A biomimetic adaptive variable stiffness robotic arm joint control method

Abstract

This invention relates to the field of robotic arm control technology and discloses a biomimetic adaptive variable stiffness robotic arm joint control method. The method includes acquiring multi-source biosignals such as electromyography (EMG), tendon tension, and joint angle signals during joint operation; constructing stiffness feature vectors based on EMG and tendon tension signals, and generating motion trajectory sequences based on joint angle signals; extracting corresponding stiffness feature vectors for each trajectory point and calculating stiffness sensitivity; constructing a biomimetic stiffness reference model containing the mapping relationship between stiffness feature vectors and joint angle signals, and dynamically adjusting the mapping relationship using stiffness sensitivity; predicting joint stiffness change sequences based on the adjusted model and real-time signals, fusing the sequence with the current joint angle signal to generate dynamic stiffness parameters; calculating a drive torque correction amount based on the dynamic stiffness parameters and a preset safety threshold range, iteratively updating the model weighting parameters until the correction amount meets the convergence condition, and outputting a drive torque command to the actuator.

Description

Technical Field

[0001] This invention relates to the field of robotic arm control technology, specifically to a biomimetic adaptive variable stiffness robotic arm joint control method. Background Technology

[0002] In fields such as industrial automation, medical rehabilitation, and service robots, robotic arms are increasingly widely used, and their performance directly affects operational accuracy and safety. Traditional robotic arm joint control often employs a fixed stiffness mode, which struggles to meet the diverse task requirements in complex dynamic environments. For example, in assembly operations, the robotic arm needs to maintain low stiffness when contacting the workpiece to avoid collision damage, while higher stiffness is required when handling heavy objects to ensure motion stability. Fixed stiffness control clearly cannot accommodate both scenarios.

[0003] Most existing variable stiffness control methods are based on preset motion trajectories or empirical models for stiffness adjustment, lacking the ability to perceive and adaptively adjust to the actual working environment and the state of the robotic arm itself in real time. Some methods attempt to optimize stiffness parameters by collecting physical signals such as joint angles and velocities, but these signals only reflect the kinematic state of the robotic arm and cannot directly relate to the intrinsic variation law of joint stiffness. In addition, due to the complex structure of robotic arm joints and the existence of factors such as nonlinear friction and elastic deformation, traditional control methods are difficult to establish accurate stiffness models, which easily leads to problems such as stiffness adjustment lag and overshoot during dynamic operations, affecting control accuracy and system stability.

[0004] With the development of biomimetic robotics, researchers have discovered that human joints can achieve efficient adaptive stiffness adjustment through the coordinated action of muscles and tendons, providing a new approach for variable stiffness control of robotic arms. However, existing biomimetic control methods often only utilize single biological signals for stiffness mapping, neglecting the correlation between multiple sources of signals such as tendon tension and joint angles, making it difficult to accurately reproduce the stiffness adjustment mechanism of human joints. Furthermore, because the sensitivity of joints to stiffness changes varies under different working scenarios, models lacking dynamic weighted adjustments cannot optimize stiffness parameters according to actual needs, limiting the adaptability of robotic arms in complex environments. Summary of the Invention

[0005] The purpose of this invention is to provide a biomimetic adaptive variable stiffness robotic arm joint control method to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a biomimetic adaptive variable stiffness robotic arm joint control method, the method comprising:

[0007] Multi-source biosignal data, including electromyography (EMG) signals, tendon tension signals, and joint angle signals, are collected during the operation of the robotic arm joints. Based on the EMG and tendon tension signals, a joint stiffness feature vector is constructed, and a joint motion trajectory sequence is generated according to the joint angle signals.

[0008] For each joint motion trajectory point in the joint motion trajectory sequence, extract the stiffness feature vector corresponding to that trajectory point and calculate the stiffness sensitivity of the current trajectory point;

[0009] A biomimetic stiffness reference model is constructed, which includes a mapping relationship between stiffness feature vectors and joint angle signals, and the mapping relationship is dynamically weighted and adjusted using the stiffness sensitivity.

[0010] Based on the adjusted biomimetic stiffness reference model and real-time acquired signals, the joint stiffness change sequence is predicted.

[0011] By fusing the joint stiffness change sequence with the current joint angle signal, dynamic stiffness parameters of the joint are generated.

[0012] The joint drive torque correction amount is calculated based on the joint dynamic stiffness parameters and the preset stiffness safety threshold range.

[0013] The weighted parameters of the biomimetic stiffness reference model are iteratively updated until the joint driving torque correction satisfies the preset convergence condition.

[0014] The final joint drive torque command is output to the robotic arm joint actuator.

[0015] Preferably, the construction of the joint stiffness feature vector includes: performing bandpass filtering on the electromyographic signal and dividing it into equal time windows; extracting the energy integral value, median frequency, and zero-crossing rate of the electromyographic signal in each time window, and simultaneously acquiring the peak value of the tendon tension signal in the corresponding time window; arranging the energy integral value, median frequency, zero-crossing rate, and peak value of the tendon tension signal in the order of the time windows to form a four-dimensional stiffness feature vector sequence.

[0016] Preferably, the calculation of the stiffness sensitivity of the current trajectory point includes: calculating the stiffness sensitivity of the current trajectory point by combining the historical stiffness feature vector distribution of adjacent trajectory points.

[0017] Preferably, the step of calculating the stiffness sensitivity of the current trajectory point by combining the historical stiffness feature vector distribution of adjacent trajectory points includes: obtaining the stiffness feature vectors of the five consecutive trajectory points preceding the current trajectory point; calculating the distribution dispersion of the stiffness feature vectors of the five consecutive trajectory points in four-dimensional space; performing negative correlation normalization on the distribution dispersion to obtain the stiffness feature stability coefficient of the current trajectory point; calculating the physical response intensity of the current trajectory point based on the absolute value of the gradient of the tendon tension signal within the corresponding time window; and using the product of the stiffness feature stability coefficient and the physical response intensity as the stiffness sensitivity of the current trajectory point.

[0018] Preferably, the predicted joint stiffness change sequence includes: performing cluster analysis on the stiffness feature vectors of the target trajectory points based on the adjusted biomimetic stiffness reference model to select a key stiffness feature subset; and predicting the joint stiffness change sequence through a long short-term memory network based on the key stiffness feature subset and the real-time acquired tendon tension signal.

[0019] Preferably, the dynamic weighted adjustment includes: extracting all historical mapping relationship records of the current trajectory point in the biomimetic stiffness reference model; arranging the stiffness sensitivity in each historical mapping relationship record in descending order to form a sensitivity sequence; calculating the proportion of the difference between the maximum and minimum values ​​in the sensitivity sequence to the total range of the sequence, as the mapping relationship confidence weight; and using the mapping relationship confidence weight to scale the newly added mapping relationship values ​​of the current trajectory point to update the mapping relationship database of the biomimetic stiffness reference model.

