Fault diagnosis method and system based on multi-scale adaptive hypergraph convolutional neural network

Abstract

The invention discloses a fault diagnosis method and system based on a multi-scale adaptive hypergraph convolutional neural network, and belongs to the technical field of fault diagnosis. The method comprises the following steps: acquiring original vibration data of mechanical equipment, and carrying out data preprocessing; constructing a fault diagnosis model based on the multi-scale adaptive hypergraph convolutional neural network; and inputting the vibration data after data preprocessing into the constructed fault diagnosis model, and outputting a fault diagnosis result. According to the method, a vibration signal is decomposed into different frequency bands, features under various scales are extracted respectively, prior knowledge is fused, local topological graphs are constructed through various weighting measures, and a plurality of local topological graphs jointly form a spatial hypergraph to serve as input of a model; a dual-channel graph convolutional network is introduced to extract features, weights of different channels are adaptively distributed through an attention mechanism, and deep fusion of the features is realized. Experiments on two reference data sets show that the method can effectively improve the fault diagnosis accuracy, and has good stability and generalization.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and more specifically to a fault diagnosis method and system based on a multi-scale adaptive hypergraph convolutional neural network. Background Art

[0002] With the rapid development of information technology, artificial intelligence has been widely used in various fields. Among them, the fault diagnosis of mechanical equipment mainly relies on intelligent methods based on deep learning. However, in actual scenarios, the scale of collected fault data is usually very limited, resulting in poor generalization ability of intelligent diagnosis models in the case of insufficient data. In addition, due to the lack of effective correlation between cross-scale features, the complementarity of multi-scale information is not fully utilized, resulting in poor diagnostic results. At the same time, fault data usually has characteristics such as high dimension, high noise and redundancy. It is difficult for the model to effectively extract discriminative features, and it fails to make full use of prior knowledge in the field, which affects the performance of deep learning models in practical applications. On the other hand, with the gradual increase in the number of layers of deep neural networks, the model is prone to gradient vanishing or gradient exploding problems, the training process becomes more difficult, and the training accuracy drops sharply, further exacerbating the phenomenon of insufficient generalization ability.

[0003] Inspired by the definition of convolutional filtering of graph signals in graph signal processing, graph convolutional networks (GCNs) can obtain the relationship between data by introducing association graphs, thereby generating graph models, modeling data structurally, and improving the generalization performance of the model. Compared with traditional CNNs, GCNs achieve information updates by aggregating features in graph structure data, and effectively extract high-dimensional redundant data. Based on the advantages of generalized networks in obtaining the relationship between signals and supplementing missing information by establishing association graphs, researchers have applied them to the field of fault diagnosis. For example, in terms of multi-scale feature fusion, Li et al. proposed a multi-receptive field GCN, which can learn information from different receptive fields and can indeed improve the model's perception of multi-scale features. However, this also increases the complexity of the model. As the number of model parameters increases, the gradient update during training may show instability at different scales, making model training more difficult and the convergence speed slower. Zhang et al. proposed a multi-scale aggregation graph neural network based on feature similarity, which enhances the feature expression ability by constructing the original feature similarity matrix and using feature extractors to achieve multi-scale feature propagation. However, due to the complexity of graph structure data, graph convolutional networks (GCN) may have difficulty in effectively obtaining global information when capturing long-distance dependencies, and the model training process may be unstable. Sun et al. proposed a multi-scale multi-channel residual network (RESNET) clustering graph convolutional neural network (MR-MCGCN), and performed feature fusion through a splicing method, thereby enhancing the ability to identify fault types. However, in the feature fusion process, it is impossible to effectively distinguish important features from unimportant features, resulting in the introduction of redundant information and low fusion efficiency, increasing the complexity of the model and the risk of overfitting. Secondly, by introducing prior knowledge to make up for data deficiency and improve model accuracy, Chen et al. combined the measurement values ​​with prior knowledge and introduced weight coefficients for model optimization, thereby converting the pre-diagnosis results into a correlation graph, and finally inputting the graph and measurement values ​​into the GCN model, solving the problems of difficulty in fusion of multi-source heterogeneous data, insufficient feature correlation, and limited model processing ability for complex structure data. However, this method failed to reasonably balance the feature weight distribution between key areas and low-quality areas when introducing prior knowledge, resulting in insufficient recognition of important features and low efficiency of feature fusion, which ultimately affected the accuracy of fault diagnosis. Hu et al. proposed a residual shrinkage prototype network based on prior knowledge, which extracted prior features from labeled source data through a denoising residual shrinkage network, enhancing the fault diagnosis feature representation in small sample scenarios. However, this method lacks systematic exploration of the effective fusion of multiple types of prior knowledge, and the extracted prior knowledge is not adaptable enough under dynamic conditions, resulting in a decrease in the model's diagnostic performance.Liao et al. used the graph convolutional layer with strong feature extraction capability of GCN as a classifier and found that the complex nonlinear relationship between dissolved gas and fault type was effective for transformer fault diagnosis, which improved the model's recognition accuracy for transformer faults. However, due to the complex nonlinear relationship and the deepening of the network layers, the convergence speed of the model slowed down and its generalization ability weakened.

