CN-122008261-A - Industrial robot motion trail obstacle avoidance method based on deep learning

CN122008261ACN 122008261 ACN122008261 ACN 122008261ACN-122008261-A

Abstract

The invention discloses a deep learning-based industrial robot motion track obstacle avoidance method, which relates to the technical field of industrial robots and comprises the steps of extracting a position sequence, a speed sequence and an acceleration sequence from track data of an industrial robot motion track, respectively generating a geometric feature vector and a dynamic feature vector through a convolutional neural network, and fusing the geometric feature vector and the dynamic feature vector to obtain an initial representation vector.

Inventors

LI JUNHONG
LI HAO
TANG SHIXIONG
LIU CHAO
QIAN JIANXING
Yao Chufeng

Assignees

广安职业技术学院

Dates

Publication Date: 20260512
Application Date: 20260415

Claims (10)

1. The obstacle avoidance method for the motion trail of the industrial robot based on deep learning is characterized by comprising the following steps of: s1, extracting a position sequence, a speed sequence and an acceleration sequence from track data of an industrial robot motion track, respectively generating a geometric feature vector and a dynamic feature vector through a convolutional neural network, and fusing the geometric feature vector and the dynamic feature vector to obtain an initial representation vector; s2, calculating the synchronization degree between local curvature change and acceleration peak value, the space coincidence density of curvature extreme points and speed abrupt points, the track torsion rate gradient and the direction consistency of acceleration change rate vector according to the initial representation vector, and judging that the geometric shape and the dynamics characteristic have a strong coupling relation if the weighted average value of the three indexes exceeds a preset threshold value; S3, extracting a geometric dynamic coupling main direction which keeps the maximum variance, a combined main component axis in the cross-modal information set and a subspace which minimizes the combined reconstruction error of geometric distortion and dynamic distortion after projection by adopting main component analysis according to the initial representation vector which is judged to be strongly coupled, and obtaining a combined characteristic representation after dimension reduction; S4, restoring the industrial robot track sample point sequence from the dimension-reduced joint characteristic representation, inputting a time sequence rule corresponding to the matching degree of the curvature zero crossing frequency and the speed zero crossing frequency captured by the long-term and short-term memory network and the matching trend of the bending section length distribution and the acceleration amplitude statistical distribution, and obtaining the track dynamics mode vector.
2. The method for avoiding obstacle on motion trajectories of industrial robots based on deep learning as set forth in claim 1, wherein S1 comprises: track data in the operation process of the industrial robot is acquired, and the track data are analyzed to obtain a position sequence, a speed sequence and an acceleration sequence; The position sequence is input into a space feature extraction convolutional neural network to generate a geometric feature vector, and the velocity sequence and the acceleration sequence are input into a dynamic feature extraction convolutional neural network to generate a dynamic feature vector; And performing dimension mapping and feature stitching on the geometric feature vector and the dynamic feature vector, and fusing to obtain an initial representation vector containing complete space-time information.
3. The method for avoiding obstacle on motion trajectories of industrial robots based on deep learning as set forth in claim 1, wherein S2 comprises: Analyzing the initial representation vector to obtain a geometric feature component and a dynamic feature component, and extracting a local curvature change sequence and an acceleration peak sequence from the geometric feature component and the dynamic feature component; Calculating the synchronous degree value of the local curvature change sequence and the acceleration peak value sequence, and determining the space coincidence density of the curvature extreme point and the speed abrupt change point according to the synchronous degree value; And calculating a direction consistency index based on the space coincidence density, and judging that the geometric shape and the dynamic characteristic have a strong coupling relation if the comprehensive coupling coefficient generated by the synchronization degree value, the space coincidence density and the direction consistency index meets the condition.
4. The method for avoiding obstacle on motion trajectories of industrial robots based on deep learning as set forth in claim 1, wherein the step S3 comprises: Acquiring an initial representation vector screened out through strong coupling judgment, and constructing a joint covariance matrix by using the initial representation vector; Decomposing the joint covariance matrix to extract a geometric power coupling principal direction retaining the maximum variance, and identifying a joint principal component axis in the cross-modal information set; constructing an initial subspace based on the geometric power coupling main direction and the combined main component axis, and optimizing to obtain a target subspace by minimizing a combined reconstruction error function; and mapping the initial representation vector to a target subspace to obtain the reduced-dimension joint feature representation.
5. The method for avoiding obstacle on motion trajectories of industrial robots based on deep learning as set forth in claim 1, wherein S4 comprises: acquiring a combination characteristic representation after dimension reduction, and reconstructing through back projection to obtain an industrial robot track sample point sequence; Calculating the frequency matching degree of the curvature and the zero crossing frequency of the speed according to the sample point sequence of the industrial robot track, and the distribution coincidence trend of the length of the bending section and the acceleration amplitude; combining the frequency matching degree with the distribution matching trend to generate a time sequence characteristic sequence, and inputting the time sequence characteristic sequence into a long-term and short-term memory network; and obtaining a hidden layer state vector output by the long-short-period memory network, and mapping the hidden layer state vector to obtain a track dynamics mode vector.
