unit of a coal mining machine. The method comprises the following steps: firstly, collecting vibration data of different fault types; randomly dividing the data into a training data set and a test data set, and setting an unbalance degree in the training data set; then, complex structure information of the vibration data of the gearbox of the cutting part of the coal mining machine can be quickly and efficiently embedded into the improved cascade width learning model by using a random hypergraph convolution mechanism, so that the discrimination of fault features learned by the improved cascade width learning model is greatly improved; meanwhile, a class specificity weight distribution strategy is designed according to proportion information between data classes, an improved cascade width learning model can pay more attention to a few class samples, the fault diagnosis performance of unbalanced data is improved, and the fault diagnosis efficiency of the unbalanced data is improved through a mode of combining a random hypergraph convolution mechanism and the class specificity weight distribution strategy. And finally, the fault diagnosis precision of the gear box of the cutting part of the coal mining machine is effectively improved.

Description

Technical Field

[0001] The invention relates to the technical field of fault diagnosis of a gear box of a cutting part of a coal mining machine, and in particular to an intelligent fault diagnosis method of a gear box of a cutting part of a coal mining machine based on ICBL. Background Art

[0002] The safe and stable operation of coal mining machines is the key to ensuring efficient coal production. The cutting gearbox is the core power transmission component of the coal mining machine. Once a failure occurs, it will seriously affect the efficiency of coal mining, cause huge economic losses, and even cause major safety accidents. Therefore, it is of great significance to diagnose the fault of the cutting gearbox of the coal mining machine.

[0003] Fault diagnosis methods based on deep learning can simulate the multi-level learning process of the human brain, adaptively mine the high-level abstract features inherent in the state monitoring signals of coal mining equipment, and autonomously construct a complex mapping relationship between monitoring data and health status, which greatly overcomes the limitations of manual feature extraction and is favored by industry and academia. However, existing diagnostic methods based on deep learning all improve the fault feature extraction performance of the model by deeply stacking multi-layer network structures, but this leads to common problems such as complex model structure, numerous parameters, and low learning efficiency. In addition, most existing deep learning models use gradient descent methods for parameter optimization, which is easy to fall into local optimal solutions.

[0004] In response to the above problems, some researchers proposed a Cascade Broad Learning System (CBLS), which constructs multiple groups of cascade feature nodes and enhanced nodes in the width direction, and extracts linear and nonlinear features of the input data at the same time. At the same time, the ridge regression algorithm is used to quickly and efficiently obtain the global optimal solution of the model. Therefore, the CBLS model has begun to emerge in the fields of time series prediction, image recognition, and life prediction. However, the application of the CBLS model to the fault diagnosis of the gearbox of the cutting part of the coal mining machine still faces the following two problems: 1) Due to the randomness and unevenness of the underground coal seam distribution, the monitoring data structure of the gearbox of the cutting part of the coal mining machine is complex and multivariate, resulting in insufficient discriminability of the extracted fault features. 2) In actual engineering applications, the gearbox of the cutting part of the coal mining machine has been in normal service for a long time, so that the data of the fault state is far less than the normal state data. The imbalanced distribution of data categories makes it easy for the model to pay insufficient attention to the minority fault data, which seriously affects the accuracy of fault diagnosis. Summary of the invention

[0005] In view of the problems existing in the above-mentioned prior art, the present invention provides an intelligent fault diagnosis method for the gearbox of the cutting part of a coal mining machine based on ICBL. By combining unbalanced training data with improved cascade width learning, the model increases its attention to minority fault data, thereby effectively improving the accuracy of fault diagnosis of the gearbox of the cutting part of the coal mining machine.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is: an ICBL-based intelligent fault diagnosis method for a gearbox of a shearer cutting part, comprising the following steps:

[0007] Step 1: Install a vibration sensor at a suitable position of the gearbox of the shearer cutting part to collect vibration data of different fault types;

[0008] Step 2: Use Hilbert marginal spectrum to preprocess the collected vibration signal, and randomly divide the preprocessed data into a training data set and a test data set, wherein the imbalance of the training samples is set in the training data set, and the imbalance refers to the value of the number of training samples in the normal state divided by the number of training samples in any other fault state in the training data set. For example, the number of training samples in the normal state is 100, and the number of training samples in any other fault state is 10, then the imbalance is 100÷10=10, that is, the larger the imbalance, the greater the difference between the number of training samples in the normal state and the number of training samples in other fault states in the training data set. Different imbalances are set by different training data sets for subsequent model training;