[0004] Therefore, how to efficiently perform fault diagnosis of mechanical equipment and improve the accuracy, stability and generalization of fault diagnosis is an urgent problem that technical personnel in this field need to solve. Summary of the invention

[0005] In view of this, the present invention provides a fault diagnosis method and system based on a multi-scale adaptive hypergraph convolutional neural network that solves at least some of the above-mentioned technical problems. The method and system can efficiently, quickly and accurately diagnose mechanical equipment faults and have good stability and generalization.

[0006] To achieve the above object, the technical solution adopted by the present invention is:

[0007] In a first aspect, an embodiment of the present invention provides a fault diagnosis method based on a multi-scale adaptive hypergraph convolutional neural network, the method comprising:

[0008] Collect the original vibration data of mechanical equipment and perform data preprocessing;

[0009] Construct a fault diagnosis model based on a multi-scale adaptive hypergraph convolutional neural network;

[0010] The vibration data after data preprocessing is input into the constructed fault diagnosis model, and the fault diagnosis result is output.

[0011] Furthermore, the data preprocessing process includes: normalizing the vibration data using a maximum and minimum normalization method, and grouping the normalized data.

[0012] Furthermore, the process of constructing a fault diagnosis model based on a multi-scale adaptive hypergraph convolutional neural network includes:

[0013] Construct a cross-scale feature map based on CNN;

[0014] An unsupervised deep convolutional autoencoder network is used to extract general features, and general features are combined with prior knowledge features to construct special features.

[0015] A local topological graph is constructed using multiple weighted metrics, and multiple local topological graphs together form a spatial hypergraph;

[0016] A dual-channel graph convolutional neural network is used to extract features, and the weights of different channels are adaptively allocated through the attention mechanism.

[0017] Furthermore, the unsupervised deep convolutional autoencoder network includes: an encoder and a decoder, wherein: the encoder includes four convolutional layers and four pooling layers, and the decoder includes four convolutional layers and four upsampling layers.

[0018] Furthermore, the prior knowledge features include: time domain features and frequency domain features.

[0019] Furthermore, the local topology graph is constructed by using multiple weighted metrics, and multiple local topology graphs together form a spatial hypergraph. The specific process includes:

[0020] The weighted measurement of Gaussian kernel function, Euclidean distance, Minkowski distance and KNN algorithm are used to construct spatially independent local graphs, and the spatially independent local graphs are combined into a topological space hypergraph, where the weighted measurement formulas of Gaussian kernel function, Euclidean distance and Minkowski distance are:

[0021]

[0022] Where x i ,x j are two sample inputs, and ζ is the width of the Gaussian kernel function.

[0023] Furthermore, a batch normalization layer is introduced between every two convolutional layers of the graph convolutional neural network, and the original ReLU activation function is replaced with LeakyReLU.

[0024] In a second aspect, an embodiment of the present invention further provides a fault diagnosis system based on a multi-scale adaptive hypergraph convolutional neural network, which applies the above-mentioned fault diagnosis method based on a multi-scale adaptive hypergraph convolutional neural network to perform fault diagnosis of mechanical equipment, and the system includes:

[0025] Data acquisition module, used to collect raw vibration data of mechanical equipment and perform data preprocessing;

[0026] A fault diagnosis model building module is used to build a fault diagnosis model based on a multi-scale adaptive hypergraph convolutional neural network;

[0027] The fault diagnosis module is used to input the vibration data after data preprocessing into the constructed fault diagnosis model and output the fault diagnosis result.

[0028] Compared with the prior art, the present invention has at least the following beneficial effects:

[0029] 1. The present invention provides a fault diagnosis method and system based on a multi-scale adaptive hypergraph convolutional neural network, which can efficiently, quickly and accurately diagnose mechanical equipment faults.

[0030] 2. The present invention extracts features at each scale and integrates prior knowledge, constructs a local topology map with multiple weighted metrics, and uses multiple local topology maps to form a spatial hypergraph as the input of the model. A dual-channel graph convolutional network is also introduced to extract features, and the weights of different channels are adaptively allocated through the attention mechanism to achieve deep fusion of features. Experiments on two benchmark data sets show that the present invention can effectively improve the accuracy of fault diagnosis and has good stability and generalization.

[0031] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0032] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0034] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0035] Figure 1 A schematic flow chart of a fault diagnosis method based on a multi-scale adaptive hypergraph convolutional neural network provided in an embodiment of the present invention.

[0036] Figure 2 A schematic diagram of a GCN network provided in an embodiment of the present invention.

[0037] Figure 3 A schematic diagram of a process for constructing a cross-scale feature map provided in an embodiment of the present invention.

[0038] Figure 4 Schematic diagram of the DCAE network feature extraction process provided by an embodiment of the present invention.

[0039] Figure 5 A schematic diagram of constructing a hypergraph provided in an embodiment of the present invention.

[0040] Figure 6 A schematic diagram of an SGCN model provided in an embodiment of the present invention.

