Wheeled robot fault diagnosis method and system based on dynamic graph convolutional network

Abstract

The invention discloses a wheeled robot fault diagnosis method and system based on a dynamic graph convolutional network, and belongs to the technical field of fault diagnosis. The method comprises the following steps: acquiring multi-channel sensor data of the wheeled robot, performing correlation analysis on the sensor data of different channels, and generating a sensor dynamic association graph; combining the sensor dynamic association diagrams corresponding to different time nodes along the time dimension, and constructing a dynamic space-time diagram; processing the dynamic space-time diagram through a trained fault diagnosis network to generate a fault diagnosis result; the fault diagnosis network alternately executes time sequence convolution and space diagram converter convolution on an input dynamic space-time diagram through a space-time diagram converter module, and integrates multi-modal time sequence information into feature representation of diagram nodes. The understanding and representation capability of the model on multi-sensor spatio-temporal data can be improved, and then the accuracy and robustness of fault diagnosis model prediction are improved; the problem that the robustness of fault diagnosis of an existing wheeled robot is low is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and particularly to a fault diagnosis method and system for a wheeled robot based on a dynamic graph convolutional network. Background Technique

[0002] The statements in this part only mention the background techniques related to the present invention and do not necessarily constitute prior art.

[0003] In recent years, robots have been widely used in many fields such as industrial manufacturing, logistics distribution, and medical surgery, effectively promoting the transformation and upgrading of traditional industries and the improvement of efficiency. As an important intelligent device, if a wheeled robot malfunctions during operation, it will not only affect the service quality but may also cause equipment damage or even casualties. Therefore, it is of great significance to carry out research on the fault diagnosis method of wheeled robots to ensure the safety and reliability of their operation.

[0004] With the continuous development of high-performance computers, deep learning technology has become the focus of extensive research and application in the field of intelligent fault diagnosis due to its excellent feature learning ability, outstanding diagnostic performance, and the advantage of low requirements for professional knowledge. In many fields such as wind power equipment, high-speed trains, and chemical equipment, fault diagnosis methods based on deep learning have been successfully implemented and achieved certain results.

[0005] Wheeled robots are usually equipped with various types of sensors and installed at different positions. However, most current deep learning-based fault diagnosis methods usually only analyze and diagnose based on a single type of sensor signal (such as vibration signal, current signal). When facing a highly coupled complex electromechanical system such as a wheeled robot, it is often difficult to achieve accurate fault classification.

[0006] In order to comprehensively and accurately evaluate the health status of a wheeled robot, it is very important to make full use of its multi-sensor data for comprehensive analysis. However, traditional deep learning-based methods are difficult to effectively model the rich correlations between multi-sensors, and thus difficult to achieve the fusion and feature extraction of multi-sensor data.

[0007] Kipf and Welling proposed the Graph Convolutional Network (GCN), which introduced a data relationship graph to model the correlation between data and opened up a new way to process data in non-Euclidean spaces. The measurement signals of multiple sensors carried by a wheeled robot can naturally form graph-structured data in a non-Euclidean space. The measurement data of a single sensor is regarded as a node feature, and the relationship between sensors is represented by edge connections. Compared with traditional deep learning-based methods, graph neural networks have successfully overcome the problem of being unable to directly analyze graph-structured data, and have the ability to effectively analyze graph data, enabling in-depth exploration of signal features and the relationships between various signals.

[0008] Miao et al. proposed an intelligent fault diagnosis method for wheeled robots based on graph convolutional networks. They constructed a data relationship graph of the wheeled robot according to the mathematical model of the wheeled robot, and used the designed spatio-temporal difference graph convolutional network to extract spatio-temporal correlation features of multi-sensor data and perform fault judgment. However, there are differences in physical characteristics and data types among the multiple sensors carried by the wheeled robot, and their deployment positions are different, resulting in differences in their responses to different faults. The correlation degree between heterogeneous sensors changes with the change of the fault mode. Therefore, a fixed sensor association graph is not sufficient to adapt to different robot states. At the same time, overly dense connections on the graph will not only increase additional computational overhead, but also amplify sensor signal noise and weaken the robustness of the model.

[0009] In addition, when existing methods extract the temporal correlation or spatial correlation of multi-sensor data, they do not consider the spatio-temporal position information of sensor node features, which causes the model to be unable to distinguish each sensor signal, thereby reducing the model's ability to model the spatio-temporal relationship between multi-sensor signals. Summary of the Invention

[0010] To solve the deficiencies of the prior art, the present invention provides a wheeled robot fault diagnosis method, system, electronic device, computer-readable storage medium, and computer program product based on a dynamic graph convolutional network. It uses a dynamic spatio-temporal graph to model the relationship between multiple heterogeneous sensors carried by the wheeled robot and constructs a graph transformer convolution mechanism to improve the accuracy and robustness of fault diagnosis prediction.

