Intelligent fault diagnosis method and system for wheeled robots based on graph convolutional networks

Abstract

The present invention proposes a wheeled robot intelligent fault diagnosis method and system based on a graph convolutional network, including: constructing a wheeled robot data relationship graph according to a wheeled robot mathematical model; based on the wheeled robot data relationship graph, converting the measurement data of the robot sensor at different time slices into a spatiotemporal graph to generate a spatiotemporal graph data set; using the spatiotemporal graph data set to train a spatiotemporal differential graph convolutional network, the spatiotemporal differential graph convolutional network calculates multi-order backward differential features of the wheeled robot data relationship graph, uses local differential characteristics to enhance the features of nodes, and uses a spatiotemporal graph convolution module to obtain spatiotemporal related features; inputs the sensor measurement values ​​of the wheeled robot to be tested into the spatiotemporal differential graph convolutional network for fault judgment, and outputs the fault classification result. The proposed differential layer and the constructed robot data relationship graph are conducive to fault classification, and the developed STDGCN has the most advanced performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis, and in particular to a method and system for intelligent fault diagnosis of a wheeled robot based on a graph convolutional network. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] With the rapid development of robotics, various wheeled robots are being applied to a wider range of fields, including home services, industry, and aerospace. However, like other electronic systems, wheeled robots are prone to failures as they age. These failures may affect the performance of the robot and even lead to serious accidents. Therefore, it is of great significance to study the fault diagnosis methods of wheeled robots.

[0004] In recent years, the development of data acquisition hardware, neural network theory and parallel computing technology has made great breakthroughs in data-driven fault diagnosis technology, especially fault diagnosis technology based on deep learning. Many fault diagnosis methods based on deep learning have been significantly applied in various fault diagnosis cases, such as gearboxes, bearings, transformers, etc. For example, Wang et al. developed a deep neural network-based framework to monitor the condition of wind turbine gearboxes. Shao et al. proposed a convolutional deep belief network for bearing fault diagnosis. Huang et al. proposed a high-speed train bogie fault diagnosis method based on one-dimensional convolutional neural network. Peng et al. proposed a multi-branch multi-scale convolutional neural network for wheelset bearing fault diagnosis.

[0005] In most of the above fault diagnosis cases, satisfactory diagnostic performance can be achieved by using only a single type of sensor signal, such as vibration signal or current signal. Unlike the above studies, a wheeled robot is a highly coupled electromechanical system with nonlinear and time-varying characteristics, and it is difficult to accurately classify its faults based on a single type of sensor signal. Wheeled robots are usually equipped with a variety of heterogeneous sensors, so multi-sensor data is needed to comprehensively evaluate their health status. However, multi-sensor data feature extraction and fusion is a difficult problem, and traditional methods find it difficult to model the relationship between sensor signals.

[0006] Graph convolutional networks (GCNs) can solve this problem. GCNs were developed by Bruna et al. and Defferrard et al. By introducing a relationship graph, GCNs can aggregate the features of the neighborhood nodes of the central node to generate new central node features. At present, GCNs have been successfully applied in applications such as flow prediction and knowledge graph embedding. Inspired by these achievements, some studies in recent years have tried to apply GCNs to fault diagnosis. Li et al. proposed a multi-receptive field GCN to realize gearbox fault diagnosis, and converted data samples into weighted graphs by calculating the cosine similarity between nodes. Liao et al. proposed a transformer fault diagnosis method based on GCN, which uses an adjacency matrix to represent the similarity measure between unknown samples and labeled samples. Zhang et al. developed a deep GCN for rolling bearing fault diagnosis, which converts the collected acoustic signals into graphs, and the weights of the edges represent the similarity between the connected nodes.

[0007] The above GCN-based methods determine the relationship graph based on the similarity or characteristics of data samples. However, the relationship graph obtained by this method cannot accurately reflect the actual relationship, which will have a negative impact on network training and even limit the diagnostic performance. In addition, these methods tend to mine spatial correlations while ignoring temporal correlations.

