Robot core part fault diagnosis method based on artificial intelligence
By using reinforcement learning and Chebishev graph convolutional network in troubleshooting combined with short-time Fourier transform and graph theory method, the problem that traditional methods are difficult to extract fault features in noise complex environments is solved, and harmonic deceleration is achieved High accuracy fault diagnosis of device vibration signal.
Patent Information
- Application Number
- CN202410957867.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-17
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-07-17
AI Technical Summary
Traditional rotary machinery fault diagnosis methods are difficult to effectively extract the fault characteristics in vibration signals. Especially in the case of complex noise, the graph convolutional network is prone to excessive smoothing in fault diagnosis, which affects the accuracy.
Using the Chebischev graph convolution network based on reinforcement learning, combined with short-time Fourier transform and graph theory method, feature extraction and graph structure modeling of the vibration signals of the harmonic reducer are performed, and the Chebischev graph convolution network is used. Adaptively adjust the network layer to improve detection accuracy.
Effectively avoid noise interference, extract topological structures, improve the accuracy of fault diagnosis, and ensure the reliability of detection results.
Smart Images

Figure CN119089337A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of fault diagnosis of harmonic reducers and relates to a method for diagnosing faults of core components of a robot based on artificial intelligence. Background Art
[0002] As one of the core components of the robot, the harmonic reducer plays a vital role in the transmission process of the robot joint arm. Traditional rotating machinery fault diagnosis methods require the analysis of vibration signals collected during operation. However, these vibration signals are often non-stationary and contain complex noise, which makes it difficult to effectively extract fault features in the signal processing process. Recent studies have shown that graph modeling can be used to extract potential structural and topological information hidden in state signals, and it has shown significant efficiency in capturing the long-term correlation of non-static signals. Among them, graph convolutional networks (GCNS) can effectively propagate and absorb information at the edge of the graph structure, and obtain the feature representation of nodes by iteratively aggregating and transforming the features of adjacent nodes; due to the powerful ability of GCNS in spatial information modeling, researchers have begun to apply it to the field of fault diagnosis. In the literature Xiaoxia Yu, Baoping Tang, and Kai Zhang. Fault diagnosis of wind turbine gearbox using anovel method of fast deep graph convolutional networks. IEEE Transactions on Instrumentation and Measurement, 70: 1–14, 2021, the vibration signal of the wind turbine gearbox is decomposed and constructed into graph data, and the GCN model is used for fault diagnosis. In the literature Xiangqu Xiao, Chaoshun Li, Jie Huang, Tian Yu, and Pak Kin Wong. Animproved graph convolutional networks for fault diagnosis of rolling bearing with limited labeled data. Measurement Science and Technology, 34(12): 125109, 2023., an improved GCN is proposed for bearing fault diagnosis with limited labeled data. Although GCN has shown good performance in mechanical fault diagnosis of graph data, over-smoothing is prone to occur during the training process of GCNs, resulting in gradient disappearance and affecting the accuracy of the final judgment. Summary of the invention
[0003] In order to solve the above technical problems, the purpose of the present invention is to provide a robot core component fault diagnosis method based on artificial intelligence.
[0004] The present invention provides a method for diagnosing faults of core components of a robot based on artificial intelligence, comprising:
[0005] Step 1: Collect the vibration signal of the harmonic reducer and pre-process it to obtain graph structure data;
[0006] Step 2: Use 70% of the graph structure data as the training sample set and 30% of the graph structure data as the test sample set;
[0007] Step 3: Establish a Chebyshev graph convolutional network;
[0008] Step 4: Input the training sample set into the Chebyshev graph convolutional network, train the network through the reinforcement learning method, and test it using the test set to obtain the final network model;
[0009] Step 5: The vibration signal of the harmonic reducer collected in real time is input into the trained Chebyshev graph convolutional network for fault diagnosis.
[0010] The present invention provides a method for diagnosing faults of robot core components based on artificial intelligence, which has the following beneficial effects:
[0011] The diagnostic method of the present invention uses the method of short-time Fourier transform plus graph theory to extract features from one-dimensional vibration signals, effectively avoiding the interference of noise and extracting the topological structure of the signal. Combining reinforcement learning with Chebyshev graph neural network can adaptively adjust the network layer and ensure the accuracy of the detection results. BRIEF DESCRIPTION OF THE DRAWINGS
[0012] Figure 1 It is a flow chart of a method for diagnosing faults of robot core components based on artificial intelligence of the present invention. DETAILED DESCRIPTION
[0013] like Figure 1 As shown, a method for diagnosing a robot core component fault based on artificial intelligence of the present invention comprises:
[0014] Step 1: Collect the vibration signal of the harmonic reducer and pre-process it to obtain the graph structure data, specifically:
[0015] Step 1.1: Collect the vibration signals of four types of harmonic reducers when they are faulty and in normal state.
