Artificial intelligence-based robot joint structure load prediction method

Abstract

The application discloses an artificial intelligence-based robot joint structure load prediction method, relates to the field of robot control, and comprises the following steps: collecting load data of each joint in a robot motion process, identifying a load transmission time sequence and a decay law between adjacent joints, calculating a load distribution ratio at different joints, predicting a load transmission trend, performing real-time compensation and correction on a prediction result, and outputting an optimized load distribution scheme; converting the load distribution scheme into joint motion parameters, adjusting a motion trajectory of each joint, and reducing local load concentration. The application realizes accurate prediction and optimized distribution of robot joint loads through an artificial intelligence algorithm, effectively solves technical problems such as inaccurate load prediction and response lag in the prior art, and improves the operation efficiency and service life of the robot.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, and more specifically, to a method for predicting the load on robot joint structures based on artificial intelligence. Background Technology

[0002] With the continuous improvement of industrial automation, data acquisition robots are increasingly widely used in industrial production. Data acquisition robots need to perform frequent movements and load transfers, which places higher demands on the stability and lifespan of the robot's joint structure. Traditional robot joint load control methods mainly rely on pre-set fixed parameters for adjustment, which cannot effectively cope with complex and changing working conditions, easily leading to excessive load on local joints, affecting the robot's operating efficiency and lifespan.

[0003] In existing technologies, load prediction for robot joint structures typically employs static modeling methods, considering only the load state at a single moment and ignoring the dynamic transfer of load between joints. This method struggles to accurately reflect the actual load distribution during robot motion, leading to significant discrepancies between predictions and reality. Due to a lack of in-depth analysis of load transfer patterns, existing methods cannot effectively identify and predict the load transfer characteristics between adjacent joints, and cannot promptly optimize and adjust joint motion parameters. Summary of the Invention

[0004] The purpose of this invention is to provide an artificial intelligence-based method for predicting the load on robot joint structures, which solves the technical problems of low accuracy, slow response speed and poor adaptability in the existing technology of robot joint load prediction. It provides an intelligent prediction method that can accurately predict the load transmission law and achieve balanced load distribution, so as to improve the operating efficiency and service life of the robot.

[0005] The robot joint structure load prediction method based on artificial intelligence provided in this embodiment of the invention includes the following steps:

[0006] Collect load data of each joint during robot movement, and identify the timing and attenuation of load transmission between adjacent joints;

[0007] Based on the transmission timing and attenuation law, the distribution ratio of the load in different joints is calculated, and a load transmission characteristic matrix is ​​generated.

[0008] Based on the load transfer feature matrix, the load transfer trend is predicted, the prediction results are compensated and corrected in real time, and the optimized load distribution scheme is output.

[0009] The load distribution scheme is converted into joint motion parameters, and the motion trajectory of each joint is adjusted to reduce the local load concentration.

[0010] Furthermore, collecting load data from each joint during robot movement and identifying the timing and attenuation patterns of load transmission between adjacent joints includes:

[0011] The load torque and angle signals of each joint of the robot are collected, and multi-band torque components are obtained through wavelet decomposition and cross-correlation function is constructed to obtain the initial phase relationship.

[0012] The load transfer direction and attenuation characteristics between adjacent joints are determined based on the angle signal. The initial phase is corrected based on the attenuation characteristics, and the load transfer timing and attenuation pattern between adjacent joints are identified.

[0013] Furthermore, the initial phase relationship is obtained by obtaining multi-band torque components through wavelet decomposition and constructing a cross-correlation function, including:

[0014] The load torque signal is decomposed into multi-band wavelet components to obtain multi-band torque components. The energy value of each frequency band torque component is calculated and converted into the weighting coefficient of the corresponding frequency band.

[0015] Construct cross-correlation functions for torque components in the same frequency band of adjacent joints, and determine the phase delay characteristics of the frequency band based on the time delay corresponding to the maximum value of the cross-correlation function;

[0016] The initial phase relationship is obtained by weighting the weighting coefficients with the phase delay characteristics of each frequency band.

[0017] Furthermore, based on the transmission timing and attenuation patterns, the load distribution ratio at different joints is calculated, generating a load transmission characteristic matrix including:

[0018] A transfer function containing coupling and attenuation components is constructed based on the transfer timing and attenuation rules.

[0019] The coupling and attenuation components in the transfer function are integrated separately, and the dynamic coupling coefficient is obtained by combining the joint motion posture.

[0020] The transfer function is modified using dynamic coupling coefficients to construct a joint load transfer relationship matrix;

[0021] Solving the coupling matrix yields the load distribution ratio, and combining the transfer function and the load distribution ratio generates the load transfer characteristic matrix.

[0022] Furthermore, the transfer function is modified using dynamic coupling coefficients, and the joint load transfer relationship matrix is ​​constructed as follows:

[0023] The initial coupling matrix is ​​divided into blocks and diagonalized based on the dynamic coupling coefficients to obtain the weighted coupling matrix.

[0024] Singular value decomposition is performed on the weighted coupling matrix to reconstruct the eigenvectors and generate coupling matrix elements;

[0025] Fill the elements of the coupling matrix into the weighted coupling matrix to construct the joint load transfer relationship matrix.

[0026] Furthermore, predicting load transfer trends based on the load transfer feature matrix includes:

[0027] A sequence of feature vectors is obtained by temporal decomposition based on the load transfer feature matrix;

[0028] The load mutation points are identified based on the changing trends of the feature vector sequence. These load mutation points are used as boundaries to divide the prediction intervals, and the load transmission trend is predicted within each prediction interval.

[0029] Furthermore, the prediction results are compensated and corrected in real time, and the optimized load distribution scheme is output, including:

[0030] Calculate the residual sequence between the predicted and measured values ​​of the load transfer trend, and decompose the residual sequence into periodic and random components;

[0031] The compensation value is calculated based on the periodic and random components. The compensation value is then fused with the predicted value of the load transmission trend to output the optimized load distribution scheme.