[0009] Step 3: construct an improved cascade broad learning model (ICBL), which includes an input layer, a cascade hypergraph convolution node, an enhancement node, and an output layer; wherein the cascade hypergraph convolution node is constructed using a random hypergraph convolution mechanism, which can fully mine the high-order complex relationships in the vibration data, thereby extracting more discriminative fault features; at the same time, according to the training data set in step 2, the imbalance of the training samples is set, a class-specific weight allocation strategy is constructed, and it is embedded in the objective function of the improved cascade broad learning model to improve the processing capability of unbalanced data, and then the model is trained using the training data set to obtain the trained improved cascade broad learning model;

[0010] Step 4: Use the test data set in step 2 to test the trained improved cascade width learning model. If the required accuracy is achieved, proceed to step 5; otherwise, repeat step 3 to continue training until the trained model reaches the required accuracy and proceed to step 5;

[0011] Step 5: After the vibration data collected by the vibration sensor in step 1 is preprocessed in step 2, it is input into the model tested in step 4. The model can perform fault diagnosis on the gearbox of the cutting part of the coal mining machine based on the input vibration data.

[0012] Furthermore, the improved cascade width learning model uses a hypergraph to fully characterize the complex multivariate structure of the gearbox vibration data (i.e., input data) of the shearer cutting part during the training process in step 3 and during the fault diagnosis process after the training in step 5, specifically:

[0013] For the sample matrix X = [x1,L,x N ], each sample x is a vertex v of the hypergraph, and the sample and its k nearest neighbor samples form a hyperedge e; therefore, the hypergraph model of the sample matrix X is expressed as in is a vertex set, ε=[e1,…,e N ] is the hyperedge set, w is the hyperedge weight set; the relationship between vertices and hyperedges in the hypergraph model is represented by the association matrix H:

[0014]

[0015] Where N k (x i ) is the sample x i The k nearest neighbor samples, the metric function between samples is selected as:

[0016] S(i,j)=exp(||x i -x j || 2 / 2t 2 )

[0017] In the formula is the heat kernel parameter; the weight of each hyperedge w(e j ) is defined as:

[0018]

[0019] In a hypergraph, the vertex degree d(v i ) represents the hyperedge weight of each vertex, and the hyperedge degree d(e j ) represents the number of vertices contained in the hyperedge, and the calculation formulas are:

[0020] d(v i )=∑ j＝1 w(e j )H(v i ,e j ),v i ∈e j

[0021] d(ej )=∑ i＝1 H(v i ,e j ),v i ∈e j

[0022] The vertex degree, hyperedge degree and hyperedge weight matrix are represented by diagonal matrices: Dv = diag(d(v1),…,d(v N )), De=diag(d(e1),…,d(e N )),Ω=diag(w(e1),…,w(e N )).

[0023] Furthermore, the cascaded hypergraph convolution nodes in step 3 are constructed using a random hypergraph convolution mechanism, specifically:

[0024] The Laplacian operator Δ of the hypergraph model G is a semidefinite matrix, and its eigenvalue decomposition Δ=UΛU T , where U=[u1,u2,…,u N ] is the eigenvector matrix, Λ=diag(λ1,λ2,…λ N ) is the diagonalized eigenvalue matrix; for the signal on the hypergraph model G Its Fourier transform is defined as The eigenvector U is the Fourier basis in the hypergraph domain, and the corresponding eigenvalue is the frequency; the frequency domain convolution of the signal x and the filter f is expressed as:

[0025] f*x=U((U T ·f)⊙(U T x))=Uf(Λ)U T x

[0026] Where * is the convolution operator, ⊙ is the element-wise product operator, and f(Λ) is the eigenvalue function of Δ;

[0027] However, due to its large amount of calculation, the computational complexity is O(|V| 2 ); using multi-order truncated Chebyshev polynomials, the formula is simplified to:

[0028]

[0029] Where Θ is the hypergraph convolution filter parameter; According to the above formula, the first group of hypergraph convolution nodes of the improved cascade width learning model is expressed as:

[0030]

[0031] The improved cascade width learning model uses a random generation method to determine the hypergraph convolution filter parameter Θ1 in the above formula, and requires the generated Θ1 to satisfy orthogonalization; this random hypergraph convolution mechanism can not only retain the structural characteristics of the original cascade width learning model to the greatest extent, but also quickly and efficiently embed hypergraph prior knowledge into the model;

[0032] Similarly, the improved cascade width learning model also cascades the hypergraph convolution nodes layer by layer; therefore, the nth group of hypergraph convolution nodes is defined as:

[0033]

[0034] Next, all the hypergraph convolution nodes are integrated to obtain the total hypergraph convolution node matrix Fn = [F1, F2, ..., F n ], and obtain m groups of feature nodes Z through nonlinear activation function m =[Z1,Z2,…,Z m ]; where the j-th group of enhanced nodes is expressed as:

[0035] Z j =ξ j (F n W hj +β hj ),j=1,2,…,m

[0036]

[0037] Where W hj and β hj are the jth group of randomly generated enhancement mapping weight matrices and biases respectively;

[0038] By integrating all cascaded hypergraph convolution nodes and enhancement nodes, the features learned by the improved cascade width learning model are expressed as:

[0039] A=[F1,F2,…,F n |Z1,Z2,…,Z m ]=[F n ,Z m ].

[0040] Furthermore, the step three constructs a class-specific weight allocation strategy and embeds it into the objective function of the improved cascade width learning model, specifically:

[0041] The class-specific weight assignment strategy assigns different weights to each class of samples; for the c-th class sample, its weight is defined as:

[0042]

[0043] Where N crepresents the number of samples in the cth category, N ave is the mean of the number of samples in each category; τ∈(0,1] is the weight scaling factor, which is used to fine-tune the weights of each category. From the above formula, χ c It is inversely proportional to the number of samples of each category, so the minority samples can get greater weight in the modeling process, increasing their attention;

[0044] By integrating the above formula, the objective function of the improved cascade width learning model is:

[0045]

[0046] Where ρ is the regularization parameter, ξ i For sample x i The training error, Y i For sample x i The true label of , W is the output weight matrix of the improved cascade width learning model, is the diagonalized weight matrix, R ii is the i-th element on its diagonal;

[0047] According to the Lagrange multiplier method, the objective function of the improved cascade width learning model is rewritten as:

[0048]

[0049] Where α i is the Lagrange multiplier, respectively for (W,ξ i ,α i ) Take the partial derivative and make it equal to 0, and get the following formula:

[0050]

[0051] The output weight matrix of the improved cascade width learning model can be directly obtained through the ridge regression algorithm without complicated gradient optimization.

[0052] From the above formula, the output weight matrix of the improved cascade width learning model is obtained as:

[0053] W=(A T RA+ρI) -1 A T RY.

[0054] Compared with the existing fault diagnosis method, the present invention has the following advantages:

[0055] (1) The present invention first constructs an improved cascade width learning model, and uses a random hypergraph convolution mechanism to quickly and efficiently embed the complex structural information of the vibration data of the gearbox of the cutting part of the coal mining machine into the improved cascade width learning model, which greatly improves the discriminability of the fault features learned by the improved cascade width learning model.

[0056] (2) The present invention designs a class-specific weight allocation strategy according to different types of imbalance in the data, which enables the improved cascade width learning model to pay more attention to minority class samples, improves the fault diagnosis performance of unbalanced data, and has higher accuracy for fault diagnosis.

[0057] (3) The present invention combines a random hypergraph convolution mechanism with a class-specific weight allocation strategy to solve the problem that traditional fault diagnosis methods cannot fully mine the complex multivariate structural information between monitoring data and have poor processing capabilities for unbalanced data, thereby effectively improving the accuracy of gearbox fault diagnosis in the cutting section of coal mining machines. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a fault diagnosis flow chart of the present invention.

[0059] Figure 2 It is a structural diagram of the improved cascade width learning model of the present invention.

[0060] Figure 3 It is a comparison chart of the diagnostic accuracy of the present invention on the data set DA.

[0061] Figure 4 It is a comparison chart of the F1-score diagnosis results of the present invention on the data sets DB~DF. DETAILED DESCRIPTION

[0062] The present invention will be further described below.