[0041] Figure 7 A schematic diagram of an adaptive ASGCN model provided in an embodiment of the present invention.

[0042] Figure 8 A schematic diagram of the MS-ASGCN fault diagnosis process provided by an embodiment of the present invention.

[0043] Fig. 9 Schematic diagram of different gear faults provided in an embodiment of the present invention.

[0044] Fig.10 A schematic diagram of the impact of different metrics on model accuracy provided by an embodiment of the present invention.

[0045] Fig.11 A schematic diagram of diagnostic results before and after introducing prior knowledge and model improvement provided in an embodiment of the present invention.

[0046] Fig.12 Schematic diagram of the impact of different scale features on the model in the dataset 1 provided in an embodiment of the present invention.

[0047] Fig.13 Schematic diagram of the impact of different scale features on the model in dataset 2 provided in an embodiment of the present invention.

[0048] Fig.14 A schematic diagram of the accuracy of different models provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. In addition, various serial numbers are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0050] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] Combination Figure 1 and Figure 8 As shown, the present invention provides a fault diagnosis method based on a multi-scale adaptive hypergraph convolutional neural network, which mainly includes:

[0052] ①. Collect the original vibration data of mechanical equipment and perform data preprocessing;

[0053] ②. Build a fault diagnosis model based on the Multi-Scale-Attention Super Graph Convolutional Network (MS-ASGCN);

[0054] ③. Input the vibration data after data preprocessing into the constructed fault diagnosis model and output the fault diagnosis results.

[0055] The following describes in detail the relevant theories and specific implementation methods of the method of the present invention:

[0056] 1. Related theories:

[0057] 1.1 Graph Neural Network:

[0058] Graph Neural Network (GNN) is a framework that uses deep learning to directly learn graph structure data. By formulating strategies on the nodes and edges in the graph, the graph structure data is converted into a normalized representation and input into the neural network for training. The graph structure can be simply represented as: G = (W, A), where W = {w 1 ,w 2 …w n}∈R n×d is a set of nodes, A∈R n×n is the adjacency matrix.

[0059] Compared with traditional neural networks, the key difference of graph neural networks (GNNs) is that they can perform convolution operations based on node connection relationships in graph structures in non-Euclidean geometric spaces, thereby solving the problem that traditional neural networks are difficult to process graph structure data, especially feature extraction and information propagation in irregular data such as complex relationship networks, social networks, and molecular structures. The features of each node in the graph convolution layer are updated by aggregating the features of its neighboring nodes, and the updated node embedding representation is generated by aggregating and transforming the neighborhood features, thereby capturing local and global patterns in the graph structure. The feature update formula for each node can usually be expressed as:

[0060]

[0061] in, represents the feature vector of node v in the k+1th layer, σ is a nonlinear activation function, W (k) represents the trainable weight matrix of the kth layer, AGG is the aggregation function used to aggregate the node features in the neighbor node set N(v) of node v.

[0062] 1.2 Graph Convolutional Neural Network:

[0063] Traditional GCN mainly includes two types: spectral domain convolution and spatial domain convolution. Spectral domain convolution is mainly based on a solid graph theory foundation. Therefore, graph convolutional networks based on spectral domain have gradually become a research hotspot. Due to the robustness of GCN models in feature representation and their ability to mine the relationship between raw data nodes, they have been widely used in the field of fault diagnosis. Figure 2 As shown in Figure 1, the core of the model is that it contains multiple independent convolutional layers, and the layers are connected by activation functions. In each convolutional layer, the spectral graph convolution operation is used to adaptively extract features from the input high-dimensional feature data. Spectral graph convolution is used in each layer to adaptively extract features from the input high-dimensional feature data, thereby gradually generating more discriminative node representations. However, the current mainstream methods still have shortcomings in modeling the association between cross-scale features. For example, a weighted ε-recursive network is used to construct a recursive graph to analyze the recursive relationship between samples. However, the recursive graph is not sensitive enough to signals, which limits the model's ability to effectively model multi-scale features. Using network parameters to construct a graph structure can easily lead to poor model generalization and limited scalability. A heterogeneous graph convolutional network (HGCN) is used, but HGCN faces the problem of high computational complexity when processing high-dimensional parameters, resulting in slow inference speed. A single-channel visible graph convolutional neural network is used, but the horizontal visible graph it constructs is difficult to fully capture global features, resulting in the loss of potential correlation information, affecting the integrity of feature representation and the representation ability of the model. To address the above problems, the present invention proposes the MS-ASGCN model, establishes a multi-channel hypergraph structure to comprehensively capture global features, and improves the integrity of feature representation and the representation ability of the model through an adaptive fusion mechanism.