[0032] Furthermore, the load distribution scheme is converted into joint motion parameters, and the motion trajectory of each joint is adjusted to reduce local load concentration, including:

[0033] Establish the mapping relationship between the target load value of each joint and the joint motion parameters in the load distribution scheme, and generate the initial parameters of joint motion;

[0034] Calculate the deviation between the actual load value and the target load value to determine the motion parameter correction amount;

[0035] The motion trajectory of each joint is adjusted based on the correction of motion parameters until the local load concentration is lower than the preset concentration threshold.

[0036] This invention also provides a technical solution for an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the aforementioned methods.

[0037] This invention also provides a technical solution: a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the steps of any of the aforementioned methods.

[0038] In this embodiment, by collecting real-time load data of each joint during motion and combining it with artificial intelligence algorithms for deep learning and analysis of load transfer patterns, the timing and attenuation characteristics of load transfer between adjacent joints can be accurately identified, enabling precise prediction of load distribution. The dynamic optimization method based on the load transfer feature matrix can adjust joint motion parameters in real time, effectively reducing local load concentration and preventing certain joints from being under high load for extended periods. A real-time compensation and correction mechanism can promptly correct prediction deviations, improving the accuracy and reliability of load prediction. This invention significantly enhances the robot's adaptability to complex working conditions, reduces joint wear, and extends the robot's service life. Attached Figure Description

[0039] Figure 1 A flowchart illustrating an artificial intelligence-based robot joint structure load prediction method provided in an embodiment of the present invention;

[0040] Figure 2 This is a diagram illustrating the nonlinear mapping relationship between load and attenuation for various joint types according to embodiments of the present invention.

[0041] Figure 3 This is a schematic diagram of the multi-joint load transfer coupling strength matrix according to an embodiment of the present invention. Detailed Implementation

[0042] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0043] like Figure 1 As shown, Figure 1 A flowchart of an artificial intelligence-based robot joint structure load prediction method provided in an embodiment of the present invention is shown. The method includes the following steps:

[0044] Collect load data of each joint during robot movement, and identify the timing and attenuation of load transmission between adjacent joints;

[0045] Based on the transmission timing and attenuation law, the distribution ratio of the load in different joints is calculated, and a load transmission characteristic matrix is ​​generated.

[0046] Based on the load transfer feature matrix, the load transfer trend is predicted, the prediction results are compensated and corrected in real time, and the optimized load distribution scheme is output.

[0047] The load distribution scheme is converted into joint motion parameters, and the motion trajectory of each joint is adjusted to reduce the local load concentration.

[0048] Furthermore, collecting load data from each joint during robot movement and identifying the timing and attenuation patterns of load transmission between adjacent joints includes:

[0049] The load torque and angle signals of each joint of the robot are collected, and multi-band torque components are obtained through wavelet decomposition and cross-correlation function is constructed to obtain the initial phase relationship.

[0050] The load transfer direction and attenuation characteristics between adjacent joints are determined based on the angle signal. The initial phase is corrected based on the attenuation characteristics, and the load transfer timing and attenuation pattern between adjacent joints are identified.

[0051] The system employs a distributed sensor network to collect load torque and angle signals from each joint of the robot. During data acquisition, high-precision torque sensors and angle encoders are installed at each joint, with a sampling frequency set to 1000Hz to ensure the capture of minute changes during joint movement. The collected data undergoes preprocessing by the controller, including noise reduction and filtering to eliminate environmental interference and sensor drift.

[0052] Wavelet decomposition was applied to the acquired torque signal to decompose the original signal into torque components in multiple frequency bands. Specifically, the db4 wavelet basis function was selected to perform a 5-level decomposition of the torque signal, obtaining detail coefficients and approximation coefficients for different frequency bands. Low-frequency components reflect the basic trend of joint load changes, while high-frequency components contain load abrupt changes and vibration information. For each frequency band signal after decomposition, a cross-correlation function was constructed to analyze the correlation of torque signals between adjacent joints. The cross-correlation function was calculated using a sliding time window with a length of 500 ms and a step size of 50 ms, reflecting the load transfer relationship between joints at different time scales. The initial phase relationship was determined by the peak position of the cross-correlation function, and the time offset of the peak position represents the time delay of load transfer.

[0053] The load transfer direction between adjacent joints is determined based on angle signals. When the angle change of the upstream joint precedes that of the downstream joint, and the two angle changes are positively correlated, the load is determined to be transferred from the upstream joint to the downstream joint. An angular velocity sequence is constructed by calculating the angle difference between adjacent time points, and the positive or negative change of the angular velocity sequence reflects the joint movement direction. When the angular velocity change of the upstream joint precedes that of the downstream joint in the same direction, it is determined to be forward transfer; otherwise, it is reverse transfer.

[0054] The attenuation characteristics of the load during transmission are analyzed. The envelope of the torque signal is extracted, and the peak torque ratio between adjacent joints is calculated to obtain the attenuation coefficient. The torque signal envelope is extracted using the moving average method with a window size of 100 ms. For rigid joint pairs, the attenuation coefficient is typically between 0.8 and 0.95; for flexible joint pairs, the attenuation coefficient is between 0.5 and 0.8. The attenuation coefficient is closely related to the joint material, structure, and connection method.

[0055] The initial phase relationship is corrected based on attenuation characteristics. The initial phase and attenuation coefficient are weighted and fused; the smaller the attenuation coefficient, the greater the phase correction. The weighting coefficient is set to 0.7, and the correction formula is to multiply the initial phase value by the sum of the attenuation coefficient and the weighting coefficient. The corrected phase more accurately reflects the time characteristics of load transmission, eliminating phase errors caused by differences in joint structure.

[0056] Identify the timing of load transfer between adjacent joints. Based on the corrected phase relationship, establish a timing diagram for load transfer between joints. A phase difference greater than a critical value (set to 30 degrees) is considered significant timing transfer, while a phase difference less than the critical value is considered approximately synchronous transfer. For joint pairs with significant timing transfer, record their transfer delay time. In a multi-joint series structure, by comparing the transfer delays of each joint pair, the primary and secondary paths of load transfer can be identified.