[0063] like Figure 1 As shown, this embodiment includes the following steps:

[0064] Step 1: Install a vibration sensor at a suitable position of the gearbox of the cutting part of the coal mining machine to collect vibration data of different fault types; the vibration data used in this embodiment is obtained through a fault simulation test bench for the gearbox of the cutting part of the coal mining machine. During the experiment, five states were set on the sun gear of the planetary gearbox, including normal state (NS), missing teeth (MT), root crack (RC), gear wear (GW) and broken tooth (BT). The operating data under five states were collected by the vibration sensor, with a sampling frequency of 48kHz and a motor speed of 40Hz. 250 samples were collected for each state, and the sample length was uniformly 8192.

[0065] Step 2: Use Hilbert marginal spectrum to preprocess the collected vibration signal, and randomly divide the preprocessed data into a training data set and a test data set, wherein the imbalance of the training samples is set in the training data set; in this embodiment, 6 different training data sets are constructed for comparative analysis, and each training data set is abbreviated as DA~DF, and the specific description is shown in Table 1. For the training data set DA, the number of training samples of the gearbox vibration data in each state is equal, so the sample distribution of the data set is balanced. The sample distribution of the training data sets DB~DF is unbalanced (that is, the normal state obviously exceeds the sample size of different fault type states), and the imbalance of the training samples is 2, 3, 6, 10 and 15, respectively. In the data sets DA~DF, the number of test samples of the gearbox vibration data in each state is 100, that is, there are 500 test samples in each test data set.

[0066] Table 1 Detailed information of DA~DF training dataset

[0067]

[0068] Step 3: Construct an improved cascade width learning model, such as Figure 2 As shown in the figure, it includes an input layer, a cascaded hypergraph convolution node, an enhancement node and an output layer; the cascaded hypergraph convolution node is constructed using a random hypergraph convolution mechanism, which can fully mine the high-order complex relationships in the vibration data, thereby extracting more discriminative fault features; at the same time, according to the training data set in step 2, the imbalance degree of the training samples is set, and a class-specific weight allocation strategy is constructed, which is embedded in the objective function of the improved cascade width learning model to improve the processing capability of unbalanced data. The specific steps are:

[0069] The improved cascade width learning model uses a hypergraph to fully characterize the complex multivariate structure of the vibration data of the gearbox of the shearer cutting part (i.e., input data) during the training process of this step and the subsequent fault diagnosis process after training in step 5, specifically:

[0070] For the sample matrix X = [x1,L,x N ], each sample x is a vertex v of the hypergraph, and the sample and its k nearest neighbor samples form a hyperedge e; therefore, the hypergraph model of the sample matrix X is expressed as in is a vertex set, ε=[e1,…,e N ] is the hyperedge set, w is the hyperedge weight set; the relationship between vertices and hyperedges in the hypergraph model is represented by the association matrix H:

[0071]

[0072] Where N k (x i ) is the sample x i The k nearest neighbor samples, the metric function between samples is selected as:

[0073] S(i,j)=exp(||x i -x j || 2 / 2t 2 )

[0074] In the formula is the heat kernel parameter; the weight of each hyperedge w(e j ) is defined as:

[0075]

[0076] In a hypergraph, the vertex degree d(v i ) represents the hyperedge weight of each vertex, and the hyperedge degree d(e j ) represents the number of vertices contained in the hyperedge, and the calculation formulas are:

[0077] d(v i )=∑ j＝1 w(e j )H(v i ,e j ),v i ∈e j

[0078] d(e j )=∑ i＝1 H(v i ,e j ),v i ∈e j

[0079] The vertex degree, hyperedge degree and hyperedge weight matrix are represented by diagonal matrices: Dv = diag(d(v1),…,d(v N )), De=diag(d(e1),…,d(e N )),Ω=diag(w(e1),…,w(e N )).