6. The method for avoiding obstacle on the basis of deep learning of industrial robot motion trajectories according to claim 1, further comprising S5 calculating a corresponding score of a path fractal dimension and a kinetic energy dissipation rate for a trajectory dynamics pattern vector and a geometry feature vector, and if the corresponding score is lower than a preset threshold, iteratively adjusting key control point positions in a sample point sequence of the industrial robot trajectory through gradient descent until the adjusted sequence meets a predefined kinematic continuity constraint, specifically comprising: Obtaining a track dynamics mode vector and a geometric feature vector, and respectively calculating a path fractal dimension and a dynamics energy dissipation rate; Calculating the mutual information value of the fractal dimension of the path and the kinetic energy dissipation rate as a corresponding relation score; if the corresponding relation score is lower than a preset threshold value, updating the position of the control point along the gradient vector of the maximized corresponding relation score so as to reconstruct a corrected industrial robot track sample point sequence; And performing third-order differentiation on the corrected industrial robot track sample point sequence to obtain a jerk sequence, and if the jerk sequence meets the smoothness limit value, determining that the adjusted sequence meets the predefined kinematic continuity constraint.
7. The method for avoiding the obstacle on the basis of the motion track of the deep learning of the industrial robot according to claim 6, further comprising the steps of S6, according to the adjusted industrial robot track sequence meeting the constraint of kinematic continuity, combining an obstacle model in an industrial robot working space, calculating interference probability caused by local curvature change and acceleration peak synchronization degree in a potential obstacle region, and searching for an alternative node sequence avoiding the interference probability region by adopting a dynamic programming algorithm if a plurality of interference probabilities are higher than a preset threshold value, so as to obtain an obstacle avoidance optimal path, wherein the method specifically comprises the following steps of: acquiring an industrial robot track sequence and an obstacle model which meet kinematic continuity constraint, and identifying potential obstacle areas positioned near the obstacle model in the track sequence; Calculating the degree of synchronization according to the local curvature change and the acceleration peak value in the potential obstacle area so as to determine the interference probability; And searching the optimal state transition path passing through the alternative node at the periphery of the potential obstacle area by using a dynamic programming algorithm if the plurality of involved probabilities are higher than a preset threshold value to obtain an obstacle avoidance optimal path.
8. The method for avoiding the obstacle on the basis of the motion trail of the industrial robot based on the deep learning of claim 7, further comprising the steps of S7, extracting the key node position with high spatial coincidence density of the curvature extreme point and the speed mutation point from the obstacle avoidance optimal path, verifying the maintenance integrity of the geometrical self-similarity and the acceleration sequence self-correlation peak value through a convolutional neural network, and determining the final motion trail representation of the industrial robot, wherein the method specifically comprises the following steps: Calculating the space coincidence density between the curvature extreme point and the speed abrupt change point in the obstacle avoidance optimization path, and intercepting a local track section according to the space coincidence density; obtaining the geometrical self-similarity of the local track section and the self-correlation peak value of the acceleration sequence, and constructing a multidimensional feature tensor according to the geometrical self-similarity and the self-correlation peak value of the acceleration sequence; Performing integrity maintaining verification on the multidimensional feature tensor through a convolutional neural network, and outputting an integrity verification result; and performing smooth reconstruction on the local track segment according to the integrity verification result, and determining the final industrial robot motion track representation.
9. The method for avoiding obstacle on the basis of deep learning of industrial robot motion trail according to claim 8, further comprising the steps of S8, calculating the alignment degree of geometric symmetry axis and the axis of the turning moment of the velocity vector and the consistency score of the matching degree of the zero-crossing frequency of curvature and the zero-crossing frequency of the velocity vector for the final industrial robot motion trail representation and the original industrial robot high-dimensional trail data, updating the joint principal component axis weight in the initial representation vector through error feedback if the consistency score is lower than a preset threshold value, and re-entering a dimension reduction processing link, wherein the method specifically comprises the following steps: Extracting the geometric symmetry axis represented by the motion trail of the final industrial robot and calculating the axis alignment degree at the moment of turning over the speed vector of the original trail data; And (5) extracting the curvature zero crossing frequency of the track representation and the speed zero crossing frequency of the original track data to calculate the frequency matching degree.
10. The method for avoiding obstacle on motion trajectories of industrial robots based on deep learning as set forth in claim 9, wherein S8 further comprises: Calculating a consistency score according to the shaft alignment degree and the frequency matching degree; if the consistency score is lower than a preset threshold value, generating an error feedback signal to update the joint principal component axis weight in the initial representation vector, and re-entering a dimension reduction processing link according to the updated joint principal component axis weight.

