Robot end pose error prediction method and equipment based on graph structure dynamic spatio-temporal evolution

Abstract

The invention belongs to the related technical field of robot milling, and discloses a robot end pose error prediction method and device based on graph structure dynamic spatio-temporal evolution, and the method comprises the steps: (1) calculating a correlation index between each joint node feature of a robot and a pose error of a robot end based on training set data; (2) calculating the embedding of each node of the robot by adopting a time convolutional network, and carrying out semantic space alignment; (3) respectively constructing a strong correlation sub-graph and a weak correlation sub-graph, and respectively updating node embedding; (4) calculating cosine similarity between node embedding, and generating a new graph adjacency matrix based on the cosine similarity; fusing the current graph adjacency matrix and the newly generated graph adjacency matrix to obtain an updated graph adjacency matrix; and (5) constructing a global graph convolutional neural network based on the node embedding and updating graph adjacency matrix in the step (2) so as to realize the prediction of the robot end pose error and improve the error prediction precision.

Description

Technical Field

[0001] This invention belongs to the field of robotic milling technology, and more specifically, relates to a method and device for predicting robot end-effector pose error based on graph structure dynamic spatiotemporal evolution. Background Technology

[0002] With their advantages of large workspace, high flexibility, and rapid reconfigurability, industrial robots have rapidly developed, becoming another new manufacturing paradigm after machine tools. However, due to limitations in their body geometry and the influence of coupling factors such as joint clearance, link flexibility, thermal deformation, and dynamic loads, machining posture errors are difficult to control precisely, severely restricting their widespread application in precision-sensitive conditions. Achieving high-precision posture error prediction has become a core challenge in overcoming the bottleneck of robot machining accuracy and improving process reliability, and is urgently needed to promote the autonomous development of intelligent manufacturing equipment.

[0003] In error modeling, existing methods mainly include analytical models based on physical mechanisms and data-driven learning methods. The former focuses on revealing the physical essence and mathematical expression of robot machining pose errors, and has good interpretability. However, its core limitation lies in the complexity of building accurate models, and its practicality and robustness are insufficient when facing highly nonlinear, strong time-varying coupling, and unknown disturbances. As for the latter, existing research methods ignore the time-varying dynamic characteristics of the machining process and lack spatial domain information understanding and interaction. Although there have been some attempts based on graph neural networks, they are limited by predefined spatial distance structures, resulting in limited model prediction accuracy. Summary of the Invention

[0004] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a method and device for predicting robot end-effector pose error based on graph structure dynamic spatiotemporal evolution, which aims to solve the problem of low prediction accuracy of robot end-effector pose error by existing methods.

[0005] To achieve the above objectives, according to one aspect of the present invention, a method for predicting robot end-effector pose error based on graph structure dynamic spatiotemporal evolution is provided, comprising the following steps: (1) The Pearson correlation coefficient between the features of each joint node of the robot and the pose error of the robot end is calculated based on the training set data, and the Pearson correlation coefficient is used as the correlation index. (2) The embeddings of each node of the robot are calculated using a temporal convolutional network, and all the obtained embeddings are aligned in semantic space; (3) Select the node with the highest correlation index as the center node of the strongly correlated subgraph, and select the edge nodes of the strongly correlated subgraph based on their similarity to the center node of the strongly correlated subgraph. Construct a strongly correlated subgraph from the center node of the strongly correlated subgraph to the edge node of the strongly correlated subgraph, including self-loops, based on the center node and edge nodes of the strongly correlated subgraph. The corresponding edge weights adopt the values ​​corresponding to the graph adjacency matrix of the current state. Select the node with the lowest correlation index as the center node of the weakly correlated subgraph, and select the nodes not covered in the strongly correlated subgraph as the edge nodes of the weakly correlated subgraph. Construct a weakly correlated subgraph from the edge nodes of the weakly correlated subgraph to the center node of the weakly correlated subgraph, and adopt the values ​​corresponding to the graph adjacency matrix of the current state for the corresponding edge weights. At the same time, update the node embeddings of the strongly correlated subgraph and the weakly correlated subgraph respectively. (4) Calculate the cosine similarity between node embeddings, and generate a new graph adjacency matrix based on the obtained cosine similarity; merge the current graph adjacency matrix and the newly generated graph adjacency matrix to obtain the updated graph adjacency matrix for the next state; (5) Construct a global graph convolutional neural network based on the node embedding and updated graph adjacency matrix calculated by the temporal convolutional neural network, thereby realizing the prediction of robot end pose error; the output of the global graph convolutional neural network is the robot end pose error.

[0006] Furthermore, the training set data includes node features and robot end-effector pose error; an initial graph adjacency matrix is ​​constructed based on the robot's initial state Euclidean space distance.