[0057] Identify the attenuation pattern of load transfer. Establish a load-attenuation mapping relationship by analyzing the changes in the attenuation coefficient under different operating conditions. Under low load conditions (torque less than 30% of the rated value), attenuation is mainly affected by the joint structure and exhibits a linear relationship. Under high load conditions (torque greater than 70% of the rated value), attenuation shows a non-linear relationship with the load and is affected by factors such as joint temperature and operating time. Establish a database of joint load and attenuation coefficients for subsequent load prediction and anomaly detection.

[0058] like Figure 2 As shown, Figure 2 This diagram illustrates the nonlinear mapping relationship between load and attenuation for various joint types in this embodiment. It shows the variation of the attenuation coefficient for the three connection types under different load ratios. This technical solution (solid red line) captures the linear relationship in the low load region (0-40%) and the nonlinear characteristics in the high load region (60-100%), while the traditional linear regression model (dashed blue line) only presents a simple linear change. The attenuation coefficient decreases from 0.92 to 0.81 for rigid connections, from 0.72 to 0.61 for semi-rigid connections, and from 0.53 to 0.38 for flexible connections.

[0059] This method achieves accurate prediction of joint structure load by deeply analyzing the load transfer characteristics of robot joints, providing an effective tool for robot health monitoring and fault diagnosis. Through wavelet decomposition and cross-correlation analysis, it accurately captures the temporal sequence and attenuation law of load transfer between joints, avoiding prediction bias caused by neglecting the time factor in traditional methods. It features strong real-time performance and good adaptability, capable of handling load variations under different robot operating conditions, and can automatically adjust the prediction model based on historical data.

[0060] Furthermore, the initial phase relationship is obtained by obtaining multi-band torque components through wavelet decomposition and constructing a cross-correlation function, including:

[0061] The load torque signal is decomposed into multi-band wavelet components to obtain multi-band torque components. The energy value of each frequency band torque component is calculated and converted into the weighting coefficient of the corresponding frequency band.

[0062] Construct cross-correlation functions for torque components in the same frequency band of adjacent joints, and determine the phase delay characteristics of the frequency band based on the time delay corresponding to the maximum value of the cross-correlation function;

[0063] The initial phase relationship is obtained by weighting the weighting coefficients with the phase delay characteristics of each frequency band.

[0064] The preprocessed load torque signal was subjected to multi-level wavelet decomposition using the sym6 wavelet basis function, with five levels of decomposition applied to the torque signal of each joint. The decomposition yielded five detail coefficients (D1-D5) and one approximation coefficient (A5), representing torque components in different frequency bands. Specifically, D1 (50-100Hz) reflects high-frequency vibrations, D2 (25-50Hz) reflects mid-to-high frequency vibrations, D3 (12.5-25Hz) reflects mid-frequency components, D4 (6.25-12.5Hz) reflects mid-to-low frequency components, D5 (3.125-6.25Hz) reflects low-frequency vibrations, and A5 (0-3.125Hz) reflects the basic load trend. For example, after wavelet decomposition, the torque signal of the third joint of a six-axis robot shows that the A5 band reflects the basic load change, accounting for about 45% of the energy; the D5 band captures the low-frequency load change, accounting for about 25% of the energy; the D4 band captures the load fluctuation, accounting for about 15% of the energy; and the D1-D3 bands reflect high-frequency vibration and noise, accounting for about 15% of the energy.

[0065] The energy values ​​of the torque components in each frequency band are calculated by summing the squares of the coefficients for each frequency band. Taking the second joint of a six-axis robot performing an assembly task as an example, the energy value of the A5 band is 18500, the D5 band is 7200, the D4 band is 3600, the D3 band is 2100, the D2 band is 1500, and the D1 band is 900, for a total energy value of 33800. Dividing the energy value of each frequency band by the total energy value yields the weighting coefficients for each band, which are 0.547 (A5), 0.213 (D5), 0.107 (D4), 0.062 (D3), 0.044 (D2), and 0.027 (D1). These weighting coefficients reflect the importance of each frequency band component in load transfer, with higher energy bands having a greater weight in load transfer characteristics.

[0066] Cross-correlation functions for the torque components of adjacent joints in the same frequency band were constructed to analyze the transmission relationship of torque signals between adjacent joints. The cross-correlation function reflects the phase relationship between the signals by calculating the similarity of the two signals at different time delays. The calculation method involves performing a sliding multiplication of the torque signal sequence of one joint with the torque signal sequence of another joint and summing the results to obtain the correlation value sequence at different sliding positions. For each frequency band component, its cross-correlation function was calculated separately. A time window of length 2048 points was used for the cross-correlation function calculation, with a window sliding step of 10 ms and a calculation range of ±500 ms time delay.

[0067] The phase delay characteristics of a frequency band are determined by the time delay corresponding to the maximum value of the cross-correlation function. The global maximum value of the cross-correlation function is identified; the time delay corresponding to this point is the phase delay characteristic of that frequency band. For example, when a six-axis robot performs a handling task, the phase delay between the second and third joints is 35ms in band A5, 28ms in band D5, 22ms in band D4, 18ms in band D3, 15ms in band D2, and 12ms in band D1. Phase delay reflects the time difference in load transmission between joints across different frequency bands; high-frequency components typically transmit faster than low-frequency components.

[0068] The initial phase relationship is obtained by weighting the weighting coefficients with the phase delay characteristics of each frequency band. The weighting method is to multiply the phase delay of each frequency band by the corresponding weighting coefficient and then sum them. The initial phase relationship comprehensively considers the load transfer characteristics of each frequency band and provides the overall timing characteristics of load transfer between adjacent nodes.