[0011] In the first aspect, the present invention provides a wheeled robot fault diagnosis method based on a dynamic graph convolutional network;

[0012] A wheeled robot fault diagnosis method based on a dynamic graph convolutional network includes:

[0013] Obtain multi-channel sensor data of a wheeled robot, perform correlation analysis on the sensor data of different channels, and generate a sensor dynamic association graph; combine the sensor dynamic association graphs corresponding to different time nodes along the time dimension to construct a dynamic spatio-temporal graph;

[0014] Process the dynamic spatio-temporal graph through a trained fault diagnosis network to generate a fault diagnosis result; among them, the fault diagnosis network alternately performs temporal convolution and spatial graph transformer convolution on the input dynamic spatio-temporal graph through a spatio-temporal graph transformer module, and integrates multi-modal temporal information into the feature representation of graph nodes.

[0015] In some embodiments, the performing correlation analysis on the sensor data of different channels and generating a sensor dynamic association graph includes:

[0016] Perform Pearson correlation analysis on the sensor data of different sensor channels within a preset time window length to obtain the correlation degree between different sensor channels;

[0017] Based on the comparison result between the correlation degree and a preset threshold, obtain an adjacency matrix to form a sensor dynamic association graph.

[0018] In some embodiments, the fault diagnosis network includes a spatio-temporal graph transformer module, a standard convolutional layer, a global average pooling layer, and a fully connected layer;

[0019] The spatio-temporal graph transformer module is used to extract spatio-temporal correlation fault features from the dynamic spatio-temporal graph by using a time embedding layer and a graph transformer layer to obtain spatio-temporal graph data features; the standard convolutional layer is used to perform a dimensionality reduction operation on the spatio-temporal graph data features, the global average pooling layer is used to integrate the spatio-temporal graph data features after dimensionality reduction in the time dimension, and the fully connected layer is used to output a fault diagnosis result according to the integrated spatio-temporal graph data features.

[0020] In some embodiments, the time embedding layer uses a gated temporal convolution mechanism to extract time-dependent information from the input dynamic spatio-temporal graph, and the graph transformer layer uses spatio-temporal position encoding to embed position information in the dynamic spatio-temporal graph into node features and adopts a cross-attention mechanism to aggregate feature information of multiple sensor channels.

[0021] In some embodiments, the alternately performing temporal convolution and spatial graph transformer convolution on the input dynamic spatio-temporal graph through the spatio-temporal graph transformer module includes:

[0022] Perform temporal convolution on the input dynamic spatio-temporal graph through the time embedding layer by using a gated temporal convolution mechanism to extract temporal features of each sensor node from the time dimension and update the dynamic spatio-temporal graph;

[0023] Perform spatio-temporal position encoding on the spatial and temporal positions of node features in the dynamic spatio-temporal graph using sine and cosine position encoding functions with different frequencies;

[0024] Add the spatio-temporal position encoding to the node features, utilize the spatio-temporal position information of the node features through the graph multi-head cross-attention mechanism, and aggregate the node features of neighboring nodes based on the connection relationships between sensor nodes.

[0025] In some embodiments, the spatio-temporal graph transformer module includes a first temporal embedding layer, a first graph transformer layer, a second temporal embedding layer, and a second graph transformer layer connected in sequence, and the input of the first graph transformer layer is fused with the output of the second graph transformer layer.

[0026] In a second aspect, the present invention provides a wheeled robot fault diagnosis system based on a dynamic graph convolutional network;

[0027] A wheeled robot fault diagnosis system based on a dynamic graph convolutional network, comprising:

[0028] A dynamic spatio-temporal graph construction module, configured to: obtain multi-channel sensor data of a wheeled robot, perform correlation analysis on the sensor data of different channels to generate a sensor dynamic association graph; combine the sensor dynamic association graphs corresponding to different time nodes along the time dimension to construct a dynamic spatio-temporal graph;

[0029] A fault diagnosis module, configured to: process the dynamic spatio-temporal graph through a trained fault diagnosis network to generate a fault diagnosis result; wherein, the fault diagnosis network alternately performs temporal convolution and spatial graph transformer convolution on the input dynamic spatio-temporal graph through the spatio-temporal graph transformer module, and integrates multi-modal temporal information into the feature representation of graph nodes.

[0030] In a third aspect, the present invention provides an electronic device;

[0031] An electronic device, comprising a memory, a processor, and a computer program stored on the memory, and the processor executes the computer program to implement the steps of the above-mentioned wheeled robot fault diagnosis method based on a dynamic graph convolutional network.

[0032] In a fourth aspect, the present invention provides a computer-readable storage medium;

[0033] A computer-readable storage medium, on which a computer program / instructions are stored, and when the computer program / instructions are executed by a processor, the steps of the above-mentioned wheeled robot fault diagnosis method based on a dynamic graph convolutional network are implemented.

[0034] In a fifth aspect, the present invention provides a computer program product;

[0035] A computer program product includes a computer program / instructions, and when the computer program / instructions are executed by a processor, the steps of the above-mentioned wheeled robot fault diagnosis method based on a dynamic graph convolutional network are implemented.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] 1. The technical solution provided by the present invention designs a strategy for constructing a sensor dynamic association graph to adapt to the changes in the robot state and constructs graph structure information related to the state category; designs a new graph Transformer convolutional mechanism, uses a cross-attention mechanism to aggregate the feature information of heterogeneous multi-sensor nodes, and embeds the position information in the spatio-temporal graph into the node features through spatio-temporal position encoding, improving the model's understanding and representation ability of multi-sensor spatio-temporal data, and further enhancing the accuracy and robustness of the fault diagnosis model prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The specification drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention.