[0007] Furthermore, after passing the information of the center node of the strongly correlated subgraph to the edge node of the strongly correlated subgraph, the information of the edge node of the strongly correlated subgraph is then passed to the center node of the strongly correlated subgraph to complete the embedding and updating of the strongly correlated subgraph node.

[0008] Furthermore, after passing the information of the edge nodes of the weakly correlated subgraph to the center node of the weakly correlated subgraph, the information of the center node of the weakly correlated subgraph is then passed to the edge nodes of the weakly correlated subgraph to update the node embedding of the weakly correlated subgraph.

[0009] Furthermore, a subgraph convolutional neural network is used to update the node embeddings of strongly correlated and weakly correlated subgraphs; after obtaining the global graph convolutional neural network, the regression loss is calculated, and the regression loss and alignment loss are weighted and merged to obtain the total loss. Backpropagation is performed based on the total loss to update the model parameters of the temporal convolutional neural network, the subgraph convolutional neural network, and the global graph convolutional neural network.

[0010] Furthermore, the formula for constructing a strongly correlated subgraph is:

[0011]

[0012]

[0013]

[0014] In the formula, and Representing samples respectively Strongly correlated nodes and The node embedding after the node has been computed by the temporal convolutional network; Indicates the first In a sample, strongly correlated nodes and nodes Euclidean distance; yes The reverse process is to sort the corresponding node indices by ascending distance value to obtain... , For its first One element, This represents the number of nodes in the strongly correlated subgraph (including strongly correlated nodes). Indicates the first A strongly correlated subgraph constructed from individual samples. These represent the strongly correlated subgraph node embedding representation, topological relations, and graph adjacency matrix, respectively.

[0015] Furthermore, the formula corresponding to the embedding update of strongly correlated subgraph nodes is:

[0016]

[0017] In the formula, For graph convolutional layers, after The activation function obtains the embeddings of each node after the central broadcast. ; This represents the embedding of each node after the edge nodes are aggregated. Used for reverse edge connection.

[0018] Furthermore, the formula for constructing a weakly correlated subgraph is:

[0019]

[0020]

[0021]

[0022] in Represents a weakly correlated subgraph. For the sample The set of node indices for strongly correlated subgraphs. For the sample Weakly correlated nodes This represents the complement operation. For the sample The index of edge nodes in the weakly correlated subgraph. An ordered list of node indices representing a weakly correlated subgraph. These represent the weakly correlated subgraph node embedding representation, topological relations, and graph adjacency matrix, respectively.

[0023] The formula for updating the embedding of weakly correlated subgraph nodes is:

[0024]

[0025] In the formula, This represents the embedding of each node after aggregating the edge node information of a weakly correlated subgraph; This indicates that the information of the central node of a weakly correlated subgraph is broadcast and then embedded by each node.

[0026] The present invention also provides a robot end-effector pose error prediction system based on graph structure dynamic spatiotemporal evolution. The system includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it performs the robot end-effector pose error prediction method based on graph structure dynamic spatiotemporal evolution as described above.

[0027] The present invention also provides a computer-readable storage medium storing machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the robot end-effector pose error prediction method based on graph structure dynamic spatiotemporal evolution as described above.

[0028] In summary, compared with the prior art, the robot end-effector pose error prediction method and device based on graph structure dynamic spatiotemporal evolution provided by the present invention have the following advantages: 1. The prediction method provided by this invention, through the synergy of temporal convolutional networks and graph convolutional networks, achieves comprehensive perception and deep fusion of spatiotemporal features during robot machining. The temporal convolutional network, with its dilated convolutional structure, can effectively capture the temporal dependencies in multi-joint motion sequences and accurately extract the dynamic characteristics of joint states evolving over time. Furthermore, the designed neural network model is not limited to traditional predefined spatial graph structures, but instead constructs a graph evolution method based on semantic space alignment constraints and a bidirectional message passing mechanism, significantly enhancing the representation ability of complex spatial coupling relationships between joints. The designed spatiotemporal dual-view learning architecture enables the model to simultaneously capture the temporal dynamics and spatial correlations of the machining system, thereby comprehensively improving the accuracy and robustness of pose error prediction.

[0029] 2. The prediction method provided by this invention performs deep fusion modeling of the multi-source temporal coupling relationship and spatial structure topology in a robotic machining system, inferring pose errors from joint machining states, which is of great significance for ensuring the overall accuracy of robotic machining. In the machining trajectory planning stage, this method can predict the error distribution under different machining toolpaths, providing prior knowledge. In the actual machining stage, on-site operators can monitor the pose error of the robot's end effector, indirectly guiding adjustments to system configuration and task parameters in subsequent machining, or designing corresponding error compensation mechanisms to achieve control and assurance of the final machining quality.

