Master-slave manipulator torque feedback stability control method and system based on deep learning

Abstract

The invention discloses a deep learning-based master-slave manipulator torque feedback stability control method and system. The method comprises the following steps of 1, generating a master-slave torque feedback original time sequence data set; step 2, re-sampling a unified timestamp, and executing unified coordinate system conversion, dimension normalization and sliding window mode segmentation to obtain a master-slave alignment time sequence state tensor; 3, discrete frequency domain transformation is executed, and feature dimension splicing is carried out to form a master-slave moment frequency-time fusion feature tensor; 4, performing interaction torque evolution through an improved FEDform model to obtain a predicted interaction torque evolution sequence; and 5, according to the predicted interaction torque evolution sequence, calculating a torque change slope sequence, a torque fluctuation peak sequence and a frequency band energy distribution sequence, and forming a torque stability control parameter set. According to the invention, through the improved FEDformer model and moment evolution prediction, stable feedback control of the master and slave manipulators is realized.

Description

Technical Field

[0001] This invention relates to the field of robotic arm control technology, and in particular to a master-slave robotic arm torque feedback stabilization control method and system based on deep learning. Background Technology

[0002] Master-slave manipulators are widely used in remote operation, hazardous environment work, precision assembly, medical assistance, and special maintenance scenarios. In these applications, operators issue posture commands and operational inputs through the master manipulator, and the slave manipulator completes the corresponding actions according to the master's commands. When in contact with the external environment, the slave manipulator feeds back the contact force and load torque information to the master, allowing the operator to perceive the slave's operational status. Existing master-slave manipulator control technologies mostly focus on position following control or basic force feedback control, typically achieving interactive control between the master and slave through master-slave position mapping, impedance control, admittance control, or fixed parameter compensation. These methods can meet basic follow-up requirements during free-space motion, but when contact, collisions, sudden load changes, and complex environmental interactions are involved, the dynamic changes in the torque feedback link become more drastic, placing higher demands on the system's timing consistency, frequency domain disturbance recognition capabilities, and the ability to predict future torque change trends.

[0003] In existing technologies, on the one hand, the master-end operation input, slave-end joint movement, end-effector execution state, environmental contact action, and feedback torque output typically originate from different sensor channels. These data differ in sampling period, timestamp accuracy, coordinate representation, and dimensional form. If only simple synchronization or direct splicing methods are used, it can easily lead to timing misalignment, inconsistent spatial mapping, and distorted feature representation, thus affecting subsequent control accuracy. On the other hand, traditional force feedback control methods often directly generate feedback control quantities based on sampled values ​​at the current moment or within a short period, lacking the ability to model the evolution trend of interactive torque over multiple future control cycles. This makes it difficult to effectively distinguish between different types of torque change components, such as low-frequency drift, periodic oscillation, and high-frequency impact. Under conditions such as contact establishment, continuous pressing, or oscillation amplification, problems such as sudden changes in feedback torque, increased fluctuations, unsmooth output, and insufficient stability can easily occur. At the same time, existing methods do not adequately establish the correlation between prediction results and control parameters. They usually cannot further extract control data such as torque change slope, fluctuation peak and frequency band energy distribution from the predicted torque sequence, resulting in a lack of targeted adjustment of parameters such as damping compensation, stiffness correction, output limiting and chattering suppression.

[0004] Therefore, how to provide a torque feedback stable control method and system for master-slave manipulators based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based master-slave manipulator torque feedback stabilization control method and system. This invention constructs a master-slave torque feedback original time series dataset, combines time series alignment, frequency-time fusion representation and an improved FEDformer model to predict the evolution trend of interactive torque, and generates a torque stabilization control parameter set, thereby realizing stable torque feedback control of the master-slave manipulator in complex contact environments.

[0006] The master-slave manipulator torque feedback stabilization control method based on deep learning according to embodiments of the present invention includes the following steps: Step 1: Obtain the sequence of joint motion states of the robotic arm, the sequence of execution states of the robotic arm end effector, and the sequence of interaction with the robotic arm environment from the robotic arm side; obtain the sequence of operation inputs of the master end effector and the sequence of execution states of the master robotic arm from the master end effector side; and sort them in time sequence to generate the original time series dataset of master-slave torque feedback. Step 2: Add a unified timestamp to each data point in the original time series dataset of master-slave torque feedback, resample, and perform unified coordinate system transformation, dimension normalization, and sliding window segmentation to obtain the master-slave aligned time series state tensor; Step 3: Perform discrete frequency domain transformation on the master-slave aligned temporal state tensor and concatenate the feature dimensions to form the master-slave torque frequency-time fusion feature tensor; Step 4: Input the master-slave torque frequency-time fusion feature tensor into the improved FEDformer model, and perform interactive torque evolution through the sequence decomposition module, multi-scale trend enhancement module, sparse frequency domain attention module and local transient fusion decoding module to obtain the predicted interactive torque evolution sequence for multiple future control cycles; Step 5: Based on the predicted interactive torque evolution sequence, calculate the torque change slope sequence, torque fluctuation peak sequence, and frequency band energy distribution sequence to form a torque stability control parameter set.

[0007] Optionally, step one specifically includes: The angular velocity signal of each joint motor is calculated from the output shaft end of each joint motor of the robot based on the change in angular displacement at adjacent sampling times. The current signal of each joint motor is collected by the current detection unit in the drive circuit. The angular acceleration signal of each joint motor is calculated based on the time sequence change of the angular velocity signal, thus forming a sequence of joint motion states of the robot. The attitude and displacement signals of the end effector relative to the preset reference coordinate system are collected from the end effector of the robot. Based on the current position of the end effector, the amount of attitude change, and the contact trigger signal between the end effector and the work object, the current free motion state, initial contact state, continuous contact state, or discontinuation contact state of the end effector are determined, forming a sequence of end effector states of the robot. By collecting contact force signals during the operation of the end effector through a torque sensor installed at the end of the robot, and calculating the load torque signal based on the lever arm relationship between the contact force signal and the mounting base of the end effector, a sequence of interaction between the robot and the environment is formed. The pose command signal of the master hand operating handle relative to the reference coordinate system of the master manipulator is acquired at the master hand operating handle, and the operation speed signal of the master hand operating handle is calculated based on the position change and attitude change of the pose command signal at adjacent sampling times to form the master end operation input sequence. The main manipulator's joint position signal and joint speed signal are collected at each joint of the main manipulator. The main manipulator's drive current signal is collected in the main manipulator's servo drive circuit. The feedback torque signal output by the controller is collected at the main manipulator's feedback execution end to form the main manipulator's execution state sequence. A unified sampling time stamp is added to the sequence of joint motion state of the robot arm, the sequence of execution state of the robot arm end effector, the sequence of interaction between the robot arm and the environment, the sequence of operation input of the master end effector, and the sequence of execution state of the master robot arm. The collected signals are then sorted in time according to a preset sampling period to generate a master-slave torque feedback raw time series dataset covering the master end operation input, slave end motion response, environmental contact action, and slave end feedback output.

[0008] Optionally, step two specifically includes: A unified timestamp is added to each data in the original time series dataset of master-slave torque feedback, and resampling is performed according to a preset control cycle. Outlier removal, missing value interpolation, and noise filtering are then performed to generate a standardized time series signal. Based on the preset motion mapping relationship between the master manipulator and the slave manipulator, the pose command signal of the master manipulator's operating handle is transformed from the master reference coordinate system to the unified mapping coordinate system, and the pose signal of the slave manipulator's end effector is mapped to the unified mapping coordinate system; the joint position signals of the master manipulator and the slave manipulator are processed to perform joint space unified representation processing to generate pose association data after coordinate unification. The contact force signal, load torque signal, joint velocity signal, current signal, and feedback torque signal are subjected to dimensional normalization processing to generate normalized feature data with unified dimensions. The standardized time-series signal, pose-related data, and normalized feature data are divided into short-time window sequences, medium-time window sequences, and long-time window sequences according to preset sliding window sizes, resulting in a master-slave aligned time-series state tensor.