[0008] In summary, the key issue in wheeled robot fault diagnosis is to use multi-sensor data to comprehensively evaluate its health status, but traditional deep learning-based methods have difficulty modeling the relationship between multi-sensor measurements. Unlike these methods, graph convolutional networks (GCNs) use graph-structured data and data relationship graphs as input, which is more effective for relationship modeling. However, existing GCN-based fault diagnosis methods have the following disadvantages:

[0009] 1) The relationship graph is obtained based on the similarity of data samples or their characteristics, and its accuracy cannot be guaranteed;

[0010] 2) Focus on extracting spatial correlations while ignoring temporal correlations. Summary of the invention

[0011] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides an intelligent fault diagnosis method for wheeled robots based on graph convolutional networks, which uses prior knowledge, i.e. a simplified mathematical model of a wheeled robot, to construct a relationship graph, uses local differential characteristics to enhance the features of nodes, and improves the performance of the fault diagnosis model by jointly capturing spatiotemporal correlations.

[0012] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0013] In the first aspect, a wheeled robot intelligent fault diagnosis method based on a graph convolutional network is disclosed, comprising:

[0014] Construct a data relationship diagram of the wheeled robot based on the wheeled robot mathematical model;

[0015] Based on the wheeled robot data relationship graph, the measurement data of the robot sensor at different time slices is converted into a spatiotemporal graph to generate a spatiotemporal graph dataset;

[0016] A spatiotemporal differential graph convolutional network is trained using a spatiotemporal graph dataset. The spatiotemporal differential graph convolutional network calculates multi-order backward differential features on the wheeled robot data relationship graph, enhances the features of nodes using local differential characteristics, and obtains spatiotemporal related features using a spatiotemporal graph convolutional module.

[0017] The sensor measurement values ​​of the wheeled robot to be tested are input into the spatiotemporal differential graph convolutional network for fault judgment, and the fault classification results are output.

[0018] As a further technical solution, when converting to a spatiotemporal graph, sensor measurements within one time slice are connected according to the relationship between the obtained sensor variables, and then each sensor measurement is connected to the same sensor measurement in consecutive time slices.

[0019] As a further technical solution, each node of the spatiotemporal graph contains a monitored value of a sensor variable.

[0020] As a further technical solution, the step of constructing a mathematical model of a wheeled robot is also included. When constructing the mathematical model of a wheeled robot:

[0021] The wheeled mobile robot is used as the experimental platform, and the sensor data containing the robot's operating status information is selected for fault diagnosis;

[0022] Sensor data includes encoders, wheel speed, motor torque, robot acceleration, and robot angular velocity;

[0023] All points on the wheeled robot rotate around the instantaneous center of rotation, ignoring the effect of friction.

[0024] As a further technical solution, in the constructed robot data relationship graph, each sensor variable is added with a self-connection edge.

[0025] As a further technical solution, the spatiotemporal differential graph convolutional network includes a differential layer: the differential layer calculates 0-D order backward differential features, and then uses the obtained multi-order backward differential features as new features of the graph nodes.

[0026] As a further technical solution, the spatiotemporal graph convolution module includes a graph convolution layer, two gated convolution layers, a residual structure and a batch normalization layer;

[0027] The graph convolution layer is used to mine the spatial dependencies of adjacent nodes, and the stacked gated convolution layer updates the features of the nodes by merging the features of the corresponding nodes on consecutive time slices;

[0028] Each gated convolutional layer consists of two standard 1D convolutional layers with the same parameters.

[0029] The residual structure is used to allow more information to pass;

[0030] The batch normalization layer normalizes the updated features to a fixed distribution, reducing internal covariate shifts.

[0031] In a second aspect, a robot fault diagnosis system is disclosed, comprising:

[0032] The relationship graph forming module is configured to: construct a wheeled robot data relationship graph according to a wheeled robot mathematical model;

[0033] The spatiotemporal graph forming module is configured to: generate a spatiotemporal graph data set by converting measurement data of the robot sensor at different time slices into a spatiotemporal graph based on the wheeled robot data relationship graph;

[0034] The spatiotemporal differential graph convolutional network training module is configured to: train the spatiotemporal differential graph convolutional network using the spatiotemporal graph data set, wherein the spatiotemporal differential graph convolutional network calculates multi-order backward differential features on the wheeled robot data relationship graph, enhances the features of the nodes using the local differential characteristics, and obtains spatiotemporal related features using the spatiotemporal graph convolution module;

[0035] The fault judgment module is configured to: input the sensor measurement values ​​of the wheeled robot to be tested into the spatiotemporal differential graph convolutional network for fault judgment, and output the fault classification result.