[0069] like Figure 3 As shown, Figure 3This diagram illustrates the coupling strength matrix of load transfer between multiple joints in this embodiment, showing the coupling strength of the load transfer relationship between the joints of the six-axis robot. A value of 1.0 on the main diagonal represents the load itself, with the strongest coupling between adjacent joints (e.g., 0.78 for the second and third joints). The lower triangular region has a higher value than the upper triangular region, indicating that the load transfer from the base to the end effector is more efficient; for example, the value for the fourth and fifth joints is 0.65, while the reverse is only 0.32.

[0070] This method achieves accurate characterization of the load transfer characteristics of robot joint structures through multi-layer wavelet decomposition and energy-weighted cross-correlation analysis, providing a solid foundation for load prediction. By decomposing the torque signal into multiple frequency bands and analyzing their energy distribution, the load transfer characteristics in different frequency domains can be comprehensively captured. The phase delay extraction method based on cross-correlation functions avoids the limitations of traditional time-domain analysis methods and can accurately identify the load transfer timing in different frequency bands. The weighted fusion strategy comprehensively considers the contribution of each frequency band to load transfer, generating a more accurate initial phase relationship.

[0071] Furthermore, based on the transmission timing and attenuation patterns, the load distribution ratio at different joints is calculated, generating a load transmission characteristic matrix including:

[0072] A transfer function containing coupling and attenuation components is constructed based on the transfer timing and attenuation rules.

[0073] The coupling and attenuation components in the transfer function are integrated separately, and the dynamic coupling coefficient is obtained by combining the joint motion posture.

[0074] The transfer function is modified using dynamic coupling coefficients to construct a joint load transfer relationship matrix;

[0075] Solving the coupling matrix yields the load distribution ratio, and combining the transfer function and the load distribution ratio generates the load transfer characteristic matrix.

[0076] The transfer function, comprising coupling and attenuation components, is constructed based on the transmission timing and attenuation rules. The method for constructing the transfer function involves converting the time delay into a phase offset and the attenuation coefficient into an amplitude reduction factor. The coupling component is represented as a time-domain delay response, achieved by time-shifting the original load signal; the offset is the time delay value determined by the transmission timing. The attenuation component is represented as an amplitude attenuation response, achieved by scaling the original load signal; the scaling factor is the attenuation coefficient determined by the attenuation rules. Taking a six-axis robot as an example, the transfer function from the second to the third joint consists of two parts: the coupling component is the response with a 30ms time delay, and the attenuation component is the response with an amplitude reduction of 0.82 times. The parameters of the transfer function are dynamically adjusted according to the robot's operating state. During high-speed motion, the time delay decreases by approximately 10%, and the attenuation coefficient increases by approximately 5%; under heavy load conditions, the time delay increases by approximately 15%, and the attenuation coefficient decreases by approximately 8%.

[0077] The coupling and attenuation components in the transfer function are integrated separately, and the dynamic coupling coefficient is obtained by combining the joint motion posture. The integration method involves accumulating and summing the coupling and attenuation components within a specified time window of 500 ms with a sliding step of 50 ms. The joint motion posture is acquired in real time through an angle encoder, including joint angle, angular velocity, and angular acceleration information. The calculation of the dynamic coupling coefficient considers the influence of joint motion state on load transfer by weighting the integration results with joint posture parameters. The weighting coefficients are learned from historical data using machine learning methods, and a random forest algorithm is used to establish the mapping relationship between posture parameters and coupling coefficients. When the six-axis robot performs a grasping task, the dynamic coupling coefficient is 0.65 when the third joint is in an extended position (angle of 75 degrees); when it is in a bent position (angle of 30 degrees), the dynamic coupling coefficient increases to 0.78, indicating that joint posture significantly affects load transfer characteristics.

[0078] The transfer function is modified using a dynamic coupling coefficient to construct a joint load transfer relationship matrix. The modification method involves multiplying the coupling and attenuation components of the transfer function by the dynamic coupling coefficient, and then recombine them to form the modified transfer function. The coupling matrix is ​​a two-dimensional array describing the load transfer relationship between all adjacent joint pairs of the robot. The rows and columns of the matrix correspond to the robot's joint numbers, and the matrix elements are the modified transfer functions for the corresponding joint pairs. For a six-axis robot, the coupling matrix is ​​a 6×6 square matrix, with the main diagonal elements being 1 (representing the joint itself) and the off-diagonal elements representing the modified transfer functions between adjacent joints. For example, the element in the second row and third column represents the modified transfer function from the second joint to the third joint, and its value is determined by the product of the original transfer function and the dynamic coupling coefficient 0.78 (for bending positions). For non-adjacent joints, such as the first and third joints, the elements of their coupling matrix are calculated by combining the transfer functions of the intermediate joints.

[0079] The load distribution ratio is obtained by solving the coupling matrix. The transfer function and the load distribution ratio are then combined to generate the load transfer characteristic matrix. The load distribution ratio is calculated by performing eigenvalue decomposition on the coupling matrix, extracting the principal eigenvectors, and normalizing these eigenvectors to obtain the load distribution ratio for each joint. When a six-axis robot performs a handling task, the load distribution ratios for each joint are: 0.15 for the first joint, 0.25 for the second joint, 0.30 for the third joint, 0.15 for the fourth joint, 0.10 for the fifth joint, and 0.05 for the sixth joint. The load transfer characteristic matrix is ​​constructed by performing tensor multiplication between the corrected transfer function and the load distribution ratio to obtain a high-dimensional matrix characterizing the load transfer properties of the entire robot. This load transfer characteristic matrix contains all the information about the load transfer between the joints in the current state of the robot and can be used for subsequent load prediction and anomaly detection.

[0080] The predictive capability of the load transfer feature matrix was verified under different operating conditions. It was found that under standard operating conditions, the load prediction error based on the feature matrix was less than 7%; under high-speed conditions, the prediction error increased to 12%; and under heavy-load conditions, the prediction error was approximately 15%. By introducing a working condition recognition module, the most suitable feature matrix parameters for the current working condition can be automatically selected, keeping the average prediction error below 10%. The feature matrix is ​​updated at a frequency of 10 Hz, which can adapt to the dynamically changing working state of the robot.

