Trajectory speed optimization method of serial robot based on deep learning

Abstract

The present invention relates to mechanical motion control technology, and in particular to a trajectory speed optimization method for a serial robot arm based on deep learning. The method comprises the following steps: performing linear interpolation processing on the robot arm to generate spatial speed data, performing pseudo-inverse data conversion on the spatial speed data to generate joint spatial speed data; performing data preprocessing on the joint spatial speed data to generate joint timestamp speed data, using the joint timestamp speed data for numerical comparison, if the joint timestamp speed data is greater than a preset value, it is marked as overspeed spatial speed data; performing inverse solution and calculation processing on the overspeed spatial speed data to generate a joint linear data set; performing trajectory prediction on the robot arm based on the joint linear data set to generate secondary joint position data. The present invention provides a trajectory speed optimization method for a serial robot arm based on deep learning to solve the problem of joint overspeed caused by non-dexterous spatial planning of a straight line trajectory.

Description

Technical Field

[0001] The present invention relates to mechanical motion control technology, and in particular to a trajectory speed optimization method of a serial robot arm based on deep learning. Background Art

[0002] Mechanical motion control is one of the important technologies in the field of industrial automation. It can be used to control the motion trajectory, speed and force of the mechanical system to achieve high-precision, high-speed and high-reliability automation generation process. At the same time, the control algorithm is an important part of mechanical motion control, which can help the mechanical system achieve accurate position and speed control and ensure the stability and accuracy of the mechanical system.

[0003] Nowadays, artificial intelligence technology is widely used in the field of mechanical motion control and has played a huge role. Artificial intelligence is a new technical science that studies and develops theories, methods, technologies and application systems for simulating, extending and expanding human intelligence. Among them, deep learning and machine vision are two important branches of artificial intelligence, which can realize the automatic perception, analysis and processing of visual information such as objects and scenes. At the same time, artificial neural networks can be used for multi-level feature extraction and learning, so as to achieve tasks such as data classification, recognition and prediction. Combining artificial intelligence technology with mechanical motion control can achieve more accurate and efficient motion control, and realize more intelligent mechanical motion control, which greatly improves the intelligence level of mechanical control systems and the accuracy and efficiency of motion control. Existing control algorithms cannot solve the problem of joint overspeed caused by linear trajectories in non-dexterous space planning. Summary of the invention

[0004] The present invention provides a trajectory speed optimization method for a serial robot arm based on deep learning, which solves the problem of joint overspeed caused by linear trajectory, circular arc trajectory and spline trajectory after non-dexterous space planning.

[0005] To achieve the above purpose, a trajectory speed optimization method of a serial robot arm based on deep learning includes the following steps:

[0006] Step S1: performing linear interpolation processing on the motion trajectory of the robot arm to generate spatial velocity data, and performing pseudo-inverse data conversion on the spatial velocity data to generate joint spatial velocity data;

[0007] Step S2: preprocessing the joint space velocity data to generate joint timestamp velocity data, using the joint timestamp velocity data to compare the values ​​with the preset rated joint velocity data, if the joint timestamp velocity data is greater than the preset rated joint velocity data, then optimizing the performance speed and marking the joint timestamp velocity data as overspeed space velocity data, otherwise outputting the joint timestamp velocity data;

[0008] Step S3: Perform inverse planning and inference optimization processing on the speed data of the overspeed space to generate a joint linear data set;

[0009] Step S4: Perform trajectory prediction processing on the joint linear data set for the robotic arm to generate secondary joint position data.

[0010] This example performs linear interpolation processing on the robot arm, sets the starting point and target position, determines the interpolation point, implements linear interpolation, generates joint space velocity data, makes the robot arm move more smoothly, improves the accuracy and movement speed of the robot arm, helps improve the accuracy of obtaining numerical values, performs numerical conversion on the space velocity data, generates joint space velocity data, improves the fit of the data, and is more conducive to the implementation of subsequent steps. Descriptive statistical analysis and processing of the joint space velocity is performed to better identify and eliminate anomalies and errors in the data. The joint space velocity is preprocessed to generate joint timestamp velocity data to ensure the real-time nature of the data. , compress the data, facilitate data storage and transmission, improve the smoothness of the data, compare the joint timestamp data with the rated joint speed, mark the data exceeding the rated value as overspeed spatial speed data, make the data more accurate, and perform subsequent optimization processing on abnormal data. The inverse kinematics method and deep learning method are used to optimize the abnormal overspeed spatial speed data, and a joint linear data set is generated, making the generated data more precise, improving the system performance, and making the data processing process intelligent. The trajectory prediction processing of the robotic arm is then performed to generate the remaining joint position data, thereby completing the speed optimization.

[0011] In one embodiment of this specification, step S1 includes the following steps:

[0012] Step S11: extracting the inertial measurement unit of the robotic arm to generate joint angle data;

[0013] Step S12: Performing rotation matrix processing on the joint angle data using a kinematic model to obtain the joint position Jacobian matrix data;

[0014] Step S13: using a pseudo-inverse processing method on the joint Jacobian matrix data to generate joint Jacobian inverse matrix data;

[0015] Step S14: Perform inverse kinematics solution processing on the joint Jacobian inverse matrix data to generate joint space velocity data.

[0016] In this example, the forward kinematics technology is used to reconstruct the length, curvature and other information of each joint of the robot arm, extract the inertial measurement unit of the robot arm, generate joint angle data, help the robot arm complete precise target point movement, and improve the accuracy of the data. Based on the joint angle data, the position and posture of the end effector of the robot arm are recorded, and each joint angle is adjusted slightly. The position and posture of the end effector of the robot arm are recalculated, and the linear regression method is used to calculate the end effector position and posture twice to generate the joint Jacobian matrix data, improve the control accuracy and efficiency of the robot arm, solve the obstacle avoidance problem of the robot arm, calculate the determinant of the Jacobian matrix, obtain the pseudo-inverse matrix of the Jacobian matrix, and use the generalized inverse algorithm to solve the pseudo-inverse matrix to generate the Jacobian inverse matrix, and then use inverse kinematics control to solve and process, generate the joint space velocity, and convert the expected velocity of the end effector into the expected velocity of the joint space, so that the robot arm can understand the situation where the number of joints is greater than the degrees of freedom, and avoid singular points and joint motion range limitations in the robot arm.

[0017] In one embodiment of the present specification, step S2 includes the following steps:

[0018] Step S21: performing time series cutting and storage processing on the joint space velocity data to generate joint time series data;

[0019] Step S22: Using a window sliding algorithm to perform moving window data processing on the joint time series data to generate frame joint space velocity data;

[0020] Step S23: using the long short-term memory network to perform model training processing on the frame joint space velocity data to generate model training set data;

[0021] Step S24: Based on the model training set data, the data of the robot arm joint space velocity is audited, the model prediction processing is performed, the error situation is feedback controlled, and the joint timestamp velocity data is generated;

[0022] Step S25: Process the joint timestamp velocity data using overspeed abnormality detection to generate overspeed spatial velocity data.

[0023] In this example, the joint space velocity data is stored as a time series, and then the time series data is cut into multiple continuous small segments to generate joint time series data, which can more effectively understand and analyze the trends and patterns of the data, provide support for decision-making, and clearly monitor the abnormalities of the joint time series data. The sliding window algorithm is used to process the moving window data to generate frame joint space velocity data, which significantly reduces the time complexity, avoids repeated calculations, and significantly improves the efficiency of the algorithm. At the same time, by adjusting the window size and moving step size to meet different needs, the hyperparameters such as the number of layers and hidden state size of the long short-term memory network are determined, and the linear rectification function and cross entropy loss function are selected to construct a frame-based The model data of the joint space velocity data uses stochastic gradient descent to adjust parameters and generate model training set data to solve the problems of gradient disappearance and gradient explosion, making the processing of long sequence data more stable and efficient. It can adaptively learn the characteristics of the input sequence, making the model training set data have stronger generalization ability and more flexibility. The model training set data is used to detect the joint space velocity of the robotic arm, and the model is used to predict the future joint space velocity data. Feedback and control are performed based on the error situation to generate joint timestamp velocity data, which significantly reduces the error of the data, makes the data more accurate, solves the problems of data deviation, overfitting and underfitting, and helps to improve the accuracy of the data.