[0064] 2. Construction of MS-ASGCN diagnostic model:

[0065] 2.1 Construction of cross-scale feature maps based on CNN:

[0066] Modeling the lack of correlation between features of different scales leads to the problem of difficulty in cross-scale feature fusion. In the embodiment of the present invention, a multi-scale feature extraction method is used to achieve efficient feature expression. By optimizing the feature combination strategy from a multi-scale perspective and selecting appropriate scales for fusion, the model's expressiveness in capturing multi-scale information is enhanced. In the initial stage, the original vibration signal is preprocessed. For the original data sequence N with a vibration signal length of M, the maximum and minimum normalization method is used to normalize the vibration data. The process can be expressed as N nol =normalize(N), where N nol is the normalized time series, and normalize(·) is the maximum and minimum normalization method.

[0067] In order to ensure that the data size at each scale is the same, the normalized data is grouped and represented as follows:

[0068]

[0069] Among them, x, y, u, z are the number of samples in each group.

[0070] In order to construct a cross-scale graph model, features are extracted from single-scale graph models and graphs are combined. First, the dimension of each input vector is expressed as:

[0071] D i =d / n (3)

[0072] Among them, D i is the dimension number of the input data, i=1,2,3…n, and n is the number of groups.

[0073] Each group then undergoes a one-dimensional convolution operation and passes through a one-dimensional convolutional network. After convolution, the dimension of each output vector is as follows:

[0074] D i ′=D i -(a-1) (4)

[0075] Among them, D i ′ is the dimension of the vector and a is the size of the convolution operation.

[0076] Finally, the total convolution dimension after convolution is as follows:

[0077] D=D′-n(a-1) (5)

[0078] Among them, D is the total dimension after convolution.

[0079] In order to ensure that the feature output dimension D is the same at each scale, the relationship between n and a is compared and the convolution kernel size in each one-dimensional convolution is adjusted. In this way, the output dimension can be kept consistent when extracting features at different scales, and a cross-scale graph can be constructed by splicing and synthesis, such as Figure 3 As shown. K It is the feature extracted by convolution operation. For each scale K, the feature extraction process can be expressed as: F k =Conv(X k ), the feature extraction process is:

[0080] F=(F k ) 2 =CNN(X K ,θ X ) (6)

[0081] Through feature hierarchical modeling and multi-scale information fusion, efficient feature representation of signals can be achieved. This method can enhance the model's perception of cross-scale information and improve the accuracy of capturing potential laws and fault diagnosis in complex systems.

[0082] 2.2 Feature map construction:

[0083] In the embodiment of the present invention, in order to solve the problem of insufficient correlation of local features of vibration signals, a priori structural information is introduced to merge with general features to generate more discernible feature representations and enhance the characterization ability of features. In the method design, it is verified based on effective indicators, and the priori knowledge information is shown in Table 1:

[0084] Table 1 Prior knowledge information

[0085]

[0086] Select some indicators as the prior knowledge feature set {p 1 ,p 2 ,..p n}.

[0087]

[0088] In the formula, p is the prior knowledge feature, mean(p) is the prior knowledge average, and std(p) is the prior knowledge standard value.

[0089] Since the unsupervised deep convolutional autoencoder (DCAE) network is an unsupervised learning method, it does not rely on a large amount of labeled data and can learn useful features even with small sample data. The present invention uses the unsupervised deep convolutional autoencoder (DCAE) network for general feature extraction, such as Figure 4 The DCAE network consists of two parts: the encoder and the decoder. The encoder consists of four convolutional layers and four pooling layers, and the decoder consists of four convolutional layers and four upsampling layers. The general features of the vibration signal are extracted through the unsupervised DCAE network {p 1 ′,p 2 ′,..p n ′}, and normalized by formula (7). In order to make up for the fact that the general features may lack the fault information contained in the prior features, the present invention combines the prior knowledge features with the general features to construct the final feature set {F 1 ,F 2 ,F 3 …F m}.

[0090] 2.3 Construction of Adaptive Hypergraph Convolutional Neural Network (ASGCN):

[0091] In the embodiment of the present invention, in order to further improve the comprehensiveness of multi-scale feature extraction, a dual-channel weight sharing GCN network is used to realize multi-channel modeling of features. Through the fusion of the attention mechanism, features with high importance scores are adaptively selected to achieve more effective feature screening and enhancement, thereby improving the feature expression ability and decision accuracy of the model in the hypergraph input scenario.

[0092] 2.3.1 Hypergraph Construction:

[0093] In order to overcome the problem that traditional graph structures are difficult to capture high-order node associations, a supergraph is used to connect multiple nodes simultaneously through hyperedges, effectively representing the complex relationships between multiple nodes, avoiding information loss or oversimplification, and achieving more comprehensive data expression. The input SuperGraph is represented as G super = {V s ,E s ,A}-, where V s =H is the node set of SuperGraph, H is the total number of samples; E s is an edge set, connecting signals of the same type in the edge set in the form of an undirected graph. m×m is the adjacency matrix between any two points. The value of the adjacency matrix is ​​1 or 0, and it is defined as:

[0094]

[0095] Furthermore, the multi-scale features and special features are fused as nodes in the sample graph data, and the hyperedge is constructed using the KNN algorithm. The Euclidean distance between nodes is calculated, and the k nodes closest to the central node are selected as the neighbors of the central node. An edge can be formed when the relationship between every two nodes is close. Thus, each local k-nearest neighbor graph (K-NNG) is obtained. The neighbors of each node are obtained by the KNN algorithm, as shown in formula (9):

[0096] N e =KNN(k i ,x i ) (9)

[0097] Among them, N e Represents x i neighbor nodes.