[0081] This method constructs a transfer function containing coupling and attenuation components and dynamically corrects it based on joint motion posture, achieving an accurate description and prediction of the load transfer characteristics of robot joint structures. It uses coupling and characteristic matrices to characterize complex multi-joint load transfer relationships, effectively solving the nonlinear coupling problem that traditional methods struggle with. Furthermore, it can dynamically adjust the load transfer model based on the robot's real-time motion state, adapting to load variations under different working conditions.

[0082] Furthermore, the transfer function is modified using dynamic coupling coefficients, and the joint load transfer relationship matrix is ​​constructed as follows:

[0083] The initial coupling matrix is ​​divided into blocks and diagonalized based on the dynamic coupling coefficients to obtain the weighted coupling matrix.

[0084] Singular value decomposition is performed on the weighted coupling matrix to reconstruct the eigenvectors and generate coupling matrix elements;

[0085] Fill the elements of the coupling matrix into the weighted coupling matrix to construct the joint load transfer relationship matrix.

[0086] The load transfer weights are calculated based on the dynamic coupling coefficients. The calculation method involves normalizing the dynamic coupling coefficients to ensure that the sum of all weights is 1. For example, the dynamic coupling coefficient of each pair of adjacent joints is divided by the sum of the coupling coefficients of all adjacent joint pairs. For the six-axis robot described above, the normalized load transfer weights are as follows: joints 1-2: 0.196, joints 2-3: 0.226, joints 3-4: 0.177, joints 4-5: 0.213, and joints 5-6: 0.188. These weights reflect the relative importance of each adjacent joint pair in the overall load transfer. A larger weight value indicates a more significant influence of that joint pair in the load transfer network.

[0087] The initial coupling matrix is ​​divided into blocks and diagonalized based on load transfer weights. The initial coupling matrix is ​​an n×n matrix (n is the number of joints), where off-diagonal elements represent the coupling relationships between joints, and diagonal elements represent the characteristics of the joints themselves. The purpose of block diagonalization is to highlight the coupling relationships between adjacent joints and weaken the indirect coupling between non-adjacent joints. The initial coupling matrix is ​​divided into multiple 2×2 sub-matrices, each sub-matrice corresponding to a pair of adjacent joints. Then, load transfer weights are applied to each sub-matrice for weighting. Taking a six-axis robot as an example, the initial 6×6 coupling matrix is ​​divided into five 2×2 sub-matrices (joints 1-2, 2-3, 3-4, 4-5, 5-6). The sub-matrix of joint 2-3 is weighted with a weight of 0.226 to obtain its contribution to the overall coupling relationship. After weighting all sub-matrices, they are recombine to obtain a weighted coupling matrix representing the load transfer characteristics.

[0088] Singular value decomposition (SVD) is performed on the weighted coupling matrix to obtain eigenvectors. SVD is a matrix factorization technique that decomposes the original matrix into the product of three matrices: a left singular matrix, a singular value diagonal matrix, and the transpose of the right singular matrix. A truncated SVD method is used, retaining the k largest singular values ​​(k is typically chosen to be half the number of joints) to reduce computational complexity and filter noise. For a six-axis robot, the first three singular values ​​are retained. SVD yields the left and right singular matrices of the weighted coupling matrix, where the column vectors of the left singular matrix are eigenvectors, representing the main load transfer patterns. For example, the obtained eigenvectors have a dimension of 6×3, with each column representing a load transfer pattern. The first eigenvector [0.35, 0.42, 0.45, 0.38, 0.32, 0.25] represents the main load transfer path, with the weight reaching its maximum at the third joint, reflecting the core role of this joint in load transfer.

[0089] The feature vectors are reconstructed according to the load transfer direction to generate coupling matrix elements reflecting the load distribution. The reconstruction process considers the load transfer direction, arranging the feature vector elements in order from upstream joint to downstream joint. For forward transfer (from joint 1 to joint 6), the feature vector elements maintain their original order; for reverse transfer (from joint 6 to joint 1), the order of the feature vector elements is reversed. The current load transfer direction is identified based on the real-time monitored torque signal, and then an appropriate reconstruction strategy is selected. For example, when performing a grasping task, forward reconstruction is used when the load is mainly transferred from the base to the end; reverse reconstruction is used when the load is mainly transferred from the end to the base. The reconstructed feature vectors are multiplied by the corresponding singular values ​​to generate coupling matrix elements reflecting the load distribution. For joints 2-3, the reconstructed coupling matrix element value is 0.38, indicating that the load transferred from joint 2 to joint 3 is 38%.

[0090] The elements of the coupling matrix are filled into the weighted coupling matrix to construct the joint load transfer relationship matrix. The filling method involves placing the reconstructed elements into their corresponding positions in the weighted coupling matrix according to the correspondence between joint pairs. For a robot with n joints, the coupling matrix is ​​an n×n square matrix, where the element (i,j) represents the load transfer ratio from joint i to joint j. For non-adjacent joint pairs, the element values ​​are calculated by multiplying the values ​​of adjacent joints, reflecting the cumulative effect of load transfer through multiple joints. For example, the load transfer ratio from joint 1 to joint 3 is the product of the transfer ratios of joints 1-2 and 2-3, i.e., 0.32×0.38=0.122. After filling all elements, a complete coupling matrix is ​​obtained, comprehensively characterizing the load transfer relationship between robot joints.

[0091] This method introduces dynamic coupling coefficients and load transfer weights, combined with matrix factorization techniques, to achieve accurate modeling of the load transfer characteristics of robot joint structures. The feature extraction method based on singular value decomposition effectively captures the main patterns of load transfer, avoiding the limitations of traditional methods in handling high-dimensional nonlinear systems. By considering a feature reconstruction strategy based on the load transfer direction, the model's adaptability to bidirectional load transfer is enhanced.