[0036] One or more of the above technical solutions have the following beneficial effects:

[0037] The spatial-temporal difference graph convolutional network (STDGCN) proposed in this paper is used for wheeled robot fault diagnosis. The network contains a difference layer that uses local difference characteristics for feature enhancement. At the same time, the spatial-temporal graph convolution module is introduced to jointly capture the spatial and temporal correlation. Multiple sets of experimental results show that STDGCN has excellent fault diagnosis performance.

[0038] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0040] Figure 1 A time-space diagram of sensor measurement conversion of a wheeled robot according to an embodiment of the present invention;

[0041] Figure 2 This is a schematic diagram of a Mecanum wheeled mobile robot according to an embodiment of the present invention;

[0042] Figure 3 A schematic diagram of a robot diagram according to an embodiment of the present invention;

[0043] Figure 4 This is a schematic diagram of the difference layering of an embodiment of the present invention;

[0044] Figure 5 Schematic diagram of a spatiotemporal graph convolution model according to an embodiment of the present invention;

[0045] Figure 6 Schematic diagram of the overall framework of the wheeled robot fault diagnosis method based on STDGCN according to an embodiment of the present invention;

[0046] Figure 7 A schematic diagram of the health status of a Mecanum wheeled mobile robot according to an embodiment of the present invention;

[0047] Figure 8 Schematic diagram of performance comparison between STGCN and STDGCN models implemented in the present invention. DETAILED DESCRIPTION

[0048] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0049] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0050] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0051] Overall concept:

[0052] The overall framework of the wheeled robot fault diagnosis method based on STDGCN proposed in this disclosure example is as follows: Figure 6 As shown in the figure. The process is summarized as follows: First, the robot data relationship graph is constructed based on the established robot mathematical model. Secondly, the spatiotemporal graph dataset is generated by converting the measurement data of the robot sensor at different time slices into a spatiotemporal graph. Then, we use the training dataset to train STDGCN. Finally, we input the test dataset and use the trained STDGCN for fault diagnosis.

[0053] Embodiment 1

[0054] This embodiment discloses a wheeled robot intelligent fault diagnosis method based on a graph convolutional network, comprising:

[0055] Firstly, a simplified mathematical model of wheeled robots is established, and the robot data relationship graph is determined based on it. Then, the multi-sensor measurements of the wheeled robot are converted into a spatiotemporal graph, and a difference layer is introduced into the spatiotemporal difference graph convolutional network (STDGCN). The difference layer calculates multi-order backward difference features and uses local difference characteristics to enhance the features of graph nodes.

[0056] In addition, it also includes the use of a spatial-temporal graph convolutional module (STGCM) to jointly capture spatial and temporal correlations, thereby achieving the simultaneous extraction of fault information in the spatial and temporal dimensions.

[0057] Finally, the proposed STDGCN is used to implement fault diagnosis. STDGCN is compared with several existing fault diagnosis methods, and the experimental results show the superiority of STDGCN.

[0058] About Graph Neural Networks:

[0059] Similar to convolutional neural network (CNN), the essential goal of GCN is to extract feature information from graph data through convolution operation. The difference between the two is that GCN processes non-Euclidean data.

[0060] Consider the convolution operation f*x in the spatial domain, where f represents the convolution kernel and x represents the input data. The Fourier transform of the graphic signal processed by the convolution operation is equal to the product of these transforms to the spectrum domain signal, so the convolution operation f*x in the spatial domain can be expressed as:

[0061]

[0062] Among them, F and F -1Denote Fourier transform and inverse Fourier transform respectively, and U denotes Fourier basis. In order to introduce the relational graph into the neural network, the Fourier basis used by GCN is calculated by the Laplacian matrix of the relational graph. The specific calculation method is as follows: Assume that L = DA is the Laplacian operator matrix of the graph, which can be normalized as Among them I N represents an identity matrix, A represents the adjacency matrix, and D is the degree matrix D ii ∈∑ j A ij Then, the Fourier basis U and the eigenvalue matrix Λ are obtained by eigenvalue decomposition:

[0063] UΛU T =L, Λ = diag([λ0,...,λ N-1 ]) (2)

[0064] According to the properties of the Laplace matrix, U is a set of orthogonal matrices that meet the basic mathematical requirements of Fourier transform. Let g θ =diag(U T f), (1) can be simplified as:

[0065] f*x=U((U T f)⊙(U T x))=Ug θ U T x (3)