[0098] Since a single metric cannot guarantee an accurate description of the underlying geometric structure of the data, and the graph constructed by a single metric average is likely to contain more noise, thus affecting the performance of the model. The present invention uses three weighted metrics, Gaussian kernel function, Euclidean distance and Minkowski distance, to calculate the feature node x i and x j The distance between.

[0099]

[0100] The adjacency matrix is ​​calculated by formula (11). The SuperGraph is formed by four spatially independent local graphs, such as Figure 5 As shown in the figure, the color of the labeled node indicates the fault type. Nodes of the same color are connected by hyperedges, and there is no connection between nodes of different colors. For two unlabeled samples, they belong to the same fault type because they are data points collected in the same time period. When GCN aggregates adjacent features, information is propagated in SuperGraph through hyperedges.

[0101] 2.3.2 Adaptive Hypergraph Convolutional Neural Network:

[0102] (1) Hypergraph Convolutional Neural Network:

[0103] Kip et al. first proposed a GCN, the principle of which is to consider the convolution operation f*x, where f and x are the feature vectors of the convolution kernel and the node respectively. The frequency domain form of the above convolution operation is expressed by Fourier transform as:

[0104] f*x=U((U T f)⊙(U T x))=Ug w U T x (12)

[0105] Among them, U and Λ are the eigenvector and eigenvalue vector of the Laplace matrix L respectively. After the eigenvalue decomposition of L, we can get L = UΛU -1 ; U T x represents the Fourier transform of x. However, due to the complexity of the training parameters, Kipf and Welling et al. used Chebyshev polynomials Then, by setting the order of the polynomial reasonably (K=1), the simplified form of the convolution operation is derived as follows;

[0106]

[0107] Where w = w 0 =-w 1 , will I N +D -0.5 AD -0.5 Expressed as The forward propagation formula of Spectral GCN can be written as follows:

[0108]

[0109] Among them, W is the training parameter matrix, X is the input data, and σ is the activation function.

[0110] In SGCN, hypergraphs are often used to represent complex structures and relationships. s 、Superedge set E s , neighbor matrix A. The normalized incidence matrix H is obtained by the vertex degree matrix D v and the hyperedge degree matrix D e After normalization, it is expressed as:

[0111]

[0112] The operation of the SGCN layer is similar to that of the traditional GCN, but is based on a hypergraph structure. Given the input feature X and the weight matrix W, the output Z of the SGCN layer can be expressed as:

[0113] Z=σ(H×W) (16)

[0114] The network model constructed by the above method has the problem of gradient vanishing and model overfitting during the gradient propagation process as the number of network layers increases. In order to improve the fluidity of the gradient, enhance the model training efficiency and increase the regularization effect, the present invention introduces a batch normalization layer (BN layer) between every two convolutional layers of the original graph convolutional neural network (GCN), and replaces the original ReLU activation function with Leaky ReLU to alleviate the gradient vanishing phenomenon and improve the nonlinear expression ability, such as Figure 6 In addition, in order to ensure that each type of data feature and local and global features can be fully mined, the attention mechanism is used to dynamically calculate the importance scores of different node features, realize adaptive weighting in the feature fusion process, and improve the ability to capture multi-scale information, such as Figure 7 shown.

[0115] (2) Adaptive graph information fusion:

[0116] Since the input data of the hypergraph has high spatial complexity, a single-channel GCN may not be able to fully capture local and global features during feature extraction, resulting in the loss of diagnostic information. Therefore, the present invention adopts a dual-channel hypergraph input method, in which the two channels share parameters but extract features independently to ensure that different types of features are fully mined. Finally, the feature information of the two channels is fused through the attention mechanism to generate the output

[20] .

[0117] In order to extract specific information in the feature space, the graph adjacency matrix A and the feature mapping matrix X are respectively input into two specific graph convolutional networks to obtain two specific embedding outputs Z T and Z F , expressed as:

[0118]

[0119] Among them, SGCN1 (·) and SGCN 2 (·) are two different convolution operations.

[0120] The hypergraph data features are calculated through two channels to obtain a score representation as shown in formula (18):

[0121]

[0122] The weights of the two channels are expressed as formula (19):

[0123]

[0124] After fusion through the attention mechanism, the output is expressed as formula (20):

[0125] Z fusion =a 1 ·Z T +a 2 ·Z F (20)

[0126] for Figure 7 The two channels have shared weight parameters and the output embedding is Z T and Z F , First, the output embedding matrices are normalized using L2 regularization and used to calculate the similarity matrix between nodes and Used to calculate the similarity matrix S between nodes T and S F .