[0009] Optionally, step three specifically includes: Perform a discrete frequency domain transformation on the master-slave aligned time-series state tensor to obtain the frequency domain representation matrix corresponding to each window; In the frequency domain representation matrix corresponding to each window, the amplitude of the low-frequency component, the amplitude of the dominant frequency component, the periodic repetition intensity value, and the energy value of the high-frequency impact component under the corresponding window are extracted and arranged according to window type, time order, and signal source to form the master-slave torque frequency-time fusion feature tensor.

[0010] Optionally, the improved FEDformer model is specifically as follows: The master-slave torque frequency-time fusion feature tensor is input into the sequence decomposition module, and the moving average calculation is performed on each feature channel in the master-slave torque frequency-time fusion feature tensor according to the time axis. In each feature channel, a sliding window of a preset length slides along the time axis moment by moment, calculates the average value of each sampled value in the current window, and arranges the moving average results corresponding to each moment in chronological order to generate the trend component subsequence corresponding to each feature channel. The trend component subsequences corresponding to each feature channel are combined according to the original feature channel order to generate the trend component input sequence; The original time series data of each feature channel in the master-slave torque frequency-time fusion feature tensor are subtracted from the corresponding trend component subsequence to generate the periodic component subsequence corresponding to each feature channel. The periodic component subsequences corresponding to each feature channel are combined according to the original feature channel order to generate the periodic component input sequence; The trend component input sequence is input into the multi-scale trend enhancement module, multiple moving average kernels of different lengths are set, and smoothing calculations are performed on the trend component input sequence using each moving average kernel to obtain trend candidate subsequences at multiple time scales. The trend component input sequence is subtracted from each trend candidate subsequence to obtain the trend residual subsequence at each time scale. A weighted average is performed on each trend candidate subsequence and trend residual subsequence to obtain the trend-enhanced output sequence. The periodic component input sequence is input into the sparse frequency domain attention module, and the periodic frequency domain representation sequence is obtained through fast Fourier transform; Based on the spectral energy values ​​corresponding to each frequency component in the periodic frequency domain representation sequence, sort them in descending order, and select the top N frequency components to form the dominant frequency index set. Based on the dominant frequency index set, the corresponding dominant frequency components are extracted from the periodic frequency domain representation sequence to generate a sparse frequency domain feature sequence. The sparse frequency domain feature sequence is input into the frequency domain self-attention computing unit. The proportion of each dominant frequency component to the sum of all dominant frequency components is used as the weight of each dominant frequency component. After weighting, the periodic enhanced output sequence is obtained. The trend enhancement output sequence and the period enhancement output sequence are input into the local transient fusion decoding module and concatenated according to the feature channel dimension to obtain the fused feature sequence; The fused feature sequence is input into the feedforward decoding network, and the predicted interaction torque evolution sequence for multiple future control cycles is output according to the preset prediction step size.

[0011] Optionally, the step of inputting the fused feature sequence into the feedforward decoding network and outputting the predicted interaction torque evolution sequence for multiple future control cycles according to a preset prediction step size specifically involves: Read the values ​​of each feature channel of the fused feature sequence at the current input time, and extract the feature vector of the last time step in the fused feature sequence in chronological order; The feature vector at the end time step is input into the mapping layer, and the channel values ​​in the feature vector at the end time step are multiplied by the corresponding preset weights to generate the initial decoded feature vector. The initial decoded feature vector is written into the decoding buffer unit as the input state at the first prediction time. Nonlinear activation is performed on each element in the initial decoded feature vector to obtain the activated feature vector. The activation feature vector is input into the pooling layer, and max pooling is performed on the activation feature vector to obtain the candidate interaction torque value corresponding to the current prediction time. The candidate interaction torque value corresponding to the current prediction time is concatenated with the activation feature vector corresponding to the current prediction time to generate the feedback input vector for the next prediction time. Determine whether the current prediction time has reached the last prediction time corresponding to the preset prediction step size; if not, use the current feedback input vector as the input state for the next prediction time. Repeat the iteration until the current prediction time reaches the last prediction time corresponding to the preset prediction step size. Arrange the candidate interaction torque values ​​corresponding to each prediction time in the prediction result sequence in chronological order to generate a candidate interaction torque time series sequence. The difference between the measured load torque value and the first candidate interactive torque value in the candidate interactive torque time series is calculated to obtain the correction offset at the first moment. The initial time-time correction offset is added to the candidate interaction torque values ​​at each time-time in the candidate interaction torque time series to complete the initial boundary alignment process; The torque difference between two adjacent prediction times after the initial boundary alignment process is completed is calculated sequentially, and the interaction torque value of the subsequent time corresponding to the torque difference exceeding the preset change threshold is adjusted to the difference between the interaction torque value of the previous time and the preset change threshold, thus completing the continuity correction process between adjacent times. The time series sequence of interactive torques, after initial boundary alignment and continuity correction, is output as the predicted interactive torque evolution sequence for multiple future control cycles.

[0012] Optionally, step five specifically includes: Read the interaction torque values ​​corresponding to each prediction time in the predicted interaction torque evolution sequence, and arrange them in chronological order of the prediction time to form the predicted torque time series sequence within the future control interval; The interaction torque values ​​corresponding to two adjacent prediction times within the future control interval are grouped together. The difference between the interaction torque value at the next prediction time and the interaction torque value at the previous prediction time is calculated respectively. The difference is then divided by the corresponding prediction time interval to obtain the torque change slope value between each adjacent prediction time. Arrange the torque change slope values ​​between each adjacent prediction time in the order of prediction time to generate a torque change slope sequence within the future control interval. The interaction torque values ​​at each prediction time point in the predicted torque time series are compared point by point. It is determined whether the interaction torque value at the current prediction time is greater than both the interaction torque value at the previous prediction time and the interaction torque value at the next prediction time. If so, the interaction torque value corresponding to the current prediction time is determined as a candidate value for the local fluctuation peak. The candidate values ​​of each local fluctuation peak are arranged in chronological order, and the candidate values ​​of local fluctuation peaks that are lower than the preset peak screening threshold are removed. The remaining candidate values ​​of local fluctuation peaks are retained as torque fluctuation peaks and arranged in the order of the predicted time to generate a torque fluctuation peak sequence in the future control interval. The predicted torque time series is subjected to frequency domain transformation to map the predicted torque time series within the future control interval into the corresponding frequency domain spectral line sequence, and the result is obtained by arranging them in order of frequency coordinates. Based on the preset frequency band division rules, the frequency domain representation results are divided into low-frequency band, mid-frequency band and high-frequency band, and the corresponding frequency components in each band are extracted respectively. The amplitude squares of each frequency component within each frequency band are accumulated to obtain the energy values ​​of the low-frequency band, the mid-frequency band, and the high-frequency band. These values ​​are then arranged in the order of the frequency bands to generate a frequency band energy distribution sequence. Read the slope values ​​of each torque change in the torque change slope sequence, take the absolute value and match it with the preset slope mapping interval to determine the feedback damping compensation amount; Read the peak values ​​of each torque fluctuation in the torque fluctuation peak sequence, and match each peak value with a preset peak mapping interval to determine the output limit value; Read the low-frequency, mid-frequency, and high-frequency band energy values ​​from the frequency band energy distribution sequence, and obtain the feedback stiffness correction, prediction feedforward gain, and chattering suppression values ​​based on the proportion of each band energy value to the total band energy. The torque stability control parameter set includes feedback damping compensation, output limiting value, feedback stiffness correction, prediction feedforward gain value, and chattering suppression value.