[0016] In the specific implementation, the vibration signal is collected using a customized experiment designed by Midea Group specifically for harmonic reducer fault testing, which includes two main parts: a load system and a data acquisition system. The load system includes a servo motor that drives the harmonic reducer, and a magnetic powder brake is used to provide the load. An angle encoder and a torque sensor are installed to accurately measure the output speed and torque. The experimental platform is designed as an open power supply system with a simple structure and high control accuracy, which is very suitable for vibration and noise signal testing of harmonic reducers. Although vibration data are collected in the X, Y and Z directions, this application only focuses on the data collected in the direction of rotation of the harmonic reducer (X axis), because the Spearman correlation value of radial vibration is significantly higher than that of axial data, and the values of X and Y are less susceptible to interference, and have the highest correlation with harmonic reducer faults. The collected vibration signals include four common harmonic reducer faults on the actual production line. Faults include: fracture failure of flexible bearings, fatigue failure of input shafts caused by long-term operation, pitting failure at the root of teeth, and tip failure of flexible wheels.
[0017] Step 1.2: Let G = {F, E, W} represent the space-time graph, where F is the set of nodes in the space-time graph, E represents the edges connecting these nodes, and W represents the weight of the edge.
[0018] Step 1.3: Use short-time Fourier transform to cut the vibration signal of the harmonic reducer into k segments for feature extraction;
[0019]
[0020] Among them, ΔM is the time difference of a segment, Δf is the frequency difference, m represents time, T(k, m) is the output of frequency kΔf and time mΔM, x[n] is the vibration signal, N is the window length, the operator * represents conjugation, and ω is the window function, specifically the Hanning window is used.
[0021]
[0022] Among them, S(k, m) is the output in the frequency interval 1≤k≤K and time 0≤m≤M. As the node feature of the spatiotemporal graph, the frequency F1, F2...F of each segment K Represents a node of a space-time graph.
[0023] Step 1.4: Use Euclidean distance to calculate the weights between nodes;
[0024]
[0025] Where Dis represents the nodes Fi and F at a specific time point m. j The Euclidean distance between them is calculated, thus completing the construction of the space-time graph.
[0026] Step 1.5: In spectral analysis, the symmetric normalized Laplace matrix is defined as follows:
[0027]
[0028] Where, L s is the symmetric normalized Laplace matrix, I is the identity matrix, D is the diagonal matrix of the graph, A is the adjacency matrix, K is the number of nodes in the space-time graph, and the real symmetric normalized Laplace matrix L is transformed by using the eigenvalue decomposition in Fourier transform. s Transformed into orthogonal basis vector U, as follows:
[0029] L s =UΛU T
[0030] In the formula, Λ=diag{λ1, λ2,…,λ K} is the inclusion matrix L s The eigenvalue matrix of all eigenvalues of each spatiotemporal graph is represented by a one-dimensional eigenvector containing the eigenvalues as P = [λ1, λ2, …, λ K ].
[0031] Step 1.6: Define the correlation graph G a = {V a , E a , A}, where V a is the node set of the association graph, E a represents the edges connecting these nodes; each node represents a specific harmonic reducer fault state; A is the adjacency matrix, which represents the connection relationship between any pair of nodes, where the value of each element in the adjacency matrix is 0 or 1, that is:
[0032]
[0033] Among them, nodes with the same fault label are connected to each other, while nodes with different fault labels remain unconnected.
[0034] Step 1.7: Transform the one-dimensional feature vector P = [λ1, λ2, …, λ K ] as the feature of the node in the association graph to complete the construction of the association graph.