[0127]

[0128] If the two output embedding similarity matrices are similar, then the following constraints apply:

[0129] L c =||S T -S F || (22)

[0130] 3.MS-ASGCN fault diagnosis algorithm process:

[0131] In the embodiment of the present invention, the diagnosis steps are mainly divided into data preprocessing, hypergraph construction, ASGCN model training and fault diagnosis. The specific diagnosis process is as follows: Figure 8 As shown:

[0132] (1) Data preprocessing: The original vibration signal is sampled and normalized. Three feature extractions are performed in different receptive fields to construct a multi-scale feature map. The general features are extracted through the DCAE model and then fused with the prior knowledge features to construct a feature map.

[0133] (2) Hypergraph construction: The hypergraph input is constructed by fusing the feature map and the multi-scale feature map, using the results calculated by multi-metric weighting, and using the KNN algorithm to determine the similarity between samples.

[0134] (3) ASGCN model training: The extracted features and the adjacency matrix of the hypergraph structure are used as input. The hypergraph convolution operation is used to dynamically update the message passing node status. The model introduces an adaptive attention mechanism to perform weighted fusion of multi-channel features and finally outputs the fused global feature representation.

[0135] (4) Fault diagnosis: Input the test set data into the trained ASGCN model, output the category probability of each sample, and obtain the category with the highest probability as the final diagnosis result.

[0136] 4. Experimental verification:

[0137] The correctness of the model is verified using the planetary gearbox WT-Planetary gearbox dataset and the MCC5-THU gearbox dataset collected from the wind turbine transmission test bench.

[0138] 4.1 Dataset 1: Planetary gearbox data collected from a wind turbine transmission test bench:

[0139] 4.1.1 Dataset Introduction:

[0140] The test bench consists of a motor, a planetary gearbox, a fixed-axis gearbox, and a load device. The vibration data is collected by a Sinocera CA-YD-1181 accelerometer, and the sampling rate of all channels is set to 48kHz. The five working states of the planetary gear are as follows: Fig. 9 shown.

[0141] This data set simultaneously collects vibration signals in the x and y directions and encoder data of the planetary gearbox input shaft. By setting the sample length to 2000 points and not performing overlapping operations, eight speed conditions (20Hz-55Hz) are considered. In this experiment, the vibration signal data in the x direction with a speed of 45Hz is selected. As shown in Table 2, the length of each sample is 1024 sampling points, the total number of samples for each state is 300, and the data set contains 4 gear fault states and 1 healthy state.

[0142] Table 2 WT-Planetary gearbox fault type information table

[0143]

[0144] 4.1.2 Diagnostic results and analysis:

[0145] To avoid the accidental nature of the experimental results, all methods were repeated 6 times. The network structure and hyperparameter settings of all GNNs are the same. The main hyperparameter settings of the model are shown in Table 3:

[0146] Table 3 Parameter settings of GNN network

[0147]

[0148] (1) The impact of different metric distance compositions on the ASGCN model:

[0149] In order to demonstrate the superiority of the weighted distance proposed in this paper, the method of the present invention is compared with three weighted metrics: Gaussian kernel function, Euclidean distance and Minkowski distance. Fig.10 As shown in the figure, it can be seen that the performance under different distance measurement methods of single or dual channels is different. In order to make the model fully learn features and overcome the errors caused by the limitations of GCN itself, 200 experimental samples were selected. First, the effective performance of the model in this paper was verified only under the auxiliary input of general features combined with prior knowledge, as shown in Fig.11 As shown in the figure, the GCN model that combines only general features and prior knowledge and is optimized (BN layer and Leaky Relu are added to each GCN layer) is better than the model without prior knowledge in terms of diagnostic effect.

[0150] In the case of weighted average distance, the ASGCN model performs best in fault diagnosis in dual channels, with an accuracy of 97.12%. In the case of Gaussian Kernel Function, Euclidean Distance, and Minkowski Distance, the accuracy reaches 95.44%, 95.67%, and 95.39%, respectively. Whether in single channel or dual channel, the quality of single method composition is worse than that of weighted average composition, indicating that the quality of constructed graph also affects the performance of the model.

[0151] The effectiveness of the proposed method is verified by constructing a hypergraph input model by fusing multi-scale features: The method of extracting features at a single scale often ignores the potential connection between signals at different scales, which will lead to weak generalization ability and poor stability of the model. However, the multi-scale feature fusion method can significantly improve this problem, making the model show higher stability and stronger generalization ability under different data sets and experimental conditions. The hypergraph input model is constructed by fusing multi-scale features with general features to verify the effectiveness of the proposed method. Fig.12 and Fig.13As shown in the figure, the DCNN, DGCN, DGAT, and MS-GCN models similar to the model in this paper were selected for comparison in the two experiments. The experimental results show that the GCN model optimized with general features and prior knowledge is significantly better than the model without prior knowledge in terms of diagnosis effect, and the accuracy is improved by about 2%. In addition, by integrating multi-scale features with general features, the limitations of single-scale feature extraction are overcome, effectively improving the performance and reliability of the model.