[0080] The cascaded hypergraph convolution node is constructed using a random hypergraph convolution mechanism, specifically:

[0081] The Laplacian operator Δ of the hypergraph model G is a semidefinite matrix, and its eigenvalue decomposition Δ=UΛU T , where U=[u1,u2,…,u N ] is the eigenvector matrix, Λ=diag(λ1,λ2,…λ N ) is the diagonalized eigenvalue matrix; for the signal on the hypergraph model G Its Fourier transform is defined as The eigenvector U is the Fourier basis in the hypergraph domain, and the corresponding eigenvalue is the frequency; the frequency domain convolution of the signal x and the filter f is expressed as:

[0082] f*x=U((U T f)⊙(U T x))=Uf(A)U T x

[0083] Where * is the convolution operator, ⊙ is the element-wise product operator, and f(Λ) is the eigenvalue function of Δ;

[0084] However, due to its large amount of calculation, the computational complexity is O(V 2 ); using multi-order truncated Chebyshev polynomials, the formula is simplified to:

[0085]

[0086] Where Θ is the hypergraph convolution filter parameter; According to the above formula, the first group of hypergraph convolution nodes of the improved cascade width learning model is expressed as:

[0087]

[0088] The improved cascade width learning model uses a random generation method to determine the hypergraph convolution filter parameter Θ1 in the above formula, and requires the generated Θ1 to satisfy orthogonalization; this random hypergraph convolution mechanism can not only retain the structural characteristics of the original cascade width learning model to the greatest extent, but also quickly and efficiently embed hypergraph prior knowledge into the model;

[0089] Similarly, the improved cascade width learning model also cascades the hypergraph convolution nodes layer by layer; therefore, the nth group of hypergraph convolution nodes is defined as:

[0090]

[0091] Next, all the hypergraph convolution nodes are integrated to obtain the total hypergraph convolution node matrix Fn = [F1, F2, ..., F n ], and obtain m groups of feature nodes Zm=[Z1,Z2,…,Z m ]; where the j-th group of enhanced nodes is expressed as:

[0092] Z j =ξ j (F n W j +βh j ),j=1,2,…,m,

[0093]

[0094] Where Wh j and β hj are the jth group of randomly generated enhancement mapping weight matrices and biases respectively;

[0095] By integrating all cascaded hypergraph convolution nodes and enhancement nodes, the features learned by the improved cascade width learning model are expressed as:

[0096] A=[F1,F2,…,F n |Z1,Z2,…,Z m ]=[F n ,Z m ].

[0097] The class-specific weight allocation strategy is constructed and embedded into the objective function of the improved cascade width learning model, specifically:

[0098] The class-specific weight assignment strategy assigns different weights to each class of samples; for the c-th class sample, its weight is defined as:

[0099]

[0100] Where N c represents the number of samples in the cth category, N ave is the mean of the number of samples in each category; τ∈(0,1] is the weight scaling factor, which is used to fine-tune the weights of each category. From the above formula, χ c It is inversely proportional to the number of samples of each category, so the minority samples can get greater weight in the modeling process, increasing their attention;

[0101] By integrating the above formula, the objective function of the improved cascade width learning model is:

[0102]

[0103] Y i -A i W=ξ i ,

[0104] Where ρ is the regularization parameter, ξ i For sample x i The training error, Y i For sample x i The true label of , W is the output weight matrix of the improved cascade width learning model, is the diagonalized weight matrix, R ii is the i-th element on its diagonal;

[0105] According to the Lagrange multiplier method, the objective function of the improved cascade width learning model is rewritten as:

[0106]

[0107] Where α i is the Lagrange multiplier, respectively for (W,ξ i ,α i ) Take the partial derivative and make it equal to 0, and get the following formula:

[0108]

[0109] The output weight matrix of the improved cascade width learning model can be directly obtained through the ridge regression algorithm without complicated gradient optimization.

[0110] From the above formula, the output weight matrix of the improved cascade width learning model is obtained as:

[0111] W=(A T RA+ρI) -1 A T RY.

[0112] After completion, the model is trained using the training data set to obtain the trained improved cascade width learning model.

[0113] Step 4: Use the test data set in step 2 to test the trained improved cascade width learning model. If the required accuracy is achieved, proceed to step 5; otherwise, repeat step 3 to continue training until the trained model reaches the required accuracy and proceed to step 5;

[0114] Step 5: After the vibration data collected by the vibration sensor in step 1 is preprocessed in step 2, it is input into the model tested in step 4. The model can perform fault diagnosis on the gearbox of the cutting part of the coal mining machine based on the input vibration data.