[0024] In one embodiment of this specification, step S3 includes the following steps:

[0025] Step S31: Generate starting joint position data based on the overspeed space speed data using an inverse motion model control method;

[0026] Step S32: Based on the joint position data, the joint axis is subjected to proportional adjustment control processing to generate joint axis overshoot data;

[0027] Step S33: comparing the data based on the joint axis overshoot data by using data alignment analysis processing to generate joint velocity extreme value data;

[0028] Step S34: obtaining joint threshold position data by using a robot arm joint position prediction method based on joint velocity extreme value data;

[0029] Step S35: obtaining a joint linear data set using a linear dynamics equation calculation formula based on the joint threshold position data;

[0030] In this example, based on the overspeed spatial speed data, the end posture matrix is ​​calculated to describe the posture matrix of the end of the robot arm in three-dimensional space. The analytical method in inverse kinematics is used to generate the starting joint position data to improve data accuracy. The structural parameters of the robot arm are more flexible to adapt to different working scenarios. Based on the joint position data, the target position and error are calculated. According to the error value, the joint is proportionally coordinated and controlled to generate a control signal. The control signal is overshoot processed to generate joint axis overshoot data, which improves the control accuracy and stability of the robot arm. The overshoot technology is used to effectively avoid the loss of control caused by position error and inertial resistance, shorten the response time of the control signal, and achieve faster motion response speed.

[0031] Preferably, the linear dynamics equation calculation formula in step S35 is as follows:

[0032]

[0033]

[0034] Among them, x i (t) refers to the joint displacement segment data, t refers to the joint movement time, t i is the motion time of the i-th joint, i refers to the joint subscript value, v i (T) refers to the velocity function value of the i-th joint, T refers to the joint motion cycle constant, a i refers to the acceleration of the i-th joint, δ is the position vector of the end of the robot arm, v i refers to the motion speed value of the i-th joint, a i (T) refers to the joint motion acceleration function value, D refers to the joint linear data set, N refers to the number of segments, and x0 refers to the original position value of the robotic arm.

[0035] The present invention provides a linear motion displacement equation calculation formula, which fully considers the joint linear data set D, the joint displacement segment data x i (t), joint movement time t, i-th joint movement time t i , joint subscript value i, i-th joint velocity function value v i (T), joint motion cycle constant T, i-th joint acceleration a i , the end position vector δ of the robot arm, the motion speed value v of the i-th joint i , joint motion acceleration function value a i (T), segment number value N, robot arm original position value x0, according to the interaction between the functions of joint motion displacement segment data, joint motion velocity, joint motion acceleration and robot arm end position vector δ, to form a functional relationship According to the interaction between the joint displacement segment data and the joint original position data, it is ensured that the position change of the robot arm at every moment is recorded, which saves computing power and makes the calculation converge quickly. In addition, since the joint cannot move normally within certain angle ranges during movement, the robot arm cannot reach a specific position. In order to avoid the dead zone problem, the dynamic positive zone correction is comprehensively considered, and the robot arm end position vector δ is introduced for positive zone calibration processing to enhance the accuracy and controllability of the robot arm, avoid the dead zone problem, and more accurately control the robot arm to operate in non-flexible space. The function is used to predict the position of the joint in the next frame to generate a joint linear data set, avoiding the collision problem of the robot arm joint during the movement process, realizing the linear motion control of the robot arm, predicting the future position of the robot arm, and realizing precise positioning and control. The Laplace inverse transform is used to adjust the abnormal parameter value in real time to realize the linear motion control of the robot arm at different speeds. By controlling the acceleration of the robot arm, the position change of the robot arm is more accurate and stable, which improves the flexibility and applicability of the algorithm.

[0036] In one embodiment of this specification, step S31 includes the following steps:

[0037] Step S311: constructing a kinematic model of the robot arm using an adaptive control algorithm based on the overspeed space speed data;

[0038] Step S312: performing coordinate system data acquisition and processing according to the robot arm kinematics model to generate a base coordinate system;

[0039] Step S313: using the sensor to perform positioning operations on the end of the manipulator based on the sensor-measured coordinates and Euler angles according to the base coordinate system, and generating posture matrix data;

[0040] Step S314: performing posture matrix orthogonalization processing on the posture matrix data to generate a rotation matrix;

[0041] Step S315: Perform joint data conversion mapping based on the rotation matrix to generate joint position data.

[0042] In this example, an adaptive control algorithm is used to construct a robot kinematic model based on the overspeed spatial speed data. The adaptive control algorithm is used to monitor the overspeed spatial speed data in real time, dynamically adjust the kinematic model parameters, and improve the kinematic accuracy and control stability of the robot. By continuously optimizing the kinematic model description, it is more in line with the actual motion law and is ready to describe the robot's motion state to avoid drift and error problems during the robot's motion. Coordinate system data collection and processing are performed according to the robot's kinematic model to improve the accuracy and stability of the collected data, adapt to different robot arm collection needs, and meet different data analysis and processing needs. Coordinate system data collection based on the robot's kinematic model can obtain high-precision data, avoid human errors and deviations, and provide a more reliable basis for subsequent data analysis and processing. Sensors are used to implement robot end positioning operations based on sensor-measured coordinates and Euler angles according to the base coordinate system to achieve high-precision positioning of the robot end and improve the robot's motion control capability and application range. Enhance the adaptive ability of the robot arm in complex environments, effectively handle the orthogonality problem in the posture matrix data, ensure that each vector of the posture matrix is ​​orthogonal to each other and of unit length, ensure the stability and certainty of the posture matrix, perform joint data conversion mapping based on the rotation matrix, eliminate errors in the movement of the robot arm, and thus improve the movement accuracy of the robot arm.

[0043] In one embodiment of the present specification, step S34 includes the following steps:

[0044] Step S341: using sensors to fit the motion trajectory data of the robotic arm to generate model raw data;

[0045] Step S342: extracting features from the original model data to generate a model feature vector;

[0046] Step S343: Perform joint threshold prediction processing using the trajectory prediction training model based on the model feature vector to generate joint threshold bit data.

[0047] In this example, sensors are used to fit the motion trajectory data of the robotic arm, and calibration is performed through a calibration program to ensure the accuracy of sensor measurement data, enhance the accuracy of fitting data, improve model accuracy, improve the accuracy of motion trajectory, reduce motion trajectory errors and make timely adjustments. Feature extraction is performed on the original data of the model, and data processing is automated to improve data processing efficiency. Through feature extraction, the data volume is reduced, storage and transmission costs are reduced, data processing accuracy is improved, and the generalization ability and performance of the model are enhanced. Data is classified through feature extraction, and data with similar features are classified into one category to improve data classification accuracy. At the same time, data support is provided for subsequent decision-making. The trajectory prediction training model is used based on the model feature vector to perform joint threshold prediction processing, thereby improving the safety and reliability of the robotic arm. It can adapt to different robotic arm types and task requirements, has a certain degree of versatility and strong scalability.

[0048] In one embodiment of this specification, step S4 includes the following steps:

[0049] Step S401: using a depth camera to collect and process point cloud data of the robotic arm to generate a point cloud data set;

[0050] Step S402: performing fitting residual sum of squares processing on the point cloud data set to generate a point cloud linear data table set;

[0051] Step S403: extracting geometric features of the robot arm based on the point cloud linear data table set to generate a training data set;

[0052] Step S404: construct a data training model using a neural network algorithm based on the training data set, perform posture prediction processing according to the data training model, and generate secondary joint position data.