[0152] (2) Experimental comparison of different models:

[0153] In order to verify the superiority of the model, under the same conditions, the parameter settings of GCN, GAT, GraphSage, TAGCN, AM-GCN, and MRF-GCN are compared as shown in Table 5. The diagnostic accuracy of each method is shown in Table 4. The maximum accuracy (Max-acc) and minimum accuracy (Min-acc) are considered, and the average accuracy (Avg-acc) is used as the overall classification accuracy evaluation indicator.

[0154] Table 4 Comparative experiments of different models

[0155] Models Max-acc(%) Min-acc(%) Avg-acc(%) Train time GCN 93.24％ 92.16％ 92.56％ 1min10s GAT 93.13％ 92.78％ 93.47％ 1min09s GraphSage 94.36％ 93.01％ 93.83％ 1min19s TAGCN 94.99％ 94.02％ 94.57％ 2min11s AM-GCN 95.26％ 94.11％ 94.83％ 2min26s MRF-GCN 96.09％ 95.21％ 95.54％ 2min34s MS-ASGCN 98.06％ 96.93％ 97.63％ 5min39s

[0156] Table 5 Model parameter settings

[0157] Number of layers Structure and parameters Shape Layer1 H0=Input(shape=X.shape[1])) 1×9 Layer2 H1=GraphConv(filter=16,ReLU)([H0]+Graph) 1×16 Layer3 H2=GraphConv(filters=8,ReLU)([H1,Graph]) 1×8 Layer3 Y=Dense(unit=7,Softmax)(H2) 1×7

[0158] There are some differences in the accuracy of different models in fault diagnosis tasks, and many models show high classification accuracy. However, MS-ASGCN performed best in terms of the average accuracy of fault diagnosis, reaching 97.63%, which is 5.47% higher than the original GCN model. This shows that MS-ASGCN has stronger capabilities in feature extraction, feature enhancement, and feature fusion, and can more accurately identify fault modes under different states.

[0159] 4.2 Dataset 2: MCC5-THU gearbox data:

[0160] 4.2.1 Dataset Introduction:

[0161] The experimental bench includes a 2.2kW three-phase asynchronous motor, a torque sensor, a two-stage parallel gearbox, a magnetic powder brake as a torque generator, and a measurement and control system. The test device is equipped with two triaxial vibration accelerometers (model TES 001V, sensitivity: 100mv / g) to measure the triaxial vibration of the motor and the gearbox intermediate shaft along the x-axis, y-axis and z-axis, with a sampling frequency of 12.8kHz. The data set collects data under two conditions: torque cycle and speed cycle. The equipment is at a speed of 1000r / pm and 2000r / pm. The load is 10Nm and 20Nm, and the sampling frequency is 12.8kHz. There are 6 types of faults in this data set. Due to the large number of data types (such as gear wear degree, speed and load conditions), the faults caused by the speed cycle used in this paper are shown in Table 6.

[0162] Table 6 Gearbox fault type information table

[0163]

[0164] Four different types of gearbox faults and two types of broken teeth and bearing inner and outer ring composite faults were collected along the x-axis direction under the speed cycle with a load and a speed of 10Nm-1000r / pm. The time domain waveform is shown in Figure 14. The sampling point is 2048 as a sample, and each fault state has 120 samples, with a total of 720 samples. The training set and the test set are divided according to the ratio of 8:2.

[0165] 4.2.2 Diagnostic results and analysis:

[0166] In order to further verify the superiority of the model in this paper, GCN, GAT, GraphSage, TAGCN, AM-GCN, MRF-GCN, etc. are used as comparison models in this experiment. All experimental parameters are consistent with Section 4.1.2, and the maximum accuracy (Max-acc), minimum accuracy (Min-acc) and average accuracy (Avg-acc) are used as evaluation indicators of the overall classification accuracy. Aiming at the actual working conditions of industrial gearboxes in complex noise environments, Gaussian white noise (-3dB to 2dB) with different signal-to-noise ratios (SNR) is added to the vibration signal. The calculation formula is shown in formula (23). Each group of experiments was repeated 7 times, and the average diagnostic accuracy was taken as the final evaluation indicator. The experimental results are shown in Fig.14 As shown, the noise resistance and robustness of the model of the present invention are fully verified.

[0167]

[0168] Among them, P signal Represents the original power signal, P noise represents the Gaussian noise power.

[0169] Through the description of the above implementation, those skilled in the art can know that the present invention proposes a fault diagnosis method based on a multi-scale adaptive hypergraph convolutional neural network, which constructs a local graph by fusing multi-scale features with prior knowledge, using three metric weights and the KNN algorithm, and four local graphs constitute a topological space hypergraph. A dual-channel adaptive graph convolutional network is used to adaptively select high-importance features under the guidance of the attention mechanism. The model can adaptively learn the weight relationship between different features, perform weighted fusion on features with high importance scores, and achieve feature screening and enhancement. The feature representation ability of the model under the hypergraph structure is effectively improved, and the recognition and classification performance of complex data patterns is enhanced. While retaining local features, this method deeply mines global information, effectively captures hidden patterns in vibration signals, solves the problem of insufficient potential relationship mining, and thus improves the generalization ability of the model. Experimental results show that the fault diagnosis accuracy of this method on two data sets is 96.5%, showing good performance and robustness.