[0013] The master-slave manipulator torque feedback stabilization control system based on deep learning according to embodiments of the present invention includes the following modules: The data acquisition module is used to acquire the sequence of joint motion states of the robotic arm from the robotic arm side, the sequence of execution states of the robotic arm end effector, and the sequence of interaction with the robotic arm environment from the robotic arm side, acquire the sequence of operation inputs of the master end effector and the sequence of execution states of the master robotic arm from the master end effector side, and sort them in time sequence to generate the original time sequence dataset of master-slave torque feedback. The timing alignment module is used to add a unified timestamp to each data in the original timing dataset of master-slave torque feedback, resample it, and perform unified coordinate system transformation, dimension normalization and sliding window segmentation to obtain the master-slave aligned timing state tensor. The fusion representation module is used to perform discrete frequency domain transformation on the master-slave aligned temporal state tensor and concatenate the feature dimensions to form the master-slave torque frequency-time fusion feature tensor. The torque prediction module is used to input the master-slave torque frequency-time fusion feature tensor into the improved FEDformer model. Through the sequence decomposition module, multi-scale trend enhancement module, sparse frequency domain attention module and local transient fusion decoding module, the interactive torque evolution is performed to obtain the predicted interactive torque evolution sequence in multiple future control cycles. The parameter generation module is used to calculate the torque change slope sequence, torque fluctuation peak sequence, and frequency band energy distribution sequence based on the predicted interactive torque evolution sequence, and form a torque stability control parameter set.

[0014] The beneficial effects of this invention are: This invention provides a deep learning-based master-slave manipulator torque feedback stabilization control method and system. Addressing issues such as temporal asynchrony, inconsistent spatial mapping, large torque feedback fluctuations, and insufficient basis for control parameter adjustment during contact operations, it establishes a complete technical chain from data acquisition, temporal alignment, frequency-time fusion representation, interactive torque prediction to stable control parameter generation. By uniformly acquiring and temporally sorting the slave manipulator joint motion state sequence, slave manipulator end-effector execution state sequence, slave manipulator environmental interaction sequence, master-end operation input sequence, and master manipulator execution state sequence, it can incorporate master-end operation behavior, slave-end motion response, environmental contact effects, and feedback execution information into a single data system, providing a consistent original foundation for subsequent modeling. By performing unified timestamp resampling, coordinate transformation, dimensional normalization, and sliding window segmentation on the original master-slave torque feedback temporal dataset, it can convert heterogeneous signals from multiple sources into a unified expression form, reducing feature offsets caused by differences in sampling frequency, coordinate system, and physical quantity scale. By performing a discrete frequency domain transformation on the master-slave aligned temporal state tensor and extracting the amplitude of the low-frequency component, the amplitude of the dominant frequency component, the periodic repetition intensity value, and the energy value of the high-frequency impact component, a master-slave torque frequency-time fusion feature tensor is formed. This tensor can simultaneously characterize the continuous variation characteristics of the interaction torque in the time domain and the oscillating distribution characteristics in the frequency domain, making the torque variation information more complete.

[0015] By employing an improved FEDformer model to perform sequence decomposition, multi-scale trend enhancement, sparse frequency domain attention processing, and fusion decoding on the master-slave torque frequency-time fusion feature tensor, continuous prediction of the interaction torque evolution sequence over multiple future control cycles can be achieved. This transforms interaction torque analysis from a single-moment response to an evolutionary analysis oriented towards future control intervals. Furthermore, by calculating the torque change slope sequence, torque fluctuation peak sequence, and frequency band energy distribution sequence from the predicted interaction torque evolution sequence, and forming a torque stability control parameter set consisting of feedback damping compensation, output limiting value, feedback stiffness correction, predicted feedforward gain, and chattering suppression value, a direct basis can be provided for subsequent control execution. This enhances the targeting, consistency, and adjustability of the master-slave manipulator's torque feedback control, thereby improving the manipulator's torque feedback stability control capability in complex contact environments. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of the master-slave manipulator torque feedback stabilization control method based on deep learning proposed in this invention; Figure 2This is a schematic diagram of the master-slave manipulator torque feedback stabilization control system based on deep learning proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1 A deep learning-based master-slave manipulator torque feedback stabilization control method includes: Step 1: Obtain the sequence of joint motion states of the robotic arm, the sequence of execution states of the robotic arm end effector, and the sequence of interaction with the robotic arm environment from the robotic arm side; obtain the sequence of operation inputs of the master end effector and the sequence of execution states of the master robotic arm from the master end effector side; and sort them in time sequence to generate the original time series dataset of master-slave torque feedback. Step 2: Add a unified timestamp to each data point in the original time series dataset of master-slave torque feedback, resample, and perform unified coordinate system transformation, dimension normalization, and sliding window segmentation to obtain the master-slave aligned time series state tensor; Step 3: Perform discrete frequency domain transformation on the master-slave aligned temporal state tensor and concatenate the feature dimensions to form the master-slave torque frequency-time fusion feature tensor; Step 4: Input the master-slave torque frequency-time fusion feature tensor into the improved FEDformer model, and perform interactive torque evolution through the sequence decomposition module, multi-scale trend enhancement module, sparse frequency domain attention module and local transient fusion decoding module to obtain the predicted interactive torque evolution sequence for multiple future control cycles; Step 5: Based on the predicted interactive torque evolution sequence, calculate the torque change slope sequence, torque fluctuation peak sequence, and frequency band energy distribution sequence to form a torque stability control parameter set.

[0019] This step enables unified organization and alignment of master-end operation, slave-end motion, environmental contact, and feedback execution information, improving the consistency and availability of heterogeneous master-slave data. By constructing frequency-time fusion features, it enhances the comprehensive characterization of torque change trends, periodic oscillation characteristics, and impact disturbance characteristics. Combined with the improved FEDformer model, it enables forward-looking prediction of future interactive torque evolution and generates a targeted set of stable control parameters, thereby improving the stability, continuity, and anti-disturbance capability of the master-slave manipulator torque feedback and reducing the risk of feedback mutations and oscillations during the contact process.

[0020] In this embodiment, step one specifically includes: The angular velocity signal of each joint motor is calculated from the output shaft end of each joint motor of the robot based on the change in angular displacement at adjacent sampling times. The angular acceleration signal of each joint motor is calculated based on the current signal of each joint motor from the current detection unit in the drive circuit and the time sequence change result of the angular velocity signal, thus forming a sequence of joint motion states of the robot. The attitude and displacement signals of the end effector relative to the preset reference coordinate system are obtained from the end effector of the robot arm. Based on the current position of the end effector, the amount of attitude change, and the contact trigger signal between the end effector and the work object, the current free motion state, initial contact state, continuous contact state, or discontinuation contact state of the end effector is determined, thus forming a sequence of end effector states from the robot arm. By setting the contact force signal of the end effector during the operation of the end effector from the torque sensor at the end of the robot, and calculating the load torque signal according to the lever arm relationship of the contact force signal relative to the mounting base of the end effector, an interaction sequence from the robot to the environment is formed. The pose command signal of the master hand operating handle relative to the reference coordinate system of the master robot is obtained at the master hand operating handle. The operation speed signal of the master hand operating handle is calculated based on the position change and attitude change of the pose command signal at adjacent sampling times, forming the master end operation input sequence. The main manipulator's joint position signal and speed signal at each joint, the main manipulator's drive current signal in the main manipulator's servo drive circuit, and the feedback torque signal output by the controller at the main manipulator's feedback execution end form the main manipulator's execution state sequence. A unified sampling time stamp is added to the sequence of joint motion state of the robot arm, the sequence of execution state of the robot arm end effector, the sequence of interaction between the robot arm and the environment, the sequence of operation input of the master end effector, and the sequence of execution state of the master robot arm. The signals are then sorted in time according to a preset sampling period to generate a master-slave torque feedback raw time series dataset covering the master end operation input, slave end motion response, environmental contact action, and slave end feedback output.