[0053] This example uses a depth camera to collect and process point cloud data for a robotic arm, generates a point cloud data set, accurately collects three-dimensional point cloud data on the surface of an object, improves data collection accuracy, avoids data deviation problems caused by measurement errors and illumination changes in traditional methods, and improves data collection efficiency. The collected point cloud data has high-precision spatial information, thereby improving the accuracy and robustness of the machine learning algorithm and improving application flexibility. The point cloud data set is fitted with the sum of squared residuals to generate a point cloud linear data table set, effectively improving the accuracy of the point cloud data table set, reducing fitting errors, reducing data complexity and calculations. The point cloud data provides an accurate motion path, improves the positioning accuracy of the robotic arm, and reduces the amount of manual programming. A data training model is constructed based on the training data set using a neural network algorithm, and posture prediction processing is performed based on the data training model. The neural network algorithm can automatically adapt to various posture conditions without manual programming to make adaptive adjustments for different posture conditions, thereby improving the adaptability of the robotic arm. The neural network algorithm constructed based on the training data set is continuously optimized according to actual conditions, has high scalability, can predict and process postures in a short time, and has high real-time performance.

[0054] In one embodiment of this specification, step S402 includes the following steps:

[0055] Step S411: performing data denoising and filtering processing based on the point cloud linear data table set to generate preprocessed data;

[0056] Step S412: performing shape analysis feature extraction operation based on the preprocessed data to generate feature point data;

[0057] Step S413: performing feature corresponding point detection processing on the feature point data to generate feature registration point data;

[0058] Step S414: Perform an integrated training operation based on the feature registration point data to generate a training data set.

[0059] This example performs data denoising and filtering based on point cloud linear data tables to generate preprocessed data. Through denoising and filtering, the noise interference of point cloud data is reduced. The linear data table set is used with high processing performance, which is suitable for large-scale point cloud data processing. Threshold filtering and Gaussian filtering are used to improve data filtering efficiency, avoid data resampling and gridding problems in traditional methods, improve data stability and reliability, perform shape analysis feature extraction operations based on preprocessed data, generate feature point data, and improve shape analysis accuracy. By performing shape analysis on preprocessed data, the data information density is improved. Cluster analysis methods are used to improve Improve the distinguishability of feature points, reduce the amount of data, improve computing efficiency, perform feature corresponding point detection on feature point data, improve the accuracy and stability of feature corresponding point detection, avoid feature point drift and matching errors, improve the registration accuracy and reconstruction quality of 3D point cloud data, and have wide applicability and flexibility. Perform integrated training operations based on feature registration point data, improve the robustness of object recognition and positioning by registering and integrating feature point data, use feature point subsets for machine learning training, reduce the scale of training data sets, and improve training efficiency. Feature point classifiers can classify according to new feature point data to achieve real-time recognition and positioning

[0060] In one embodiment of this specification, step S412 includes the following steps:

[0061] Step S421: Generate image scaling data using an image scale normalization processing method based on the preprocessed data;

[0062] Step S422: based on the image scaling data, enhancing the image contrast and generating a balanced histogram;

[0063] Step S423: performing image denoising operation using Gaussian filtering based on the equalized histogram to generate denoised image data;

[0064] Step S424: performing a local feature extraction operation based on the denoised image data using a linear back-projection algorithm to generate a local feature vector;

[0065] Step S425: performing feature encoding operation based on the local feature vector using the visual word bag model to generate feature encoding data;

[0066] Step S426: Perform dimensionality reduction processing based on the feature coding data using principal component analysis to generate feature point data.

[0067] This example uses an image scale normalization processing method based on preprocessed data to generate image scaling data, eliminate the differences in images at different scales, and improve the accuracy and robustness of image processing. Through preprocessing and normalization processing, the calculation amount and time complexity of subsequent computer vision tasks are greatly reduced, and the calculation efficiency is improved. The scaled data is contrast enhanced to improve image clarity and visibility, reduce noise and interference in the image, improve the image signal-to-noise ratio, improve the image grayscale distribution, enhance the image visual impact, optimize image data representation, and improve the accuracy and reliability of the image in algorithm processing. Based on the balanced histogram, Gaussian filtering is used to perform image denoising operations, effectively eliminate Gaussian noise in the image, improve image quality, and facilitate subsequent image processing and analysis operations. The Gaussian filtering algorithm is used for image denoising, which can retain image edge information, avoid blurring and distortion, and improve image clarity and fidelity. The system provided by the present invention can automatically perform image processing operations. Improve processing efficiency, facilitate high-quality image processing and analysis, use linear back-projection algorithm to perform local feature extraction based on denoised image data, effectively extract key information from the image, and make subsequent operations more accurate and reliable. Use visual word bag model to perform feature encoding on local feature vectors, convert high-dimensional local feature vectors into low-dimensional visual vocabulary representation, reduce the number of feature vector bits, reduce the amount of calculation, cluster local features through clustering algorithm, better reflect important image features, eliminate scale differences between images, improve feature vector comparability, use principal component analysis method to perform dimensionality reduction based on feature coding data, reduce data redundant information, improve computing efficiency, perform dimensionality reduction on feature coding data, map data to low-dimensional space, visualize data, reduce data complexity, reduce the burden of storage and transmission, improve algorithm accuracy and robustness, and improve algorithm performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 A schematic diagram of a process flow of trajectory joint speed optimization based on deep learning provided by an embodiment of the present invention;

[0069] Figure 2 A schematic diagram of a process for calculating joint space velocity data provided by an embodiment of the present invention;

[0070] Figure 3 A schematic diagram of a process of generating joint timestamp velocity data by a model based on deep learning provided by an embodiment of the present invention;

[0071] Figure 4 A schematic diagram of a flow chart of calculating a joint linear data set based on machine vision provided by an embodiment of the present invention;

[0072] Figure 5 A schematic diagram of a flow chart for implementing the machine vision-based prediction of secondary joint position data provided in one embodiment of the present invention.

[0073] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0074] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0075] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0076] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0077] An embodiment of the present application provides a trajectory speed optimization method for a serial robot arm based on deep learning.

[0078] To achieve this, please refer to Figures 1 to 5 , a trajectory speed optimization method of a serial robot arm based on deep learning, the method comprising the following steps:

[0079] Step S1: performing linear interpolation processing on the robot arm to generate spatial velocity data, and performing pseudo-inverse data conversion on the spatial velocity data to generate joint spatial velocity data;

[0080] Step S2: preprocessing the joint space velocity data to generate joint timestamp velocity data, using the joint timestamp velocity data to compare the values ​​with the preset rated joint velocity data, if the joint timestamp velocity data is greater than the preset rated joint velocity data, then optimizing the performance speed and marking the joint timestamp velocity data as overspeed space velocity data, otherwise outputting the joint timestamp velocity data;

[0081] Step S3: Perform inverse planning and inference optimization processing on the speed data of the overspeed space to generate a joint linear data set;

[0082] Step S4: Perform trajectory prediction processing on the joint linear data set for the robotic arm to generate secondary joint position data.

[0083] This example performs linear interpolation processing on the robot arm, sets the starting point and target position, determines the interpolation point, implements linear interpolation, generates joint space velocity data, makes the robot arm move more smoothly, improves the accuracy and movement speed of the robot arm, helps improve the accuracy of obtaining numerical values, performs numerical conversion on the space velocity data, generates joint space velocity data, improves the fit of the data, and is more conducive to the implementation of subsequent steps. Descriptive statistical analysis and processing of the joint space velocity is performed to better identify and eliminate anomalies and errors in the data. The joint space velocity is preprocessed to generate joint timestamp velocity data to ensure the real-time nature of the data. , compress the data, facilitate data storage and transmission, improve the smoothness of the data, compare the joint timestamp data with the rated joint speed, mark the data exceeding the rated value as overspeed spatial speed data, make the data more accurate, and perform subsequent optimization processing on abnormal data. The inverse kinematics method and deep learning method are used to optimize the abnormal overspeed spatial speed data, and a joint linear data set is generated, making the generated data more precise, improving the system performance, and making the data processing process intelligent. The trajectory prediction processing of the robotic arm is then performed to generate the remaining joint position data, thereby completing the speed optimization.