[0092] Furthermore, predicting load transfer trends based on the load transfer feature matrix includes:

[0093] A sequence of feature vectors is obtained by temporal decomposition based on the load transfer feature matrix;

[0094] The load mutation points are identified based on the changing trends of the feature vector sequence. These load mutation points are used as boundaries to divide the prediction intervals, and the load transmission trend is predicted within each prediction interval.

[0095] Furthermore, the prediction results are compensated and corrected in real time, and the optimized load distribution scheme is output, including:

[0096] Calculate the residual sequence between the predicted and measured values ​​of the load transfer trend, and decompose the residual sequence into periodic and random components;

[0097] The compensation value is calculated based on the periodic and random components. The compensation value is then fused with the predicted value of the load transmission trend to output the optimized load distribution scheme.

[0098] The process involves time-series decomposition based on the load transfer feature matrix to obtain a sequence of feature vectors. The time-series decomposition employs a sliding window technique with a window length of 2 seconds and a sliding step size of 200 ms. Within each window, eigenvalue decomposition is applied to the load transfer feature matrix to extract principal eigenvectors. Eigenvalue decomposition is achieved by calculating the eigenvalues ​​and eigenvectors of the feature matrix, selecting the eigenvector with the largest eigenvalue as the principal eigenvector. The principal eigenvector reflects the main load transfer pattern within the current time window. When performing a handling task, the extracted principal eigenvector has a dimension of 6×1, such as [0.32, 0.45, 0.58, 0.42, 0.35, 0.22], indicating that the third joint bears the largest proportion of the load. As the sliding window moves, a series of principal eigenvectors are obtained, forming a feature vector sequence that reflects the evolution of the load transfer pattern over time. The sampling rate of the feature vector sequence is 5 Hz, sufficient to capture the dynamic characteristics of load changes in most industrial applications.

[0099] Load abrupt change points are identified based on the changing trends of feature vector sequences. A load abrupt change point refers to a point in time where the load transfer pattern changes significantly, typically corresponding to a transition in the robot's task phase or a sudden change in the external environment. The identification method involves calculating the Euclidean distance between adjacent feature vectors; when the distance exceeds a preset threshold, it is determined to be a load abrupt change point. The threshold is set to three times the standard deviation of normal fluctuations, i.e., 0.15. For example, during the task transition from gripping to handling, the Euclidean distance between adjacent feature vectors jumps from 0.05 to 0.22, significantly exceeding the threshold, and is therefore identified as a load abrupt change point. The load abrupt change points are used as boundaries to divide prediction intervals, with each pair of adjacent abrupt change points forming a prediction interval. Within each prediction interval, the load transfer pattern is relatively stable, making it suitable to use a unified prediction model. For a six-axis robot performing an assembly task, four load abrupt change points were identified during a 30-second operation, dividing the entire process into five prediction intervals, corresponding to the five phases of preparation, gripping, movement, assembly, and release.

[0100] The load transfer trend is predicted within each prediction interval. The prediction method employs a Long Short-Term Memory (LSTM) network, which is particularly suitable for processing time-series data. The LSM network structure consists of an input layer, two hidden layers (64 neurons each), and an output layer. The input is the feature vectors from the past 10 time points, and the output is the predicted feature vector values ​​for the next 3 time points. Network training uses the Adam optimizer with a learning rate of 0.001 and a batch size of 32. Taking the third prediction interval (movement phase) as an example, the LSM network model predicts that the load proportion at the third joint gradually decreases from 0.58 to 0.49, while the load proportion at the second joint increases from 0.45 to 0.52, reflecting the trend of load transfer from the third joint to the second joint during the movement. The mean absolute error between the predicted load transfer trend and the measured value is 0.035, meeting the accuracy requirements for industrial applications.

[0101] The residual sequence between the predicted and measured values ​​of the load transfer trend is calculated. The residual calculation method involves subtracting the predicted eigenvector from the measured eigenvector to obtain the residual vector, and then combining the residual vectors at consecutive time points to form the residual sequence. Taking the fourth prediction interval (assembly stage) of a six-axis robot as an example, the residual sequence of the third joint is [-0.02, 0.03, -0.04, 0.03, -0.02], exhibiting certain fluctuations. The residual sequence is decomposed into periodic and random components. The decomposition method employs seasonal decomposition techniques. First, the main frequency components in the residual sequence are identified through Fast Fourier Transform. Then, based on the frequency characteristics, the residual is divided into periodic variations (periodic components) and irregular fluctuations (random components). For the above residual sequence, the decomposed periodic components are [-0.01, 0.02, -0.03, 0.02, -0.01], and the random components are [-0.01, 0.01, -0.01, 0.01, -0.01]. The periodic component reflects the regular load fluctuations during the robot's motion, while the random component reflects unpredictable external disturbances and system noise.

[0102] Compensation values ​​are calculated based on periodic and random components. The calculation method involves applying an autoregressive model to predict the periodic components and calculating the mean and variance of the random components to construct confidence intervals. The autoregressive model has an order of 3, and the model parameters are estimated using the least squares method. The predicted value of the periodic component is added to the mean of the random component to obtain the comprehensive compensation value. For the third joint of a six-axis robot during the assembly stage, the compensation value sequence is [0.02, -0.03, 0.02, -0.01, 0.01]. The compensation value is then fused with the predicted value of the load transfer trend by direct addition to obtain the optimized load distribution scheme. The optimized load distribution scheme more accurately reflects the actual load condition, providing a reliable basis for adjusting the robot's control strategy and monitoring its health status.

[0103] Dedicated predictive model libraries are established for different types of industrial robots and various application scenarios. These libraries contain multiple pre-trained Long Short-Term Memory (LSTM) networks, automatically selecting the most suitable predictive model based on robot type, load characteristics, and task type. An online learning mechanism is incorporated to enable the predictive models to adapt to performance changes and wear conditions during long-term robot use, maintaining stable predictive accuracy. Prediction results are displayed in real-time through a visual interface, supporting engineers in intuitive analysis and decision-making.