[0030] 3. This invention provides a graph structure dynamic evolution strategy. By constructing a mapping relationship between robot processing temporal features and pose errors, a correlation-driven spatiotemporal graph evolution neural network model is established, which realizes accurate and reliable prediction of robot pose errors during processing. This method can infer robot end-effector pose errors based on robot processing data, thereby providing key data support for error compensation control.

[0031] 4. The prediction method provided by this invention can provide important support for robot body structure optimization and critical component health management. By analyzing the correlation indicators and graph adjacency matrix obtained from model learning, it provides criteria for identifying critical joints and weak links with the greatest correlation to end-effector pose error; while the evolution law of the graph structure during processing describes the impact of working condition parameters such as load and pose on accuracy characteristics. This deeply mined physical information can guide robot configuration optimization and the selection and maintenance strategies of critical components, realizing the improvement of inherent accuracy potential from the design stage and the implementation of predictive health maintenance from the operation and maintenance stage, comprehensively ensuring the long-term accuracy stability and service life of the robot processing system. Attached Figure Description

[0032] Figure 1 This is a flowchart of a robot end-effector pose error prediction method based on graph structure dynamic spatiotemporal evolution provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of the robot's end effector stiffness ellipsoid; Figure 3 This is a schematic diagram illustrating the division of the time-based sliding window; Figure 4 This is a schematic diagram illustrating the calculation of correlation indicators; Figure 5 This is a schematic diagram of the embedding computation and semantic alignment of each node based on the temporal convolutional network; Figure 6 This is a schematic diagram illustrating the calculation of the dynamic evolution strategy and error of the graph structure; Figure 7 (a), (b), and (c) in the diagram are detail images of the strongly / weakly correlated subgraph convolutional network and the global graph convolutional network, respectively. Figure 8 This is a flowchart of the inference application based on the spatiotemporal graph evolutionary neural network model. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0034] Please see Figure 1 This invention provides a method for predicting robot end-effector pose error based on dynamic spatiotemporal evolution of graph structure. The prediction method breaks through the bottleneck of traditional methods being limited by the insufficient applicability of predefined spatial graph structures to complex working conditions. By comprehensively sensing the relationship between changes in processing state and spatial structure, it deeply explores the error transmission mechanism and provides accuracy assurance for robotized intelligent manufacturing.

[0035] The prediction method mainly includes the following steps: Step 1: Plan the source domain processing task and construct training set data, which includes node features and robot end-effector pose error.

[0036] In one implementation, based on an established robot machining and measurement experimental platform and part model, and relying on industrial CAM software, the robot source domain machining task is planned, and experimental process data is collected. The actual pose and error of the robot tool end are calculated, and a training dataset is constructed after time sliding window partitioning.

[0037] Taking the Stäubli TX2-90L robot as an example, using a Leica AT960LR laser tracker and an end effector six-dimensional force sensor, the specific process is as follows: S11, Training Dataset Construction. Based on the integrated robot machining-measurement system, the directly obtainable machining process data includes: the robot's six joint angle sequences. Robot joint torque Terminal six-dimensional force and system sampling period .

[0038] Calculate the robot joint velocity sequence based on the following formula. , acceleration sequence :

[0039]

[0040] In the formula, The total number of samples, Identify the six joints of the robot.

[0041] The diagonal matrix of robot joint stiffness can be obtained through stiffness identification:

[0042] To characterize the stiffness characteristics of the robot's end effector in different directions, an end effector stiffness ellipsoid is introduced, such as... Figure 2 As shown, the principal axis lengths of the stiffness ellipsoid are calculated as follows:

[0043]

[0044]

[0045] in For the first Jacobian matrix in the base coordinate system of each sample robot Here is the end stiffness matrix. for Eigenvectors. Solving the above equation yields the principal direction stiffness values ​​of the robot's end effector ellipsoid. .

[0046] The initial feature set for the source domain task is constructed as follows:

[0047] For the sake of simplicity, the source domain set identifier will be omitted below. The maximum-minimum normalization method is applied to each feature column as follows:

[0048] in These represent calculating the maximum and minimum values ​​of the sequence, respectively. At this point, the feature dataset of the processing procedure can be obtained. .

[0049] S12, Calculation of tool end-effector pose error. For each sample... First, according to the robot's forward kinematics, the theoretical representation of the flange in the robot's base coordinate system can be described as follows:

[0050]

[0051] in Let be the positive kinematics function of the robot. .