[0084] In the embodiment of the present invention, reference Figure 1 The above is a schematic flow chart of the steps of a trajectory speed optimization method of a serial robot arm based on deep learning of the present invention. In this example, the steps of the trajectory speed optimization method of a serial robot arm based on deep learning include:

[0085] Step S1: performing linear interpolation processing on the motion trajectory of the robot arm to generate spatial velocity data, and performing pseudo-inverse data conversion on the spatial velocity data to generate joint spatial velocity data;

[0086] In an embodiment of the present invention, the joint angle data of the robotic arm is obtained through an inertial measurement unit, and the distance and time that the robotic arm needs to move are calculated based on the target end point position and the current end point position of the robotic arm. The joint angle data is processed by linear interpolation and rotation matrix processing is performed in combination with the kinematic model to generate joint Jacobian matrix data. The joint Jacobian matrix data is processed based on inverse kinematics, including pseudo-inverse processing methods and inverse kinematics solution processing, to generate joint space velocity data.

[0087] Step S2: preprocessing the joint space velocity data to generate joint timestamp velocity data, using the joint timestamp velocity data to compare the values ​​with the preset rated joint velocity data, if the joint timestamp velocity data is greater than the preset rated joint velocity data, then optimizing the performance speed and marking the joint timestamp velocity data as overspeed space velocity data, otherwise outputting the joint timestamp velocity data;

[0088] In an embodiment of the present invention, the joint space velocity data is preprocessed, including time series cutting and storage processing, sliding window algorithm, model training processing based on long short-term memory network, and model prediction processing, so as to generate joint timestamp velocity data.

[0089] Step S3: Perform inverse planning and inference optimization processing on the speed data of the overspeed space to generate a joint linear data set;

[0090] In an embodiment of the present invention, the overspeed space speed data is subjected to inverse planning, inference and optimization processing, including an inverse motion model control method, proportional harmonic control processing, data alignment analysis processing, a deep learning-based robotic arm joint position prediction method and linear kinematics equations, thereby generating a joint linear data set.

[0091] Step S4: Perform trajectory prediction processing on the joint linear data set for the robotic arm to generate secondary joint position data.

[0092] In an embodiment of the present invention, the joint linear data set is processed using trajectory prediction based on image recognition and deep learning, several features are extracted from the point cloud data set acquired by the depth camera to generate a training data set, a neural network algorithm is used to perform model training on the training data set, and posture prediction processing is performed to generate secondary joint position data.

[0093] Preferably, step S1 comprises the following steps:

[0094] Step S11: extracting the inertial measurement unit of the robotic arm to generate joint angle data;

[0095] Step S12: Performing rotation matrix processing on the joint angle data using a kinematic model to obtain the joint position Jacobian matrix data;

[0096] Step S13: using a pseudo-inverse processing method on the joint Jacobian matrix data to generate joint Jacobian inverse matrix data;

[0097] Step S14: Perform inverse kinematics solution processing on the joint Jacobian inverse matrix data to generate joint space velocity data.

[0098] The present invention utilizes forward kinematics technology to reconstruct information such as the length and curvature of each joint of the robotic arm, extracts the inertial measurement unit of the robotic arm, generates joint angle data, helps the robotic arm complete precise target point movement, and improves the accuracy of the data. Based on the joint angle data, the position and posture of the end effector of the robotic arm are recorded, each joint angle is slightly adjusted, the position and posture of the end effector of the robotic arm are recalculated, and the end effector position and posture are calculated twice using the linear regression method to generate joint Jacobian matrix data, thereby improving the control accuracy and efficiency of the robotic arm and solving the obstacle avoidance problem of the robotic arm. The determinant of the Jacobian matrix is ​​calculated to obtain the pseudo-inverse matrix of the Jacobian matrix and the generalized inverse algorithm is used to solve the pseudo-inverse matrix to generate the Jacobian inverse matrix, which is then solved using inverse kinematics control to generate the joint space velocity, convert the desired velocity of the end effector into the desired velocity of the joint space, so that the robotic arm understands the situation where the number of joints is greater than the degrees of freedom, and avoids the problems of singular points and joint motion range limitation in the robotic arm.

[0099] As an example of the present invention, refer to Figure 2 ,for Figure 1 Detailed implementation steps of step S1 in the flowchart, in this example, step S1 includes:

[0100] Step S11: extracting the inertial measurement unit of the robotic arm to generate joint angle data;

[0101] In the embodiment of the present invention, the sampling frequency, sensor sensitivity and calibration parameters of the inertial measurement unit of the robotic arm are configured, and the inertial measurement unit is used to obtain the joint angle data during the movement of the robotic arm.

[0102] Step S12: Performing rotation matrix processing on the joint angle data using a kinematic model to obtain the joint position Jacobian matrix data;

[0103] In an embodiment of the present invention, based on the Denavit-Hartenber parameters of the robotic arm, the rotation matrix of each joint is calculated, the position and posture of the end effector of the robotic arm are calculated according to the joint angle data and the rotation matrix, and the Jacobian matrix of the end effector of the robotic arm, that is, the joint Jacobian matrix, is calculated according to the position and posture of the end effector.

[0104] Step S13: using a pseudo-inverse processing method on the joint Jacobian matrix data to generate joint Jacobian inverse matrix data;

[0105] In the embodiment of the present invention, the pseudo-inverse moment of the joint Jacobian matrix is ​​calculated, and the pseudo-inverse matrix is ​​calculated using the least squares method, thereby generating the joint Jacobian inverse matrix.

[0106] Step S14: Perform inverse kinematics solution processing on the joint Jacobian inverse matrix data to generate joint space velocity data.

[0107] In an embodiment of the present invention, the target joint angle is calculated using the inverse kinematics method based on the target end effector position and posture, the error value is calculated based on the current joint angle data, and the difference is multiplied by the joint Jacobian inverse matrix to generate the joint space velocity data.

[0108] Preferably, step S2 comprises the following steps:

[0109] Step S21: performing time series cutting and storage processing on the joint space velocity data to generate joint time series data;

[0110] Step S22: Using a window sliding algorithm to perform moving window data processing on the joint time series data to generate frame joint space velocity data;

[0111] Step S23: using the long short-term memory network to perform model training processing on the frame joint space velocity data to generate model training set data;

[0112] Step S24: Based on the model training set data, the data of the robot arm joint space velocity is audited, the model prediction processing is performed, the error situation is feedback controlled, and the joint timestamp velocity data is generated;

[0113] Step S25: Process the joint timestamp velocity data using overspeed abnormality detection to generate overspeed spatial velocity data.

[0114] In this example, the joint space velocity data is stored as a time series, and then the time series data is cut into multiple continuous small segments to generate joint time series data, which can more effectively understand and analyze the trends and patterns of the data, provide support for decision-making, and clearly monitor the abnormalities of the joint time series data. The sliding window algorithm is used to process the moving window data to generate frame joint space velocity data, which significantly reduces the time complexity, avoids repeated calculations, and significantly improves the efficiency of the algorithm. At the same time, by adjusting the window size and moving step size to meet different needs, the hyperparameters such as the number of layers and hidden state size of the long short-term memory network are determined, and the linear rectification function and cross entropy loss function are selected to construct a frame-based The model data of the joint space velocity data uses stochastic gradient descent to adjust parameters and generate model training set data to solve the problems of gradient disappearance and gradient explosion, making the processing of long sequence data more stable and efficient. It can adaptively learn the characteristics of the input sequence, making the model training set data have stronger generalization ability and more flexibility. The model training set data is used to detect the joint space velocity of the robotic arm, and the model is used to predict the future joint space velocity data. Feedback and control are performed based on the error situation to generate joint timestamp velocity data, which significantly reduces the error of the data, makes the data more accurate, solves the problems of data deviation, overfitting and underfitting, and helps to improve the accuracy of the data.

[0115] As an example of the present invention, refer to Figure 3 As shown, Figure 1 FIG. 1 is a flow chart showing the detailed implementation steps of step S2 in FIG. 1 . In this example, step S2 includes:

[0116] Step S21: performing time series cutting and storage processing on the joint space velocity data to generate joint time series data;

[0117] In an embodiment of the present invention, the robot arm joint space velocity data is cut into a number of time series data according to time, each time series data includes a timestamp and corresponding joint space velocity data, and is stored as joint time series data.