[0021] In this embodiment, step two specifically includes: The master-slave torque feedback original timing dataset includes the acquisition sequence of robot joint motion state, acquisition sequence of robot end execution state, acquisition sequence of robot environmental interaction, acquisition sequence of master operation input and acquisition sequence of master robot execution state. Each data path is appended with a unified timestamp, and resampling is performed on each data path based on a preset control cycle to generate multi-channel timing signals with a unified sampling frequency. The resampled multi-channel time-series signals are subjected to outlier detection according to a preset threshold range. Data points that exceed the corresponding threshold range are identified as outlier sampling points and are removed. For the data gaps formed after the removal of outlier sampling points and the missing sampling points generated during the original acquisition process, interpolation compensation is performed based on the same type of signal data at adjacent time points to generate a time-series signal with missing data compensation. Denoising and filtering processes are performed on the timing signals after missing compensation. Specifically, smoothing filtering is performed on joint position data, end effector displacement data, and pose command data; low-pass filtering is performed on contact force data, load torque data, and feedback output torque data; and timing de-glitch processing is performed on joint velocity data, current data, and angular acceleration data to generate a denoised and standardized timing signal. Based on the preset motion mapping relationship between the master manipulator and the slave manipulator, the master manipulator handle pose command data is transformed from the master reference coordinate system to the unified mapping coordinate system. The master manipulator joint position data and the slave manipulator joint position data are processed to achieve a unified joint space representation. The slave manipulator end effector pose data is transformed from the end effector Cartesian space to the unified mapping coordinate system to generate pose association data after coordinate unification. The maximum, minimum and mean values ​​of the corresponding signals in the preset time interval are extracted from the contact force data, load torque data, joint speed data, current data and feedback output torque data respectively. The data are then mapped to a unified numerical interval according to the preset normalization rules to generate normalized feature data with unified dimensions. After completing unified timestamp processing, resampling processing, outlier removal processing, missing value interpolation processing, noise reduction filtering processing, coordinate system transformation processing, and dimension normalization processing, the time-domain data of each channel are divided in a sliding window manner. Specifically, a short-time window sequence is obtained by dividing the data with a first preset window length, a medium-time window sequence is obtained by dividing the data with a second preset window length, and a long-time window sequence is obtained by dividing the data with a third preset window length. The first preset window length is less than the second preset window length, and the second preset window length is less than the third preset window length. The short-time window sequence, medium-time window sequence, and long-time window sequence are stacked in chronological order to generate a master-slave aligned temporal state tensor containing master-end operation input features, slave-end motion response features, environmental contact interaction features, and master-end feedback execution features.

[0022] In this embodiment, step three specifically includes: A discrete frequency domain transformation is performed on the master-slave aligned time-series state tensor. The discrete frequency domain transformation is to map the master-slave aligned time-series state tensor in each window to the corresponding frequency domain spectral line sequence, and arrange them in order of frequency coordinates to obtain the frequency domain representation matrix corresponding to each window. In the frequency domain representation matrix corresponding to each window, the set of low-frequency lines is extracted according to the preset low-frequency band range, and the cumulative result of the amplitude corresponding to each frequency point in the set of low-frequency lines is calculated to obtain the amplitude of the low-frequency component under the corresponding window. In the frequency domain representation matrix corresponding to each window, the frequency components are sorted from largest to smallest according to their spectral amplitude. A preset number of frequency components are selected as the dominant frequency components, and the amplitude corresponding to the dominant frequency components is extracted to obtain the amplitude of the dominant frequency components under the corresponding window. In the frequency domain representation matrix corresponding to each window, the occurrence frequency, amplitude change and frequency position offset of the dominant frequency component in consecutive adjacent windows are statistically analyzed, and the repetition ratio of each dominant frequency component in the preset time interval is calculated based on the statistical results to obtain the periodic repetition intensity value under the corresponding window. In the frequency domain representation matrix corresponding to each window, a set of high-frequency lines is extracted according to the preset high-frequency band range, and the squared amplitude results corresponding to each frequency point in the set of high-frequency lines are accumulated to obtain the energy value of the high-frequency impulse component under the corresponding window. The amplitudes of low-frequency components, dominant frequency components, periodic repetition intensity values, and high-frequency impact components corresponding to short-time window sequences, the amplitudes of low-frequency components, dominant frequency components, periodic repetition intensity values, and high-frequency impact components corresponding to medium-time window sequences, and the amplitudes of low-frequency components, dominant frequency components, periodic repetition intensity values, and high-frequency impact components corresponding to long-time window sequences are arranged according to window type, time order, and signal source, respectively, to generate the master-slave torque frequency-time fusion feature tensor.

[0023] In this embodiment, the improved FEDformer model is specifically as follows: The master-slave torque frequency-time fusion feature tensor is input into the sequence decomposition module, and the moving average calculation is performed on each feature channel in the master-slave torque frequency-time fusion feature tensor according to the time axis. In each feature channel, a sliding window of a preset length slides along the time axis moment by moment, calculates the average value of each sampled value in the current window, and arranges the moving average results corresponding to each moment in chronological order to generate the trend component subsequence corresponding to each feature channel. The trend component subsequences corresponding to each feature channel are combined according to the original feature channel order to generate the trend component input sequence; The original time series data of each feature channel in the master-slave torque frequency-time fusion feature tensor are subtracted from the corresponding trend component subsequence to generate the periodic component subsequence corresponding to each feature channel. The periodic component subsequences corresponding to each feature channel are combined according to the original feature channel order to generate the periodic component input sequence; The trend component input sequence is input into the multi-scale trend enhancement module, multiple moving average kernels of different lengths are set, and smoothing calculations are performed on the trend component input sequence using each moving average kernel to obtain trend candidate subsequences at multiple time scales. The trend component input sequence is subtracted from each trend candidate subsequence to obtain the trend residual subsequence at each time scale. A weighted average is performed on each trend candidate subsequence and trend residual subsequence to obtain the trend-enhanced output sequence. The periodic component input sequence is input into the sparse frequency domain attention module, and the periodic frequency domain representation sequence is obtained through fast Fourier transform; Based on the spectral energy values ​​corresponding to each frequency component in the periodic frequency domain representation sequence, sort them in descending order, and select the top N frequency components to form the dominant frequency index set. Based on the dominant frequency index set, the corresponding dominant frequency components are extracted from the periodic frequency domain representation sequence to generate a sparse frequency domain feature sequence. The sparse frequency domain feature sequence is input into the frequency domain self-attention computing unit. The proportion of each dominant frequency component to the sum of all dominant frequency components is used as the weight of each dominant frequency component. After weighting, the periodic enhanced output sequence is obtained. The trend enhancement output sequence and the period enhancement output sequence are input into the local transient fusion decoding module and concatenated according to the feature channel dimension to obtain the fused feature sequence; The fused feature sequence is input into the feedforward decoding network, which outputs the predicted interaction torque evolution sequence for multiple future control cycles according to a preset prediction step size. Specifically, this includes: Read the values ​​of each feature channel of the fused feature sequence at the current input time, and extract the feature vector of the last time step in the fused feature sequence in chronological order; The feature vector at the end time step is input into the mapping layer, and the channel values ​​in the feature vector at the end time step are multiplied by the corresponding preset weights to generate the initial decoded feature vector. The initial decoded feature vector is written into the decoding buffer unit as the input state at the first prediction time. Nonlinear activation is performed on each element in the initial decoded feature vector to obtain the activated feature vector. The activation feature vector is input into the pooling layer, and max pooling is performed on the activation feature vector to obtain the candidate interaction torque value corresponding to the current prediction time. The candidate interaction torque value corresponding to the current prediction time is concatenated with the activation feature vector corresponding to the current prediction time to generate the feedback input vector for the next prediction time. Determine whether the current prediction time has reached the last prediction time corresponding to the preset prediction step size; if not, use the current feedback input vector as the input state for the next prediction time. Repeat the iteration until the current prediction time reaches the last prediction time corresponding to the preset prediction step size. Arrange the candidate interaction torque values ​​corresponding to each prediction time in the prediction result sequence in chronological order to generate a candidate interaction torque time series sequence. The difference between the measured load torque value and the first candidate interactive torque value in the candidate interactive torque time series is calculated to obtain the correction offset at the first moment. The initial time-time correction offset is added to the candidate interaction torque values ​​at each time-time in the candidate interaction torque time series to complete the initial boundary alignment process; The torque difference between two adjacent prediction times after the initial boundary alignment process is completed is calculated sequentially, and the interaction torque value of the subsequent time corresponding to the torque difference exceeding the preset change threshold is adjusted to the difference between the interaction torque value of the previous time and the preset change threshold, thus completing the continuity correction process between adjacent times. The time series sequence of interactive torques, after initial boundary alignment and continuity correction, is output as the predicted interactive torque evolution sequence for multiple future control cycles.