[0104] This method achieves high-precision prediction of robot joint loads through deep learning-based temporal analysis and residual compensation techniques. The feature vector sequence extracted by temporal decomposition comprehensively captures the dynamic changes in load transfer patterns, while the mutation point identification technique accurately divides the prediction interval, providing a reliable foundation for a segmented prediction strategy. The prediction model based on a long short-term memory network fully utilizes the temporal dependencies in historical data, effectively overcoming the nonlinearity and long-term dependency problems that traditional prediction methods struggle to handle. Residual decomposition and compensation mechanisms further improve prediction accuracy, particularly in handling periodic load fluctuations. By fusing predicted and compensated values, the final output load distribution scheme reflects both long-term trends and short-term fluctuations, providing comprehensive and reliable load information for robot control decisions.

[0105] Furthermore, the load distribution scheme is converted into joint motion parameters, and the motion trajectory of each joint is adjusted to reduce local load concentration, including:

[0106] Establish the mapping relationship between the target load value of each joint and the joint motion parameters in the load distribution scheme, and generate the initial parameters of joint motion;

[0107] Calculate the deviation between the actual load value and the target load value to determine the motion parameter correction amount;

[0108] The motion trajectory of each joint is adjusted based on the correction of motion parameters until the local load concentration is lower than the preset concentration threshold.

[0109] The process involves obtaining the target load values ​​for each joint in the load distribution scheme and establishing a mapping relationship between the target load values ​​and joint motion parameters. This mapping relationship is learned from historical running data using a machine learning algorithm, employing a gradient boosting decision tree model to construct a non-linear mapping between the target load values ​​and joint motion parameters. Joint motion parameters include four dimensions: joint angle, angular velocity, angular acceleration, and running time. The training dataset contains 5000 samples, each containing complete motion parameters and corresponding load data. The model configuration includes 100 decision trees with a maximum depth of 5, a learning rate of 0.1, and uses mean squared error as the loss function. After model training, the prediction accuracy on the test dataset reached 92%, validating the model's reliability. Taking the third joint as an example, when the angle is set to 60 degrees, the angular velocity is 30 degrees / s, and the angular acceleration is 15 degrees / s... 2 When the running time is 2 seconds, the predicted load value is 28.5 N·m, which is close to the target load value of 30 N·m.

[0110] Initial parameters for joint motion are generated based on the established mapping relationship. The generation method uses a reverse lookup technique, where a given target load value is used to deduce the corresponding combination of motion parameters. Considering that a single load value may correspond to multiple sets of motion parameters, additional constraints are introduced: minimizing energy consumption, minimizing runtime, or achieving the smoothest motion trajectory. When performing the handling task, minimizing energy consumption is chosen as the constraint, resulting in the following initial motion parameters for the third joint: angle range 45-75 degrees, maximum angular velocity 25 degrees / s, and maximum angular acceleration 12 degrees / s². 2 The runtime is 2.2 seconds. These parameters are expected to generate a target load close to 30 N·m while maintaining low energy consumption. Initial motion parameters are generated for all six joints, forming a complete initial motion trajectory.

[0111] Actual load data for each joint was collected under initial parameters. A high-precision torque sensor was used during data acquisition, with a sampling frequency of 1000Hz to ensure the capture of instantaneous load changes. A moving average filter was applied to the raw data with a window size of 50ms to eliminate high-frequency noise. Taking the trial run data of a six-axis robot as an example, the average actual load of the third joint was 33.2 N·m, with a peak of 36.5 N·m and a trough of 29.8 N·m. The deviation between the actual load and the target load was calculated by subtracting the target load from the actual load. For the third joint, the average deviation was 3.2 N·m, indicating that the actual load was slightly higher than the target value.

[0112] The correction amounts for the motion parameters are determined based on the deviation values. The correction amounts are calculated using a proportional-integral-derivative (PID) control strategy, with the proportional coefficient set to 0.6, the integral coefficient to 0.1, and the derivative coefficient to 0.2. For a deviation of 3.2 N·m in the third joint, the calculated angular velocity correction is -2.5 degrees / s², and the angular acceleration correction is -1.5 degrees / s². 2 The sign of the correction amount indicates whether the corresponding parameter is increased or decreased, and the magnitude of the value reflects the adjustment range. For joints with larger deviations, such as the third joint, the correction range is relatively large; for joints with smaller deviations, such as the sixth joint (deviation of only 0.5 N·m), the correction range is relatively small, with an angular velocity correction of -0.4 degrees / s.

[0113] The correction values ​​are superimposed on the initial motion parameters to adjust the motion trajectory of each joint. The superposition method is to directly add the correction values ​​to the initial parameters to obtain the updated motion parameters. For the third joint, the updated maximum angular velocity is 22.5 degrees / s² (25-2.5), and the maximum angular acceleration is 10.5 degrees / s². 2 (12-1.5). The angle range and runtime remain unchanged because these parameters are typically determined by task requirements and are not involved in load adjustment. The updated motion parameters are applied to the robot control system to generate a new motion trajectory.

[0114] Calculate the local load concentration to assess the uniformity of load distribution. Local load concentration is defined as the proportion of the total load of any three adjacent joints to the total load of all joints. The calculation method is to add the actual loads of the three adjacent joints and then divide by the total load of all joints. For example, the local load concentration of the second, third, and fourth joints is (28.5+33.2+16.8) / (16.2+28.5+33.2+16.8+9.5+4.8)=0.72, indicating that these three joints bear 72% of the total load. The preset concentration threshold is 0.65. When the local load concentration exceeds this threshold, it is determined that the load distribution is uneven and further adjustments are needed.

