Machine learning based resistance servo adjustment method and system

Abstract

The application relates to a machine learning-based resistance servo adjustment method and system, which comprises the following steps: acquiring multi-source signals and motion state data collected by a target device, and performing feature fusion to obtain motion feature data; acquiring historical motion performance data of the target device, and performing intention recognition on the motion feature data to obtain a target resistance value and an associated confidence parameter; performing feedforward compensation control and closed-loop error correction on the target resistance value and the associated confidence parameter according to a preset dynamic resistance prediction model to obtain a resistance output instruction. The application can fully consider motion features, real-time states and safety boundaries, so that resistance adjustment is more accurate.

Description

Technical Field

[0001] This invention relates to the technical field of resistance servo adjustment, and in particular to a resistance servo adjustment method and system based on machine learning. Background Technology

[0002] The main limitation of existing resistance servo adjustment technology lies in its weak multi-source signal fusion capability, making it difficult to comprehensively capture the user's personalized motion characteristics. The intent recognition process relies on static historical data and cannot dynamically adapt to changes in the user's real-time motion state, resulting in insufficient matching accuracy of the resistance output. Furthermore, existing control strategies typically employ a single closed-loop feedback mechanism, lacking feedforward compensation and dynamic boundary adjustment capabilities. This makes the system slow to respond to rapidly changing motion demands, and may even exceed safe limits, affecting the safety and effectiveness of training. Summary of the Invention

[0003] The main objective of this invention is to provide a resistance servo adjustment method and system based on machine learning, which can dynamically adapt to the user's real-time state changes and solve the problem of lag in intention recognition in traditional methods.

[0004] To achieve the above objectives, the present invention provides a resistance servo adjustment method based on machine learning, comprising: Acquire multi-source signals and motion state data collected by the target device, and perform feature fusion to obtain motion feature data; Acquire historical motion performance data of the target device, perform intent recognition with the motion feature data, and obtain the target resistance value and associated confidence parameters; Based on the preset dynamic resistance prediction model, feedforward compensation control and closed-loop error correction are performed on the target resistance value and the associated confidence parameter to obtain the resistance output command.

[0005] Furthermore, the step of acquiring multi-source signals and motion state data collected by the target device and performing feature fusion to obtain motion feature data includes: The multi-source signal is segmented and transformed to obtain a hybrid feature vector; Based on the motion state data, motion coupling calculation is performed on the hybrid feature vector to obtain the enhanced feature matrix; The mechanical parameters of the target device are obtained, and the enhanced feature matrix is ​​subjected to torque conversion to obtain dynamic feature data; The dynamic feature data is fused, matched, and recombined according to a preset weight allocation strategy to obtain the motion feature data.

[0006] Further, the step of acquiring historical motion performance data of the target device and performing intent recognition with the motion feature data to obtain the target resistance value and associated confidence parameters includes: The historical motion performance data is periodically segmented and identified to obtain historical force value feature segments; Based on preset action type labels, the historical force value feature segments are matched using conventional patterns to obtain conventional motion pattern data. The motion feature data is dynamically normalized and matched with the conventional motion pattern data to obtain the pattern matching similarity. The pattern matching similarity and the motion feature data are used to perform multi-mode drag prediction to obtain the initial drag prediction value; The initial resistance prediction value is compensated based on the motion characteristic data to obtain the target resistance value; The association confidence parameter is obtained by performing an association confidence calculation based on the pattern matching similarity and the target resistance value.

[0007] Further, the step of performing feedforward compensation control and closed-loop error correction on the target resistance value and the associated confidence parameter based on a preset dynamic resistance prediction model to obtain a resistance output command includes: The target resistance value and the associated confidence parameter are input into the dynamic resistance prediction model, and feature segmentation and resistance feature coupling are performed to obtain coupled resistance features; The dynamic feedforward compensation amount is obtained by performing nonlinear activation mapping on the coupled resistance characteristics through the feedforward compensation layer of the dynamic resistance prediction model. The target resistance value is predicted by the closed-loop correction layer of the dynamic resistance prediction model to obtain the closed-loop error correction amount. The composite resistance command is obtained by constructing commands and compensating dead zones for the dynamic feedforward compensation and the closed-loop error correction through the command output layer of the dynamic resistance prediction model. Based on the motion characteristic data, the composite resistance command is dynamically boundary-checked and incrementally updated to obtain the resistance output command.

[0008] Further, the step of performing nonlinear activation mapping on the coupled resistance characteristics through the feedforward compensation layer of the dynamic resistance prediction model to obtain the dynamic feedforward compensation amount includes: The feedforward compensation layer performs feature extraction and vector compensation on the coupling resistance features to obtain a compensation feature vector; The compensation feature vector is subjected to piecewise linear mapping calculation to obtain preliminary compensation values; The amplitude range of the initial compensation value is adjusted based on the correlation confidence parameter to obtain the adjusted compensation value. The deviation offset correction is applied to the adjusted compensation value to obtain the dynamic feedforward compensation amount.