[0118] Step S22: Using a window sliding algorithm to perform moving window data processing on the joint time series data to generate frame joint space velocity data;

[0119] In this example, the window size and step size parameters are defined, and the joint time series data is processed by moving the window through the window sliding algorithm. The data in each window is subjected to feature extraction, including the mean, variance, etc. The extracted feature data is stored in the specified data structure to generate the frame joint space velocity data.

[0120] Step S23: using the long short-term memory network to perform model training processing on the frame joint space velocity data to generate model training set data;

[0121] In this example, a long short-term memory network model is constructed, and the frame joint space velocity data is used as the model input to train the model. During the training process, the cross-validation method is used to optimize the model to generate model training set data.

[0122] Step S24: Based on the model training set data, the data of the robot arm joint space velocity is audited, the model prediction processing is performed, the error situation is feedback controlled, and the joint timestamp velocity data is generated;

[0123] In this example, the spatial velocity data of the robot arm joints is used as the model input, and the model training set data is used for prediction processing. Error analysis is performed on the prediction results. If the prediction results do not meet the preset error requirements, feedback control is performed and the model is optimized by adjusting the model parameters to generate joint timestamp velocity data.

[0124] Step S25: Process the joint timestamp velocity data using overspeed abnormality detection to generate overspeed spatial velocity data.

[0125] In this example, the joint timestamp speed data is compared with the preset speed reference data speed by a speed threshold. If the joint speed is greater than the speed reference data, it is marked as overspeed spatial speed data.

[0126] Preferably, step S3 comprises the following steps:

[0127] Step S31: Generate starting joint position data based on the overspeed space speed data using an inverse motion model control method;

[0128] Step S32: Based on the joint position data, the joint axis is subjected to proportional adjustment control processing to generate joint axis overshoot data;

[0129] Step S33: comparing the data based on the joint axis overshoot data by using data alignment analysis processing to generate joint velocity extreme value data;

[0130] Step S34: obtaining joint threshold position data by using a robot arm joint position prediction method based on joint velocity extreme value data;

[0131] Step S35: obtaining a joint linear data set using a linear dynamics equation calculation formula based on the joint threshold position data;

[0132] In this example, based on the overspeed spatial speed data, the end posture matrix is ​​calculated to describe the posture matrix of the end of the robot arm in three-dimensional space. The analytical method in inverse kinematics is used to generate the starting joint position data to improve data accuracy. The structural parameters of the robot arm are more flexible to adapt to different working scenarios. Based on the joint position data, the target position and error are calculated. According to the error value, the joint is proportionally coordinated and controlled to generate a control signal. The control signal is overshoot processed to generate joint axis overshoot data, which improves the control accuracy and stability of the robot arm. The overshoot technology is used to effectively avoid the loss of control caused by position error and inertial resistance, shorten the response time of the control signal, and achieve faster motion response speed.

[0133] As an example of the present invention, refer to Figure 4 As shown, Figure 1 FIG. 1 is a flow chart showing the detailed implementation steps of step S3 in FIG. 1 . In this example, step S3 includes:

[0134] Step S31: Generate starting joint position data based on the overspeed space speed data using an inverse motion model control method;

[0135] In this example, the forward kinematics method is used to solve the joint space position of the robot end effector according to the position and posture of the robot end effector, and the inverse kinematics method is used to solve the joint space position of the robot according to the target position and the current position error, thereby generating the starting joint position data.

[0136] Step S32: Based on the joint position data, the joint axis is subjected to proportional adjustment control processing to generate joint axis overshoot data;

[0137] In this example, the joint position error is calculated based on the starting joint position data and the target joint position data, and the proportional-integral controller is used to calculate the adjustment amount of the joint position error to obtain the output of the joint position control. The joint position output control signal is used to control the motion of each joint of the robotic arm, thereby generating joint axis overshoot data.

[0138] Step S33: comparing the data based on the joint axis overshoot data by using data alignment analysis processing to generate joint velocity extreme value data;

[0139] In this example, the joint axis overshoot data and the overspeed spatial speed data are time-aligned to obtain the corresponding timestamps, and the timestamps are cut to obtain the joint axis overshoot data and the overspeed spatial speed data in different time periods. The data in each time period are compared to generate the joint speed extreme value data.

[0140] Step S34: obtaining joint threshold position data by using a robot arm joint position prediction method based on joint velocity extreme value data;

[0141] In this example, the joint velocity extreme value data is used to calculate the predicted value of the joint position of the robot at a certain moment in the future, and the predicted value is compared with the target joint position to generate joint threshold position data.

[0142] Step S35: obtaining a joint linear data set using a linear dynamics equation calculation formula based on the joint threshold position data;

[0143] In this example, according to the robot arm dynamics model, the joint threshold position data is input into the model, and the linear dynamics equation calculation formula is used to generate the joint linear data set.

[0144] Preferably, the linear dynamics equation calculation formula in step S35 is as follows:

[0145]

[0146]

[0147] Among them, x i (t) refers to the joint displacement segment data, t refers to the joint movement time, t i is the motion time of the i-th joint, i refers to the joint subscript value, v i (T) refers to the velocity function value of the i-th joint, T refers to the joint motion cycle constant, a i refers to the acceleration of the i-th joint, δ is the position vector of the end of the robot arm, vi refers to the motion speed value of the i-th joint, a i (T) refers to the joint motion acceleration function value, D refers to the joint linear data set, N refers to the number of segments, and x0 refers to the original position value of the robotic arm.

[0148] The present invention provides a linear dynamics equation calculation formula, which fully considers the joint linear data set D, the joint displacement segment data x i (t), joint movement time t, i-th joint movement time t i , joint subscript value i, i-th joint velocity function value v i (T), joint motion cycle constant T, i-th joint acceleration a i , the end position vector δ of the robot arm, the motion speed value v of the i-th joint i , joint motion acceleration function value a i (T), segment number value N, robot arm original position value x0, according to the interaction between the functions of joint motion displacement segment data, joint motion velocity, joint motion acceleration and robot arm end position vector δ, to form a functional relationship According to the interaction between the joint displacement segment data and the joint original position data, it is ensured that the position change of the robot arm at every moment is recorded, which saves computing power and makes the calculation converge quickly. In addition, since the joint cannot move normally within certain angle ranges during movement, the robot arm cannot reach a specific position. In order to avoid the dead zone problem, the dynamic positive zone correction is comprehensively considered, and the robot arm end position vector δ is introduced for positive zone calibration processing to enhance the accuracy and controllability of the robot arm, avoid the dead zone problem, and more accurately control the robot arm to operate in non-flexible space. The function is used to predict the position of the joint in the next frame to generate a joint linear data set, avoiding the collision problem of the robot arm joint during the movement process, realizing the linear motion control of the robot arm, predicting the future position of the robot arm, and realizing precise positioning and control. The Laplace inverse transform is used to adjust the abnormal parameter value in real time to realize the linear motion control of the robot arm at different speeds. By controlling the acceleration of the robot arm, the position change of the robot arm is more accurate and stable, which improves the flexibility and applicability of the algorithm.

[0149] Preferably, step S31 includes the following steps:

[0150] Step S311: constructing a kinematic model of the robot arm using an adaptive control algorithm based on the overspeed space speed data;

[0151] Step S312: performing coordinate system data acquisition and processing according to the robot arm kinematics model to generate a base coordinate system;

[0152] Step S312: using the sensor to perform positioning operations on the end of the manipulator based on the sensor-measured coordinates and Euler angles according to the base coordinate system, and generating posture matrix data;

[0153] Step S313: performing posture matrix orthogonalization processing on the posture matrix data to generate a rotation matrix;

[0154] Step S314: Perform joint data conversion mapping based on the rotation matrix to generate joint position data.