[0115] In order to verify the accuracy of the method of the present invention in diagnosing the fault of the gearbox of the cutting part of the coal mining machine, various existing methods in the industry are selected, including the Cascade Broad Learning System (CBLS); Weighted Broad Learning System (WBLS); Multi-scale Convolutional Neural Networks (MCNN); Hypergraph Neural Nerworks (HGNN); Multiresolution Hypergraph Neural Network (MHNN). Then, the improved cascade width learning model (ICBL) constructed in this embodiment and the above-mentioned existing methods are trained and the fault diagnosis accuracy is measured; in order to avoid the occasional comparison results, the training samples and test samples of the DA~DF data sets are randomly selected, and each model is repeatedly run 10 times on each data set. Figure 3 are the results of 10 tests of each model on the test data set DB. Figure 3 It can be seen that the improved cascade width learning model (ICBL) of the present invention achieves the highest fault diagnosis accuracy in each test, and even reaches 100% in the 2nd and 10th tests. Figure 4 The F1-score of each model on the training dataset DB~DF is shown. Figure 4 It can be seen that as the data imbalance increases, the F1-score of all models decreases to a certain extent, while the F1-score of the improved cascade width learning model decreases more slowly, which shows that data imbalance has less impact on the improved cascade width learning model (ICBL) of the present invention. Figure 3 and Figure 4 The experimental results verify the superiority of the improved cascade width learning model (ICBL) of the present invention for unbalanced data fault diagnosis, that is, the accuracy is significantly better than other existing methods.

[0116] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. An ICBL-based intelligent fault diagnosis method for a gearbox of a coal mining machine cutting part, characterized in that: The following steps are involved: Step 1: Install a vibration sensor at a suitable position of the gearbox of the shearer cutting part to collect vibration data of different fault types; Step 2: preprocess the collected vibration signal using Hilbert marginal spectrum, and randomly divide the preprocessed data into a training data set and a test data set, wherein the imbalance degree of training samples is set in the training data set; Step 3: construct an improved cascade width learning model, which includes an input layer, a cascade hypergraph convolution node, an enhancement node and an output layer; wherein the cascade hypergraph convolution node is constructed using a random hypergraph convolution mechanism, which can fully mine the high-order complex relationships in the vibration data, thereby extracting more discriminative fault features; at the same time, according to the training data set in step 2, the imbalance of the training samples is set, a class-specific weight allocation strategy is constructed, and it is embedded in the objective function of the improved cascade width learning model to improve the processing capability of unbalanced data, and then the model is trained using the training data set to obtain the trained improved cascade width learning model; Step 4: Use the test data set in step 2 to test the trained improved cascade width learning model. If the required accuracy is achieved, proceed to step 5; otherwise, repeat step 3 to continue training until the trained model reaches the required accuracy and proceed to step 5; Step 5: After the vibration data collected by the vibration sensor in step 1 is preprocessed in step 2, it is input into the model tested in step 4. The model can perform fault diagnosis on the gearbox of the cutting part of the coal mining machine based on the input vibration data.

2. According to claim 1, the ICBL-based intelligent fault diagnosis method for the gearbox of the cutting part of the coal mining machine is characterized in that: The improved cascade width learning model uses a hypergraph to fully characterize the complex multivariate structure of the vibration data of the gearbox of the shearer cutting part during the training process of step 3 and the fault diagnosis process after the training of step 5, specifically: For the sample matrix X = [x1,L,x N ], each sample x is a vertex v of the hypergraph, and the sample and its k nearest neighbor samples form a hyperedge e; Therefore, the hypergraph model of the sample matrix X is expressed as in is a vertex set, ε=[e1,…,e N ] is the hyperedge set, w is the hyperedge weight set; the relationship between vertices and hyperedges in the hypergraph model is represented by the association matrix H: Where N k (x i ) is the sample x i The k nearest neighbor samples, the metric function between samples is selected as: S(i,j)=exp(||x i -x j || 2 / 2t 2 ) In the formula is the heat kernel parameter; the weight of each hyperedge w(e j ) is defined as: In a hypergraph, the vertex degree d(v i ) represents the hyperedge weight of each vertex, and the hyperedge degree d(e j ) represents the number of vertices contained in the hyperedge, and the calculation formulas are: d(v i )=∑ j＝1 w(e j )H(v i ,e j ),v i ∈e j d(e j )=∑ i＝1 H(v i ,e j ),v i ∈e j The vertex degree, hyperedge degree and hyperedge weight matrix are represented by diagonal matrices: Dv = diag(d(v1),…,d(v N )), De=diag(d(e1),…,d(e N )),Ω=diag(w(e1),…,w(e N )).