[0020] Preferably, the step of selecting the key stiffness feature subset includes: obtaining all stiffness feature vector samples at the target trajectory point; dividing all stiffness feature vector samples into multiple feature clusters using a fuzzy clustering algorithm; calculating the intra-class average stiffness sensitivity of each feature cluster; selecting feature clusters whose intra-class average stiffness sensitivity is greater than a preset sensitivity threshold; and merging all stiffness feature vectors of the selected feature clusters to form the key stiffness feature subset.

[0021] Preferably, the predicted joint stiffness change sequence includes: inputting a subset of key stiffness features into the input layer of a pre-trained long short-term memory network; fusing the tendon tension signal difference values ​​of three historical time steps in the hidden layer of the long short-term memory network; extracting the temporal dependence features of joint stiffness step by step through three hidden units; and outputting the predicted joint stiffness values ​​and their confidence levels for the next five time steps to form a joint stiffness change sequence.

[0022] Preferably, the generation of joint dynamic stiffness parameters includes: extracting the stiffness prediction value corresponding to the highest confidence from the joint stiffness change sequence; obtaining the differential value of the current joint angle signal; inputting the stiffness prediction value and the differential value of the joint angle signal into a preset stiffness dynamic equation; solving the stiffness dynamic equation to obtain the instantaneous stiffness coefficient of the joint and its rate of change, and merging them into joint dynamic stiffness parameters.

[0023] Preferably, the calculation of the joint drive torque correction includes: comparing the instantaneous stiffness coefficient in the joint dynamic stiffness parameters with the upper boundary value of the stiffness safety threshold range; when the instantaneous stiffness coefficient exceeds the upper boundary value, calculating the percentage value of the excess portion; determining the amount of downward adjustment of the joint drive torque based on the product of the percentage value and the stiffness change rate; comparing the instantaneous stiffness coefficient with the lower boundary value of the stiffness safety threshold range; when the instantaneous stiffness coefficient is lower than the lower boundary value, calculating the percentage value of the insufficient portion; and determining the amount of upward adjustment of the joint drive torque based on the product of the percentage value and the stiffness change rate.

[0024] Compared with the prior art, the beneficial effects of the present invention are:

[0025] By collecting multi-source biological signals and constructing a feature system that includes electromyography (EMG), tendon tension, and joint angle signals, we can more comprehensively reflect the intrinsic variation patterns of joint stiffness, overcoming the limitations of traditional single-signal control. The fusion and utilization of multi-source signals makes the construction of stiffness feature vectors more closely aligned with the human joint's regulatory mechanisms, providing a richer information foundation for subsequent stiffness prediction and parameter generation.

[0026] By calculating the stiffness sensitivity for each trajectory point in the joint motion trajectory sequence, the response characteristics of the joint to stiffness changes under different motion states can be accurately captured. Dynamically weighting the mapping relationship of the biomimetic stiffness reference model based on the stiffness sensitivity allows the model to adaptively optimize according to the needs of actual working scenarios, avoiding the problem of insufficient adjustment accuracy of a fixed model in complex environments, and making stiffness adjustment more targeted and flexible.

[0027] Based on the adjusted model and real-time signals, the system predicts the joint stiffness change sequence and integrates the current joint angle signal to generate dynamic stiffness parameters, achieving closed-loop linkage from signal perception to parameter generation. This linkage mechanism can quickly respond to changes in the robotic arm's motion state, update stiffness parameters in a timely manner, reduce the lag in stiffness adjustment, and enable the joint to maintain appropriate stiffness characteristics at different motion stages, thereby improving the real-time performance and accuracy of control.

[0028] By comparing the dynamic stiffness parameters with a preset safety threshold range to calculate the drive torque correction, and iteratively updating the model weighted parameters until the convergence condition is met, a continuously optimized control closed loop is formed. This process can continuously correct model errors, offset the effects of nonlinear factors on the robotic arm joints, enhance the system's resistance to changes in structural parameters and external disturbances, and help maintain the stability of the control process.

[0029] The final output joint drive torque command can be directly applied to the actuator, ensuring that the robotic arm maintains both motion stability and good compliance during contact operations in various working scenarios. This control effect, which balances stability and compliance, expands the application range of the robotic arm, making it more practical in fields with high requirements for stiffness adjustment, such as precision assembly and rehabilitation assistance. Attached Figure Description

[0030] Figure 1 This is a timing diagram of the biomimetic adaptive variable stiffness robotic arm joint control method described in this invention.

[0031] Figure 2 A flowchart for constructing joint stiffness feature vectors;

[0032] Figure 3 This is a graph showing the results of signal acquisition and processing.

[0033] Figure 4 A flowchart for detailed calculation of stiffness sensitivity;

[0034] Figure 5 Figure showing stiffness characteristics and sensitivity analysis;

[0035] Figure 6 A flowchart for dynamically weighted adjustment of mapping relationships;

[0036] Figure 7 A flowchart for selecting a subset of key stiffness features. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] Please see Figure 1 This invention provides a biomimetic adaptive variable stiffness robotic arm joint control method, the method comprising:

[0039] High-precision biosensors synchronously acquire electromyographic (EMG) signals, tendon tension signals, and joint angle signals during the operation of the robotic arm joints. EMG signals are acquired using a surface electrode array at a sampling rate of 1 kHz, tendon tension signals are measured using an embedded fiber Bragg grating sensor, and joint angle signals are recorded by a high-resolution optical encoder. The signal acquisition system ensures time alignment through hardware-synchronized triggering.

[0040] Electromyographic (EMG) and tendon tension signals are preprocessed to construct a four-dimensional stiffness feature vector. Joint angle signals are Kalman filtered to generate a smooth joint motion trajectory sequence. For each point in the trajectory sequence, the stiffness sensitivity of the current point is calculated by analyzing the historical stiffness feature distribution. A biomimetic stiffness reference model incorporating a stiffness-angle mapping relationship is established, and the mapping relationship is dynamically weighted using stiffness sensitivity. Based on the adjusted model and real-time signals, a long short-term memory (LSTM) network is used to predict the joint stiffness changes over the next five time steps. The prediction results and the current differential value of the joint angle are input into the stiffness dynamic equation to obtain dynamic parameters containing instantaneous stiffness coefficients and their rates of change. The dynamic parameters are compared with a preset safety threshold to calculate the drive torque correction. The model parameters are iteratively updated until the correction converges. Finally, the optimized torque command is output to the joint actuator.