[0024] The improved FEDformer model proposed in this step shares similarities with the traditional FEDformer model in that both are geared towards time series prediction tasks and generally follow the basic processing approach of "sequence decomposition—feature enhancement—decoding output." Both first decompose the input sequence into trend and periodic components, then model the time series information of different properties separately, and finally output the prediction results for future periods through a decoding network. Both retain the core idea of ​​FEDformer—utilizing frequency domain information to process long-term time series data—emphasizing the ability to improve the extraction of periodic variation patterns through frequency domain representation, and reducing the computational burden and fitting difficulty of directly predicting complex time series data through decompositional modeling. From the structure presented in this step, the improved model still retains the framework of processing the trend and periodic input sequences separately, and also retains the basic approach of using Fast Fourier Transform to form a periodic frequency domain representation sequence, and then filtering and weighting the frequency components. Therefore, in terms of the core idea of ​​the model, the long-sequence prediction objective, the trend and periodic decoupling mechanism, and the frequency domain modeling method, it has significant consistency with the traditional FEDformer, essentially still belonging to the time series prediction model with decompositional frequency domain enhancement as its core. The difference lies in the fact that the improved model in this step does not merely remain at the level of conventional decomposition and frequency domain attention modeling of trend and seasonal terms in the traditional FEDformer. Instead, it makes targeted adjustments to the input construction method, trend enhancement method, frequency domain weight calculation method, and decoding method around the master-slave robot interaction torque prediction scenario. In the trend processing section, this scheme explicitly employs multiple moving average kernels of varying lengths to perform multi-scale smoothing on the trend component input sequence. It further calculates the trend candidate subsequence and trend residual subsequence, then performs a weighted average, enhancing the hierarchical characterization of trend changes across different time scales. In the period processing section, instead of uniformly modeling all frequency components, this scheme first sorts them in descending order based on spectral energy values, selecting the top N frequency components to form a dominant frequency index set. Then, it uses the proportion of each dominant frequency component to the sum of all dominant frequency components as a weight for weighted processing, improving the focusing ability on key dominant frequency bands. In the output stage, this scheme inputs the trend-enhanced output sequence and the period-enhanced output sequence into the local transient fusion decoding module. A fused feature sequence is formed by splicing feature channels, and a feedforward decoding network is used to iteratively generate candidate interaction torque values ​​time-by-time. Simultaneously, it adds initial boundary alignment processing and continuity correction processing, making the prediction results more closely match the temporal characteristics of the interaction torque. These features demonstrate a substantial difference from the traditional FEDformer.

[0025] The improvements are beneficial because the model can more effectively adapt to the complex characteristics of interactive torque sequences in master-slave manipulators, which simultaneously exhibit trend changes, periodic fluctuations, and local abrupt changes. Through multi-scale trend enhancement processing, trend candidate subsequences and trend residual subsequences at different time scales are utilized simultaneously, which helps improve the ability to express information about slow drift, continuous changes, and local offsets, making trend prediction no longer limited to a single time scale. By filtering the periodic frequency domain representation sequence according to spectral energy, retaining only the top N dominant frequency components and weighting them proportionally, the model's ability to identify major oscillatory components and key periodic features is improved, reducing the interference of low-contribution frequency components on the prediction results. By iteratively outputting candidate interactive torque values ​​time-by-time through local transient fusion decoding and feedforward decoding networks, and adding initial boundary alignment and continuity correction processing at the output, the predicted sequence maintains better initial consistency and continuity between adjacent time steps compared to the measured load torque, reducing abrupt jumps in the prediction results at boundary positions and between adjacent time steps. Overall, these improvements enable the model to better predict the evolution sequence of interactive torques over multiple future control cycles, thus providing a more reliable data foundation for calculating the subsequent torque change slope, torque fluctuation peak, and frequency band energy distribution.

[0026] In this embodiment, step five specifically includes: Read the interaction torque values ​​corresponding to each prediction time in the predicted interaction torque evolution sequence, and arrange them in chronological order of the prediction time to form the predicted torque time series sequence within the future control interval; The interaction torque values ​​corresponding to two adjacent prediction times within the future control interval are grouped together. The difference between the interaction torque value at the next prediction time and the interaction torque value at the previous prediction time is calculated respectively. The difference is then divided by the corresponding prediction time interval to obtain the torque change slope value between each adjacent prediction time. Arrange the torque change slope values ​​between each adjacent prediction time in the order of prediction time to generate a torque change slope sequence within the future control interval. The interaction torque values ​​at each prediction time point in the predicted torque time series are compared point by point. It is determined whether the interaction torque value at the current prediction time is greater than both the interaction torque value at the previous prediction time and the interaction torque value at the next prediction time. If so, the interaction torque value corresponding to the current prediction time is determined as a candidate value for the local fluctuation peak. The candidate values ​​of each local fluctuation peak are arranged in chronological order, and the candidate values ​​of local fluctuation peaks that are lower than the preset peak screening threshold are removed. The remaining candidate values ​​of local fluctuation peaks are retained as torque fluctuation peaks and arranged in the order of the predicted time to generate a torque fluctuation peak sequence in the future control interval. The predicted torque time series is subjected to frequency domain transformation to map the predicted torque time series within the future control interval into the corresponding frequency domain spectral line sequence, and the result is obtained by arranging them in order of frequency coordinates. Based on the preset frequency band division rules, the frequency domain representation results are divided into low-frequency band, mid-frequency band and high-frequency band, and the corresponding frequency components in each band are extracted respectively. The amplitude squares of each frequency component within each frequency band are accumulated to obtain the energy values ​​of the low-frequency band, the mid-frequency band, and the high-frequency band. These values ​​are then arranged in the order of the frequency bands to generate a frequency band energy distribution sequence. Read the slope values ​​of each torque change in the torque change slope sequence, take the absolute value and match it with the preset slope mapping interval to determine the feedback damping compensation amount; Read the peak values ​​of each torque fluctuation in the torque fluctuation peak sequence, and match each peak value with a preset peak mapping interval to determine the output limit value; Read the low-frequency, mid-frequency, and high-frequency band energy values ​​from the frequency band energy distribution sequence, and obtain the feedback stiffness correction, prediction feedforward gain, and chattering suppression values ​​based on the proportion of each band energy value to the total band energy. The torque stability control parameter set includes feedback damping compensation, output limiting value, feedback stiffness correction, prediction feedforward gain value, and chattering suppression value.