[0155] In this example, an adaptive control algorithm is used to construct a robot kinematic model based on the overspeed spatial speed data. The adaptive control algorithm is used to monitor the overspeed spatial speed data in real time, dynamically adjust the kinematic model parameters, and improve the kinematic accuracy and control stability of the robot. By continuously optimizing the kinematic model description, it is more in line with the actual motion law, and the robot motion state is prepared to avoid drift and error problems during the movement of the robot. Coordinate system data collection and processing are performed according to the robot kinematic model to improve the accuracy and stability of the collected data, adapt to different robot arm collection needs, and meet different data analysis and processing needs. Coordinate system data collection based on the robot kinematic model can obtain high-precision data, avoid human errors and deviations, and provide a more reliable foundation for subsequent data analysis and processing. Sensors are used to implement robot end positioning operations based on sensor-measured coordinates and Euler angles according to the base coordinate system, to achieve high-precision positioning of the robot end and improve the motion control capability and application scope of the robot. Enhance the adaptive ability of the robot arm in complex environments, effectively handle the orthogonality problem in the posture matrix data, ensure that each vector of the posture matrix is ​​orthogonal to each other and of unit length, ensure the stability and certainty of the posture matrix, perform joint data conversion mapping based on the rotation matrix, eliminate errors in the movement of the robot arm, and thus improve the movement accuracy of the robot arm.

[0156] As an example of the present invention, refer to Figure 4 As shown, Figure 4 FIG. 1 is a flow chart showing the detailed implementation steps of step S31 in FIG. 1 . In this example, step S31 includes:

[0157] Step S311: constructing a kinematic model of the robot arm using an adaptive control algorithm based on the overspeed space speed data;

[0158] In this example, the kinematic equations are established according to the geometric structure and degrees of freedom of the robot arm, and are converted into an optimization problem for solving the inverse kinematics problem. The inverse kinematics problem is continuously solved using the reference model adaptive control algorithm, thereby constructing the dynamic motion model of the robot arm.

[0159] Step S312: performing coordinate system data acquisition and processing according to the robot arm kinematics model to generate a base coordinate system;

[0160] In this example, based on the kinematic model of the robotic arm, the transformation relationship of the coordinate system is established so as to accurately describe the position and posture of the robotic arm in three-dimensional space. The position and posture of the robotic arm are measured using sensors, the position and posture data of the end of the robotic arm are obtained, and the base coordinate system is generated.

[0161] Step S313: using the sensor to perform positioning operations on the end of the manipulator based on the sensor-measured coordinates and Euler angles according to the base coordinate system, and generating posture matrix data;

[0162] In this example, the pose matrix of the end of the robot arm relative to the base coordinate system is calculated based on the position and attitude data measured by the sensor. According to the definition of Euler angles, the pose matrix is ​​decomposed into a rotation matrix and a translation vector to generate the attitude matrix data.

[0163] Step S314: performing posture matrix orthogonalization processing on the posture matrix data to generate a rotation matrix;

[0164] In this example, the attitude matrix is ​​orthogonalized to satisfy the properties of an orthogonal matrix. The three column vectors of the attitude matrix are orthogonalized using the Schmidt orthogonal algorithm and normalized to obtain a rotation matrix.

[0165] Step S315: Perform joint data conversion mapping based on the rotation matrix to generate joint position data.

[0166] In this example, according to the kinematic equation of the robot arm and the coordinate system transformation relationship, the posture matrix of the end of the robot arm is converted into joint angles, the position of each joint is calculated, and the joint position data is generated.

[0167] Preferably, step S34 includes the following steps:

[0168] Step S341: using sensors to fit the motion trajectory data of the robotic arm to generate model raw data;

[0169] Step S342: extracting features from the original model data to generate a model feature vector;

[0170] Step S343: Perform joint threshold prediction processing using the trajectory prediction training model based on the model feature vector to generate joint threshold bit data.

[0171] In this example, sensors are used to fit the motion trajectory data of the robotic arm, and calibration is performed through a calibration program to ensure the accuracy of sensor measurement data, enhance the accuracy of fitting data, improve model accuracy, improve the accuracy of motion trajectory, reduce motion trajectory errors and make timely adjustments. Feature extraction is performed on the original data of the model, and data processing is automated to improve data processing efficiency. Through feature extraction, the data volume is reduced, storage and transmission costs are reduced, data processing accuracy is improved, and the generalization ability and performance of the model are enhanced. Data is classified through feature extraction, and data with similar features are classified into one category to improve data classification accuracy. At the same time, data support is provided for subsequent decision-making. The trajectory prediction training model is used based on the model feature vector to perform joint threshold prediction processing, thereby improving the safety and reliability of the robotic arm. It can adapt to different robotic arm types and task requirements, has a certain degree of versatility and strong scalability.

[0172] As an example of the present invention, refer to Figure 4 As shown, Figure 4 FIG. 1 is a flow chart showing the detailed implementation steps of step S34 in FIG. 1 , wherein step S34 in this example includes:

[0173] Step S341: using sensors to fit the motion trajectory data of the robotic arm to generate model raw data;

[0174] In this example, the robot arm motion trajectory data collected by the sensor is collected and recorded to obtain raw data, which is then filtered and preprocessed to remove noise and outliers to obtain a clear data set. The data set is then subjected to linear regression fitting to generate model raw data.

[0175] Step S342: extracting features from the original model data to generate a model feature vector;

[0176] In this example, data analysis and feature extraction are performed on the original data of the model to extract key feature information from the data set, including velocity, acceleration, and displacement. The feature information is processed and calculated using Simcenter Amesim software to generate a model feature vector.

[0177] Step S343: Perform joint threshold prediction processing using the trajectory prediction training model based on the model feature vector to generate joint threshold bit data.

[0178] In this example, the support vector machine is used to train the model feature vector, the trained model is used to predict the data, and the prediction results are verified and adjusted to generate joint threshold position data.

[0179] Preferably, step S4 comprises the following steps:

[0180] Step S401: using a depth camera to collect and process point cloud data of the robotic arm to generate a point cloud data set;

[0181] Step S402: performing fitting residual sum of squares processing on the point cloud data set to generate a point cloud linear data table set;

[0182] Step S403: extracting geometric features of the robot arm based on the point cloud linear data table set to generate a training data set;

[0183] Step S404: construct a data training model using a neural network algorithm based on the training data set, perform posture prediction processing according to the data training model, and generate secondary joint position data.

[0184] This example uses a depth camera to collect and process point cloud data for a robotic arm, generates a point cloud data set, accurately collects three-dimensional point cloud data on the surface of an object, improves data collection accuracy, avoids data deviation problems caused by measurement errors and illumination changes in traditional methods, and improves data collection efficiency. The collected point cloud data has high-precision spatial information, thereby improving the accuracy and robustness of the machine learning algorithm and improving application flexibility. The point cloud data set is fitted with the sum of squared residuals to generate a point cloud linear data table set, effectively improving the accuracy of the point cloud data table set, reducing fitting errors, reducing data complexity and calculations. The point cloud data provides an accurate motion path, improves the positioning accuracy of the robotic arm, and reduces the amount of manual programming. A data training model is constructed based on the training data set using a neural network algorithm, and posture prediction processing is performed based on the data training model. The neural network algorithm can automatically adapt to various posture conditions without manual programming to make adaptive adjustments for different posture conditions, thereby improving the adaptability of the robotic arm. The neural network algorithm constructed based on the training data set is continuously optimized according to actual conditions, has high scalability, can predict and process postures in a short time, and has high real-time performance.

[0185] As an example of the present invention, refer to Figure 5 As shown, Figure 1 Detailed implementation step flow chart of step S4 in FIG. 1 , in this example, step S4 includes:

[0186] Step S401: using a depth camera to collect and process point cloud data of the robotic arm to generate a point cloud data set;

[0187] In this example, a depth camera is used to scan the edge of the robotic arm, control the position and posture of the end of the robotic arm, collect comprehensive point cloud data, clean, filter, and align the point cloud data to generate a point cloud data set.