[0039] Figure 1 It is a schematic flowchart of the wheeled robot fault diagnosis method based on a dynamic graph convolutional network provided by an embodiment of the present invention;

[0040] Figure 2 It is a schematic flowchart of the construction of the sensor dynamic association graph provided by an embodiment of the present invention;

[0041] Figure 3 It is a schematic flowchart of the construction of the dynamic spatio-temporal graph provided by an embodiment of the present invention;

[0042] Figure 4 It is a schematic network architecture diagram of the spatio-temporal graph transformer module provided by an embodiment of the present invention;

[0043] Figure 5 It is a schematic network architecture diagram of the graph transformer layer provided by an embodiment of the present invention;

[0044] Figure 6 It is an example diagram of the classification accuracy of different methods in the wheeled robot dataset provided by an embodiment of the present invention;

[0045] Figure 7 It is an example diagram of sensor signals under different fault modes provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] It should be noted that the following detailed description is exemplary and is intended to provide further illustration of the present invention. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention pertains.

[0047] It should be noted that the terms used herein are merely for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units need not be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices.

[0048] In the case of no conflict, the embodiments in the present invention and the features in the embodiments may be combined with each other.

[0049] Embodiment 1

[0050] Wheeled robots are usually equipped with various types of sensors and installed at different positions. The correlation degree between heterogeneous sensors changes with the change of the failure mode. The existing methods obtain a fixed sensor association graph through prior knowledge, which is not sufficient to adapt to different robot states. Moreover, when the existing methods extract the temporal correlation or spatial correlation of multi-sensor data, they do not consider the spatio-temporal position information of the sensor node features, which results in the model being unable to distinguish each sensor signal, thereby reducing the model's ability to model the spatio-temporal relationship between multi-sensor signals.

[0051] Therefore, this embodiment provides a fault diagnosis method for wheeled robots based on a dynamic graph convolutional network, constructs a dynamic spatio-temporal graph from the multi-sensor data of the robot, and constructs a fault diagnosis network introducing a spatio-temporal graph transformer module (spatio-temporal graph Transformer module).

[0052] Next, in combination with Figures 1-7 , a fault diagnosis method for wheeled robots based on a dynamic graph convolutional network disclosed in this embodiment will be described in detail. The fault diagnosis method for wheeled robots based on a dynamic graph convolutional network includes:

[0053] S1. Obtain the multi-channel sensor data of the wheeled robot, perform correlation analysis on the sensor data of different channels to generate a sensor dynamic association graph; combine the sensor dynamic association graphs corresponding to different time nodes along the time dimension to construct a dynamic spatio-temporal graph.

[0054] Here, sensors of different channels represent different sensor nodes.

[0055] Suppose there are N sensor nodes installed at different positions of a wheeled robot, and their measurement signals form an original measurement sample X ∈ R N×L (multi-channel sensor data), where L represents the total timestamp of sensor measurements; As an implementation, the specific process of S1 is as follows:

[0056] S101. Preprocess the original measurement sample.

[0057] Specifically, the maximum-minimum normalization method is used to standardize the original measurement sample to eliminate the interference caused by the differences in the dimensions and numerical ranges of different sensor data; A time-sliding window of appropriate size is used to segment the original measurement sample after the standardization process to ensure that each node in the sensor dynamic association graph has sufficient data features.

[0058] Here, the length and moving stride of the time-sliding window are set to F and S respectively, and the data segmented by the sliding window is used as the node feature of the graph, so as to obtain the feature set of the nodes in the sensor dynamic association graph Among them, represents the node feature of the i-th sensor channel at the t-th time window.

[0059] S102. Perform Pearson correlation analysis on the sensor data of different sensor channels within a preset time window length to obtain the correlation degree between different sensor channels; Based on the comparison result of the correlation degree and the preset threshold, obtain the adjacency matrix to form a sensor dynamic association graph.

[0060] In order to construct a sensor dynamic association graph that adapts to the changes of the robot system at different moments, this embodiment adopts a graph construction method based on similarity analysis; Specifically, the Pearson correlation coefficient is used to measure the correlation degree between sensor nodes, and further screening is carried out on this basis to ensure the sparsity of the graph, so as to generate the connections between sensor nodes.

[0061] The Pearson correlation coefficient is a statistical index used to measure the strength and direction of the linear relationship between two continuous variables. It quantifies their correlation by calculating the covariance and standard deviation between the variables. The specific calculation process is as follows:

[0062]

[0063] Among them, X and Y are two continuous variables, corresponding to s in the following text i and s j .

[0064] After calculating the sensor data of the current H time window lengths using the Pearson correlation coefficient, the lower-degree node connections are filtered by setting a threshold, and this operation can be expressed as:

[0065]

[0066] where s i and s j respectively represent the historical measurement data of the sensor data of sensor nodes i and j with the current H time serial lengths, and ò represents the correlation threshold.