[0041] Example 1: See Figure 2 Electromyography (EMG) signal acquisition employed an eight-channel surface electrode array with an electrode spacing of 20 mm to optimize signal spatial resolution. The signal acquisition hardware utilized a 24-bit analog-to-digital converter, maintaining a sampling rate strictly at 1 kHz and an input impedance greater than 1 GΩ to minimize signal attenuation. Raw signal transmission used shielded twisted-pair cable, with a differential amplifier at the end of the line, achieving a common-mode rejection ratio (CMRR) of 110 dB. During signal conditioning, a cascaded fourth-order Butterworth bandpass filter was implemented, controlling passband ripple to within 0.5 dB and stopband attenuation to no less than 40 dB. The filtered signal was then processed by a hardware-implemented 50 Hz notch filter to eliminate power frequency interference, with a notch depth greater than 60 dB.

[0042] The time window is divided using a sliding window mechanism with a fixed window length of 100ms and 50% overlap between adjacent windows. Window boundary alignment is smoothed using a Hanning window function to reduce spectral leakage. Within each processing window, signal analysis comprises three parallel computational flows. The energy integration calculation flow first performs full-wave rectification on the signal, then uses the trapezoidal numerical integration method to calculate the absolute area value, with the integration step size consistent with the sampling interval. The median frequency calculation is achieved using a 256-point Fast Fourier Transform, achieving a spectral resolution of approximately 4Hz. The power spectrum is calculated using the periodogram method, and the median frequency is located using a binary search algorithm. Zero-crossing rate detection is set with a hysteresis threshold of ±50μV, and a hardware comparator is used for real-time counting.

[0043] Tendon tension signal acquisition utilizes a fiber Bragg grating sensor array embedded in a bionic tendon, achieving a wavelength demodulation accuracy of 1 pm. Signal sampling is strictly synchronized with electromyographic (EMG) signals, with microsecond-level alignment achieved via hardware trigger signals. Peak detection employs a sliding extremum method, with the detection window width matched to the EMG processing window. Signal preprocessing includes three levels of moving average filtering with window widths of 5 ms, 10 ms, and 20 ms to eliminate noise interference across different frequency bands. The deviation between the tension peak timestamp and the center point of the EMG feature extraction window is controlled within ±2 ms.

[0044] The feature vectors are constructed using a time-aligned four-dimensional data structure. Each feature dimension undergoes independent normalization, with the normalization range dynamically adjusted based on historical data. Energy integral values ​​are linearly normalized after logarithmic transformation, the median frequency is linearly scaled, the zero-crossing rate retains its original count value, and the peak tendon tension is piecewise normalized. The feature vector sequence is stored in a circular buffer, with separate write and read pointers for lock-free concurrent access. The buffer data is arranged chronologically, with each record containing a timestamp accurate to the microsecond.

[0045] Joint angle signal processing employs a 17-bit absolute optical encoder, achieving an angle resolution of 0.001 degrees. Signal acquisition utilizes a quadruple frequency multiplication circuit for subdivision, with orthogonal decoding error less than one count. Raw angle data undergoes a five-point cubic smoothing filter, followed by noise suppression using a parameter-adjustable Kalman filter. The filtering parameters are adaptively adjusted based on joint movement speed, increasing the process noise covariance at low speeds and the observation noise covariance at high speeds. Trajectory sequence generation uses cubic spline interpolation, with the interpolation node density synchronized with the electromyography processing window to ensure time alignment accuracy within 1 ms.

[0046] Multi-source signal synchronization employs a distributed clock architecture. The main controller generates a 1PPS synchronization pulse, and each acquisition node achieves clock synchronization via the IEEE 1588 precise time protocol, with an overall system clock deviation of less than 10μs. Data timestamps are hardware-level marked, with sampling time recorded at the sensor end. Network transmission uses a deterministic Ethernet protocol, with maximum transmission latency controlled within 100μs. During signal alignment processing, any potential fixed delays are pre-calibrated and compensated, while dynamic delays are corrected using interpolation.

[0047] The signal quality control module monitors the data integrity of each channel in real time. The electromyography (EMG) signal quality index includes three indicators: signal-to-noise ratio, baseline drift, and saturation detection. Tendon tension signal quality assessment primarily focuses on the wavelength offset of the fiber optic grating and the integrity of the reflection spectrum. Joint angle signal quality is verified through encoder zero-position signals and counting continuity. If any channel exhibits quality anomalies, the system automatically switches to a redundant backup channel or activates a data compensation algorithm. The compensation algorithm builds an autoregressive model based on historical data to predict data trends during missing time periods.

[0048] The feature vector sequence is stored in a columnar format, with each feature dimension stored separately to improve access efficiency. Data persistence uses a circular log structure, where the latest data overwrites the oldest. Memory management employs a pre-allocation strategy to avoid latency caused by dynamic memory allocation. The feature data access interface provides two modes: time range query and nearest neighbor query, supporting multi-threaded concurrent reading. Data compression uses a lossy floating-point compression algorithm, with a compression ratio controlled within 4:1 to ensure feature accuracy.

[0049] The real-time processing pipeline adopts a producer-consumer pattern design. Signal acquisition runs as a producer thread in the real-time operating system kernel space, while feature extraction runs as a consumer thread in user space. Inter-thread communication is achieved through a lock-free circular queue, with the queue depth determined based on the worst-case processing latency. A processing latency monitoring module tracks the processing time of each data packet in real time and dynamically adjusts thread priorities to balance system load. Exception handling mechanisms include a watchdog timer and heartbeat detection to ensure that real-time requirements are met.

[0050] The system calibration module executes an automatic calibration process upon startup. Electromyography (EMG) channel calibration includes gain calibration, offset calibration, and frequency response testing. Tendon tension calibration includes static load testing and dynamic frequency response testing. Joint angle calibration performs zero-point calibration and rotation direction verification. Calibration data is stored in non-volatile memory and is automatically recalibrated periodically; the calibration cycle is automatically adjusted based on changes in ambient temperature. An online calibration function allows for background calibration of specific channels while the system is running, without affecting normal data acquisition.

[0051] The hardware acceleration module accelerates key computations using an FPGA. The filter bank employs a polyphase structure for parallel processing of all channel data. The Fast Fourier Transform (FFT) uses a radix-2 pipelined implementation, performing one butterfly operation per clock cycle. The numerical integrator uses an accumulator structure, supporting continuous, uninterrupted computation. The peak detection circuit uses a comparator array for parallel processing, keeping the detection latency within 5 clock cycles. All hardware acceleration modules interact with the main processor via a standard AXI bus, with data transfer using DMA.