[0052] Therefore, the theoretical pose of the robot tool tip in the base coordinate system is described as follows:

[0053]

[0054] In the formula, and For the first The theoretical position and orientation of the end of each sample tool. This represents the tool in the robot flange coordinate system. Let be a rotation matrix. The position coordinate vector, Let be the transformation function from the homogeneous transformation matrix to the Euler angle.

[0055] Subsequently, based on the measurement results from the laser tracker, the description of the T-Mac reflector mounted on the robot in the world coordinate system can be obtained. Furthermore, the true pose description of the tool in the robot's base coordinate system is obtained as follows:

[0056]

[0057] In the formula, and These are descriptions of the world coordinate system in the robot's base coordinate system and the tool in the end-effector T-Mac coordinate system, respectively. and The first The actual position and orientation of the tool during the machining process for each sample. All the homogeneous transformation matrices mentioned above can be obtained through the prior calibration of the machining system coordinate system.

[0058] Therefore, the six-dimensional pose error label of the robot's end-effector during the machining process can be obtained as follows:

[0059] The dataset features and labels are concatenated to obtain the source domain training set data. :

[0060] S13, Time Sliding Window Partitioning. To enable the model to effectively capture the dynamic temporal characteristics of the machining process, time sliding window partitioning is used to serialize the continuously acquired robot state data, enhancing the model's ability to represent the dynamic characteristics of the machining process and the accuracy of error prediction. Figure 3 As shown. (The following is a list of numbers / items / etc.) Column features and the first The first of the dimensional tags The sample sequences were divided using a time sliding window and calculated as follows:

[0061] In the formula, , To set the length of the time window.

[0062] Step 2: Calculate the Pearson correlation coefficient between the features of each joint node of the robot and the pose error of the robot's end effector based on the training set data, and use the Pearson correlation coefficient as the correlation index; at the same time, construct the initial graph adjacency matrix based on the Euclidean space distance of the robot's initial state.

[0063] Based on the physical correlation between joint motion and end-effector pose error during robot machining, to quantify the influence of each joint node on the error, the Pearson correlation coefficient between the characteristics of each robot joint node (joint angle, joint velocity, joint acceleration, joint torque) and the six-dimensional pose error of the robot's end effector is calculated as an indicator of node correlation. Figure 4 As shown; then the first The joint in the first The correlation index for each sample is defined as follows:

[0064] In the formula, For the first Joints in a sample Correlation indicators; This indicates the calculation of covariance. Indicates the standard deviation of the sequence; superscript The indicator has been divided into time-based sliding window segments; The first time window after time sliding window division Column sequence characteristics; The first time window after time sliding window division 3D pose error sequence.

[0065] Subsequently, the corresponding joint numbers are sorted in descending order according to the correlation index. This process is described as follows:

[0066] In the formula, Sort the data in descending order and return the index; For the first The joint index is obtained by sorting the samples in descending order based on their correlation index. For the first Correlation indices for each joint in each sample.

[0067] The robot pose is defined as follows during the initial processing state:

[0068] Based on the principle of spatial proximity, an initial graph adjacency matrix is ​​established based on the Euclidean spatial distances of each joint in the robot's initial processing posture:

[0069]

[0070]

[0071] in, , representing the node sequence number; Represents a node Description in the robot's base coordinate system; Let be a rotation matrix. The coordinates of the nodes; In order to be in Calculation of joint angle to joint The positive kinematic function, Represents the initial node distance matrix; Use inverse maximum-min normalization to make larger elements approach 0.1 and smaller elements approach 1; This is the initial graph adjacency matrix.

[0072] Step 3: Use a temporal convolutional network to calculate the embeddings of each node of the robot, and align all the obtained embeddings in the semantic space.

[0073] Robotic processing exhibits significant time-dependent characteristics. Temporal Convolutional Networks (TCNs), through their causal dilated convolution mechanism, can obtain rich receptive fields without causing information leakage. This can be described as follows:

[0074] in Represents the convolution operator. As the expansion factor, according to As the number of layers increases exponentially, it is used to expand the receptive field. It is the size of the filter. The first one in the filter Element-wise operator functions. It is the complete input sequence The front of the middle There are [number] elements. The length of the effective time series historical data is [number]. However, larger receptive fields require deeper network structures.

[0075] To prevent gradient vanishing in temporal convolutional networks, residual blocks and ResNet-like weight normalization are added to the backbone structure. The output features can be represented as:

[0076] in, It is an activation function, used , This is the model training parameter matrix. Similarly, after several convolutional layers, a fully connected layer is added. Obtain the output mapping.