[0067] When the correlation degree between sensor nodes exceeds the threshold, the connection relationship is considered valid; otherwise, it is considered invalid. By calculating all sensor channels in the above way, the adjacency matrix A is obtained. According to the Top-k sorting mechanism, the first k nearest neighbors of each node are selected, that is, the final adjacency matrix representing the connectivity of the sensor association graph can be obtained by the following formula:

[0068]

[0069] where represents the sparse adjacency matrix.

[0070] By restricting the maximum number of neighbors of each node, the sparsity of the graph can be guaranteed, and sparsity is a good inductive bias for learning graph datasets. At the same time, this approach will also make the sensor association graphs under different health states different, so that the graph structure information related to the state category can be obtained.

[0071] After the above operations, the static sensor association graph of the t-th time window and its node feature set

[0072] Combined with Figure 2 , S = {s1, s2,... s N} represents the multi-channel heterogeneous sensor data collected from the wheeled robot system, and N represents the number of sensors.

[0073] S103. The sensor association graphs of different time windows have graph structures that change with the change of the robot state, forming the final dynamic spatio-temporal graph

[0074] Combined with Figure 3 , the sensor association graphs of different time windows have graph structures that change with the change of the robot state.

[0075] S2. Process the dynamic spatio-temporal graph through the trained fault diagnosis network to generate the fault diagnosis result.

[0076] In this embodiment, the fault diagnosis network includes two sequentially connected Spatial-temporal graph Transformer modules (STGTMs), a standard convolutional layer, a global average pooling layer, and a fully connected layer; the STGTM includes a first temporal embedding layer, a first graph transformer layer (graph Transformer convolutional layer), a second temporal embedding layer, and a second graph transformer layer that are sequentially connected, and the input of the first graph transformer layer is fused with the output of the second graph transformer layer.

[0077] Here, the spatial-temporal graph transformer module can also be referred to as the spatial-temporal graph Transformer module.

[0078] First, stack two STGTMs to capture the spatio-temporal correlation fault features of multi-sensor data; then, connect a 1×1 convolutional layer with an activation function after the STGTM to reduce the data dimension and the number of parameters; then, the graph readout is realized through a global average pooling layer (GAP), which integrates the spatio-temporal graph data features in the time dimension; finally, the features of each sensor node are concatenated and then input into the fully connected layer to realize fault classification.

[0079] As an implementation, S2 includes:

[0080] S201. Process the input dynamic spatio-temporal graph through two stacked STGTMs. The specific process is as follows:

[0081] S2011. Use the gated temporal convolutional mechanism through the first temporal embedding layer to perform temporal convolution on the input dynamic spatio-temporal graph and update the dynamic spatio-temporal graph.

[0082] Specifically, the temporal embedding layer includes 1 gated convolutional layer, and the gated convolutional layer includes two conventional one-dimensional convolutional layers with the same parameters. The data processing process of the gated convolutional layer is expressed as:

[0083] y gated =(K1 * x gated + b1) ⊙ σ(K2 * x gated + b2);

[0084] In the formula, K1 and K2 represent convolutional kernels, b1 and b2 represent bias terms, σ represents the sigmoid activation function, x gated and y gated represent the input and output of the gated convolutional layer, respectively.

[0085] In this embodiment, the stride of the convolution operation in the time embedding layer in the time dimension is set to 2, which compresses the length of the spatio-temporal graph in the time dimension while updating the features of the nodes, and this helps to reduce the computational overhead of the subsequent graph Transformer convolution layer.

[0086] S2012. Process the updated dynamic spatio-temporal graph through the first graph Transformer convolution layer. The specific process is as follows:

[0087] (1) Perform spatio-temporal position encoding on the spatial and temporal positions of the node features in the dynamic spatio-temporal graph using sine and cosine position encoding functions with different frequencies.

[0088] To encode the spatial and temporal positions of the node features in the static spatio-temporal graph, this embodiment designs a spatio-temporal position encoding applicable to the spatio-temporal graph. Exemplarily, for each d-dimensional node feature, two sets of d / 2-dimensional position embeddings, including spatial embeddings and temporal embeddings, are calculated using sine and cosine position encoding functions with different frequencies. The calculation process of the position embedding can be expressed as:

[0089]

[0090] where p represents the position of the node feature in time or space, c is the dimension index of the embedding, and d e represents the dimension of the position embedding.

[0091] Finally, the spatial embedding and the temporal embedding are concatenated to obtain the d-dimensional spatio-temporal position embedding (i.e., position encoding) of the node feature, and the position encoding of node i is added to the input node feature to obtain h i , and the calculation process is expressed as:

[0092] h i = f i + δ i ;

[0093] where d m is the input channel dimension, and L is the input time dimension.

[0094] (2) Process the dynamic spatio-temporal graph with embedded spatio-temporal position encoding through graph multi-head cross-attention:

[0095] Specifically, first, perform a one-dimensional convolution operation on each node feature in parallel to obtain the corresponding key vector and query vector of the cross-attention, and the specific definitions are as follows:

[0096] Q g,i = W g,q * hi +b g,q ;

[0097] K g,i = W g,k *h i +b g,k ;

[0098] where \(g\in[1, H]\) represents the serial number of the attention head, and * represents the convolution operation. are the query vector and key vector of the \(g\) -th attention head. represents the weight of the convolution kernel. represents the bias.