[0066] In the process of graph convolution, the eigenvalue decomposition of the Laplacian matrix is ​​an important step, but when the graph is large, the eigenvalue decomposition of the Laplacian matrix consumes very high computational resources. To solve this problem, Defferrard et al. proposed using the Chebyshev polynomial T k To approximate g θ , so (3) can be alternatively expressed as:

[0067]

[0068] Where θ is the polynomial coefficient, λ max represents the maximum eigenvalue of the Laplace matrix. The Chebyshev polynomial can be defined as: T K (x) = 2xT K-1 (x)-T K-2 (x), T0(x) = 1, T1(x) = x. Using Chebyshev polynomials to approximate the solution is equivalent to using convolution kernels to extract the 0-K order neighbor information around the center of each node in the graph.

[0069] Kipf and Welling proposed a method to further simplify the Chebyshev polynomial. Let λmax =2 and K=1, which means the first-order neighborhood node features of the aggregation center node. In this way, (4) can be simplified to:

[0070]

[0071] By setting the parameter θ = θ0 = -θ1, we can get the single parameter formula In order to facilitate network training during back propagation, the parameters W and D are usually re-normalized as and Finally, the convolution operation f*x in GCN can be defined as:

[0072]

[0073] In this implementation example, the multi-sensor measurements of the wheeled robot contain rich relational information, and the temporal and spatial correlations of different sensor signals are modeled. Figure 1 As shown in FIG. 1 , the sensor measurements of a wheeled robot can be defined as a spatiotemporal graph, which consists of two steps. First, the sensor measurements within a time slice are connected according to the relationship between the sensor variables. Then each sensor measurement is connected to the same sensor measurement in consecutive time slices. The disclosed embodiment first starts with constructing a robot data relationship graph, i.e., modeling the relationship between sensor variables, and then introduces the proposed STDGCN for wheeled robot fault diagnosis in detail.

[0074] A time slice refers to a sampling moment. In the embodiment of the present disclosure, 18-dimensional sensor measurement values ​​are collected at one sampling moment.

[0075] The sensor measurement values ​​within a sampling moment are connected according to the "relationship". In the embodiment of the present disclosure, it is proposed to use the prior knowledge of the robot system, that is, the mathematical model of the robot to characterize this relationship. For example, in the mathematical model, the robot acceleration is affected by the motor torque, and the motor torque is connected to the robot acceleration.

[0076] The purpose of connecting each sensor measurement to the same sensor measurement in consecutive time slices is to build graph-structured data. Graph-structured data is composed of nodes and edges. Each measurement is considered as a node, and edges are built by connecting measurements. Edges are used to represent a specific relationship between two objects.

[0077] The input of the graph neural network is: graph structure data and data relationship graph. The graph structure data of the embodiment of the present disclosure is spatiotemporal graph data. In the spatiotemporal graph data set of the embodiment of the present disclosure, each sample size is 100*18, 100 is the number of time slices, and 18 is the number of nodes. The data of each time slice (1*18) is connected according to the data relationship graph.

[0078] Establishment of relationship diagram:

[0079] The basis for using the GCN-based method is to obtain a relationship graph. The disclosed embodiment proposes to establish a relationship graph based on the system prior knowledge, that is, a mathematical model of a wheeled robot. The disclosed embodiment uses a Mecanum wheeled mobile robot as an experimental platform and selects sensor data containing robot operation status information for fault diagnosis. The sensor variables used are shown in Table 1. The subscript numbers in the variables represent the motor number labels of the wheeled robot, and the subscript letters represent the coordinate axes.

[0080] Table 1. Sensor variables

[0081]

[0082] The purpose of the present disclosure is to use the mathematical model of the wheeled robot to reflect the relationship between sensor variables. Therefore, a simplified mathematical model that ignores the influence of environmental factors is required.

[0083] Assume that the wheeled robot is rigid, all points on the robot rotate around the instantaneous center of rotation, and ignore the influence of friction. Let (x, y, z) be the coordinate axes of the Mecanum wheeled mobile robot with the geometric center as the origin. Figure 2 Schematic diagram of a Mecanum wheeled mobile robot showing the motor driving force F acting on the wheel rollers Ti , where F Ti Divided into roller effective driving force F i and the ineffective slip force S i . F ix and F iy are the driving forces in the x and y directions respectively.

[0084] according to Figure 2 The established mathematical model is shown in (7)~(15), and the meanings of the undefined variables and functions in the equations are shown in Table 2.