[0115] If the local load concentration exceeds the preset concentration threshold, the correction amount is repeatedly calculated and the motion trajectory is updated. For joint groups with concentrated load, the correction coefficient is increased to accelerate load transfer; for joint groups with lower load, the correction coefficient is decreased to avoid excessive load transfer. In the example above, the local load concentration of the second, third, and fourth joints is 0.72, exceeding the threshold of 0.65, requiring further adjustment. The load proportion of the first, fifth, and sixth joints is increased, while the load proportion of the second, third, and fourth joints is decreased. Specifically, the angular velocity and angular acceleration of the second, third, and fourth joints are further reduced by 5%, while the angular velocity and angular acceleration of the first, fifth, and sixth joints are increased by 5%. After adjustment, the actual load data is re-collected, and the local load concentration is recalculated. This process is iterated until the local load concentration is below the preset threshold. Usually, 2-3 iterations are sufficient to achieve the target load distribution.

[0116] The load prediction and adjustment process is performed in real time, with a system response time of less than 100ms, enabling it to adapt to dynamically changing working environments. Dedicated parameter adjustment templates are established for different robot models and task types to improve adjustment efficiency. The system also possesses self-learning capabilities, continuously optimizing the mapping model and adjustment strategies by recording historical adjustment data, thereby improving its long-term operational adaptability.

[0117] By establishing a mapping relationship between the target load and joint motion parameters, and dynamically adjusting the motion trajectory based on actual load feedback, precise control and balanced distribution of robot joint load are achieved. The mapping model constructed using machine learning technology accurately captures the nonlinear relationship between load and motion parameters, providing a reliable basis for initial parameter generation. The parameter correction mechanism based on a proportional-integral-derivative strategy effectively addresses model prediction errors and external disturbances, ensuring that the actual load closely approximates the target value. Local load concentration assessment and iterative adjustment strategies effectively avoid overloading of single or a few joints, extending the overall lifespan of the robot.

[0118] This invention also provides a technical solution for an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the aforementioned methods.

[0119] This invention also provides a technical solution: a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the steps of any of the aforementioned methods.

[0120] The specific embodiments described above are preferred embodiments of the present invention and are not intended to limit the specific scope of the present invention. The scope of the present invention includes, but is not limited to, these specific embodiments. All equivalent changes made in accordance with the shape and structure of the present invention are within the protection scope of the present invention.

Claims

1. A method for predicting the load on robot joint structures based on artificial intelligence, characterized in that, Includes the following steps: Collect load data of each joint during robot movement, and identify the timing and attenuation of load transmission between adjacent joints; Based on the transmission timing and attenuation law, the distribution ratio of the load in different joints is calculated, and a load transmission characteristic matrix is ​​generated. Based on the load transfer feature matrix, the load transfer trend is predicted, the prediction results are compensated and corrected in real time, and the optimized load distribution scheme is output. The load distribution scheme is converted into joint motion parameters, and the motion trajectory of each joint is adjusted to reduce the local load concentration. Collecting load data for each joint during robot movement and identifying the timing and attenuation patterns of load transfer between adjacent joints includes: The load torque and angle signals of each joint of the robot are collected, and multi-band torque components are obtained through wavelet decomposition and cross-correlation function is constructed to obtain the initial phase relationship. The load transfer direction and attenuation characteristics between adjacent joints are determined based on the angle signal. The initial phase is corrected based on the attenuation characteristics, and the load transfer timing and attenuation pattern between adjacent joints are identified. The multi-band torque components are obtained through wavelet decomposition, and the initial phase relationship is obtained by constructing a cross-correlation function, including: The load torque signal is decomposed into multi-band wavelet components to obtain multi-band torque components. The energy value of each frequency band torque component is calculated and converted into the weighting coefficient of the corresponding frequency band. Construct cross-correlation functions for torque components in the same frequency band of adjacent joints, and determine the phase delay characteristics of the frequency band based on the time delay corresponding to the maximum value of the cross-correlation function; The initial phase relationship is obtained by weighting the weighting coefficients with the phase delay characteristics of each frequency band. Based on the transmission timing and attenuation pattern, the load distribution ratio at different joints is calculated, generating a load transmission characteristic matrix including: A transfer function containing coupling and attenuation components is constructed based on the transfer timing and attenuation rules. The coupling and attenuation components in the transfer function are integrated separately, and the dynamic coupling coefficient is obtained by combining the joint motion posture. The transfer function is modified using dynamic coupling coefficients to construct a joint load transfer relationship matrix; Solve the coupling matrix to obtain the load distribution ratio, and combine the transfer function and the load distribution ratio to generate the load transfer characteristic matrix; Predicting load transfer trends based on the load transfer feature matrix includes: A sequence of feature vectors is obtained by temporal decomposition based on the load transfer feature matrix; The load mutation point is identified based on the changing trend of the feature vector sequence. The load mutation point is used as the boundary to divide the prediction interval, and the load transmission trend is predicted within each prediction interval. The prediction results are compensated and corrected in real time, and the optimized load distribution scheme is output, including: Calculate the residual sequence between the predicted and measured values ​​of the load transfer trend, and decompose the residual sequence into periodic and random components; The compensation value is calculated based on the periodic component and the random component. The compensation value is then fused with the predicted value of the load transmission trend to output the optimized load distribution scheme. Transforming the load distribution scheme into joint motion parameters and adjusting the motion trajectory of each joint to reduce local load concentration includes: Establish the mapping relationship between the target load value of each joint and the joint motion parameters in the load distribution scheme, and generate the initial parameters of joint motion; Calculate the deviation between the actual load value and the target load value to determine the motion parameter correction amount; The motion trajectory of each joint is adjusted based on the correction of motion parameters until the local load concentration is lower than the preset concentration threshold.

2. The method according to claim 1, characterized in that, The transfer function is modified using dynamic coupling coefficients, and the joint load transfer relationship matrix is ​​constructed as follows: The initial coupling matrix is ​​divided into blocks and diagonalized based on the dynamic coupling coefficients to obtain the weighted coupling matrix. Singular value decomposition is performed on the weighted coupling matrix to reconstruct the eigenvectors and generate coupling matrix elements; Fill the elements of the coupling matrix into the weighted coupling matrix to construct the joint load transfer relationship matrix.

3. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the method as described in any one of claims 1 to 2.

4. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 2.

Images