[0077] In this implementation, seven temporal convolutional neural networks are instantiated in parallel to extract the temporal embeddings of each node and project them into the output. The KL divergence alignment loss is used to constrain the embeddings of heterogeneous node 7 to a unified semantic space with those of other nodes. Specifically: Seven temporal convolutional neural networks are implemented in parallel to extract the temporal characteristics of each node. The corresponding formula is as follows:

[0078]

[0079] In the formula, the end point is node 7, and its features are heterogeneous to joints 1-6 of the robot body; Indicates the first Initial sequence features of each node, These are the joint angle, joint velocity, joint acceleration, joint torque, end-effector six-dimensional force, and three-dimensional stiffness sequence characteristics after time-window partitioning. This represents the temporal embedding of TCN computation.

[0080] The output of the temporal convolutional network Access to projection layer After one layer With linear layer Then, the output of the projection layer is obtained. And using KL divergence Embedding of heterogeneous node 7 with other nodes into a unified semantic space, aligning the loss. The calculation is as follows:

[0081]

[0082] in For node sequence number, The dimension of the node embedding; the overall process of calculating the temporal embedding of TCN nodes, projection layer mapping, and semantic space constraint alignment loss is as follows: Figure 5 As shown.

[0083] Step 4: Select the node with the highest correlation index as the center node of the strongly correlated subgraph, and select edge nodes of the strongly correlated subgraph based on their similarity to the center node. Construct a strongly correlated subgraph from the center node to the edge node of the strongly correlated subgraph, including self-loops, based on the center node and edge nodes of the strongly correlated subgraph. The corresponding edge weights are adopted from the graph adjacency matrix values ​​of the current state. Select the node with the lowest correlation index as the center node of the weakly correlated subgraph, and select the nodes not covered in the strongly correlated subgraph as edge nodes of the weakly correlated subgraph. Construct a weakly correlated subgraph from the edge nodes to the center node of the weakly correlated subgraph, including self-loops, with the corresponding edge weights adopted from the graph adjacency matrix values ​​of the current state. At the same time, update the node embeddings of the strongly correlated and weakly correlated subgraphs using center broadcast-edge aggregation and edge aggregation-center broadcast message passing mechanisms, respectively.

[0084] After passing the information of the center node of the strongly correlated subgraph to the edge node of the strongly correlated subgraph, the information of the edge node of the strongly correlated subgraph is then passed to the center node of the strongly correlated subgraph to complete the embedding and updating of the strongly correlated subgraph node.

[0085] After passing the information of the edge nodes of the weakly correlated subgraph to the center node of the weakly correlated subgraph, the information of the center node of the weakly correlated subgraph is then passed to the edge nodes of the weakly correlated subgraph to update the node embedding of the weakly correlated subgraph.

[0086] In one implementation, a subgraph convolutional neural network is used to update the node embeddings of strongly correlated and weakly correlated subgraphs. The specific steps are as follows: S51, Graph structure established.

[0087] The robot machining process diagram is structured as follows:

[0088] In the formula, Indicates node embedding. This represents the initial graph adjacency matrix, which is iteratively updated during model training. This indicates that the seven nodes are connected to each other (including self-loops).

[0089] S52, Construction of strongly correlated subgraphs.

[0090] Based on the calculated correlation index results, for each sample, the node with the highest correlation index is selected as the center node of the strongly correlated subgraph:

[0091] The center node of a strongly correlated subgraph is the central node, and its embedding is calculated using Euclidean distance. Based on the homogeneity assumption that similar nodes are more likely to establish effective connections, and considering the similarity embeddings of each node, the similarity value is taken before... Each node is an edge node of a strongly correlated subgraph. A path is established from the center node to... One-way edges are established for each node, and self-loops are created for each node to construct a strongly correlated subgraph. The edge weights are based on the graph adjacency matrix of the current state, and the corresponding formula is:

[0092]

[0093]

[0094]

[0095] In the formula, and Representing samples respectively Strongly correlated nodes and The node embedding after the node has been computed by the temporal convolutional network; Indicates the first In a sample, strongly correlated nodes and nodes Euclidean distance; yes The reverse process is to sort the corresponding node indices by ascending distance value to obtain... , For its first One element, This represents the number of nodes in the strongly correlated subgraph (including strongly correlated nodes). Indicates the first A strongly correlated subgraph constructed from individual samples. These represent the strongly correlated subgraph node embedding representation, topological relations, and graph adjacency matrix, respectively.

[0096] S53, a strongly correlated subgraph broadcast-aggregation message passing mechanism.