[0085]

[0086] F Ti =T i r motor ,i=1,2,3,4 (8)

[0087] ai =F i / M,i=x,y (9)

[0088]

[0089]

[0090]

[0091]

[0092]

[0093] Table 2. Meaning of some variables and functions

[0094]

[0095] After establishing the mathematical model, an example is used to illustrate how to determine the relationship diagram based on the mathematical model. Considering (7)-(9), the following derivation can be obtained:

[0096]

[0097]

[0098] From equations (15) and (16), we can see that variable a x and a y It is affected by four motor torque variables T1, T2, T3 and T4. Therefore, each motor torque variable is connected to variable a x and a y , that is, T1, T2, T3 and T4 are a x and a y The remaining relationships can also be modeled in the same way. According to (7) to (15), the robot data relationship graph is constructed, as follows: Figure 3 As shown in the figure. In the constructed robot data relationship graph, a self-connection edge is added to each sensor variable. The graph convolutional neural network generates new node features by aggregating the features of the nodes connected to it. The purpose of adding a self-connection edge is to consider the characteristics of the node itself when generating new node features.

[0099] Spatiotemporal Differential Graph Convolutional Network (STDGCN): The core ideas of STDGCN are summarized as follows: 1) Use differential layers to enhance the features of graph nodes; 2) Use the spatiotemporal graph convolution module to obtain spatiotemporal related features.

[0100] Differential layer:

[0101] After constructing the robot data relationship graph, the sensor measurements of the wheeled robot are converted into a spatiotemporal graph. Each node contains the monitored value of a sensor variable. Since the graph nodes only contain digital features, the node features are extended by providing additional information. To achieve this goal, the technical solution of the present disclosure proposes a feature enhancement method based on a difference layer. The designed difference layer calculates 0-D order backward difference features, and then uses the obtained multi-order backward difference features as new features of the graph nodes.

[0102] The calculation process of the difference layer is as follows: Figure 4 As shown. For the slice x of the space-time graph at time t (t) The q node in the node has the following characteristics: The new node features calculated by differential layer can be defined as:

[0103]

[0104] in, represents the calculated D-th order backward difference feature, || represents the concatenation operation, Represents the generated new node features. When , (17) will satisfy the case of t≤D. In order to make the characteristic dimensions of the nodes the same, set

[0105] Spatio-Temporal Graph Convolutional Module (STGCM):

[0106] The multi-sensor measurements of the wheeled robot can be converted into a spatiotemporal graph, and the fault information is contained in the historical features of the node and its neighboring nodes. Modeling only the temporal or spatial dependencies will lead to unstable performance, because both temporal and spatial correlations are crucial for fault diagnosis. Therefore, in order to capture the spatiotemporal characteristics of the fault at the same time, a Figure 5 The STGCM shown.

[0107] STGCM consists of a graph convolution layer, two gated convolution layers, a residual structure, and a batch normalization layer. The graph convolution layer is used to mine the spatial dependencies of adjacent nodes, and the stacked gated convolution layers update the features of nodes by merging the features of corresponding nodes on consecutive time slices. Each gated convolution layer contains two standard one-dimensional convolution layers with the same parameters. The mathematical model of the gated convolution layer can be defined as:

[0108] y gated =(K1*x gated +b1)⊙σ(K2*x gated +b2) (18)

[0109] Where K1 and K2 are convolution kernels, b1 and b2 are biases, and σ is a sigmoid function. gated and gatedare the input and output of the gated convolutional layer. The purpose of adding the residual architecture in STGCM is to allow more information to pass through. In addition, the batch normalization layer (BN) is used to normalize the updated features to a fixed distribution, which effectively reduces the internal covariance shift.

[0110] Spatiotemporal Graph Convolution Module

[0111] The output of the spatiotemporal graph convolution module is a feature map of size 18×100×64, see Table 4. After the spatiotemporal graph convolution module (STGCM), a gated convolution layer is connected to compress the dimension of node features. Then, a global average pooling layer (GAP) is used to aggregate features in the time dimension. Finally, the features of all graph nodes are concatenated and input into two fully connected layers (FC), and the fault diagnosis results are obtained by the fully connected layers.

[0112] Experimental verification:

[0113] To verify the effectiveness of the proposed difference layer and the designed robot data relationship graph. Then STDGCN is compared with several existing fault diagnosis methods.