[0027] refer to Figure 2 A deep learning-based master-slave manipulator torque feedback stabilization control system includes the following modules: The data acquisition module is used to acquire the sequence of joint motion states of the robotic arm from the robotic arm side, the sequence of execution states of the robotic arm end effector, and the sequence of interaction with the robotic arm environment from the robotic arm side, acquire the sequence of operation inputs of the master end effector and the sequence of execution states of the master robotic arm from the master end effector side, and sort them in time sequence to generate the original time sequence dataset of master-slave torque feedback. The timing alignment module is used to add a unified timestamp to each data in the original timing dataset of master-slave torque feedback, resample it, and perform unified coordinate system transformation, dimension normalization and sliding window segmentation to obtain the master-slave aligned timing state tensor. The fusion representation module is used to perform discrete frequency domain transformation on the master-slave aligned temporal state tensor and concatenate the feature dimensions to form the master-slave torque frequency-time fusion feature tensor. The torque prediction module is used to input the master-slave torque frequency-time fusion feature tensor into the improved FEDformer model. Through the sequence decomposition module, multi-scale trend enhancement module, sparse frequency domain attention module and local transient fusion decoding module, the interactive torque evolution is performed to obtain the predicted interactive torque evolution sequence in multiple future control cycles. The parameter generation module is used to calculate the torque change slope sequence, torque fluctuation peak sequence, and frequency band energy distribution sequence based on the predicted interactive torque evolution sequence, and form a torque stability control parameter set.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to a confined space maintenance work area in a remote operation and maintenance experimental center for intelligent equipment. This work area features multiple narrow maintenance positions, damped knobs, plug-in interfaces, press-fit locking structures, and rigid mounting brackets. The site is characterized by limited viewing angles, significant differences in the stiffness of the contact objects, tortuous operation paths, and severe obstruction by local structures. Operators are located in an isolated control room and remotely control the system via a master robotic arm. The robotic arm enters the work area to perform actions such as plugging, pressing, turning, positioning, and fine-tuning. Because the maintenance objects include both hard metal contact components and elastic coverings and semi-locking structures, the master-end input actions are relatively small while the slave-end contact state changes rapidly during actual operation. Traditional master-slave control methods are prone to problems such as master-end lag, sudden jumps in feedback at the moment of contact, amplified feedback jitter during continuous pressing, and inconsistent torque changes during small displacement operations. This makes it difficult for operators to accurately determine whether the slave end has completed contact, insertion, or locking, and repeated corrections further increase contact impact and operation time. This type of problem is particularly evident in precise remote operation scenarios. Especially when the plug end is close to the mating hole, the clamping end is close to the limit edge, and the screwing end is gradually approaching the tightening threshold, if the multi-source signals on both the master and slave sides cannot be aligned uniformly, or if there is a lack of prediction of the interaction torque changes in the short term, the feedback from the master end is prone to fluctuations. This can cause the operator to misinterpret the system oscillation as a change in environmental resistance, thereby reducing the stability and reliability of the operation.

[0029] In this embodiment, the angular displacement, angular velocity, current, and angular acceleration signals are continuously collected from the motors of each joint of the slave manipulator. Simultaneously, the end effector collects attitude, displacement, and contact state signals. Contact force signals are obtained through a torque sensor installed at the slave manipulator end, and the load torque signal is calculated by combining the lever arm relationship of the end effector mounting base. Simultaneously, the master manipulator collects the master hand handle posture commands, operating speed, master manipulator joint positions, joint velocities, drive current, and feedback torque signals. All data are appended with a unified sampling timescale and then sorted in time sequence to form the original timing dataset for master-slave torque feedback. In actual operation, the control cycle is set to 10ms, each continuous operation lasts for 2 hours, and each test round includes three types of tasks: insertion, pressing and positioning, and tightening. A total of 120 complete operation samples are completed. Subsequently, the original time series data is subjected to unified timestamp resampling, outlier removal, missing value imputation, and noise reduction filtering. Coordinate unification is completed based on the motion mapping relationship between the master and slave manipulators. Then, the contact force, load torque, joint velocity, current, and feedback torque are normalized. The processed data is divided into short-time window sequences, medium-time window sequences, and long-time window sequences to generate master-slave aligned time series state tensors. Then, a discrete frequency domain transformation is performed on the tensor to extract the amplitude of the low-frequency component, the amplitude of the dominant frequency component, the periodic repetition intensity value, and the energy value of the high-frequency impact component. These are then concatenated with the time-domain features to form a master-slave torque frequency-time fusion feature tensor. This tensor is then input into the improved FEDformer model to output the predicted interactive torque evolution sequence over multiple future control cycles. Finally, the torque change slope sequence, torque fluctuation peak sequence, and frequency band energy distribution sequence are calculated to form a torque stability control parameter set consisting of feedback damping compensation, output limiting value, feedback stiffness correction, predicted feedforward gain value, and chattering suppression value. This set is then used for master-slave torque feedback stability control.

[0030] In practical applications, the operator guides the robotic arm towards the target interface using the master hand's operating handle. When the end effector moves in free space, the system primarily receives pose input from the master and joint motion response data from the slave. As the end effector approaches the interface edge, subtle changes occur in the contact force and load torque signals. The short-time window sequence in the master-slave alignment time-series state tensor records the rapid transition characteristics before and after contact establishment, the medium-time window sequence reflects the changing trend during continuous contact, and the long-time window sequence reflects the overall evolution of the master input and slave torque feedback throughout the entire task. The improved FEDformer model first performs sequence decomposition on the master-slave torque frequency-time fusion feature tensor to obtain trend component input sequences and periodic component input sequences. Then, multi-scale trend enhancement processes the slowly accumulating torque changes during contact establishment, and sparse frequency domain attention processes the repetitive oscillation bands and dominant frequency components. Finally, during the local transient fusion decoding process, it outputs the interactive torque changes over multiple future control cycles. Based on these results, the controller updates the stable control parameter set in real time, so that the master feedback no longer relies solely on the current single-moment sample value, but is adjusted in conjunction with the torque evolution trend in the future short period of time. When performing insertion, the operator can clearly feel the differences between different stages such as contact establishment, guide fit, resistance increase and locking completion; when tightening, the operator can distinguish between continuous damping increase and short-term jumps caused by local jamming, avoiding the application of excessive position input due to misjudgment; when pressing and positioning, the master feedback is smoother, reducing repeated trial movements and eliminating frequent rebounds after slight collisions at the robotic arm end.

[0031] To verify the practical effect of the method of the present invention, three comparative schemes were tested under the same location, the same master-slave manipulator platform, the same operator group, and the same task set. Comparative scheme A adopted the traditional fixed parameter force feedback control method, which only provides feedback based on the current measured load torque, without multi-source unified timing alignment and future torque prediction; the scheme of the present invention was tested using the complete method described in the claims. In each round of testing, 40 sets of insertion tasks, 40 sets of pressing and positioning tasks, and 40 sets of tightening and fastening tasks were completed continuously within the same working period, for a total of 120 sets. Statistical indicators included average task completion time, peak error of feedback at the moment of contact, mean square error of feedback torque during the stable contact stage, peak-to-peak value of master-end feedback oscillation, insertion success rate, first-time positioning success rate, tightening completion rate, number of pullback corrections after contact, number of accidental collisions, and operator subjective stability score. The test results are shown in the table below.

[0032] Table 1. Statistical comparison of torque feedback stability control effects between master and slave robotic arms As shown in Table 1, the proposed solution exhibits superior results in all three typical tasks. The average completion time for the entire task decreased from 25.2s under fixed-parameter force feedback control to 18.6s, indicating a significant reduction in the time spent by the operator in contact judgment and action correction. The peak error of the feedback at the moment of contact decreased from 0.94 N·m to 0.38 N·m, and the peak-to-peak value of the feedback oscillation at the master end decreased from 1.31 N·m to 0.44 N·m, demonstrating that the proposed solution can suppress torque abrupt changes and oscillation amplification during the contact establishment and sustained pressing phases. The mean square error of the feedback torque during the stable contact phase decreased to 0.19 N·m, indicating a higher consistency between the torque information sensed by the master end and the actual contact state at the slave end. The average high-frequency impact energy decreased significantly, reflecting that the frequency-time fusion characteristics and the improved FEDformer model can more effectively distinguish between transient impacts and continuous resistance changes, reducing ineffective chattering feedback. The number of post-contact pullback corrections and the number of accidental collisions decreased significantly, indicating that the operator does not need to confirm the contact state through multiple trials, thereby reducing unnecessary reciprocating actions. The improved consistency is also reflected in the insertion success rate, first-time positioning success rate, and tightening completion rate. The proposed solution achieves 98.3%, 96.7%, and 97.5% respectively, indicating that in precision operation scenarios, the torque stability control parameter set provided by the system enables the master-end feedback to better match the actual operation state, improving the quality of action completion. The feedback performance drift after 2 hours of continuous operation is only 2.7%, lower than the two comparative solutions, indicating that the proposed method can maintain a relatively stable output state even over longer operation periods.