[0188] Step S402: performing fitting residual sum of squares processing on the point cloud data set to generate a point cloud linear data table set;

[0189] In this example, the least squares method is used to process the fitting residuals of the point cloud data set to eliminate noise and inaccuracy in the data, and the residual sum of squares is calculated to evaluate the degree of model fitting and extract linear data, thereby generating a point cloud linear data table set.

[0190] Step S403: extracting geometric features of the robot arm based on the point cloud linear data table set to generate a training data set;

[0191] In this example, the point cloud linear data set is processed to extract the geometric features of the robotic arm, including distance, angle, length, curvature, etc., so as to extract the required posture information of the robotic arm. The geometric features are used as input data and the robotic arm posture information is used as output data to generate a training data set.

[0192] Step S404: construct a data training model using a neural network algorithm based on the training data set, perform posture prediction processing according to the data training model, and generate secondary joint position data.

[0193] In this example, a data training model is constructed based on the training data set using the convolutional neural network algorithm. The network structure, number of layers, number of nodes, and activation function are determined. The training set is input into the neural network for training. The network parameters are continuously adjusted. The back propagation algorithm is used to improve the accuracy and stability of the neural network. The model generalization ability and prediction accuracy are checked. According to the trained neural network model, the input posture data is predicted and processed to generate secondary joint position data.

[0194] Preferably, step S402 includes the following steps:

[0195] Step S411: performing data denoising and filtering processing based on the point cloud linear data table set to generate preprocessed data;

[0196] Step S412: performing shape analysis feature extraction operation based on the preprocessed data to generate feature point data;

[0197] Step S413: performing feature corresponding point detection processing on the feature point data to generate feature registration point data;

[0198] Step S414: Perform integrated training operations based on feature registration point data to generate a training data set. This example performs data denoising and filtering based on the point cloud linear data table set to generate preprocessed data. The denoising and filtering process reduces the noise interference of the point cloud data. The linear data table set is used to have high processing performance and is suitable for large-scale point cloud data processing. Threshold filtering and Gaussian filtering are used to improve data filtering efficiency, avoid data resampling and gridding problems in traditional methods, and improve data stability and reliability. Shape analysis feature extraction operations are performed based on preprocessed data to generate feature point data and improve shape analysis accuracy. Shape analysis is performed on preprocessed data to improve data information density. Cluster analysis methods are used to improve The feature point discrimination reduces the amount of data and improves the computing efficiency. The feature point data is processed for feature corresponding point detection to improve the accuracy and stability of feature corresponding point detection, avoid feature point drift and matching errors, and improve the registration accuracy and reconstruction quality of 3D point cloud data. It has wide applicability and flexibility. It performs integrated training operations based on feature registration point data. By registering and integrating feature point data, it improves the robustness of object recognition and positioning. It uses feature point subsets for machine learning training to reduce the size of training data sets and improve training efficiency. The feature point classifier can classify according to new feature point data to achieve real-time recognition and positioning.

[0199] As an example of the present invention, refer to Figure 5 As shown, Figure 5 FIG. 4 is a flow chart showing the detailed implementation steps of step S402 in FIG. 4 . In this example, step S402 includes:

[0200] Step S411: performing data denoising and filtering processing based on the point cloud linear data table set to generate preprocessed data;

[0201] In this example, the point cloud library is used to perform denoising filtering on the point cloud linear data table set to remove noise and outliers in the original data. The denoising filtering algorithm used is a voxel-based filtering algorithm to generate preprocessed data.

[0202] Step S412: performing shape analysis feature extraction operation based on the preprocessed data to generate feature point data;

[0203] In this example, the scale-invariant feature transformation process is used to perform shape analysis feature extraction operations on the preprocessed data to extract key feature points in the data, including information such as point position, direction, and scale, thereby generating feature point data.

[0204] Step S413: performing feature corresponding point detection processing on the feature point data to generate feature registration point data;

[0205] In this example, the nearest neighbor algorithm is used to perform feature corresponding point detection on the extracted feature point data, the best matching point is found in the image for each feature point, the Hamming distance algorithm is used to calculate the distance between feature descriptors, and the feature points are matched into pairs using a matching algorithm based on the distance information. After matching, the paired points are filtered and screened using a random sampling consensus algorithm, and the incorrectly matched points are deleted, thereby generating feature registration point data.

[0206] Step S414: Perform an integrated training operation based on the feature registration point data to generate a training data set.

[0207] In this example, the point cloud library is used to merge multiple feature registration point data into a large feature point data set, and the feature point data set is randomly divided into a training set and a validation set using a random partitioning method. The iterative closest point algorithm is used to perform feature point registration on each feature point in the training, and the point cloud data is aligned. The decision tree is used to train the model on the registered feature point data set. After the training is completed, the model is tested using the validation set to evaluate the model performance and generalization ability, thereby generating a training data set.

[0208] Preferably, step S412 includes the following steps:

[0209] Step S421: Generate image scaling data using an image scale normalization processing method based on the preprocessed data;

[0210] Step S422: based on the image scaling data, enhancing the image contrast and generating a balanced histogram;

[0211] Step S423: performing image denoising operation using Gaussian filtering based on the equalized histogram to generate denoised image data;

[0212] Step S424: performing a local feature extraction operation based on the denoised image data using a linear back-projection algorithm to generate a local feature vector;

[0213] Step S425: performing feature encoding operation based on the local feature vector using the visual word bag model to generate feature encoding data;

[0214] Step S426: Perform dimensionality reduction processing based on the feature coding data using principal component analysis to generate feature point data.

[0215] This example uses an image scale normalization processing method based on preprocessed data to generate image scaling data, eliminate the differences in images at different scales, and improve the accuracy and robustness of image processing. Through preprocessing and normalization processing, the calculation amount and time complexity of subsequent computer vision tasks are greatly reduced, and the calculation efficiency is improved. The scaled data is contrast enhanced to improve image clarity and visibility, reduce noise and interference in the image, improve the image signal-to-noise ratio, improve the image grayscale distribution, enhance the image visual impact, optimize image data representation, and improve the accuracy and reliability of the image in algorithm processing. Based on the balanced histogram, Gaussian filtering is used to perform image denoising operations, effectively eliminate Gaussian noise in the image, improve image quality, and facilitate subsequent image processing and analysis operations. The Gaussian filtering algorithm is used for image denoising, which can retain image edge information, avoid blurring and distortion, and improve image clarity and fidelity. The system provided by the present invention can automatically perform image processing operations. Improve processing efficiency, facilitate high-quality image processing and analysis, use linear back-projection algorithm to perform local feature extraction based on denoised image data, effectively extract key information from the image, and make subsequent operations more accurate and reliable. Use visual word bag model to perform feature encoding on local feature vectors, convert high-dimensional local feature vectors into low-dimensional visual vocabulary representation, reduce the number of feature vector bits, reduce the amount of calculation, cluster local features through clustering algorithm, better reflect important image features, eliminate scale differences between images, improve feature vector comparability, use principal component analysis method to perform dimensionality reduction based on feature coding data, reduce data redundant information, improve computing efficiency, perform dimensionality reduction on feature coding data, map data to low-dimensional space, visualize data, reduce data complexity, reduce the burden of storage and transmission, improve algorithm accuracy and robustness, and improve algorithm performance.

[0216] As an example of the present invention, refer to Figure 5 As shown, Figure 5 FIG. 1 is a flow chart showing the detailed implementation steps of step S4 in FIG. 1 , wherein step S402 in this example includes:

[0217] Step S421: Generate image scaling data using an image scale normalization processing method based on the preprocessed data;

[0218] In this example, preprocessing data is obtained, the target image size and scale are determined, the ratio between the original image size and the target image size is calculated, the scaling ratio is determined, the original image is scaled using the scaling ratio to generate the target image, and the target image is normalized to have the same scale, thereby generating image scaling data.

[0219] Step S422: based on the image scaling data, enhancing the image contrast and generating a balanced histogram;

[0220] In this example, a contrast enhancement operation is performed on the image scaling data, the image histogram is calculated, the pixel distribution of the image is obtained, the histogram is equalized, and the equalized histogram is enhanced to enhance the image contrast and highlight features, thereby generating a balanced histogram.