[0097] The original features of the center node of a strongly correlated subgraph have a greater correlation with the pose error and contain more usable potential information. Through broadcasting from the center node, the information of the strongly correlated node is passed to the neighboring nodes of the subgraph, completing the embedding update of the neighboring nodes. The corresponding formula is:

[0098] In the formula, For graph convolutional layers, after The activation function obtains the embeddings of each node after the central broadcast. In strongly correlated subgraphs, the embeddings of neighboring nodes and the embedding of the central node exhibit certain similarities. Edge aggregation can be used to effectively integrate neighboring node embeddings into the central node embedding, thus updating the central node embedding and refining the extraction of effective information. This can be described as follows:

[0099] In the formula, This represents the embedding of each node after the edge nodes are aggregated. Used for reverse edge connection.

[0100] S54, Construction of weakly correlated subgraphs.

[0101] The selection of weakly correlated nodes follows the same process as S52, choosing the node with the lowest correlation index in the sample, i.e.:

[0102] Using the selected weakly correlated node as the center node, select nodes that are not covered in the strongly correlated subgraph, establish one-way edges to the center node, and establish self-loops for each node to construct the weakly correlated subgraph. The edge weights adopt the corresponding values ​​of the adjacency matrix of the current state graph. The construction process of the weakly correlated subgraph is described as follows:

[0103]

[0104]

[0105]

[0106] in Represents a weakly correlated subgraph. For the sample The set of node indices for strongly correlated subgraphs. For the sample Weakly correlated nodes This represents the complement operation. For the sample The index of edge nodes in the weakly correlated subgraph. An ordered list of node indices representing a weakly correlated subgraph. These represent the weakly correlated subgraph node embedding representation, topological relations, and graph adjacency matrix, respectively.

[0107] S55, weakly correlated subgraph aggregation-broadcast message passing mechanism.

[0108] The original features of weakly correlated nodes have low correlation with pose errors, resulting in limited potentially usable information. Firstly, by aggregating edge node information, the embedding information of neighboring nodes is passed to the central node:

[0109] in, This indicates the embedding of each node after aggregating the edge node information of the weakly correlated subgraph.

[0110] Subsequently, through central broadcasting, the embedding of the central node carrying information about all neighboring nodes is effectively transmitted to each edge node, completing message transmission with the weakly correlated central node as a bridge. At the same time, the influence of the central node is weakened, and the embedding of each edge node is updated. This process can be described as follows:

[0111] in, This indicates that the information of the central node of a weakly correlated subgraph is broadcast and then embedded by each node.

[0112] S56, Node Embedding Update Merge.

[0113] The message passing processes implemented by S53 and S55 have completed the embedding update on each subgraph and are merged according to the following formula:

[0114] in, Indicates the node sequence number.

[0115] Step 5: Calculate the cosine similarity between node embeddings, and generate a new graph adjacency matrix based on the obtained cosine similarity; merge the current graph adjacency matrix and the newly generated graph adjacency matrix to obtain the updated graph adjacency matrix for the next state.

[0116] Based on the updated node embeddings, calculate the cosine similarity between the node embeddings and generate a new graph adjacency matrix:

[0117]

[0118] in, Indicates the first In a sample, the node Normalized embedding, Represents the nodes of the newly generated graph adjacency matrix arrive The right to the side.

[0119] Subsequently, the current graph adjacency matrix is ​​merged. Adjacency matrix of new graph generated based on embedding The updated graph adjacency matrix for the next state is obtained. .

[0120]

[0121] in The weighting coefficient controls the graph update speed by adjusting the fusion ratio between the current graph adjacency matrix and the generated new graph adjacency matrix.

[0122] Furthermore, the overall framework for strong / weak subgraph message passing and graph structure updates is as follows: Figure 6 As shown, Figure 7 Images (a), (b), and (c) illustrate the implementation details of the strong and weak subgraph reciprocal message passing mechanism and the global graph convolutional network, respectively. The central broadcast and edge aggregation mechanisms are respectively implemented using... and express.

[0123] Step 6: Construct a global graph convolutional neural network based on the node embeddings and updated graph adjacency matrix obtained by the temporal convolutional neural network, thereby realizing the prediction of robot end-effector pose error; the output of the global graph convolutional neural network is the robot end-effector pose error.

[0124] The temporal convolutional neural network, the subgraph convolutional neural network, and the global graph convolutional neural network constitute a spatiotemporal graph evolutionary neural network model. After obtaining the global graph convolutional neural network, the regression loss is calculated, and the regression loss and alignment loss are weighted and combined to obtain the total loss. Backpropagation is performed based on the total loss to update the model parameters of the temporal convolutional neural network, the subgraph convolutional neural network, and the global graph convolutional neural network.

[0125] During the training of the spatiotemporal graph evolutionary neural network model, if the predetermined number of training iterations is reached, the training is stopped; otherwise, the process proceeds to step three until the training termination condition is met.