[0033] Further analysis of individual task performance revealed that in the insertion task, traditional fixed-parameter force feedback control often resulted in significant main-end jumps when the pin just touched the edge of the guide hole. Operators frequently needed to briefly retract and realign, with an average of 4 to 5 noticeable corrections per 10 tasks. Under the same conditions, the present invention's main-end feedback more smoothly reflects the guide contact and resistance rise process, making it easier for operators to distinguish between "misaligned hard contact" and "normal contact sliding in along the guide surface," thus increasing the insertion success rate to 98.3%. In the pressing and positioning task, as the end contact surface gradually aligns with the target support surface, the traditional main-end feedback exhibits periodic fluctuations. Operators easily misinterpret minor oscillations as incomplete alignment and continue applying input, leading to overpressure followed by retraction and correction. The present invention can predict the evolution trend of the interactive torque over multiple control cycles and pre-calculate feedback damping compensation and chatter suppression values, ensuring a smoother resistance change at the main end after alignment, increasing the first-time positioning success rate to 96.7%. In tightening operations, traditional methods are poor at distinguishing between continuous damping changes and short-term jamming changes, often resulting in operators overcompensating or undercompensating before and after locking. This invention addresses this by fusing and modeling the amplitudes of low-frequency components, dominant frequency components, and high-frequency impact components. This allows the system to more accurately identify tightening trends and transient anomalies during tightening, achieving a tightening completion rate of 97.5% and significantly reducing overshoot.

[0034] The deep learning-based master-slave manipulator torque feedback stabilization control method and system described in this invention addresses problems such as asynchronous master-slave heterogeneous data, complex interactive torque changes, discontinuous contact feedback, and difficulties in oscillation suppression during remote precision operations. It establishes a complete control process from generating the original time-series dataset, alignment processing, frequency-time fusion representation, interactive torque evolution prediction to the formation of stable control parameters. In practical applications, this method enables more stable and continuous feedback from the master manipulator, more accurate operator judgment of the slave manipulator's contact state, shorter operation time, reduced accidental collisions and corrective actions, and maintains good control consistency and torque feedback quality even under long-term operation, demonstrating excellent engineering applicability.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A master-slave manipulator torque feedback stabilization control method based on deep learning, characterized in that, include: Step 1: Obtain the sequence of joint motion states of the robotic arm, the sequence of execution states of the robotic arm end effector, and the sequence of interaction with the robotic arm environment from the robotic arm side; obtain the sequence of operation inputs of the master end effector and the sequence of execution states of the master robotic arm from the master end effector side; and sort them in time sequence to generate the original time series dataset of master-slave torque feedback. Step 2: Add a unified timestamp to each data point in the original time series dataset of master-slave torque feedback, resample, and perform unified coordinate system transformation, dimension normalization, and sliding window segmentation to obtain the master-slave aligned time series state tensor; Step 3: Perform discrete frequency domain transformation on the master-slave aligned temporal state tensor and concatenate the feature dimensions to form the master-slave torque frequency-time fusion feature tensor; Step 4: Input the master-slave torque frequency-time fusion feature tensor into the improved FEDformer model, and perform interactive torque evolution through the sequence decomposition module, multi-scale trend enhancement module, sparse frequency domain attention module and local transient fusion decoding module to obtain the predicted interactive torque evolution sequence for multiple future control cycles; Step 5: Based on the predicted interactive torque evolution sequence, calculate the torque change slope sequence, torque fluctuation peak sequence, and frequency band energy distribution sequence to form a torque stability control parameter set.

2. The deep learning-based master-slave manipulator torque feedback stabilization control method according to claim 1, characterized in that, Step one specifically involves: The angular velocity signal of each joint motor is calculated from the output shaft end of each joint motor of the robot based on the change in angular displacement at adjacent sampling times. The current signal of each joint motor is collected by the current detection unit in the drive circuit. The angular acceleration signal of each joint motor is calculated based on the time sequence change of the angular velocity signal, thus forming a sequence of joint motion states of the robot. The attitude and displacement signals of the end effector relative to the preset reference coordinate system are collected from the end effector of the robot. Based on the current position of the end effector, the amount of attitude change, and the contact trigger signal between the end effector and the work object, the current free motion state, initial contact state, continuous contact state, or discontinuation contact state of the end effector are determined, forming a sequence of end effector states of the robot. By collecting contact force signals during the operation of the end effector through a torque sensor installed at the end of the robot, and calculating the load torque signal based on the lever arm relationship between the contact force signal and the mounting base of the end effector, a sequence of interaction between the robot and the environment is formed. The pose command signal of the master hand operating handle relative to the reference coordinate system of the master manipulator is acquired at the master hand operating handle, and the operation speed signal of the master hand operating handle is calculated based on the position change and attitude change of the pose command signal at adjacent sampling times to form the master end operation input sequence. The main manipulator's joint position signal and joint speed signal are collected at each joint of the main manipulator. The main manipulator's drive current signal is collected in the main manipulator's servo drive circuit. The feedback torque signal output by the controller is collected at the main manipulator's feedback execution end to form the main manipulator's execution state sequence. A unified sampling time stamp is added to the sequence of joint motion state of the robot arm, the sequence of execution state of the robot arm end effector, the sequence of interaction between the robot arm and the environment, the sequence of operation input of the master end effector, and the sequence of execution state of the master robot arm. The collected signals are then sorted in time according to a preset sampling period to generate a master-slave torque feedback raw time series dataset covering the master end operation input, slave end motion response, environmental contact action, and slave end feedback output.

3. The deep learning-based master-slave manipulator torque feedback stabilization control method according to claim 1, characterized in that, Step two specifically involves: A unified timestamp is added to each data in the original time series dataset of master-slave torque feedback, and resampling is performed according to a preset control cycle. Outlier removal, missing value interpolation, and noise filtering are then performed to generate a standardized time series signal. Based on the preset motion mapping relationship between the master manipulator and the slave manipulator, the pose command signal of the master manipulator's operating handle is transformed from the master reference coordinate system to the unified mapping coordinate system, and the pose signal of the slave manipulator's end effector is mapped to the unified mapping coordinate system. Perform joint space unified representation processing on the joint position signals of the master manipulator and the slave manipulator to generate pose association data with unified coordinates; The contact force signal, load torque signal, joint velocity signal, current signal, and feedback torque signal are subjected to dimensional normalization processing to generate normalized feature data with unified dimensions. The standardized time-series signal, pose-related data, and normalized feature data are divided into short-time window sequences, medium-time window sequences, and long-time window sequences according to preset sliding window sizes, resulting in a master-slave aligned time-series state tensor.

4. The deep learning-based master-slave manipulator torque feedback stabilization control method according to claim 1, characterized in that, Step three specifically involves: Perform a discrete frequency domain transformation on the master-slave aligned time-series state tensor to obtain the frequency domain representation matrix corresponding to each window; In the frequency domain representation matrix corresponding to each window, the amplitude of the low-frequency component, the amplitude of the dominant frequency component, the periodic repetition intensity value, and the energy value of the high-frequency impact component under the corresponding window are extracted and arranged according to window type, time order, and signal source to form the master-slave torque frequency-time fusion feature tensor.