[0052] See Figure 3The diagram illustrates the key processes of signal acquisition and processing. The top two images show the raw electromyography (EMG) signals and the EMG signals after bandpass filtering and 50Hz notch filtering, respectively. It is evident that the filtering effectively removed power frequency interference and baseline drift, preserving the effective signals reflecting muscle activity. The third image shows the raw data and smoothing results of the tendon tension signal. The three-level moving average filtering effectively suppressed high-frequency noise while retaining the main trend of tension changes. The bottom image shows the joint angle signal; the raw data, after five-point cubic smoothing filtering and adaptive Kalman filtering, yielded a smooth motion trajectory sequence. These processes lay the data foundation for subsequent feature extraction and stiffness analysis.

[0053] Example 2: See Figure 4 The stiffness sensitivity calculation module employs a multi-dimensional spatiotemporal feature analysis method to process the stiffness eigenvectors of continuous trajectory points. The system maintains a first-in, first-out (FIFO) eigenvector buffer with a depth of five trajectory points, each corresponding to a four-dimensional eigenvector. When a new trajectory point arrives, the buffer is automatically updated, discarding the oldest data point. The distribution characteristics of the eigenvectors in four-dimensional space are evaluated through eigenvalue analysis of the covariance matrix. For the current trajectory point, the Mahalanobis distance between the eigenvectors of its five consecutive points in four-dimensional space is calculated; this distance reflects the compactness of the eigenvector distribution. The distance is calculated using the following formula:

[0054]

[0055] in: : indicates Mahalanobis distance. : Represents the four-dimensional stiffness eigenvector of the current trajectory point. : Represents the mean vector of historical eigenvectors. : Represents the covariance matrix of historical eigenvectors, : Represents the inverse of the covariance matrix, used to eliminate feature scale differences and perform decorrelation processing. : Represents the deviation vector between the current feature vector and the mean vector, reflecting the offset of the current feature relative to the historical average level. : Represents the transpose operation of a matrix or vector, converting a column vector into a row vector for matrix multiplication.

[0056] The physical response intensity analysis of tendon tension signals employs the sliding differential window method. The system is configured with an analysis window of 50 ms width, and the first-order difference of the tension signal within the window is calculated in 10 ms increments. The difference calculation uses the central difference method, where the difference value at each sampling point is determined by the weighted average of the two preceding and following sampling points. The maximum absolute value of the difference is recorded during the window sliding process and used as the baseline value of the physical response intensity for that time period. The response intensity value is normalized to the range of 0 to 1 after a logarithmic transformation, with a lower threshold set during the transformation process to avoid zero values.

[0057] The stability coefficient and physical response intensity are fused using a product-weighted method. The two parameters are multiplied and then nonlinearly mapped using a hyperbolic tangent function. The output is then linearly scaled to obtain the final stiffness sensitivity value. The scaling factor is dynamically adjusted based on the statistical characteristics of historical data to keep the dynamic range of the sensitivity value within the system's defined operating range. Real-time monitoring is implemented during the calculation process, and an exception handling procedure is automatically activated when the input parameters exceed the normal range.

[0058] Historical data processing employs an exponential decay weighted strategy. The system maintains a time decay factor for each trajectory point, which decreases exponentially over time. The initial weight of the most recent trajectory point is set to 1.0, and the decay coefficient for each subsequent trajectory point is 0.7. A weighted average method is used to calculate the mean of the eigenvectors, and the covariance matrix calculation also considers the weighting factor. This approach allows recent data to have a greater impact on sensitivity calculations while retaining the reference value of more distant data.

[0059] The feature vector cache management employs a circular buffer structure. The buffer is implemented as a fixed-length array, using pointers to cyclically access memory. Pointer operations utilize atomic instructions to ensure thread safety and avoid multi-task access conflicts. When the buffer is full, it automatically overwrites the oldest data and triggers a recalculation of sensitivity. The system maintains a version number flag, which is incremented with each buffer update to check data consistency during the calculation process.

[0060] The covariance matrix is ​​calculated using a numerically stable iterative algorithm. The algorithm incrementally updates matrix elements with each buffer update, avoiding redundant calculations of all historical data. Matrix inversion is performed using Cholesky decomposition, with a small regularization term added during the decomposition process to ensure numerical stability. Eigenvalue analysis employs the Jacobi rotation method, with an upper limit of 20 iterations to ensure real-time performance. Condition number monitoring is implemented during matrix operations, automatically activating backup calculation methods when the matrix approaches singularity.

[0061] The sliding window for calculating physical response intensity is implemented as an overlapping circular buffer. When the window slides, only the difference value of newly added data points is calculated, reusing some calculation results from previous windows. The difference calculation uses fixed-point arithmetic to optimize performance, and intermediate results are cached in a dedicated register set. Maximum value detection uses branch prediction technology to optimize comparison operations, and hardware features accelerate the extreme value search process. The window sliding is strictly synchronized with the electromyography signal processing, with a time deviation controlled within 1ms.

[0062] The nonlinear transformation module employs a combination of lookup tables and linear interpolation. The values ​​of the transformation function at discrete points are pre-calculated and stored in on-chip memory. During actual calculations, continuous output is obtained through neighbor-to-neighbor interpolation, with interpolation errors controlled within 0.1%. Lookup table updates are performed asynchronously in a background thread, without affecting the real-time calculation process. Transformation parameters can be adjusted online, with a smooth transition to avoid output jumps.

[0063] Real-time performance is ensured through a multi-stage pipeline architecture. The sensitivity calculation task is divided into five stages: data preparation, distance calculation, intensity analysis, fusion calculation, and result output. Each stage runs on an independent processing unit, with intermediate results passed through a FIFO queue. The pipeline depth is determined based on the worst-case processing latency, ensuring real-time output even during peak periods. Task scheduling employs a static priority strategy, with the sensitivity calculation task set to the second-highest priority level in the system.

[0064] The anomaly handling mechanism comprises two levels: input validation and output validation. Input validation checks the numerical range of the feature vectors and the correlation between their dimensions, rejecting data points that are clearly anomalous. Output validation compares the current sensitivity value with the historical trend; when the deviation exceeds three standard deviations, a review process is triggered. The review process recalculates using a longer time window and requests manual intervention when necessary. The system records all anomaly events and their processing results for subsequent analysis and optimization.

[0065] The parameter adaptation module continuously monitors the system's operational status. Based on the statistical characteristics of historical sensitivity values, it dynamically adjusts various thresholds and coefficients during the calculation process. The adjustment strategy is based on a reinforcement learning framework, progressively optimizing parameter configurations without affecting real-time performance. Detailed logs are recorded during the adjustment process, supporting rollback to previous stable versions. Parameters are stored in non-volatile memory, automatically restoring the most recent valid configuration after a system restart.

[0066] For hardware acceleration, key computing modules are deployed on programmable logic devices. Mahalanobis distance calculation is implemented using a fourth-order parallel pipeline, processing one dimension of data per clock cycle. Covariance matrix operations employ a systolic array structure, with strict alignment between data flow and computation flow. Nonlinear transformations are implemented using dedicated function approximation units, supporting rapid calculation of various common mathematical functions. Hardware modules are connected to the main processor via a high-speed bus, and data transmission employs a double-buffering mechanism to avoid pauses.