[0099] Then, at each time step of the sensor dynamic association graph, the graph multi - head cross - attention calculation is expressed as:

[0100]

[0101] where \(d\) h is the hidden dimension size of each attention head, \(j\) represents the neighbor node, and \(l\) represents the current time step. It should be noted that the calculation range of the attention is kept within the spatial 1 - hop neighborhood to reduce the computational complexity.

[0102] After obtaining the multi - head attention weights at each time step, message aggregation can be performed to obtain the learned node features, and the calculation process is as follows:

[0103] V g,i = W g,v h i +b g,v ;

[0104]

[0105] where is the attention output at the \(l\) -th time step, || represents concatenation, \(N(i)\) is the set of neighbor nodes, is the value vector of the \(g\) -th attention head obtained by linear projection.

[0106] Finally, by concatenating in the time dimension, the final graph multi - head attention output is obtained

[0107] In addition, the graph Transformer convolution mechanism adopts skip connections in the structure to avoid the over - smoothing phenomenon of graph message propagation. At the same time, the long - distance skip connections combine the information of shallow and deep features, which helps to improve the expression ability of the model and makes the network training more stable.

[0108] S2013. The output of the first graph Transformer convolutional layer is sequentially processed through the second temporal embedding layer and the second graph Transformer convolutional layer, and the final output is fused with the input of the first graph Transformer convolutional layer as the final output of STGTM.

[0109] Considering that the multi-sensor data collected by wheeled robots has different modalities or physical characteristics, the STGTM module can effectively integrate this multi-modal temporal information into the feature representation of graph nodes by alternating between temporal convolution and spatial graph Transformer convolution. Based on the spatiotemporal position information of sensor node features, graph Transformer convolution is used to further integrate spatiotemporal information and achieve unified modeling. This approach not only enhances the accuracy and robustness of the model for wheeled robot fault diagnosis, but also ensures effective interaction and comprehensive analysis between different modal data.

[0110] S202. Perform dimensionality reduction on the spatiotemporal graph data features output through the above processing through a standard convolutional layer, integrate the reduced spatiotemporal graph data features in the time dimension through a global average pooling layer, and output the fault diagnosis results based on the integrated spatiotemporal graph data features through a fully connected layer.

[0111] Next, we experimentally verify the advanced nature of the wheeled robot fault diagnosis method based on dynamic graph convolutional networks described in this embodiment. The specific process is as follows:

[0112] Step 1: Use the self-developed Mecanum wheeled robot as the experimental platform to build a data set.

[0113] The self-developed Mecanum wheeled robot is driven by four brushless DC motors and equipped with a variety of heterogeneous sensors, including four encoders and an inertial measurement unit. The experimental platform can collect sensor variables including motor torque, encoders, wheel speed, robot acceleration, and angular velocity. To conduct wheeled robot fault diagnosis research, the Mecanum wheeled robot experimental platform was designed to simulate five common failure modes: broken gearbox teeth, failed shock absorber springs, missing wheel rollers, worn motor gearboxes, and loose shock absorbers.

[0114] Table 1 Failure modes and locations of the Mecanum wheeled robot

[0115]

[0116] To eliminate the influence of the dimensional and numerical range differences of different sensor data on subsequent analysis, the maximum-minimum normalization method is used to standardize the original data, so that its value range is regularized to the interval [0, 1]. The sliding window method is used to perform sample division operations on the normalized sensor time series data. Specifically, both the window size and the moving step size are set to 100, and finally 4011 samples are generated from the data, and the sample dimensions are all To train and evaluate the model performance, 70% of these samples are randomly assigned as the training set for model training and parameter optimization; while the remaining 30% are used as the test set to verify the performance and generalization ability of the model.

[0117] Step 2: Train the fault diagnosis network and evaluate the fault diagnosis performance of the model.

[0118] In this experiment, all methods are implemented using Python 3.9 and the Pytorch 2.2.2 architecture. It should be noted that the proposed DSTGTN model uses the DGL software package to construct a dynamic spatio-temporal graph and perform efficient graph-based message passing to train the graph neural network. All fault diagnosis models are trained using the Adam optimizer and the cross-entropy loss function. The batch size and the initial learning rate are set to 64 and 0.001 respectively, and the training iteration times for each model are 200 times to ensure sufficient convergence. To reduce the influence of randomness, in the comparative experiments of different models, each model is independently experimented 10 times. Four metrics, namely Maximum accuracy (Max-acc), Minimum accuracy (Min-acc), Average accuracy (Avg-acc), and F1-score, are used to comprehensively measure the fault diagnosis performance of the model.

[0119] Table 2 lists the detailed structure of the DSTGTN model and the specific sizes of the output feature maps of each layer of the network. To balance the computational efficiency and diagnostic performance of the model, two STGTMs are designed in the proposed DSTGTN model, which contain 64 and 128 convolutional kernels respectively in the time dimension. In addition, the threshold t and the maximum number of neighbors M for constructing the dynamic spatio-temporal graph are set to 0.10 and 5 respectively, and the influence of the selection of these two parameters on the model performance will be discussed in detail in the subsequent experimental section.