5. The deep learning-based master-slave manipulator torque feedback stabilization control method according to claim 1, characterized in that, The improved FEDformer model is specifically as follows: The master-slave torque frequency-time fusion feature tensor is input into the sequence decomposition module, and the moving average calculation is performed on each feature channel in the master-slave torque frequency-time fusion feature tensor according to the time axis. In each feature channel, a sliding window of a preset length slides along the time axis moment by moment, calculates the average value of each sampled value in the current window, and arranges the moving average results corresponding to each moment in chronological order to generate the trend component subsequence corresponding to each feature channel. The trend component subsequences corresponding to each feature channel are combined according to the original feature channel order to generate the trend component input sequence; The original time series data of each feature channel in the master-slave torque frequency-time fusion feature tensor are subtracted from the corresponding trend component subsequence to generate the periodic component subsequence corresponding to each feature channel. The periodic component subsequences corresponding to each feature channel are combined according to the original feature channel order to generate the periodic component input sequence; The trend component input sequence is input into the multi-scale trend enhancement module, multiple moving average kernels of different lengths are set, and smoothing calculations are performed on the trend component input sequence using each moving average kernel to obtain trend candidate subsequences at multiple time scales. The trend component input sequence is subtracted from each trend candidate subsequence to obtain the trend residual subsequence at each time scale. A weighted average is performed on each trend candidate subsequence and trend residual subsequence to obtain the trend-enhanced output sequence. The periodic component input sequence is input into the sparse frequency domain attention module, and the periodic frequency domain representation sequence is obtained through fast Fourier transform; Based on the spectral energy values ​​corresponding to each frequency component in the periodic frequency domain representation sequence, sort them in descending order, and select the top N frequency components to form the dominant frequency index set. Based on the dominant frequency index set, the corresponding dominant frequency components are extracted from the periodic frequency domain representation sequence to generate a sparse frequency domain feature sequence. The sparse frequency domain feature sequence is input into the frequency domain self-attention computing unit. The proportion of each dominant frequency component to the sum of all dominant frequency components is used as the weight of each dominant frequency component. After weighting, the periodic enhanced output sequence is obtained. The trend enhancement output sequence and the period enhancement output sequence are input into the local transient fusion decoding module and concatenated according to the feature channel dimension to obtain the fused feature sequence; The fused feature sequence is input into the feedforward decoding network, and the predicted interaction torque evolution sequence for multiple future control cycles is output according to the preset prediction step size.

6. The deep learning-based master-slave manipulator torque feedback stabilization control method according to claim 5, characterized in that, The process of inputting the fused feature sequence into the feedforward decoding network and outputting the predicted interaction torque evolution sequence for multiple future control cycles according to a preset prediction step size is as follows: Read the values ​​of each feature channel of the fused feature sequence at the current input time, and extract the feature vector of the last time step in the fused feature sequence in chronological order; The feature vector at the end time step is input into the mapping layer, and the channel values ​​in the feature vector at the end time step are multiplied by the corresponding preset weights to generate the initial decoded feature vector. The initial decoded feature vector is written into the decoding buffer unit as the input state at the first prediction time. Nonlinear activation is performed on each element in the initial decoded feature vector to obtain the activated feature vector. The activation feature vector is input into the pooling layer, and max pooling is performed on the activation feature vector to obtain the candidate interaction torque value corresponding to the current prediction time. The candidate interaction torque value corresponding to the current prediction time is concatenated with the activation feature vector corresponding to the current prediction time to generate the feedback input vector for the next prediction time. Determine whether the current prediction time has reached the last prediction time corresponding to the preset prediction step size; if not, use the current feedback input vector as the input state for the next prediction time. Repeat the iteration until the current prediction time reaches the last prediction time corresponding to the preset prediction step size. Arrange the candidate interaction torque values ​​corresponding to each prediction time in the prediction result sequence in chronological order to generate a candidate interaction torque time series sequence. The difference between the measured load torque value and the first candidate interactive torque value in the candidate interactive torque time series is calculated to obtain the correction offset at the first moment. The initial time-time correction offset is added to the candidate interaction torque values ​​at each time-time in the candidate interaction torque time series to complete the initial boundary alignment process; The torque difference between two adjacent prediction times after the initial boundary alignment process is completed is calculated sequentially, and the interaction torque value of the subsequent time corresponding to the torque difference exceeding the preset change threshold is adjusted to the difference between the interaction torque value of the previous time and the preset change threshold, thus completing the continuity correction process between adjacent times. The time series sequence of interactive torques, after initial boundary alignment and continuity correction, is output as the predicted interactive torque evolution sequence for multiple future control cycles.

7. The deep learning-based master-slave manipulator torque feedback stabilization control method according to claim 1, characterized in that, Step five specifically involves: Read the interaction torque values ​​corresponding to each prediction time in the predicted interaction torque evolution sequence, and arrange them in chronological order of the prediction time to form the predicted torque time series sequence within the future control interval; The interaction torque values ​​corresponding to two adjacent prediction times within the future control interval are grouped together. The difference between the interaction torque value at the next prediction time and the interaction torque value at the previous prediction time is calculated respectively. The difference is then divided by the corresponding prediction time interval to obtain the torque change slope value between each adjacent prediction time. Arrange the torque change slope values ​​between each adjacent prediction time in the order of prediction time to generate a torque change slope sequence within the future control interval. The interaction torque values ​​at each prediction time point in the predicted torque time series are compared point by point. It is determined whether the interaction torque value at the current prediction time is greater than both the interaction torque value at the previous prediction time and the interaction torque value at the next prediction time. If so, the interaction torque value corresponding to the current prediction time is determined as a candidate value for the local fluctuation peak. The candidate values ​​of each local fluctuation peak are arranged in chronological order, and the candidate values ​​of local fluctuation peaks that are lower than the preset peak screening threshold are removed. The remaining candidate values ​​of local fluctuation peaks are retained as torque fluctuation peaks and arranged in the order of the predicted time to generate a torque fluctuation peak sequence in the future control interval. The predicted torque time series is subjected to frequency domain transformation to map the predicted torque time series within the future control interval into the corresponding frequency domain spectral line sequence, and the result is obtained by arranging them in order of frequency coordinates. Based on the preset frequency band division rules, the frequency domain representation results are divided into low-frequency band, mid-frequency band and high-frequency band, and the corresponding frequency components in each band are extracted respectively. The amplitude squares of each frequency component within each frequency band are accumulated to obtain the energy values ​​of the low-frequency band, the mid-frequency band, and the high-frequency band. These values ​​are then arranged in the order of the frequency bands to generate a frequency band energy distribution sequence. Read the slope values ​​of each torque change in the torque change slope sequence, take the absolute value and match it with the preset slope mapping interval to determine the feedback damping compensation amount; Read the peak values ​​of each torque fluctuation in the torque fluctuation peak sequence, and match each peak value with a preset peak mapping interval to determine the output limit value; Read the low-frequency, mid-frequency, and high-frequency band energy values ​​from the frequency band energy distribution sequence, and obtain the feedback stiffness correction, prediction feedforward gain, and chattering suppression values ​​based on the proportion of each band energy value to the total band energy. The torque stability control parameter set includes feedback damping compensation, output limiting value, feedback stiffness correction, prediction feedforward gain value, and chattering suppression value.

8. A deep learning-based master-slave manipulator torque feedback stabilization control system, executing the deep learning-based master-slave manipulator torque feedback stabilization control method according to any one of claims 1 to 7, characterized in that, Includes the following modules: The data acquisition module is used to acquire the sequence of joint motion states of the robotic arm from the robotic arm side, the sequence of execution states of the robotic arm end effector, and the sequence of interaction with the robotic arm environment from the robotic arm side, acquire the sequence of operation inputs of the master end effector and the sequence of execution states of the master robotic arm from the master end effector side, and sort them in time sequence to generate the original time sequence dataset of master-slave torque feedback. The timing alignment module is used to add a unified timestamp to each data in the original timing dataset of master-slave torque feedback, resample it, and perform unified coordinate system transformation, dimension normalization and sliding window segmentation to obtain the master-slave aligned timing state tensor. The fusion representation module is used to perform discrete frequency domain transformation on the master-slave aligned temporal state tensor and concatenate the feature dimensions to form the master-slave torque frequency-time fusion feature tensor. The torque prediction module is used to input the master-slave torque frequency-time fusion feature tensor into the improved FEDformer model. Through the sequence decomposition module, multi-scale trend enhancement module, sparse frequency domain attention module and local transient fusion decoding module, the interactive torque evolution is performed to obtain the predicted interactive torque evolution sequence in multiple future control cycles. The parameter generation module is used to calculate the torque change slope sequence, torque fluctuation peak sequence, and frequency band energy distribution sequence based on the predicted interactive torque evolution sequence, and form a torque stability control parameter set.

Images