[0067] System resource management employs a strategy combining static allocation and dynamic adjustment. Memory requirements for sensitive computation tasks are pre-analyzed and determined, and a dedicated memory pool is allocated to avoid the overhead of dynamic allocation. CPU core binding technology ensures that computational tasks are not interfered with by other processes. A cache prefetch strategy optimizes data access patterns and reduces the impact of memory latency. When the system load is high, the computational precision is automatically reduced to maintain real-time performance; once the load decreases, the standard precision mode is restored.

[0068] A time synchronization mechanism ensures that the calculation process is aligned with the robotic arm's control cycle. Sensitivity calculations are completed within a specified time window and output is triggered by a hardware timer. Sufficient time margin is reserved before the start of the control cycle to handle worst-case calculation delays. The system clock synchronization employs a hybrid scheme, combining a network time protocol and a local high-precision clock source to maintain microsecond-level time consistency. A clock drift compensation algorithm runs continuously to eliminate long-term accumulated time errors.

[0069] See Figure 5 This diagram illustrates the process of stiffness feature extraction and sensitivity calculation. The first image shows the changing trend of the four-dimensional stiffness feature vector, including electromyographic energy integral, median frequency, zero-crossing rate, and peak tendon tension. These features comprehensively reflect the stiffness characteristics of the joint. The middle image shows the Mahalanobis distance of the feature vector, evaluating the feature stability of the current point by analyzing the distribution dispersion of the feature vectors of five consecutive trajectory points. The bottom image shows the calculated stiffness sensitivity, which integrates the feature stability coefficient and physical response intensity, reflecting the sensitivity of the current joint position to stiffness changes. High-sensitivity regions correspond to critical inflection points in joint motion; these points are crucial to the stability of the control system and require more accurate stiffness prediction and torque adjustment.

[0070] Example 3: See Figure 6 The biomimetic stiffness reference model is constructed using a graph database as its core storage structure. Nodes represent the multidimensional feature states of trajectory points, and edges represent the transition relationships between states. Each node contains a triplet of spatial coordinates, a stiffness feature vector, and a timestamp. Node attributes extend to store the stiffness sensitivity values ​​from the ten most recent accesses. Edge attributes record the transition frequency and average transition time, forming a weighted directed graph structure. The following dynamic weighting formula is used during model updates:

[0071]

[0072] in, This represents the transition weight from node i to j. It is the stiffness sensitivity recorded during the k-th visit. The corresponding timestamp, The current system time. The time decay factor is fixed at 0.05. It represents the number of historical visits. It includes all historical sensitivity values. This weighting calculation makes the impact of recent high-sensitivity records more significant on the model.

[0073] The graph database employs spatial partitioning indexes to accelerate queries, dividing the joint workspace into several cubic grids. Each grid maintains a local subgraph, connecting adjacent grids via a skip list structure. During a query, the spatial grid is first located, and then a nearest neighbor search is performed in the local subgraph, with the search radius adaptively adjusted based on the current movement speed. Node insertion utilizes a batch commit strategy, performing a batch write every five new trajectory points to reduce disk I / O overhead.

[0074] The dynamic weighted adjustment process employs a two-stage strategy. The first stage retrieves all historical mapping records for the current trajectory point and arranges them in descending order of sensitivity to form a priority queue. A binary heap data structure is used for queue construction, and the time complexity of insertion and deletion operations is controlled to the logarithmic level. In the second stage, when calculating weights, time decay compensation is applied to the records in the queue, with the compensation coefficient having a negative exponential relationship with the record's timeliness. New mapping relationships are integrated into the existing model using a weighted moving average method, while the weights of older data decay proportionally according to the Fibonacci sequence.

[0075] A hybrid inference approach is employed in the stiffness prediction stage. The input layer of the Long Short-Term Memory (LSTM) network receives a subset of key stiffness features, with feature selection based on density reachability analysis. The network's hidden layers are designed with residual connection structures, each containing gated recurrent units and a temporal attention mechanism. The tendon tension signal differential values ​​are injected into the network through an auxiliary input channel and cross-fused with the main feature flow in the third hidden layer. The network output layer uses a hybrid density architecture to simultaneously predict stiffness values ​​and their probability distributions.

[0076] Model parameter updates employ an online learning strategy. Training data streams are managed through a circular buffer with a capacity set to the 200 most recent trajectory points. Parameter update triggers include: the deviation between new data and model predictions exceeding three standard deviations, or the mean sensitivity of five consecutive trajectory points exceeding a threshold. The training process utilizes an elastic weight consolidation algorithm to retain prior knowledge of important parameters. The learning rate is dynamically adjusted based on model convergence, employing a cosine annealing scheduling strategy.

[0077] Real-time performance is guaranteed through reserved computing resources. Model inference tasks utilize two dedicated CPU cores, and memory allocation employs big-page technology to reduce TLB loss. Graph database queries utilize an asynchronous prefetching mechanism, preloading potentially accessed neighboring nodes. Network inference employs half-precision floating-point operations, and critical paths are vectorized using the SIMD instruction set. Processing latency is monitored in real-time, tracking the time consumption of each stage, and automatically switching to a simplified model when time limits are exceeded.

[0078] The anomaly recovery mechanism includes a dual path of model rollback and data repair. Model version management employs snapshot technology, automatically saving checkpoints hourly. Data repair initiates a local recalculation process for anomalous trajectory points, using cubic spline interpolation to fill in missing data. When persistent anomalies are detected, the system switches to a standby inference mode based on a physics model until the main model completes online retraining.

[0079] For hardware acceleration, graph traversal operations are processed in parallel via GPU. Node attribute access is optimized by merging memory transactions and using cache line prefetching to hide latency. Matrix operations are deployed on tensor cores, employing a tiled strategy to improve computational density. Network inference is executed through a dedicated AI accelerator, using weight quantization to reduce data transfer volume. Hardware resource scheduling takes temperature management into account, dynamically adjusting clock frequency to balance performance and heat dissipation.

[0080] The system monitoring module achieves end-to-end observability. Performance counters collect throughput and latency metrics for each processing stage in real time. Resource usage is continuously monitored via kernel-level probes, including CPU utilization, memory pressure, and I / O latency. Data quality audit logs record characteristic distribution changes and anomaly markers for each trajectory point. All monitoring data is aggregated into a time-series database, supporting sliding window analysis and long-term trend prediction.

[0081] Configuration management uses a declarative interface, with model parameters and algorithm hyperparameters defined through configuration files. Runtime modifications are implemented via a hot-reload mechanism, eliminating the need for a system restart. Parameter validation includes range checks and logical consistency verification; invalid configurations are automatically rolled back to safe default values. The version control system records all configuration changes, supporting difference comparisons and restoration of historical versions.