[0126] In one implementation, a global graph convolutional neural network (GCN) is constructed based on the embedded and updated graph adjacency matrix of each node. This network is used to predict end-effector pose errors on the training set and to calculate the regression loss. The process is described as follows:

[0127]

[0128]

[0129]

[0130] The GCN network uses a global average pooling strategy for computation. , GCN depth; Indicates the prediction result. Labels for the source domain dataset. This indicates the calculation of the mean squared error.

[0131] The regression loss and alignment loss are weighted and combined to obtain the total loss. Backpropagation is performed to update the spatiotemporal graph evolution neural network model.

[0132]

[0133] in To lose weight, The regularization coefficient is . These are the model parameters.

[0134] Simultaneously using mean absolute error Root mean square error As evaluation metrics for model performance, the corresponding calculation formulas are as follows:

[0135]

[0136] in, For predicted values, This is the actual value.

[0137] When applying task reasoning: Target domain task planning and test set construction: First, plan the target domain processing task and collect the raw data of the robot's processing. Then, apply the source domain task features to process the target domain task data and obtain a data structure usable by the spatiotemporal graph evolutionary neural network model. This process is similar to S1 and can be described as follows:

[0138]

[0139] in, Identify the target domain. For the constructed test task dataset, This represents the number of samples in the test set.

[0140] Application of robot end-effector pose error inference based on spatiotemporal graph evolutionary neural network model: After model training, the spatiotemporal graph evolutionary neural network model is saved as PTH and ONNX format files, enabling inference applications and model deployment compatible with Python and other computing platforms and environments. Furthermore, the strong / weak subgraph module is only used during training to calculate and update the graph adjacency matrix; in inference applications, the trained graph adjacency matrix is ​​used directly, thus eliminating the need for this component. The inference process is illustrated below. Figure 8 As shown.

[0141] Specifically, the spatiotemporal graph evolutionary neural network model is loaded first, along with the newly trained graph adjacency matrix and other necessary parameters; then, the temporal convolutional network of each node is applied to calculate the temporal embedding of each node.

[0142]

[0143] in, Number the nodes; Using a new graph adjacency matrix, a global graph convolutional network is applied to predict samples. The robot end-effector pose error is calculated as follows:

[0144] This invention defines the problem of end-effector pose error prediction in robot machining as a neural network learning problem based on dynamic spatiotemporal evolution of graph structures. The method establishes a correlation-driven spatiotemporal graph evolution neural network model by constructing a complex mapping relationship between the temporal motion features of robot multi-joints and the end-effector pose error. Specifically, the model first uses a temporal convolutional network to extract the temporal features of each joint motion sequence and ensures the consistency of the feature semantic space through constraint alignment. Then, based on pre-calculated correlation indices, strongly correlated subgraphs and weakly correlated subgraphs are constructed, and bidirectional reciprocal message passing mechanisms of center broadcast-edge aggregation and edge aggregation-center broadcast are used respectively to achieve collaborative updates of node embeddings. Furthermore, the cosine similarity of node embeddings is used to calculate the dynamic evolution graph structure, and finally, a global graph convolutional network is used to achieve accurate regression prediction of pose errors. This method, through the deep integration of physical priors and data-driven approaches, overcomes the limitations of traditional predefined spatial graph neural networks in adapting to complex machining conditions, providing reliable technical support for high-precision robot machining.

[0145] The present invention also provides a robot end-effector pose error prediction system based on graph structure dynamic spatiotemporal evolution. The system includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it performs the robot end-effector pose error prediction method based on graph structure dynamic spatiotemporal evolution as described above.

[0146] The present invention also provides a computer-readable storage medium storing machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the robot end-effector pose error prediction method based on graph structure dynamic spatiotemporal evolution as described above.

[0147] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting robot end-effector pose error based on graph structure dynamic spatiotemporal evolution, characterized in that, The steps are as follows: (1) The Pearson correlation coefficient between the features of each joint node of the robot and the pose error of the robot end is calculated based on the training set data, and the Pearson correlation coefficient is used as the correlation index. (2) The embeddings of each node of the robot are calculated using a temporal convolutional network, and all the obtained embeddings are aligned in semantic space; (3) Select the node with the highest correlation index as the center node of the strongly correlated subgraph, and select the edge nodes of the strongly correlated subgraph based on their similarity to the center node of the strongly correlated subgraph. Construct a strongly correlated subgraph from the center node of the strongly correlated subgraph to the edge node of the strongly correlated subgraph, including self-loops, based on the center node and edge nodes of the strongly correlated subgraph. The corresponding edge weights adopt the values ​​corresponding to the graph adjacency matrix of the current state. Select the node with the lowest correlation index as the center node of the weakly correlated subgraph, and select the nodes not covered in the strongly correlated subgraph as the edge nodes of the weakly correlated subgraph. Construct a weakly correlated subgraph from the edge nodes of the weakly correlated subgraph to the center node of the weakly correlated subgraph, and adopt the values ​​corresponding to the graph adjacency matrix of the current state for the corresponding edge weights. At the same time, update the node embeddings of the strongly correlated subgraph and the weakly correlated subgraph respectively. (4) Calculate the cosine similarity between node embeddings, and generate a new graph adjacency matrix based on the obtained cosine similarity; merge the current graph adjacency matrix and the newly generated graph adjacency matrix to obtain the updated graph adjacency matrix for the next state; (5) Construct a global graph convolutional neural network based on the node embedding and updated graph adjacency matrix calculated by the temporal convolutional neural network, thereby realizing the prediction of robot end pose error; the output of the global graph convolutional neural network is the robot end pose error.