[0114] Dataset Description:

[0115] A Mecanum wheeled mobile robot is used as the experimental platform. The robot is driven by four brushless DC motors and equipped with a variety of sensors, including four encoders and an inertial measurement unit. Figure 7 The five common faults shown in the figure are diagnosed, and Table 3 records the fault labels, fault locations, and corresponding fault modes. The Mecanum wheeled mobile robot operates in different health states to collect fault diagnosis data sets. The speed of the wheeled robot is set to 0.5m / s, and multi-sensor data is collected with a sampling frequency of 100Hz. Then, in order to eliminate the influence of data dimension and value range, we use the maximum and minimum normalization method to normalize the raw sensor data to the range of [0,1]. In addition, we use a sliding window to divide the normalized sensor data into multiple samples, and set the step size and window length to 100. 70% of the samples are randomly selected as the training set, and the remaining 30% of the samples are used as the test set. Finally, a total of 4011 samples are generated, and the dimension of each sample is

[0116] Figure 7 Among them, (a) broken teeth of motor gearbox, (b) worn motor gearbox, (c) missing rollers, (d) failed shock absorber spring, and (e) loose shock absorber.

[0117] Table 3. Mecanum

[0118]

[0119] The STDGCN model consists of two STGCMs. The first STGCM contains 32 convolution kernels in the time dimension, and the second STGCM contains 64 convolution kernels in the time dimension. In addition, the parameter D of the difference layer is set to 4, and the effect of this parameter on the model performance will be discussed in the experiment. The detailed structure of STDGCN is shown in Table 4.

[0120] Table 4. STDGCN model detailed structure

[0121]

[0122] In this paper, all models are implemented using Pytorch 1.9.1. In the training phase, we use the cross entropy loss function and the Adam optimizer, and the learning rate is fixed to 0.001. We also set the batch size to 64 and train each model for 200 epochs. All models are tested 5 times to reduce the impact of randomness on the results. The three performance indicators of maximum accuracy (Max), minimum accuracy (Min) and average accuracy (Aver) are used as evaluation indicators to evaluate the diagnostic performance of the 5 trials.

[0123] Effectiveness of differential layer: In order to study the effectiveness of differential layer and the influence of parameter D on the results, we constructed 6 versions of STDGCN, namely STDGCN_1, STDGCN_2, STDGCN_3, ..., STDGCN_6 and a STGCN without differential layer. The last number in the STDGCN model name indicates the value of parameter D. The other settings of the constructed 7 models remain unchanged. The experimental results are shown in Figure 8 shown.

[0124] from Figure 8 It can be seen that the difference layer can effectively improve the performance of STDGCN. The Max and Aver of the STDGCN model are about 1% higher than those of the STGCN model. Due to the use of multi-order backward difference features for feature enhancement, STDGCN can learn better fault representation and obtain higher diagnostic accuracy.

[0125] In addition, by comparing these STDGCN models, we found that when the value of parameter D increases from 1 to 4, the Aver of STDGCN also increases. This shows that the model can accumulate the effects of multi-order backward differential features, thereby continuously improving the diagnostic performance. However, the Aver of STDGCN_5 and STDGCN_6 is slightly lower than that of STDGCN_4. The possible reason is that more differential features may distract the model. Therefore, the STDGCN_4 version is selected as the final model.

[0126] Effectiveness of the robot data relationship graph: To verify the effectiveness of the robot data relationship graph proposed in this paper. Two cases are considered in the experiment, using the robot data relationship graph for training or using the adaptive relationship graph for training. The adaptive relationship graph is implemented by a trainable weight matrix. We established four models in the experiment, namely ASTGCN, STGCN, ASTDGCN and STDGCN, among which ASTGCN and ASTDGCN are trained with adaptive relationship graphs. The experimental results are shown in Table 5.

[0127] Table 5. Comparison results between ASTGCN and STDGCN

[0128]

[0129] From the comparison results, it can be seen that using the robot data relationship graph for training can improve the accuracy of diagnosis while reducing the standard deviation. The Aver of STDGCN is 1.51% higher than that of ASTDGCN, and the Aver of STGCN is 0.94% higher than that of ASTGCN. This is because the designed robot data relationship graph utilizes prior knowledge, enabling the model to explicitly mine spatial correlations. In contrast, the use of an adaptive relationship graph increases the uncertainty of network training, so ASTGCN and ASTDGCN are unstable for spatiotemporal graph modeling. This shows that the designed robot data relationship graph is conducive to fault diagnosis.