[0221] Step S423: performing image denoising operation using Gaussian filtering based on the equalized histogram to generate denoised image data;

[0222] In this example, Gaussian filtering is used to perform image denoising on the equalized histogram. By performing a convolution operation on each pixel in the image, each element in the convolution kernel represents the distance between the element and the target pixel. The closer the distance, the greater the weight of the element, and the farther the distance, the smaller the weight of the element. In this way, the image can be smoothed to generate denoised image data.

[0223] Step S424: performing a local feature extraction operation based on the denoised image data using a linear back-projection algorithm to generate a local feature vector;

[0224] In this example, the image data is projected to form a projection image, and the projection image is reconstructed into a three-dimensional stereo image using a linear back-projection algorithm. In the reconstructed three-dimensional stereo image, based on the local region of interest, a local feature descriptor extraction algorithm is used to obtain the local feature vector of each region, thereby generating a local feature vector.

[0225] Step S425: performing feature encoding operation based on the local feature vector using the visual word bag model to generate feature encoding data;

[0226] In this example, accelerated robust feature descriptors are extracted from local feature vectors, the descriptors are clustered into a single visual vocabulary, the local features in the image are mapped to the closest visual vocabulary, the frequency of occurrence of each vocabulary is calculated, and these frequencies are organized into a vector to generate a feature encoding vector.

[0227] Step S426: Perform dimensionality reduction processing based on the feature coding data using principal component analysis to generate feature point data.

[0228] In this example, feature coding data is read, normalized so that features of each dimension have the same importance, principal component dimensionality reduction is performed on the normalized feature coding data to obtain low-dimensional feature point data, and matching and recognition operations are performed on the low-dimensional feature point data to generate feature point data.

[0229] Therefore, no matter from which point of view, the embodiments should be regarded as illustrative and non-restrictive, and the scope of the present invention is limited by the appended claims rather than the above description, so it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any attached figure mark in the claims should not be regarded as limiting the claims involved.

[0230] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0231] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A trajectory speed optimization method for a serial robot arm based on deep learning, characterized in that: The method comprises the following steps: Step S1: performing linear interpolation processing on the motion trajectory of the robot arm to generate spatial velocity data, and performing pseudo-inverse data conversion on the spatial velocity data to generate joint spatial velocity data; Step S2: preprocessing the joint space velocity data to generate joint timestamp velocity data, using the joint timestamp velocity data to compare the values ​​with the preset rated joint velocity data, if the joint timestamp velocity data is greater than the preset rated joint velocity data, then optimizing the performance speed and marking the joint timestamp velocity data as overspeed space velocity data, otherwise outputting the joint timestamp velocity data; Step S3: Perform inverse planning and inference optimization processing on the speed data of the overspeed space to generate a joint linear data set. Step S3 includes: Step S31: Generate starting joint position data based on the overspeed space speed data using an inverse motion model control method; Step S32: Based on the joint position data, the joint axis is subjected to proportional adjustment control processing to generate joint axis overshoot data; Step S33: comparing the data based on the joint axis overshoot data by using data alignment analysis processing to generate joint velocity extreme value data; Step S34: obtaining joint threshold position data by using a robot arm joint position prediction method based on joint velocity extreme value data; Step S35: obtaining a joint linear data set using a linear dynamics equation calculation formula based on the joint threshold position data; Step S4: Perform trajectory prediction processing on the robotic arm according to the joint linear data set to generate secondary joint position data.

2. The trajectory speed optimization method of a serial robot arm based on deep learning according to claim 1 is characterized in that: Step S1 includes: Step S11: extracting the inertial measurement unit of the robotic arm to generate joint angle data; Step S12: Performing rotation matrix processing on the joint angle data using a kinematic model to obtain the joint position Jacobian matrix data; Step S13: using a pseudo-inverse processing method on the joint Jacobian matrix data to generate joint Jacobian inverse matrix data; Step S14: Perform inverse kinematics solution processing on the joint Jacobian inverse matrix data to generate joint space velocity data.

3. The trajectory speed optimization method of a serial robot arm based on deep learning according to claim 1, characterized in that: The specific steps of step S2 are: Step S21: performing time series cutting and storage processing on the joint space velocity data to generate joint time series data; Step S22: Using a window sliding algorithm to perform moving window data processing on the joint time series data to generate frame joint space velocity data; Step S23: using the long short-term memory network to perform model training processing on the frame joint space velocity data to generate model training set data; Step S24: Based on the model training set data, the data of the robot arm joint space velocity is audited, the model prediction processing is performed, the error situation is feedback controlled, and the joint timestamp velocity data is generated; Step S25: Process the joint timestamp velocity data using overspeed abnormality detection to generate overspeed spatial velocity data.

4. The trajectory speed optimization method of a serial robot arm based on deep learning according to claim 1, characterized in that: Step S31 includes: Step S311: constructing a kinematic model of the robot arm using an adaptive control algorithm based on the overspeed space speed data; Step S312: performing coordinate system data acquisition and processing according to the robot arm kinematics model to generate a base coordinate system; Step S312: using the sensor to perform positioning operations on the end of the manipulator based on the sensor-measured coordinates and Euler angles according to the base coordinate system, and generating posture matrix data; Step S313: performing posture matrix orthogonalization processing on the posture matrix data to generate a rotation matrix; Step S314: Perform joint data conversion mapping based on the rotation matrix to generate joint position data.

5. The trajectory speed optimization method of a serial robot arm based on deep learning according to claim 1, characterized in that: Step S34 includes: Step S341: using sensors to fit the motion trajectory data of the robotic arm to generate model raw data; Step S342: extracting features from the original model data to generate a model feature vector; Step S343: Perform joint threshold prediction processing using the trajectory prediction training model based on the model feature vector to generate joint threshold bit data.

6. The trajectory speed optimization method of a serial robot arm based on deep learning according to claim 1, characterized in that: The linear dynamics equation calculation formula of step S35 is specifically: in, It refers to the joint displacement segment data. It refers to the joint movement time. It is worth Joint movement time, It refers to the joint subscript value. It refers to The joint velocity function value, T refers to the joint motion cycle constant, It refers to The joint accelerations, The end position vector of the robot arm, It refers to The joint movement speed value, It refers to the value of the joint motion acceleration function. is a joint linear dataset, refers to the segment quantity value, Refers to the original position value of the robot arm.

7. The trajectory speed optimization method of a serial robot arm based on deep learning according to claim 1, characterized in that: Step S4 includes: Step S401: using a depth camera to collect and process point cloud data of the robotic arm to generate a point cloud data set; Step S402: performing fitting residual sum of squares processing on the point cloud data set to generate a point cloud linear data table set; Step S403: extracting geometric features of the robot arm based on the point cloud linear data table set to generate a training data set; Step S404: construct a data training model using a neural network algorithm based on the training data set, perform posture prediction processing according to the data training model, and generate secondary joint position data.

8. The trajectory speed optimization method of a serial robot arm based on deep learning according to claim 7, characterized in that: Step S402 includes: Step S411: performing data denoising and filtering processing based on the point cloud linear data table set to generate preprocessed data; Step S412: performing shape analysis feature extraction operation based on the preprocessed data to generate feature point data; Step S413: performing feature corresponding point detection processing on the feature point data to generate feature registration point data; Step S414: Perform an integrated training operation based on the feature registration point data to generate a training data set.

9. The trajectory speed optimization method of a serial robot arm based on deep learning according to claim 8, characterized in that: Step S412 includes: Step S421: Generate image scaling data using an image scale normalization processing method based on the preprocessed data; Step S422: based on the image scaling data, enhancing the image contrast and generating a balanced histogram; Step S423: performing image denoising operation using Gaussian filtering based on the equalized histogram to generate denoised image data; Step S424: performing a local feature extraction operation based on the denoised image data using a linear back-projection algorithm to generate a local feature vector; Step S425: performing feature encoding operation based on the local feature vector using the visual word bag model to generate feature encoding data; Step S426: Perform dimensionality reduction processing based on the feature coding data using principal component analysis to generate feature point data.

Images