[0082] The security mechanism includes a two-tiered protection system of data encryption and access control. Data in persistent storage is encrypted using the AES-256 algorithm, with the key protected by a secure chip. Sensitive data in memory is protected by a memory encryption engine. Access control is based on a role-based access control model, and operation audit logs record all critical system events. The communication channel uses two-way TLS authentication to prevent man-in-the-middle attacks.

[0083] The time synchronization subsystem maintains the temporal consistency of model updates. The distributed clock uses a hybrid PTP / NTP protocol for synchronization, with clock deviation controlled within 50 microseconds. Event timestamps consist of both logical and physical clock components, and a causal consistency model is used to handle out-of-order events. Scheduled tasks utilize priority queues to ensure timely processing of highly sensitive trajectory points.

[0084] Fault tolerance is achieved through redundant computation and heartbeat detection. Critical computational tasks are executed in parallel on heterogeneous processors, and a voting mechanism is used to determine the final result. The heartbeat detection interval is set to 100 milliseconds, and a failover process is triggered upon timeout. Persistent storage employs erasure coding technology, ensuring that data integrity is not affected by a single disk failure. In the event of a power outage, supercapacitors maintain critical operations.

[0085] Example 4: See Figure 7 The selection process for the key stiffness feature subset employs an improved fuzzy C-means clustering algorithm, processing all feature vector samples within a 0.35-meter radius around the current trajectory point. During system initialization, 3-5 cluster centers are preset, and the optimal number of clusters is determined through iterative optimization. Each iteration includes two stages: membership update and center point recalculation. The convergence condition is set to a membership change of less than 0.01 between two consecutive iterations. Dynamic weight allocation is used for each dimension of the feature vectors during clustering, with weight coefficients automatically adjusted based on the feature variance. Table 1 shows the distribution of three feature clusters in a specific clustering process.

[0086] Table 1: Distribution of three feature clusters in a certain actual clustering process.

[0087]

[0088] The Long Short-Term Memory (LSTM) network employs a three-layer architecture. The input layer processes standardized data from a subset of key features. The standardization process uses a moving window statistic, with the window size set to the most recent 200 samples. The first hidden layer contains 32 gated recurrent units, with the weight matrices of the input, forget, and output gates of each unit independently initialized. The second hidden layer, with 16 units, uses a bidirectional structure, concatenating the forward and backward processing flows at the output. The third hidden layer, with 8 units, introduces residual connections, with the weight coefficients of the skip connection paths fixed at 0.3. The tendon tension signal differential values ​​are injected into the network through an auxiliary input port and cross-fused with the main feature flow in the second hidden layer.

[0089] The network training employs a phased strategy, with the initial phase using historical datasets for pre-training. Training data augmentation includes random time shifts and feature value perturbations, with the shift range controlled within ±10ms and the perturbation amplitude not exceeding 20% ​​of the feature standard deviation. The online learning phase utilizes mini-batch gradient descent, with the batch size dynamically adjusted based on system load, ranging from 16 to 64. The gradient clipping threshold is set to 1.0, and the optimizer is selected as stochastic gradient descent with Nesterov momentum. The initial learning rate is 0.001, decaying by 5% after every 1000 updates.

[0090] The stiffness prediction process incorporates Monte Carlo dropout uncertainty estimation. During network inference, 50 forward propagations are activated, with 20% of neuron connections randomly dropped each time. The prediction results are statistically analyzed, including the mean and standard deviation of stiffness values ​​over five time steps, with the time step interval aligned with the sampling period of the control system. The output layer employs a hybrid density network structure, with each time step output containing three parameters from a Gaussian distribution: mean, variance, and weight coefficients. The final predicted value is obtained through a weighted summation, with the weight coefficients automatically learned by the network.

[0091] Real-time optimization measures include computation graph optimization and memory access optimization. Network parameters are stored in contiguous memory blocks and preloaded via memory prefetching instructions. Matrix multiplication operations employ block partitioning, with block sizes matching the CPU cache line length. Activation function calculations use an approximate implementation to reduce instruction cycles while maintaining accuracy. Parallel computing tasks are accelerated using OpenMP multithreading, with threads bound to specific CPU cores to avoid scheduling overhead.

[0092] The anomaly detection module monitors the distribution shift of the network input. A sliding window is used to calculate the Mahalanobis distance of the feature vectors, with the window size consistent with the clustering analysis range. When a sustained shift is detected, local fine-tuning of the network parameters is automatically triggered. The fine-tuning process is limited to the last two layers of the network, with the learning rate reduced to one-tenth of its normal value. In cases of severe anomalies, a safety mode is activated, replacing the main network with a pre-trained simplified model.

[0093] The data persistence solution employs a differential backup strategy. Complete network parameters are saved hourly, and incremental changes are recorded every five minutes. The storage format uses binary protocol buffer encoding, achieving a compression rate of approximately 35% of the original data. CRC checks are performed before backup data is written to non-volatile memory, and error detection automatically triggers a retry mechanism. The version control system maintains a complete history of parameter updates, supporting model rollback to any point in time.

[0094] The hardware accelerator integration solution incorporates FPGA and GPU collaborative processing. The FPGA handles feature subset preprocessing and clustering calculations, achieving deterministic, low-latency processing. The GPU accelerates network inference by performing batch matrix operations using tensor cores. Data transmission is handled via PCIeGen4 channels, employing zero-copy technology to reduce memory duplication. The power management unit dynamically adjusts the accelerator voltage and frequency, optimizing energy efficiency while meeting real-time requirements.

[0095] The system monitoring interface displays the feature clustering distribution and network prediction results in real time. Visualization components include a 3D feature space scatter plot and a time series curve of the predicted values. User interaction supports manual adjustment of clustering parameters and network hyperparameters; modifications take effect immediately but require explicit confirmation to be persistent. The historical data playback function allows for re-analysis of any time period and supports comparison of the effects of different parameter settings.

[0096] Configuration management employs a layered architecture, with system-level parameters and algorithm-level parameters stored separately. Network-related parameters, including the number of layers, number of cells, and dropout rate, are stored in dedicated configuration files. Clustering algorithm parameters, such as the fuzziness index and convergence threshold, are grouped into the feature processing configuration group. All configuration changes are logged in detail, including the modification time, operator, and reason. Configuration rollback operations automatically generate a difference report, listing all affected functional modules.

[0097] The fault-tolerance mechanism is designed to handle various possible anomalies. When input data is missing, interpolation is performed to complete it based on nearest neighbor features. If clustering fails to converge, a feature subset is generated using default parameters. When network prediction times out, the last valid result is returned and marked as the estimate. Hardware fault detection is achieved through a heartbeat signal; faulty components are automatically isolated, and backup equipment is notified to take over. In the event of a power failure, supercapacitors ensure that critical memory data is preserved.