[0130] Comparison with existing fault diagnosis methods:

[0131] To verify the superiority of the proposed method, we compared STDGCN with seven existing fault diagnosis methods, including RNN, GRU, GCN, LSTM, Resnet, DRSN, and MA1DCNN. The constructed RNN and GRU have two hidden layers, each with 128 units, a time step of 100, and an input dimension of 18. The GCN implemented in the experiment has two graph convolutional layers, and Resnet uses the Resnet-18 architecture. In addition, the parameter settings of the models LSTM, DRSN, and MA1DCNN are consistent with their articles. The experimental results are recorded in Table 6, and the best results are shown in bold. Based on these results, some observations can be made.

[0132] Table 6 Comparison results of STDGCN and other methods

[0133]

[0134] First, the maximum, minimum, and variance of RNN are worse than those of other comparison methods. Since RNN has the problem of gradient vanishing and its structure is simple, it is difficult to process long time series, resulting in poor diagnostic performance.

[0135] Secondly, the three CNN-based models Resnet, DRSN, and MA1DCNN have an Aver value of less than 90% on the wheeled robot fault diagnosis dataset, which is a poor performance. This may be because these CNN-based models are not suitable for data fusion of wheeled robots. Specifically, in the experiment, the measurement data of 18 sensors are directly input into the CNN-based model. The traditional convolution kernel generates new features by aggregating the features of all channels. In this way, important fault information in some channels is affected by irrelevant and redundant information in other channels, which has a negative impact on the model performance.

[0136] Third, the GRU, LSTM, and GCN methods outperform the CNN method. The averages of these three methods reach 93.23%, 92.04%, and 90.34%, respectively. A possible explanation is that they extract features better than the CNN method. GRU and LSTM use a gating mechanism to learn temporal dynamic features from the input, and GCN aggregates adjacent node features based on the Laplacian matrix of the relationship graph. However, these models only consider spatial or temporal correlations, which are not optimal for fault diagnosis of wheeled robots.

[0137] Finally, STDGCN achieved the best diagnostic results among all the tested methods. The maximum, minimum and average values ​​of STDGCN were 96.84%, 95.93% and 96.36% respectively, with a standard deviation of 0.29%. The direct reasons for the high accuracy of STDGCN are: 1) the prior knowledge of wheeled robots is utilized, so that STDGCN can stably exploit the spatial correlation of sensor measurements; 2) STDGCN adds a difference layer, which enables it to capture more useful information; 3) STGCM introduced in STDGCN can model spatiotemporal characteristics and dynamic correlations.

[0138] This example uses the prior knowledge of the robot system to construct a relationship graph, so that the GCN-based model can accurately capture the spatial correlation between sensor measurements; a difference layer is proposed to calculate multi-order backward difference features to enhance the features of graph nodes; a spatiotemporal convolution module is introduced to model the spatiotemporal correlation at the same time. This paper conducts ablation studies and comparative experiments based on the Mecanum wheeled mobile robot dataset. The experimental results show that the proposed difference layer and the constructed robot data relationship graph are conducive to fault classification, and the developed STDGCN has the most advanced performance.

[0139] Embodiment 2

[0140] The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the program.

[0141] Embodiment 3

[0142] The purpose of this embodiment is to provide a computer-readable storage medium.

[0143] A computer-readable storage medium stores a computer program, which executes the steps of the above method when executed by a processor.

[0144] Embodiment 4

[0145] The purpose of this embodiment is to provide a robot fault diagnosis system, comprising:

[0146] The relationship graph forming module is configured to: construct a wheeled robot data relationship graph according to a wheeled robot mathematical model;

[0147] The spatiotemporal graph forming module is configured to: generate a spatiotemporal graph data set by converting measurement data of the robot sensor at different time slices into a spatiotemporal graph based on the wheeled robot data relationship graph;

[0148] The spatiotemporal differential graph convolutional network training module is configured to: train the spatiotemporal differential graph convolutional network using the spatiotemporal graph data set, wherein the spatiotemporal differential graph convolutional network calculates multi-order backward differential features on the wheeled robot data relationship graph, enhances the features of the nodes using the local differential characteristics, and obtains spatiotemporal related features using the spatiotemporal graph convolution module;

[0149] The fault judgment module is configured to: input the sensor measurement values ​​of the wheeled robot to be tested into the spatiotemporal differential graph convolutional network for fault judgment, and output the fault classification result.