[0120] Table 2 Detailed Structure of DSTGTN Model

[0121]

[0122]

[0123] Step 3: Comparison of experimental results and analysis.

[0124] To prove the advantages of the proposed model, DSTGTN is compared with a variety of deep learning models, including 1DCNN, CWCNN-FA, Bi-LSTM, LSTM-CNN, GAT, MSTGCN, and three advanced multi-sensor fusion models: STDGCN, MMHGAT, and LiconvFormer. The comparison models are introduced as follows:

[0125] 1) 1DCNN: The one-dimensional convolutional neural network (1DCNN) contains two one-dimensional convolutional layers with 32 and 128 convolutional kernels respectively.

[0126] 2) CWCNN-FA: The channel-wise convolutional neural network with feature augmentation (CWCNN-FA) independently extracts heterogeneous fault features from multi-channel inputs and is suitable for fault diagnosis of wheeled mobile robots.

[0127] 3) Bi-LSTM: The bidirectional long short-term memory model can capture context information in both forward and backward directions of sequence data simultaneously. The constructed Bi-LSTM has 2 hidden layers with 32 and 128 hidden units respectively.

[0128] 4) LSTM-CNN: Combines the long short-term memory network and the convolutional neural network, and has powerful complex time series modeling capabilities. The parameter settings are the same as those in the reference.

[0129] 5) GAT: The graph attention network (GAT) introduces self-attention mechanism to dynamically assign different importance weights to each node. The constructed GAT model is composed of 2 stacked graph attention layers.

[0130] 6) MSTGCN: A graph neural network fault diagnosis method driven by multi-scale multi-source information fusion for fault diagnosis of wheeled robots.

[0131] 7) LiconvFormer: By combining separable multi-scale convolution and broadcast self-attention mechanism, it takes into account both the computational efficiency of the model and the accuracy of feature extraction. The detailed settings of the model are the same as those in the reference.

[0132] 8) STDGCN: The Spatial-temporal Difference Graph Convolutional Network (STDGCN) constructs a sensor association graph based on prior knowledge and uses a spatial-temporal difference graph convolutional network to jointly capture spatio-temporal correlations, showing superior performance in the fault diagnosis of wheeled service robots.

[0133] 9) MMHGAT: The Multi-sensor Multi-Head GAT (MMHGAT) converts the original signals of multiple sensors into multiple subgraphs and dynamically fuses the deep feature representations of multi-sensor signals.

[0134] Table 3 shows the specific experimental results of the method proposed in this embodiment and all comparison models on the wheeled robot dataset. Among them, the best results of each index are shown in bold.

[0135] Table 3 Comparison results of DSTGTN and other methods

[0136]

[0137] Based on the above experimental results, the following conclusions can be drawn:

[0138] 1) Several traditional models based only on Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs) did not achieve good results, and the Avg-acc index on the wheeled robot fault diagnosis dataset was lower than 94%. For CNN-based models, including 1DCNN and CWCNN-FA, when processing multi-sensor data, all sensor channels are stacked and processed with fixed convolutional kernels; this operation ignores the correlations between sensor channels, which is an important factor hindering the performance improvement of these models. For RNN-based models, including Bi-LSTM and LSTM-CNN, they can capture long-term dependence information from time series data and extract more effective time features based on the gating mechanism; however, these models only focus on temporal correlations and have the same problem as CNN-based models.

[0139] 2) Models based on graph neural networks (GNNs), including GAT and MMHGAT, perform well in diagnostic tasks. Although the GAT model focuses on capturing the spatial correlation between multiple sensors, it is limited by the lack of features of a single sensor, resulting in limited final effect, with an Avg-acc of 94.39%. In addition, as the number of graph convolution layers increases, GAT may encounter over-smoothing problems. MMHGAT constructs a graph based on the features of sensor data as nodes, and mines deep fault features by fusing feature subgraphs of multiple sensors, achieving good performance with an Avg-acc of 95.40%. However, when faced with a larger number of sensors and longer time series lengths, MMHGAT suffers from feature dimension explosion and a sharp increase in computational complexity. This limitation hinders further improvement of its diagnostic effect.

[0140] 3) LiConvFormer combines the Transformer with a CNN and offers lightweight improvements. This model employs separable multi-scale convolutions to extract features from multiple local receptive fields while leveraging broadcast self-attention to globally capture key fine-grained features of the signal. While LiConvFormer excels in feature extraction, it is limited in its multi-sensor correlation modeling. It extracts temporal correlations solely by stacking multiple channels, which limits its performance in wheeled robot fault diagnosis tasks, resulting in an average accuracy of 93.99%, which falls short of optimal results.

[0141] 4) STDGCN constructs a sensor association graph based on prior knowledge of the robot body, and has the best diagnostic effect among all the compared models, achieving an average accuracy of 96.58%. Compared with STDGCN, the DSTGTN proposed in this embodiment still has an improvement of 1.77%, and is about 5% higher in accuracy than most of the compared methods. What DSTGTN and STDGCN have in common is that they can both extract time- and space-related features. In addition, the direct reasons why DSTGTN achieves better diagnostic performance are: ① DSTGTN constructs a dynamic spatiotemporal graph to adapt to different robot health states; ② DSTGTN designs spatiotemporal position encoding, which considers the spatiotemporal position information of sensor node features in the attention mechanism, thereby improving the model's understanding and representation of spatiotemporal data.