[0009] Furthermore, the command output layer of the dynamic resistance prediction model constructs commands and performs dead-zone compensation on the dynamic feedforward compensation and the closed-loop error correction to obtain the composite resistance command: The initial composite resistance command is obtained by proportionally superimposing the dynamic feedforward compensation amount and the closed-loop error correction amount and constructing the command output layer. The initial composite resistance command is adjusted for command stability to obtain the adjusted resistance parameters; Nonlinear dead zone identification is performed on the adjusted resistance parameters to obtain the dead zone compensation range; Based on the dead zone compensation range and the adjusted resistance parameters, the initial composite resistance command is segmented and compensated to obtain a compensated resistance command. The composite resistance command is obtained by constraining the command boundary and coordinating command synchronization with the compensation resistance command. Based on the motion characteristic data, the composite resistance command is dynamically boundary-checked and incrementally updated to obtain the resistance output command.

[0010] Furthermore, the step of performing dynamic boundary verification and incremental parameter updates on the composite resistance command based on the motion characteristic data to obtain the resistance output command includes: Based on the motion characteristic data, the composite resistance command is subjected to safety boundary verification and mechanical protection limiting to obtain the mechanical limiting value; Incremental parameter decomposition is performed on the mechanical limit value to obtain torque adjustment component and power adjustment component; Based on the historical motion performance data, the torque adjustment component is dynamically range-matched to obtain the optimized torque component. The power adjustment component is corrected by an attenuation factor to obtain the fatigue adaptive component; The optimized torque component and the fatigue adaptive component are incrementally updated with parameters and commands are constructed to obtain the resistance output command.

[0011] Further, the step of performing safety boundary verification and mechanical protection limiting on the composite resistance command based on the motion characteristic data to obtain a mechanical limit value includes: Based on the motion characteristic data, the composite resistance command is segmented by angular velocity threshold to obtain the angular velocity safety boundary; The composite resistance command is gradient-limited based on preset mechanical rigidity data to obtain the command protection threshold. The safety boundary of the angular velocity and the command protection threshold are checked to obtain the primary mechanical constraint value; Based on the motion characteristic data, the primary mechanical constraint value is corrected for mechanical fatigue accumulation to obtain the secondary mechanical constraint value; The secondary mechanical constraint value is mechanically protected and limited to obtain the mechanical limit value.

[0012] The present invention also provides a machine learning-based drag servo control system, applied to any one of the machine learning-based drag servo control methods described above, comprising: The identification module is used to acquire multi-source signals and motion state data collected by the target device, and perform feature fusion to obtain motion feature data; The parsing module is used to acquire the historical motion performance data of the target device, perform intent recognition with the motion feature data, and obtain the target resistance value and associated confidence parameters. The processing module is used to perform feedforward compensation control and closed-loop error correction on the target resistance value and the associated confidence parameter according to a preset dynamic resistance prediction model, so as to obtain a resistance output command.

[0013] The present invention provides a resistance servo adjustment method and system based on machine learning, which has the following beneficial effects: By integrating multi-source signals and motion state data collected from the target device, the system can comprehensively capture the user's personalized motion characteristics, overcoming the incomplete feature extraction problem caused by a single data source in traditional methods, and significantly improving the adaptability of resistance adjustment. Combining historical motion performance data with real-time motion feature data for intent recognition and outputting associated confidence parameters allows resistance value prediction to not only be based on static patterns but also dynamically adapt to changes in the user's real-time state, effectively solving the problem of lagging intent recognition in traditional methods. Dynamic boundary verification and incremental parameter updates for composite resistance commands based on motion feature data ensure that resistance output can adapt to the user's real-time needs while automatically adjusting to a safe range, avoiding overtraining or insufficient protection problems caused by traditional fixed-boundary methods. The generated resistance output commands fully consider the user's motion characteristics, real-time state, and safety boundaries, making resistance adjustment more precise and flexible, thereby improving the effectiveness of the training or rehabilitation process and the user experience. Attached Figure Description

[0014] Figure 1 This is a flowchart of a resistance servo adjustment method based on machine learning provided for this invention; Figure 2 This is a structural diagram of a resistance servo adjustment system based on machine learning provided for this invention.

[0015] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0017] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments.

[0018] Reference Figure 1 As shown, the present invention provides a resistance servo adjustment method based on machine learning, comprising: Step S101: Acquire multi-source signals and motion state data collected by the target device, and perform feature fusion to obtain motion feature data; Step S201: Obtain historical motion performance data of the target device, perform intent recognition with motion feature data, and obtain the target resistance value and associated confidence parameters; Step S301: Based on the preset dynamic resistance prediction model, feedforward compensation control and closed-loop error correction are performed on the target resistance value and associated confidence parameters to obtain the resistance output command.