[0150] The steps involved in the apparatuses of the above embodiments 2, 3 and 4 correspond to the method embodiment 1, and the specific implementation methods can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0151] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0152] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. An intelligent fault diagnosis method for wheeled robots based on graph convolutional networks, characterized in that: include: Construct a data relationship diagram of the wheeled robot based on the wheeled robot mathematical model; Based on the wheeled robot data relationship graph, the measurement data of the robot sensor at different time slices is converted into a spatiotemporal graph to generate a spatiotemporal graph dataset; A spatiotemporal differential graph convolutional network is trained using a spatiotemporal graph dataset. The spatiotemporal differential graph convolutional network calculates multi-order backward differential features on the wheeled robot data relationship graph, enhances the features of nodes using local differential characteristics, and obtains spatiotemporal related features using a spatiotemporal graph convolutional module. The sensor measurement values ​​of the wheeled robot to be tested are input into the spatiotemporal differential graph convolutional network for fault judgment, and the fault classification results are output.

2. The intelligent fault diagnosis method for wheeled robots based on graph convolutional networks as claimed in claim 1, characterized in that: When converted to a spatiotemporal graph, sensor measurements within one time slice are connected according to the relationship between the obtained sensor variables, and then each sensor measurement is connected to the same sensor measurement in the consecutive time slices.

3. The intelligent fault diagnosis method for wheeled robots based on graph convolutional networks as claimed in claim 1, characterized in that: Each node of the spatiotemporal graph contains the monitored value of a sensor variable.

4. The method for intelligent fault diagnosis of wheeled robots based on graph convolutional networks as claimed in claim 1, characterized in that: It also includes the steps of constructing a mathematical model of a wheeled robot. When constructing a mathematical model of a wheeled robot: The wheeled mobile robot is used as the experimental platform, and the sensor data containing the robot's operating status information is selected for fault diagnosis; Sensor data includes encoders, wheel speed, motor torque, robot acceleration, and robot angular velocity; All points on the wheeled robot rotate around the instantaneous center of rotation, ignoring the effect of friction.

5. The intelligent fault diagnosis method for wheeled robots based on graph convolutional networks as claimed in claim 1, characterized in that: In the constructed robot data relationship graph, each sensor variable is added with a self-connection edge.

6. The intelligent fault diagnosis method for wheeled robots based on graph convolutional networks as claimed in claim 1, characterized in that: The spatiotemporal differential graph convolutional network includes a differential layer: the differential layer calculates 0-D order backward differential features, and then uses the obtained multi-order backward differential features as new features of the graph nodes.

7. The intelligent fault diagnosis method for wheeled robots based on graph convolutional networks as claimed in claim 1, characterized in that: The spatiotemporal graph convolution module includes a graph convolution layer, two gated convolution layers, a residual structure and a batch normalization layer; The graph convolution layer is used to mine the spatial dependencies of adjacent nodes, and the stacked gated convolution layer updates the features of the nodes by merging the features of the corresponding nodes on consecutive time slices; Each gated convolutional layer consists of two standard one-dimensional convolutional layers with the same parameters; The residual structure is used to allow more information to pass; The batch normalization layer normalizes the updated features to a fixed distribution, reducing internal covariate shifts.

8. A robot fault diagnosis system, characterized in that: include: The relationship graph forming module is configured to: construct a wheeled robot data relationship graph according to a wheeled robot mathematical model; The spatiotemporal graph forming module is configured to: generate a spatiotemporal graph data set by converting measurement data of the robot sensor at different time slices into a spatiotemporal graph based on the wheeled robot data relationship graph; The spatiotemporal differential graph convolutional network training module is configured to: train the spatiotemporal differential graph convolutional network using the spatiotemporal graph data set, wherein the spatiotemporal differential graph convolutional network calculates multi-order backward differential features on the wheeled robot data relationship graph, enhances the features of the nodes using the local differential characteristics, and obtains spatiotemporal related features using the spatiotemporal graph convolution module; The fault judgment module is configured to: input the sensor measurement values ​​of the wheeled robot to be tested into the spatiotemporal differential graph convolutional network for fault judgment, and output the fault classification result.

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the method described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method described in any one of claims 1 to 7 are performed.

Images