[0035] In the specific implementation, vibration signals of all fault types were collected under five different load conditions, namely no load, 20 N·m, 40 N·m, 60 N·m and 80 N·m. The experiments were divided into two categories, with constant speed and load under constant conditions and constant speed and load under variable conditions in the range of 0-80 N·m. The acquisition system samples the signals at a frequency of 25600 Hz, and each sampling lasts for 1 second. For each case, 10 sets of samples were collected. Samples were randomly selected from the vibration data of each case, for a total of 250 samples, and each sample was cut into 50 fragments. The cut samples were feature extracted and nodes generated using the methods proposed in steps 1.2-1.7. During the synthesis process, an association graph identical to the generated label was formed, and then the association graphs of each label were combined to classify the nodes.
[0036] Step 2: Use 70% of the graph structure data as the training sample set and 30% of the graph structure data as the test sample set.
[0037] Step 3: Establish a Chebyshev graph convolutional network, specifically:
[0038] Step 3.1: Graph convolution is represented as:
[0039] Y=g θ (L s )A
[0040] Y=Ug θ (Λ)U T A
[0041] Among them, g θ is the convolution kernel, Y is the output of the graph convolution, since g in the ordinary graph convolution θ The calculation of (Λ) is complex and consumes a lot of computing power. Therefore, the Chebyshev polynomial expansion is introduced to solve this problem. It is found that by θ (Λ) can be effectively approximated by expanding it into Chebyshev polynomials:
[0042] T k (Λ)=2ΛT k-1 (Λ)-T k-2 (Λ)
[0043] T0(Λ)=1,T1(Λ)=Λ
[0044]
[0045] Among them, θ k is the vector of Chebyshev coefficients, is the normalized form of Λ;
[0046]
[0047] Among them, λ max is the largest element of Λ, The value range of the elements in is [-1, 1], so the graph convolution is calculated as follows:
[0048]
[0049] Step 3.2: The graph convolution operation is considered as a filtering process, which retains the dimension of the input signal. By multiplying the graph structure data with the trainable weight matrix, the dimension of the feature can be adjusted, that is:
[0050] Z=Cheb(A,W (b) )=YW (b)
[0051] Among them, Z is the output of the b-th graph convolutional layer, W (b) is the trainable weight matrix of the b-th graph convolutional layer, Cheb(·) represents the Chebyshev convolution function; the construction of the graph convolutional architecture is expressed as the following transition equation:
[0052]
[0053] Among them, h b is the feature vector of layer b, and a is the number of Chebyshev graph convolutional network layers determined by the reinforcement learning agent.
[0054] Step 4: Input the training sample set into the Chebyshev graph convolutional network, train the network through the reinforcement learning method, and test it with the test set to obtain the final network model, which is as follows:
[0055] Step 4.1: Establish the training process as a Markov decision process, with the basic elements including state, action and reward;
[0056] State: At time t, the properties of a node in the current association graph are extracted:
[0057] s t ={λ1,λ2,…,λ K}
[0058] Action: The number of Chebyshev graph convolutional network layers decided by the reinforcement learning agent:
[0059] a t =a
[0060] Reward function: The reward function is defined as, at time t, action a t The effect of network classification after execution;
[0061]
[0062] Among them, M(s t ,a t ) is the accuracy of the graph convolutional network on the training set at time t, L is a hyperparameter used to define the window size of historical performance to be used as a benchmark reference, and β is a hyperparameter that determines the strength of the reward signal;
[0063] is the average of the most recent L classification accuracies, used to encourage the reinforcement learning agent to achieve better performance compared to the most recent L time steps;
[0064] Step 4.2: Select the starting node of the association graph, whose node features are regarded as the current state s t Input to the reinforcement learning agent;
[0065] Step 4.3: The agent generates action a t , represents the number of layers of the Chebyshev graph convolutional network used to identify the node;
[0066] Step 4.4: The Chebyshev graph convolutional network gives the node classification result, and the reward value R is calculated using the classification accuracy t , guiding the reinforcement learning agent to make reasonable network choices, and using the experience replay mechanism during training to ensure the long-term effectiveness of training.
[0067] Step 5: The vibration signal of the harmonic reducer collected in real time is input into the trained Chebyshev graph convolutional network for fault diagnosis.
[0068] Deep reinforcement learning algorithms can be used to optimize the above Markov decision process. Given that the action space of the aggregation process is discrete, where any action is a finite positive integer, we adopt a deep Q-learning algorithm to address this challenge.