3. According to claim 2, the ICBL-based intelligent fault diagnosis method for the gearbox of the cutting part of the coal mining machine is characterized in that: The cascaded hypergraph convolution nodes in step 3 are constructed using a random hypergraph convolution mechanism, specifically: The Laplacian operator Δ of the hypergraph model G is a semidefinite matrix, and its eigenvalue decomposition Δ=UΛU T , where U=[u1,u2,…,u N ] is the eigenvector matrix, Λ=diag(λ1,λ2,…λ N ) is the diagonalized eigenvalue matrix; for the signal on the hypergraph model G Its Fourier transform is defined as The eigenvector U is the Fourier basis in the hypergraph domain, and the corresponding eigenvalue is the frequency; the frequency domain convolution of the signal x and the filter f is expressed as: f*x=U((U T f)⊙(U T x))=Uf(Λ)U T x Where * is the convolution operator, ⊙ is the element-wise product operator, and f(Λ) is the eigenvalue function of Δ; However, due to its large amount of calculation, the computational complexity is O(V 2 ); using multi-order truncated Chebyshev polynomials, the formula is simplified to: Where Θ is the hypergraph convolution filter parameter; According to the above formula, the first group of hypergraph convolution nodes of the improved cascade width learning model is expressed as: The improved cascade width learning model uses a random generation method to determine the hypergraph convolution filter parameter Θ1 in the above formula, and requires the generated Θ1 to satisfy orthogonalization; this random hypergraph convolution mechanism can not only retain the structural characteristics of the original cascade width learning model to the greatest extent, but also quickly and efficiently embed hypergraph prior knowledge into the model; Similarly, the improved cascade width learning model also cascades the hypergraph convolution nodes layer by layer; therefore, the nth group of hypergraph convolution nodes is defined as: Next, all the hypergraph convolution nodes are integrated to obtain the total hypergraph convolution node matrix F n =[F1,F2,…,F n ], and obtain m groups of feature nodes Z through nonlinear activation function m =[Z1,Z2,…,Z m ]; where the j-th group of enhanced nodes is expressed as: Where W hj and β hj are the jth group of randomly generated enhancement mapping weight matrices and biases respectively; By integrating all cascaded hypergraph convolution nodes and enhancement nodes, the features learned by the improved cascade width learning model are expressed as: A=[F1,F2,…,F n |Z1,Z2,…,Z m ]=[F n ,Z m ]。 4. According to claim 1, the ICBL-based intelligent fault diagnosis method for the gearbox of the cutting part of the coal mining machine is characterized in that: The step three constructs a class-specific weight allocation strategy and embeds it into the objective function of the improved cascade width learning model, specifically: The class-specific weight assignment strategy assigns different weights to each class of samples; for the c-th class sample, its weight is defined as: Where N c represents the number of samples in the cth category, N ave is the mean of the number of samples in each category; τ∈(0,1] is the weight scaling factor, which is used to fine-tune the weights of each category. From the above formula, χ c It is inversely proportional to the number of samples of each category, so the minority samples can get greater weight in the modeling process, increasing their attention; By integrating the above formula, the objective function of the improved cascade width learning model is: Where ρ is the regularization parameter, ξ i For sample x i The training error, Y i For sample x i The true label of , W is the output weight matrix of the improved cascade width learning model, is the diagonalized weight matrix, R ii is the i-th element on its diagonal; According to the Lagrange multiplier method, the objective function of the improved cascade width learning model is rewritten as: Where α i is the Lagrange multiplier, respectively for (W,ξ i ,α i ) Take the partial derivative and make it equal to 0, and get the following formula: The output weight matrix of the improved cascade width learning model can be directly obtained through the ridge regression algorithm without complicated gradient optimization. From the above formula, the output weight matrix of the improved cascade width learning model is obtained as: W=(A T RA+ρI) -1 A T RY.

Images