[0170] Furthermore, the present invention also provides a fault diagnosis system based on a multi-scale adaptive hypergraph convolutional neural network, which is applied to a fault diagnosis method based on a multi-scale adaptive hypergraph convolutional neural network in the above embodiment to perform fault diagnosis of mechanical equipment, and the system includes:

[0171] Data acquisition module, used to collect raw vibration data of mechanical equipment and perform data preprocessing;

[0172] A fault diagnosis model building module is used to build a fault diagnosis model based on a multi-scale adaptive hypergraph convolutional neural network;

[0173] The fault diagnosis module is used to input the vibration data after data preprocessing into the constructed fault diagnosis model and output the fault diagnosis result.

[0174] The implementation principle and technical effects of the system provided in the embodiment of the present invention are the same as those of the aforementioned method embodiment. For the sake of brief description, for matters not mentioned in the system embodiment, reference can be made to the corresponding contents in the aforementioned method embodiment, which will not be repeated here.

[0175] In addition, an embodiment of the present invention also provides a storage medium on which one or more programs readable by a computing device are stored, and the one or more programs include instructions. When the instructions are executed by the computing device, the computing device executes a fault diagnosis method based on a multi-scale adaptive hypergraph convolutional neural network in the above-mentioned embodiment.

[0176] In the embodiment of the present invention, the storage medium may be, for example, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples of storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, and any suitable combination thereof.

[0177] Those skilled in the art should understand that the embodiments of the present invention can be provided as methods, systems or computer program products, etc. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0178] It should be noticed that the word "comprising" does not exclude the presence of steps or components not listed in a claim.The invention can be implemented by means of hardware comprising several distinct elements, and by means of a suitably programmed computer.

[0179] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0180] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A fault diagnosis method based on a multi-scale adaptive hypergraph convolutional neural network, characterized in that: The method includes: Collect the original vibration data of mechanical equipment and perform data preprocessing; Construct a fault diagnosis model based on a multi-scale adaptive hypergraph convolutional neural network; The vibration data after data preprocessing is input into the constructed fault diagnosis model, and the fault diagnosis result is output.

2. A fault diagnosis method based on a multi-scale adaptive hypergraph convolutional neural network according to claim 1, characterized in that: The data preprocessing process includes: normalizing the vibration data using the maximum and minimum normalization method, and grouping the normalized data.

3. A fault diagnosis method based on a multi-scale adaptive hypergraph convolutional neural network according to claim 1, characterized in that: The fault diagnosis model is constructed based on a multi-scale adaptive hypergraph convolutional neural network, and the construction process includes the following steps: Construct a cross-scale feature map based on CNN; An unsupervised deep convolutional autoencoder network is used to extract general features, and general features are combined with prior knowledge features to construct special features. A local topological graph is constructed using multiple weighted metrics, and multiple local topological graphs together form a spatial hypergraph; A dual-channel graph convolutional neural network is used to extract features, and the weights of different channels are adaptively allocated through the attention mechanism.

4. A fault diagnosis method based on a multi-scale adaptive hypergraph convolutional neural network according to claim 3, characterized in that: The unsupervised deep convolutional autoencoder network includes: an encoder and a decoder, wherein: the encoder includes four convolutional layers and four pooling layers, and the decoder includes four convolutional layers and four upsampling layers.

5. The fault diagnosis method based on a multi-scale adaptive hypergraph convolutional neural network according to claim 3 is characterized in that: The prior knowledge features include: time domain features and frequency domain features.

6. A fault diagnosis method based on a multi-scale adaptive hypergraph convolutional neural network according to claim 3, characterized in that: The local topology graph is constructed by using multiple weighted metrics, and multiple local topology graphs together form a spatial hypergraph. The specific process includes: The weighted measurement of Gaussian kernel function, Euclidean distance, Minkowski distance and KNN algorithm are used to construct spatially independent local graphs, and the spatially independent local graphs are combined into a topological space hypergraph, where the weighted measurement formulas of Gaussian kernel function, Euclidean distance and Minkowski distance are: In the formula, x i ,x j are two sample inputs, and ζ is the width of the Gaussian kernel function.

7. A fault diagnosis method based on a multi-scale adaptive hypergraph convolutional neural network according to claim 3, characterized in that: A batch normalization layer is introduced between every two convolutional layers of the graph convolutional neural network, and the original ReLU activation function is replaced with LeakyReLU.

8. A fault diagnosis system based on a multi-scale adaptive hypergraph convolutional neural network, characterized in that: A fault diagnosis method based on a multi-scale adaptive hypergraph convolutional neural network as described in any one of claims 1 to 7 is applied to perform fault diagnosis of mechanical equipment, the system comprising: Data acquisition module, used to collect raw vibration data of mechanical equipment and perform data preprocessing; A fault diagnosis model building module is used to build a fault diagnosis model based on a multi-scale adaptive hypergraph convolutional neural network; The fault diagnosis module is used to input the vibration data after data preprocessing into the constructed fault diagnosis model and output the fault diagnosis result.

Images