[0069] For the Q-learning algorithm, its action value function Q under strategy π is π (s,a) is defined as:
[0070] Q π (s,a)=Επ[R t s t =s,a t =a;π]
[0071] Among them, Q π (s,a) is used to score the action, E π is the expected return under strategy π, s t is the current state, a t is the current action;
[0072] In deep reinforcement learning algorithms, a deep neural network is used to approximate the action-value function Q in the policy π. π (s,a):
[0073] Q(s,a;θ)≈Q π (s,a)
[0074] Among them, θ is the parameter of the deep neural network model;
[0075] In the process of interaction between the agent and the environment, the greedy algorithm is used to generate a series of t , a t , R t ,s t+1 ) experience. These experience samples are stored in the experience pool. Randomly extracting small batches of experience data from the experience pool to iteratively update the agent's parameters can accelerate the learning process of the reinforcement learning agent.
[0076] Using mean square error as loss function, the parameters of deep neural network model are modified:
[0077] L(θ)=Επ[(yQ(s,a;θ)) 2 ]
[0078] Where y is the target action value function, obtained by the following formula:
[0079] y=R t +γmaxQ(s t+1 ,a t+1 ;θ)
[0080] Among them, γ is a hyperparameter. The smaller the value of γ, the more the agent pays attention to the current reward.
[0081] Finally, we use the gradient descent algorithm to update L(θ) to obtain the optimal Q(s,a;θ) to infer the optimal policy. Therefore, the gradient update is calculated as follows:
[0082]
[0083] In the above-mentioned reinforcement learning and graph neural network algorithms, the ramp reducer fault diagnosis process directly controls the Chebyshev graph convolution base layer by the trained reinforcement learning network, regards the node features as states and inputs them to the reinforcement learning agent, outputs the modification of the number of Chebyshev graph convolution network layers, and finally the Chebyshev graph convolution network gives the detection results.
[0084] The above description is only a preferred embodiment of the present invention and is not intended to limit the concept of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for diagnosing faults of robot core components based on artificial intelligence, characterized in that: include: Step 1: Collect the vibration signal of the harmonic reducer and pre-process it to obtain graph structure data; Step 2: Use 70% of the graph structure data as the training sample set and 30% of the graph structure data as the test sample set; Step 3: Establish a Chebyshev graph convolutional network; Step 4: Input the training sample set into the Chebyshev graph convolutional network, train the network through the reinforcement learning method, and test it using the test set to obtain the final network model; Step 5: The vibration signal of the harmonic reducer collected in real time is input into the trained Chebyshev graph convolutional network for fault diagnosis.
2. The method for diagnosing core robot parts faults based on artificial intelligence according to claim 1, characterized in that: The step 1 is specifically as follows: Step 1.1: Collect vibration signals of four types of harmonic reducers when they are faulty and vibration signals of normal states; Step 1.2: Let G = {F, E, W} represent the space-time graph, where F is the set of nodes in the space-time graph, E represents the edges connecting these nodes, and W represents the weight of the edge; Step 1.3: Use short-time Fourier transform to cut the vibration signal of the harmonic reducer into k segments for feature extraction; Where ΔM is the time difference of a segment, Δf is the frequency difference, m represents time, T(k, m) is the output of frequency kΔf and time mΔM, x[n] is the vibration signal, N is the window length, the operator * represents conjugation, and ω is the window function, specifically the Hanning window is used; Among them, S(k, m) is the output in the frequency interval 1≤k≤K and time 0≤m≤M. As the node feature of the spatiotemporal graph, the frequency F1, F2...F of each segment K Represents a node of a space-time graph; Step 1.4: Use Euclidean distance to calculate the weights between nodes; Where Dis represents the nodes Fi and F at a specific time point m. j The Euclidean distance between them completes the construction of the space-time graph; Step 1.5: In spectral analysis, the symmetric normalized Laplace matrix is defined as follows: Where, L s is the symmetric normalized Laplace matrix, I is the identity matrix, D is the diagonal matrix of the graph, A is the adjacency matrix, K is the number of nodes in the space-time graph, and the real symmetric normalized Laplace matrix L is transformed by using the eigenvalue decomposition in Fourier transform. s Transformed into orthogonal basis vector U, as follows: L s =UΛU T In the formula, Λ=diag{λ1, λ2,…,λ K } is the inclusion matrix L s The eigenvalue matrix of all eigenvalues of each spatiotemporal graph is represented by a one-dimensional eigenvector containing the eigenvalues as P = [λ1, λ2, …, λ K ]; Step 1.6: Define the correlation graph G a = {V a , E a , A}, where V a is the node set of the association graph, E a represents the edges connecting these nodes; each node represents a specific harmonic reducer fault state; A is the adjacency matrix, which represents the connection relationship between any pair of nodes, where the value of each element in the adjacency matrix is 0 or 1, that is: Among them, nodes with the same fault label are connected to each other, while nodes with different fault labels remain unconnected; Step 1.7: Transform the one-dimensional feature vector P = [λ1, λ2, …, λ K ] as the feature of the node in the association graph to complete the construction of the association graph.