[0142] Wheeled robots face interference from ambient noise during actual operation, and the collected sensor signals are easily contaminated. Therefore, it is necessary to verify the noise immunity of the proposed DSTGTN. To simulate the signals collected in real industrial practice, this experiment added Gaussian white noise with different signal-to-noise ratios (SNRs), including 5, 3, 0, -3, and -5 dB, to the original multi-sensor signals. The specific definition of SNR is as follows:

[0143]

[0144] Among them, Psignal represents the signal energy, and Pnoise represents the noise energy.

[0145] As can be seen from the above formula, the smaller the signal-to-noise ratio, the greater the interference of noise on the original signal. It should be noted that when the SNR is set to -5dB, the energy of the noise is approximately 3 times that of the original signal energy. The fault diagnosis results under different signal-to-noise ratios are shown in Table 4. It can be found that when the signal-to-noise ratio is set to 5dB, the accuracy of these methods is relatively high; as the signal-to-noise ratio of the data further decreases, the diagnostic effects of these methods all decline to varying degrees. It is worth noting that under all signal-to-noise ratio conditions, the F1-score of the proposed DSTGTN is significantly higher than the other nine methods. This is because the method proposed in this chapter constructs a sparse sensor dynamic association graph and designs a graph Transformer convolution mechanism, which can suppress the mutual interference between irrelevant sensor channels under noise and improve the ability of the model to focus on effective features in noise. Therefore, DSTGTN has good robustness under environmental noise interference.

[0146] Table 4 Anti-noise experimental results of different methods on the wheeled robot dataset

[0147]

[0148]

[0149] Step 4. Ablation experiment results and analysis

[0150] To verify the effectiveness of the dynamic spatio-temporal graph construction method and the graph Transformer convolution mechanism proposed in this embodiment, ablation experiments are carried out on the wheeled robot dataset. In the experiment, two architectures, DSTGTN-PKG and DSTGTN-GCN, are constructed for comparison with the proposed method. These two models have a similar structure to DSTGTN. Among them, the only difference between DSTGTN-PKG and DSTGTN is that it uses a fixed sensor association graph constructed based on the physical prior knowledge of the robot, and the only difference between DSTGTN-GCN and DSTGTN is that DSTGTN-GCN only uses conventional graph convolution. In the experiment, the three models are tested under the conditions of no noise and 0dB noise, and the results are shown in Table 5.

[0151] Table 5 Comparison results of DSTGTN and ablation models

[0152]

[0153] It can be found that compared with DSTGTN-PKG, the average accuracy of DSTGTN under noise-free and 0 dB noise conditions has increased by 3.20% and 12.24% respectively. This is because the constructed sensor dynamic association graph contains graph structure information related to the fault category. By aggregating the information on the graph, this structural information is embedded into the features, enabling the model to extract discriminative features even under the interference of noise. At the same time, the average accuracy of DSTGTN under noise-free and 0 dB noise conditions is 1.50% and 2.21% higher than that of DSTGTN-GCN respectively, indicating that the graph Transformer convolution mechanism can capture more spatio-temporal correlation features of the system state and enhance the model's ability to model spatio-temporal relationships.

[0154] Embodiment 2

[0155] This embodiment discloses a wheeled robot fault diagnosis system based on a dynamic graph convolutional network, including:

[0156] A dynamic spatio-temporal graph construction module, configured to: obtain multi-channel sensor data of the wheeled robot, perform correlation analysis on the sensor data of different channels to generate a sensor dynamic association graph; combine the sensor dynamic association graphs corresponding to different time nodes along the time dimension to construct a dynamic spatio-temporal graph;

[0157] A fault diagnosis module, configured to: process the dynamic spatio-temporal graph through a trained fault diagnosis network to generate a fault diagnosis result; wherein, the fault diagnosis network alternately performs temporal convolution and spatial graph transformer convolution on the input dynamic spatio-temporal graph through a spatio-temporal graph transformer module, and integrates multi-modal temporal information into the feature representation of the graph nodes.

[0158] It should be noted here that the above dynamic spatio-temporal graph construction module and fault diagnosis module correspond to the steps in Embodiment 1. The examples and application scenarios implemented by the above modules and the corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1. It should be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.

[0159] Embodiment 3

[0160] Embodiment 3 of the present invention provides an electronic device, including a memory, a processor, and computer instructions stored on the memory and running on the processor. When the computer instructions are run by the processor, the steps of the above-mentioned wheeled robot fault diagnosis method based on a dynamic graph convolutional network are completed.

[0161] Embodiment 4

[0162] Embodiment 4 of the present invention provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps of the above-mentioned wheeled robot fault diagnosis method based on a dynamic graph convolutional network are completed.

[0163] Embodiment 5

[0164] Embodiment 5 of the present invention provides a computer program product, including a computer program / instructions. When the computer program / instructions are executed by a processor, the steps of the above-mentioned wheeled robot fault diagnosis method based on a dynamic graph convolutional network are implemented.