2. The robot end-effector pose error prediction method based on graph structure dynamic spatiotemporal evolution as described in claim 1, characterized in that: The training set data includes node features and robot end-effector pose error; an initial graph adjacency matrix is ​​constructed based on the robot's initial state Euclidean space distance.

3. The robot end-effector pose error prediction method based on graph structure dynamic spatiotemporal evolution as described in claim 1, characterized in that: After passing the information of the center node of the strongly correlated subgraph to the edge node of the strongly correlated subgraph, the information of the edge node of the strongly correlated subgraph is then passed to the center node of the strongly correlated subgraph to complete the embedding and updating of the strongly correlated subgraph node.

4. The robot end-effector pose error prediction method based on graph structure dynamic spatiotemporal evolution as described in claim 1, characterized in that: After passing the information of the edge nodes of the weakly correlated subgraph to the center node of the weakly correlated subgraph, the information of the center node of the weakly correlated subgraph is then passed to the edge nodes of the weakly correlated subgraph to update the node embedding of the weakly correlated subgraph.

5. The robot end-effector pose error prediction method based on graph structure dynamic spatiotemporal evolution as described in claim 1, characterized in that: A subgraph convolutional neural network is used to update the embeddings of each node in strongly correlated and weakly correlated subgraphs. After obtaining the global graph convolutional neural network, the regression loss is calculated, and the regression loss and alignment loss are weighted and merged to obtain the total loss. Backpropagation is performed based on the total loss to update the model parameters of the temporal convolutional neural network, the subgraph convolutional neural network, and the global graph convolutional neural network.

6. The robot end-effector pose error prediction method based on graph structure dynamic spatiotemporal evolution as described in claim 1, characterized in that: The formula for constructing a strongly correlated subgraph is: In the formula, and Representing samples respectively Strongly correlated nodes and The node embedding after the node has been computed by the temporal convolutional network; Indicates the first In a sample, strongly correlated nodes and nodes Euclidean distance; yes The reverse process is to sort the corresponding node indices by ascending distance value to obtain... , For its first One element, This represents the number of nodes in the strongly correlated subgraph, including strongly correlated nodes. Indicates the first A strongly correlated subgraph constructed from individual samples. These represent the strongly correlated subgraph node embedding representation, topological relations, and graph adjacency matrix, respectively.

7. The robot end-effector pose error prediction method based on graph structure dynamic spatiotemporal evolution as described in claim 6, characterized in that: The formula for updating the embedding of strongly correlated subgraph nodes is: In the formula, For graph convolutional layers, after The activation function obtains the embeddings of each node after the central broadcast. ; This represents the embedding of each node after the edge nodes are aggregated. Used for reverse edge connection.

8. The robot end-effector pose error prediction method based on graph structure dynamic spatiotemporal evolution as described in claim 1, characterized in that: The formula for constructing a weakly correlated subgraph is: in Represents a weakly correlated subgraph. For the sample The set of node indices for strongly correlated subgraphs. For the sample Weakly correlated nodes This represents the complement operation. For the sample The index of edge nodes in the weakly correlated subgraph. An ordered list of node indices representing a weakly correlated subgraph. These represent the weakly correlated subgraph node embedding representation, topological relations, and graph adjacency matrix, respectively. The formula for updating the embedding of weakly correlated subgraph nodes is: In the formula, This represents the embedding of each node after aggregating the edge node information of a weakly correlated subgraph; This indicates that the information of the central node of a weakly correlated subgraph is broadcast and then embedded by each node.

9. A robot end-effector pose error prediction system based on graph structure dynamic spatiotemporal evolution, characterized in that: The system includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it performs the robot end-effector pose error prediction method based on graph structure dynamic spatiotemporal evolution as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the robot end-effector pose error prediction method based on graph structure dynamic spatiotemporal evolution as described in any one of claims 1-8.

Images