[0098] Time synchronization ensures the timing accuracy of feature processing and prediction output. Network input data is appended with timestamps accurate to microseconds, and the output prediction value inherits the input time reference. Distributed clock synchronization employs a hybrid PTP / NTP protocol, controlling master-slave clock skew within 100 microseconds. Time skew monitoring is performed in real-time at each stage of the processing pipeline, triggering a data realignment process if a threshold is exceeded.

[0099] Example 5: The generation process of joint dynamic stiffness parameters adopts a hybrid calculation method based on a physical model. The system extracts the predicted value with the highest confidence from the joint stiffness change sequence. This predicted value, after time alignment processing, is used as the initial condition input into the dynamic equation. The joint angle differential calculation adopts the fifth-order central difference method, and the sampling interval is strictly matched with the control system cycle. Before the differential operation, the original angle data is filtered by moving average. The window width is adaptively adjusted according to the current motion speed. A narrower window is used during fast motion to maintain the response speed, and a wider window is used during slow motion to improve the signal-to-noise ratio.

[0100] The stiffness dynamic equations are constructed in three main parts: inertia, damping, and stiffness. The inertia coefficients are obtained through offline system identification, considering the range of rotational inertia variations under joint loads. The damping term employs a nonlinear model, including first and second-order velocity terms, with coefficients determined through frequency domain analysis. The stiffness term is modeled as a continuous piecewise function of displacement, with inflection points pre-defined based on material properties. The equations are solved using the variable-step Runge-Kutta method, with the initial step size set to one-tenth of the control period. Error control is achieved through local truncation error estimation.

[0101] The extraction process of instantaneous stiffness coefficients involves two dimensions: frequency domain analysis and time domain analysis. The time series output by the solver undergoes a short-time Fourier transform to identify the amplitude at the dominant frequency component as the basic stiffness index. Time domain analysis calculates the local slope of the displacement-force curve and uses the least squares method to fit the linear region near the current operating point. Finally, the instantaneous stiffness coefficient is a weighted average of the frequency domain amplitude and the time domain slope, with the weighting coefficients dynamically adjusted according to signal quality indicators.

[0102] The rate of change of stiffness is calculated using phase-locked amplification (PLA). A narrowband filter is placed near the dominant frequency to extract in-phase and quadrature components. An analytic signal is constructed using Hilbert transform, and the instantaneous phase derivative is calculated as the basic estimate of the rate of change. The auxiliary channel uses direct numerical differentiation, and the result is cross-validated with the phase derivative estimate after Butterworth low-pass filtering. The effective rate of change is the smaller of the two estimates, and the sign is determined by trend consistency detection.

[0103] The torque correction is calculated using a dual-threshold hysteresis control strategy. The safety threshold range is dynamically updated based on the fatigue characteristics of the joint material and the current load. The upper threshold is set to 80% of the instantaneous strength limit, and the lower threshold is 120% of the minimum stiffness required for static equilibrium. The hysteresis interval width is proportional to the rate of change of stiffness; the interval is appropriately widened to avoid frequent switching during rapid changes. Corrections exceeding the upper threshold use a square root function to handle overshoot, while corrections below the lower threshold use a cubic function to amplify the deviation signal.

[0104] The PID controller achieves a smooth transition in torque correction. The proportional term coefficient is inversely proportional to the current motion speed, reducing the proportional action to avoid overshoot at high speeds. The integral term employs an anti-saturation design, limiting cumulative error to a preset range. The derivative term acts on the rate of change of stiffness rather than the error signal, enhancing the predictability of trend changes. The controller output passes through a rate limiting module, with the maximum rate of change set according to the actuator characteristics. Finally, the torque command is superimposed on the basic control quantity to form a complete joint drive signal.

[0105] The real-time monitoring system continuously tracks the changing trends of dynamic parameters. Instantaneous stiffness coefficients and rates of change are evaluated for stability using a moving window statistical analysis method. An early warning signal is triggered when the standard deviation of the data within the window exceeds a threshold, indicating a possible system anomaly. A historical data buffer records complete parameter changes for the most recent 30 seconds, supporting backtracking analysis of the causes of anomalies. The monitoring interface displays parameter values ​​and trend curves in real time, highlighting warnings when key indicators exceed safe limits.

[0106] The fault recovery mechanism includes a multi-level response strategy. For transient stiffness anomalies, it first attempts to recalculate data from the most recent three periods. In cases of persistent anomalies, it switches to a backup stiffness estimation algorithm based on simplified linear assumptions. In the event of a severe fault, a safety protection procedure is activated, gradually reducing the joint stiffness to a preset safe value. All recovery operations are logged in detail, including anomaly characteristics, handling measures, and result verification.

[0107] The parameter adaptation module optimizes the calculation process based on operational data. Dynamic equation coefficients undergo offline re-identification monthly, using accumulated operational data as a training set. Parameters such as the filter window width and PID coefficients are calculated in real time and adjusted weekly using an optimization algorithm. The adaptation process employs a conservative strategy, with each adjustment not exceeding 10% of the current value. The adjustment results are only applied to actual control after simulation verification.

[0108] In terms of hardware implementation, the core algorithms are deployed on a real-time computing unit. This unit adopts a multi-core processor architecture, with critical tasks bound to dedicated cores. Memory access is managed through static allocation, reducing the uncertainty caused by dynamic allocation. Computational acceleration uses SIMD instructions to handle vector operations, and the floating-point unit is configured to round to the nearest integer. Timed interrupt services ensure strict timeliness of the control cycle, with jitter controlled at the microsecond level.

[0109] The time synchronization system maintains timing consistency across all processing stages. Sensor data is appended with high-precision timestamps, achieving an accuracy of hundreds of nanoseconds. The processing pipeline implements latency compensation, recording the start and end times of each process step. Network transmission employs a time-sensitive network protocol, prioritizing the scheduling of critical messages. The global clock is synchronized via a combination of satellite timing and atomic clocks, with leap seconds handled using a smooth transition algorithm.

[0110] The configuration management system maintains version control for all adjustable parameters. Parameter modifications are implemented through a transaction mechanism, ensuring atomic updates of a set of related parameters. A rationality check is performed before each modification to prevent obviously contradictory parameter combinations. Historical version archiving supports rapid rollback, preserving a snapshot of the current running state during the rollback process. The parameter export format uses standardized encoding, compatible with various analysis tools.

[0111] The safety protection design covers three levels: electrical, mechanical, and software. The electrical system monitors for abnormal current and voltage, and immediately cuts off the power supply in case of overload. Mechanical limit switches provide hardware-level position protection and operate independently of the control system. The software watchdog periodically checks the status of critical processes, and triggers a system reset if there is no response within a timeout period. All safety mechanisms perform self-checks periodically, and system startup is prohibited if a self-check fails.