Description

Industrial robot motion trail obstacle avoidance method based on deep learning Technical Field The invention relates to the technical field of industrial robots, in particular to an obstacle avoidance method for a motion trail of an industrial robot based on deep learning. Background With the rapid development of intelligent manufacturing and flexible production lines, industrial robots are increasingly used in complex working environments such as welding, assembly, carrying and polishing. In the industrial sites where multiple devices cooperate, stations are densely arranged and dynamic obstacles frequently appear, the robot needs to plan safe, smooth and efficient motion tracks in real time in the process of executing motion tasks so as to avoid collision with workpieces, fixtures or other devices. Therefore, how to realize reliable track obstacle avoidance under the condition of meeting the constraint conditions of kinematics and dynamics becomes a key technical problem for improving the autonomous running capability and the production efficiency of the industrial robot. The existing industrial robot track obstacle avoidance method is mostly based on geometric space modeling and collision detection mechanisms, path correction is achieved by calculating the minimum distance between a robot and an obstacle or constructing a safety envelope area, or simple kinematic constraints such as speed, acceleration and the like are introduced on the basis of a traditional track planning algorithm so as to improve track smoothness. However, such methods are usually focused on obstacle avoidance determination at the geometric space level, or only optimize dynamic indexes independently, lack of systematic analysis of coupling relation between track geometry and dynamic behaviors, and have difficulty in comprehensively describing real motion characteristics of robots in complex industrial environments. From the data structure point of view, the motion trail of the industrial robot essentially belongs to high-dimensional time series data, and comprises multi-order derivative information such as position, speed, acceleration and the like. The curvature change, the torsion rate distribution and the like of the track in space reflect the geometric shape characteristics of the track, and the acceleration change trend and the time evolution rule reflect the dynamics execution state. In the actual motion process, a significant coupling relationship often exists between the geometric features and the dynamic features, namely, the curvature mutation can cause abnormal acceleration peak value or acceleration change rate, and the dynamic constraint in turn limits the range of the achievable geometric path. If the two types of information cannot be effectively fused in the track representation process, track representation distortion is easy to cause, so that a planning result is theoretically feasible, and the problems of vibration, impact or excessively high energy consumption and the like are generated in actual implementation. In addition, when the existing method processes high-dimensional track data, a fixed feature extraction or artificial experience parameter modeling mode is mostly adopted, automatic learning capability on deep structural features of the track is lacked, and self-adaptive optimization is difficult to realize in a complex environment. Meanwhile, in the track adjustment and obstacle avoidance decision process, a feedback mechanism for the consistency of the overall track structure is often lacking, and when the track is subjected to local correction for many times, the problem of unbalance of the overall structure or damage of dynamics continuity can occur, so that the running stability and safety of the robot are affected. Disclosure of Invention The invention aims to provide an industrial robot motion track obstacle avoidance method based on deep learning, which solves the problems in the prior art. The technical scheme includes that S1, a position sequence, a speed sequence and an acceleration sequence are extracted from track data of an industrial robot motion track, a geometric feature vector and a dynamic feature vector are respectively generated through a convolutional neural network and are fused to obtain an initial representation vector, S2, the synchronous degree between local curvature change and acceleration peak values, the spatial coincidence density of curvature extremum points and speed mutation points and the directional consistency of track torsion gradient and acceleration change rate vector are calculated for the initial representation vector, if the weighted average value of three indexes exceeds a preset threshold value, a strong coupling relation exists between the geometric shape and the dynamic feature, S3, the initial representation vector which is judged to be strong-coupled is adopted, a main component analysis is adopted to extract a geometric dynamic coupling direction which kee