[0165] The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0166] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device that implements the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0167] These computer program instructions can also be loaded onto a computer or other programmable data processing device, and a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0168] In the above embodiments, the descriptions of the various embodiments have their own emphases. For parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0169] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A fault diagnosis method for a wheeled robot based on a dynamic graph convolutional network, characterized in that Including: Obtain multi-channel sensor data of a wheeled robot, perform correlation analysis on the sensor data of different channels, and generate a sensor dynamic association graph; Combine the sensor dynamic association graphs corresponding to different time nodes along the time dimension to construct a dynamic spatio-temporal graph; Process the dynamic spatio-temporal graph through a trained fault diagnosis network to generate a fault diagnosis result; Among them, the fault diagnosis network alternately performs temporal convolution and spatial graph transformer convolution on the input dynamic spatio-temporal graph through a spatio-temporal graph transformer module, and integrates multi-modal temporal information into the feature representation of graph nodes.

2. The wheeled robot fault diagnosis method based on a dynamic graph convolutional network according to claim 1, wherein, The performing correlation analysis on the sensor data of different channels to generate a sensor dynamic association graph includes: Perform Pearson correlation analysis on the sensor data of different sensor channels within a preset time window length to obtain the correlation degree between different sensor channels; Based on the comparison result between the correlation degree and a preset threshold, obtain an adjacency matrix to form a sensor dynamic association graph.

3. The wheeled robot fault diagnosis method based on a dynamic graph convolutional network according to claim 1, characterized in that, The fault diagnosis network includes a spatio-temporal graph transformer module, a standard convolutional layer, a global average pooling layer, and a fully connected layer; The spatio-temporal graph transformer module is used to extract spatio-temporal correlation fault features from the dynamic spatio-temporal graph by using a time embedding layer and a graph transformer layer to obtain spatio-temporal graph data features; The standard convolutional layer is used to perform a dimensionality reduction operation on the spatio-temporal graph data features, the global average pooling layer is used to integrate the spatio-temporal graph data features after dimensionality reduction in the time dimension, and the fully connected layer is used to output a fault diagnosis result according to the integrated spatio-temporal graph data features.

4. The method for diagnosing faults of a wheeled robot based on a dynamic graph convolutional network according to claim 3, wherein The time embedding layer uses a gated temporal convolution mechanism to extract time-dependent information from the input dynamic spatio-temporal graph, and the graph transformer layer embeds the position information in the dynamic spatio-temporal graph into node features by using spatio-temporal position encoding, and uses a cross-attention mechanism to aggregate the feature information of multiple sensor channels.

5. The wheeled robot fault diagnosis method based on a dynamic graph convolutional network according to claim 1, characterized in that The alternately performing temporal convolution and spatial graph transformer convolution on the input dynamic spatio-temporal graph through the spatio-temporal graph transformer module includes: Perform temporal convolution on the input dynamic spatio-temporal graph through the time embedding layer by using a gated temporal convolution mechanism, extract the temporal features of each sensor node from the time dimension, and update the dynamic spatio-temporal graph; Use sine and cosine position encoding functions with different frequencies to perform spatio-temporal position encoding on the spatial and time positions of the node features in the dynamic spatio-temporal graph; Add the spatio-temporal position encoding to the node features, utilize the spatio-temporal position information of the node features through a graph multi-head cross-attention mechanism, and aggregate the node features of neighbor nodes based on the connection relationship between sensor nodes.

6. The wheeled robot fault diagnosis method based on a dynamic graph convolutional network according to claim 1, characterized in that, The spatio-temporal graph transformer module includes a first time embedding layer, a first graph transformer layer, a second time embedding layer, and a second graph transformer layer connected in sequence, and the input of the first graph transformer layer is fused with the output of the second graph transformer layer.

7. A wheeled robot fault diagnosis system based on a dynamic graph convolutional network, characterized in that, Including: A dynamic spatio-temporal graph construction module, configured to: Obtain multi-channel sensor data of a wheeled robot, perform correlation analysis on the sensor data of different channels, and generate a sensor dynamic association graph; Combine the sensor dynamic association graphs corresponding to different time nodes along the time dimension to construct a dynamic spatio-temporal graph; The fault diagnosis module is configured to: process the dynamic spatio-temporal graph through a trained fault diagnosis network to generate a fault diagnosis result; wherein, the fault diagnosis network alternately performs temporal convolution and spatial graph transformer convolution on the input dynamic spatio-temporal graph through a spatio-temporal graph transformer module, and integrates multi-modal temporal information into the feature representation of graph nodes.

8. An electronic device, comprising a memory, a processor, and a computer program stored on the memory, characterized in that, The processor executes the computer program to implement the steps of the wheeled robot fault diagnosis method based on the dynamic graph convolutional network according to any one of claims 1-6.

9. A computer-readable storage medium having computer programs / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, the steps of the wheeled robot fault diagnosis method based on the dynamic graph convolutional network according to any one of claims 1-6 are implemented.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, the steps of the wheeled robot fault diagnosis method based on the dynamic graph convolutional network according to any one of claims 1-6 are implemented.

Images