[0112] The user interface provides parameter visualization and interactive functions. Real-time data display includes instantaneous stiffness coefficient curves, rate of change histograms, and torque correction trend graphs. The parameter adjustment interface includes sliders and numerical input boxes, supporting fine-tuning. The history playback function can load data for any time period, and the playback speed is adjustable. A complete audit log is recorded for all interface operations, including operation content and execution results.

[0113] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0114] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A biomimetic adaptive variable stiffness robotic arm joint control method, characterized in that, include: Collect multi-source biosignal data during the operation of the robotic arm joints, including electromyographic signals, tendon tension signals, and joint angle signals; A joint stiffness feature vector is constructed based on the electromyographic signal and tendon tension signal, and a joint motion trajectory sequence is generated based on the joint angle signal. For each joint motion trajectory point in the joint motion trajectory sequence, extract the stiffness feature vector corresponding to that trajectory point and calculate the stiffness sensitivity of the current trajectory point; A biomimetic stiffness reference model is constructed, which includes a mapping relationship between stiffness feature vectors and joint angle signals, and the mapping relationship is dynamically weighted and adjusted using the stiffness sensitivity. Based on the adjusted biomimetic stiffness reference model and real-time acquired signals, the joint stiffness change sequence is predicted. By fusing the joint stiffness change sequence with the current joint angle signal, dynamic stiffness parameters of the joint are generated. The joint drive torque correction amount is calculated based on the joint dynamic stiffness parameters and the preset stiffness safety threshold range. The weighted parameters of the biomimetic stiffness reference model are iteratively updated until the joint driving torque correction satisfies the preset convergence condition. The final joint drive torque command is output to the robotic arm joint actuator. The calculation of the stiffness sensitivity of the current trajectory point includes: calculating the stiffness sensitivity of the current trajectory point by combining the historical stiffness feature vector distribution of adjacent trajectory points; The calculation of the stiffness sensitivity of the current trajectory point by combining the historical stiffness feature vector distribution of adjacent trajectory points includes: obtaining the stiffness feature vectors of the five consecutive trajectory points preceding the current trajectory point; calculating the distribution dispersion of the stiffness feature vectors of the five consecutive trajectory points in four-dimensional space; performing negative correlation normalization on the distribution dispersion to obtain the stiffness feature stability coefficient of the current trajectory point; calculating the physical response intensity of the current trajectory point based on the absolute value of the gradient of the tendon tension signal within the corresponding time window; and using the product of the stiffness feature stability coefficient and the physical response intensity as the stiffness sensitivity of the current trajectory point.

2. The biomimetic adaptive variable stiffness robotic arm joint control method according to claim 1, characterized in that, The construction of the joint stiffness feature vector includes: performing bandpass filtering on the electromyographic signal and dividing it into equal time windows; extracting the energy integral value, median frequency, and zero-crossing rate of the electromyographic signal in each time window, and simultaneously acquiring the peak value of the tendon tension signal in the corresponding time window; arranging the energy integral value, median frequency, zero-crossing rate, and peak value of the tendon tension signal in the order of the time windows to form a four-dimensional stiffness feature vector sequence.

3. The biomimetic adaptive variable stiffness robotic arm joint control method according to claim 1, characterized in that, The predicted joint stiffness change sequence includes: performing cluster analysis on the stiffness feature vectors of the target trajectory points based on the adjusted biomimetic stiffness reference model to select a key stiffness feature subset; and predicting the joint stiffness change sequence through a long short-term memory network based on the key stiffness feature subset and the real-time acquired tendon tension signal.

4. The biomimetic adaptive variable stiffness robotic arm joint control method according to claim 1, characterized in that, The dynamic weighted adjustment includes: extracting all historical mapping relationship records of the current trajectory point in the biomimetic stiffness reference model; arranging the stiffness sensitivity in each historical mapping relationship record in descending order to form a sensitivity sequence; calculating the proportion of the difference between the maximum and minimum values ​​in the sensitivity sequence to the total range of the sequence, as the mapping relationship confidence weight; and using the mapping relationship confidence weight to scale the newly added mapping relationship values ​​of the current trajectory point and update the mapping relationship database of the biomimetic stiffness reference model.

5. The biomimetic adaptive variable stiffness robotic arm joint control method according to claim 3, characterized in that, The process of selecting the key stiffness feature subset includes: obtaining all stiffness feature vector samples at the target trajectory point; dividing all stiffness feature vector samples into multiple feature clusters using a fuzzy clustering algorithm; calculating the intra-class average stiffness sensitivity of each feature cluster; selecting feature clusters whose intra-class average stiffness sensitivity is greater than a preset sensitivity threshold; and merging all stiffness feature vectors of the selected feature clusters to form the key stiffness feature subset.

6. The biomimetic adaptive variable stiffness robotic arm joint control method according to claim 3, characterized in that, The predicted joint stiffness change sequence includes: inputting a subset of key stiffness features into the input layer of a pre-trained long short-term memory network; fusing the tendon tension signal difference values ​​of three historical time steps in the hidden layer of the long short-term memory network; extracting the temporal dependence features of joint stiffness step by step through three hidden units; and outputting the predicted joint stiffness values ​​and their confidence levels for the next five time steps to form a joint stiffness change sequence.

7. The biomimetic adaptive variable stiffness robotic arm joint control method according to claim 1, characterized in that, The generation of joint dynamic stiffness parameters includes: extracting the stiffness prediction value corresponding to the highest confidence from the joint stiffness change sequence; obtaining the differential value of the current joint angle signal; inputting the stiffness prediction value and the differential value of the joint angle signal into a preset stiffness dynamic equation; solving the stiffness dynamic equation to obtain the instantaneous stiffness coefficient of the joint and its rate of change, and merging them into joint dynamic stiffness parameters.

8. The biomimetic adaptive variable stiffness robotic arm joint control method according to claim 1, characterized in that, The calculation of the joint drive torque correction includes: comparing the instantaneous stiffness coefficient in the joint dynamic stiffness parameters with the upper boundary value of the stiffness safety threshold range; when the instantaneous stiffness coefficient exceeds the upper boundary value, calculating the percentage value of the excess portion; determining the amount of downward adjustment of the joint drive torque based on the product of the percentage value and the stiffness change rate; comparing the instantaneous stiffness coefficient with the lower boundary value of the stiffness safety threshold range; when the instantaneous stiffness coefficient is lower than the lower boundary value, calculating the percentage value of the insufficient portion; and determining the amount of upward adjustment of the joint drive torque based on the product of the percentage value and the stiffness change rate.

Images