3. The method for diagnosing core robot parts faults based on artificial intelligence as claimed in claim 1, characterized in that: The faults in step 1.1 include: fracture fault of the flexible bearing, fatigue fault of the input shaft caused by long-term operation, pitting corrosion fault at the tooth root and flexible wheel tip fault.
4. The method for diagnosing core robot parts faults based on artificial intelligence as claimed in claim 2, characterized in that: The step 3 is specifically as follows: Step 3.1: Graph convolution is represented as: Y=g θ (L s )A Y=Ug θ (Λ)U T AND Among them, g θ is the convolution kernel, Y is the output of the graph convolution, since g in the ordinary graph convolution θ The calculation of (Λ) is complex and consumes a lot of computing power. Therefore, the Chebyshev polynomial expansion is introduced to solve this problem. It is found that by θ (Λ) can be effectively approximated by expanding it into Chebyshev polynomials: T k (Λ)=2ΛT k-1 (L)-T k-2 (L) T0(Λ)=1,T1(Λ)=Λ Among them, θ k is the vector of Chebyshev coefficients, is the normalized form of Λ; Among them, λ max is the largest element of Λ, The value range of the elements in is [-1, 1], so the graph convolution is calculated as follows: Step 3.2: The graph convolution operation is considered as a filtering process, which retains the dimension of the input signal. By multiplying the graph structure data with the trainable weight matrix, the dimension of the feature can be adjusted, that is: Z=Cheb(A,W (b) )=YES (b) Among them, Z is the output of the b-th graph convolutional layer, W (b) is the trainable weight matrix of the b-th graph convolutional layer, Cheb(·) represents the Chebyshev convolution function; the construction of the graph convolutional architecture is expressed as the following transition equation: h 1 =Relu(Cheb(A,W (0) )) h b=a =Relu(Cheb(A,W (0) )) Among them, h b is the feature vector of layer b, and a is the number of Chebyshev graph convolutional network layers determined by the reinforcement learning agent.
5. The method for diagnosing core robot parts faults based on artificial intelligence as claimed in claim 4, characterized in that: The step 4 is specifically as follows: Step 4.1: Establish the training process as a Markov decision process, with the basic elements including state, action and reward; State: At time t, the properties of a node in the current association graph are extracted: s t ={λ1,λ2,…,λ K } Action: The number of Chebyshev graph convolutional network layers decided by the reinforcement learning agent: a t =a Reward function: The reward function is defined as, at time t, action a t The effect of network classification after execution; Among them, M(s t ,a t ) is the accuracy of the graph convolutional network on the training set at time t, L is a hyperparameter used to define the window size of historical performance to be used as a benchmark reference, and β is a hyperparameter that determines the strength of the reward signal; is the average of the most recent L classification accuracies, used to encourage the reinforcement learning agent to achieve better performance compared to the most recent L time steps; Step 4.2: Select the starting node of the association graph, whose node features are regarded as the current state s t Input to the reinforcement learning agent; Step 4.3: The agent generates action a t , represents the number of layers of the Chebyshev graph convolutional network used to identify the node; Step 4.4: The Chebyshev graph convolutional network gives the result of node classification, and the reward value R is calculated using the classification accuracy t , guiding the reinforcement learning agent to make reasonable network choices, and using the experience replay mechanism during training to ensure the long-term effectiveness of training.
Citation Information
Patent Citations
Nuclear power circulating water pump planetary gearbox detection method based on space-time diagram
CN115541232A
Power distribution network reactive power optimization scheduling method and system based on improved multi-agent
CN117977617A
Electroencephalogram feature extraction method based on time sequence diagram structure learning
CN118121211A
Multi-agent federated reinforcement learning-based vehicle-road collaborative control system and method under complex intersection
WO2024016386A1