[0019] Based on the steps described above, the detailed process is as follows: Step S101: The target device integrates a multi-source biosensor array that continuously collects electromyography (EMG) signals, joint angle data, and three-dimensional acceleration information. The EMG sensors capture muscle activation potentials at a preset sampling frequency, while the inertial measurement unit tracks the limb's spatial trajectory in real time. Motion state data is recorded by an optical encoder, showing the displacement of the device's motion mechanism. The feature fusion process performs temporal alignment and noise filtering on the raw signals. The EMG signals are rectified using full-wave rectification, and the root mean square value is calculated to extract muscle activation intensity features. The joint angle sequence is processed using cubic spline interpolation to generate a continuous motion trajectory function, from which joint angular acceleration is derived to quantify motion stability. Force sensor data is combined with the user's body mass parameters for biomechanical standardization to eliminate the influence of individual physiological differences.

[0020] All single-modal features are input into the feature correlation matrix, and covariance coupling operations are performed to reveal the dynamic correlation between muscle exertion and joint movement. Principal component analysis is then used to reduce the dimensionality of the high-dimensional fused features. The output low-dimensional core feature set is then Gaussian normalized and scaled to generate a standardized biomechanical feature vector, which serves as motion feature data characterizing the user's movement properties. This data stream is updated 50 times per second, providing a stable input source for subsequent processing.

[0021] Step S201: The historical performance database stores the user's peak muscle strength, power output curves, and fatigue recovery parameters from past training cycles. The intent recognition module first performs dynamic time warping on the historical force curves, then clusters them based on action type labels to generate a typical motion pattern template library. The motion trajectory coordinates in the real-time motion feature data are compared with the template library for similarity calculation, outputting a pattern matching similarity score. Simultaneously, the real-time muscle fatigue coefficient is calculated based on the median frequency decay rate of the electromyographic signal and the range of motion decay of the joints. The intent prediction network receives both pattern matching similarity and fatigue coefficient inputs, and outputs an initial resistance prediction value through a fully connected neural network.

[0022] The predicted value is corrected for response delay by the servo system to generate a time-domain calibrated resistance value. Sensor confidence-weighted fusion is then applied to the calibrated value: the electromyography sensor has a weight of 0.5, the inertial measurement unit has a weight of 0.3, and the force sensor has a weight of 0.2. The weighted average is calculated to obtain the target resistance value, and 1 minus the weighted error rate is used as the associated confidence parameter. These dual output parameters constitute the decision-making basis of the dynamic control system.

[0023] Step S301: The dynamic resistance prediction model's feedforward compensation layer employs a dual-path processing: when the confidence parameter is greater than 0.8, superlinear gain scaling is used to scale the static component of the target resistance value; when the confidence parameter is less than 0.5, sublinear suppression is enabled. Phase lead compensation is applied to the transient component of the target resistance value to correct for actuator response lag. The two components are nonlinearly superimposed using a hyperbolic tangent function to generate the initial feedforward quantity.

[0024] The closed-loop correction layer extracts the real-time error signal of the motor speed and performs noise robust filtering through an improved sliding mode differentiator to generate the disturbance rejection speed error. The model inertia compensation module combines the torsional stiffness parameters of the transmission chain and the rotor acceleration data to calculate the dynamic inertia compensation based on a two-mass spring model. The aforementioned disturbance rejection speed error and inertia compensation are input into a multi-modal PID controller, which dynamically allocates the weights of the proportional, integral, and derivative terms according to the error interval to generate a feedback correction component. Finally, the feedforward and feedback quantities are decoupled and recombined using a singular value decomposition fusion algorithm to output a composite resistance command.

[0025] The dynamic boundary verification module first analyzes the joint torque-angle relationship curve in the motion feature data to generate a dynamic physiological safety boundary that changes in real time with body position. When the rate of change of the composite resistance command exceeds a threshold, an emergency protection zone is triggered, and mechanical constraints are applied based on the equipment's maximum load parameters. A boundary conflict detector compares the physiological safety boundary with the emergency protection zone and uses least-squares optimization to find the conflict correction solution. The verified resistance value is then input into the motion optimizer, which adjusts the trajectory continuity based on the smoothness coefficient.

[0026] The incremental update engine decomposes the verification value into torque and power components: the former is matched with the user's historical best power curve to generate optimized torque; the latter is combined with the fatigue coefficient to correct the attenuation factor. The two are fused through a feature weight matrix to generate an incremental adjustment parameter set. Finally, a time delay compensation observer is used to offset the mechanism's response lag, synchronizing the output command with the user's actions to generate a resistance output command to drive the execution unit.

[0027] The present invention provides a resistance servo adjustment method and system based on machine learning, which has the following beneficial effects: By integrating multi-source signals and motion state data collected from the target device, the system can comprehensively capture the user's personalized motion characteristics, overcoming the incomplete feature extraction problem caused by a single data source in traditional methods, and significantly improving the adaptability of resistance adjustment. Combining historical motion performance data with real-time motion feature data for intent recognition and outputting associated confidence parameters allows resistance value prediction to not only be based on static patterns but also dynamically adapt to changes in the user's real-time state, effectively solving the problem of lagging intent recognition in traditional methods. Dynamic boundary verification and incremental parameter updates for composite resistance commands based on motion feature data ensure that resistance output can adapt to the user's real-time needs while automatically adjusting to a safe range, avoiding overtraining or insufficient protection problems caused by traditional fixed-boundary methods. The generated resistance output commands fully consider the user's motion characteristics, real-time state, and safety boundaries, making resistance adjustment more precise and flexible, thereby improving the effectiveness of the training or rehabilitation process and the user experience.

[0028] In one embodiment, multi-source signals and motion state data collected by the target device are acquired, and feature fusion is performed to obtain motion feature data, including: Multi-source signals, including electromyography (EMG) signals, joint angular velocity data, and three-dimensional acceleration information, are simultaneously captured at a preset sampling frequency. The EMG sensor array is attached to the target muscle movement point, and the raw potential signals are acquired using a bipolar electrode configuration. After high-pass filtering to eliminate motion artifacts, the signals are processed by an instrumentation amplifier with a gain of up to a thousand times.

[0029] Segmented analysis, based on motion cycle event triggers, divides a continuous signal stream into independent motion cycle segments by real-time identification of heel switch signals or dead-point positions of the device's motion mechanism. The transformation calculation process performs full-wave rectification on the electromyographic signals and calculates the root mean square value within the time window, converting it into a muscle activation intensity feature sequence. Joint angle data is converted to Euler angles through quaternion calculation, and angular acceleration in the flexion-extension plane is extracted as a quantitative indicator of motion stability. Raw accelerometer data is transformed to a local limb coordinate system, and vector amplitudes are calculated to obtain motion intensity characteristics.

[0030] Each single-modal feature set constitutes an initial feature vector group. The maximum mutual information coefficient method is used to test the correlation between features, and redundant features with correlation exceeding a threshold are removed to form an intermediate feature set. This feature set undergoes frequency domain energy redistribution using a wavelet packet decomposition and reconstruction algorithm, extracting the signal energy proportion within a specific frequency band as a supplement to the frequency domain features. Finally, the time-domain features and frequency-domain features are concatenated and combined, mapped to a unified numerical range by a feature scaler, generating a hybrid feature vector representing the current motion state. The dimension of this vector remains within a preset feature number range to ensure computational efficiency in subsequent processing.

[0031] The motion monitoring system captures the displacement of moving parts of the device using a high-precision optical encoder, and combines this with the three-dimensional angular velocity data output from the gyroscope to calculate the composite vector of linear velocity and angular velocity in the motion plane. This composite vector is used to construct a limb motion dynamics model and solve for the Coriolis acceleration components and centrifugal force effect.

[0032] After receiving the mixed feature vector input, the motion coupling computation establishes a biomechanical correlation mapping model: partial correlation analysis is performed on muscle activation intensity features and corresponding joint kinematic parameters to calculate the net correlation coefficient after eliminating interference from other variables; the temporal feature vector is extended to the three-dimensional motion space through tensor product operations, and the effective component ratio of muscle contractile force in the motion direction is solved by integration along the motion trajectory path. Dynamic synchronous coupling achieves data fusion by constructing the constraint relationship between feature variables and motion equations and solving for the optimal solution of the objective function.

[0033] The enhancement process introduces a motion stability compensation coefficient, calculated from the variance derivative of joint angular acceleration, to adaptively weight and suppress highly volatile data in the feature vector. The spatial enhancement module performs feature projection transformation in a six-dimensional motion space (3D displacement + 3D rotation) based on the real-time calculated limb pose matrix, generating a high-dimensional feature tensor containing orthogonal motion components. Finally, a principal direction feature extraction algorithm compresses the data dimensionality, retaining core components whose cumulative contribution rate exceeds a preset threshold, forming an enhanced feature matrix where row vectors correspond to time series and column vectors represent feature attributes. The number of rows in this matrix matches the number of sampling points, and the number of columns is limited to the preset maximum number of feature channels, ensuring computational feasibility for subsequent processing.

[0034] The target equipment's inherent mechanical parameter library stores core physical characteristics such as the transmission system's reduction ratio, the actuator's rotor inertia, and the transmission chain's torsional stiffness coefficient. The mechanical parameter retrieval module indexes a subset of parameters based on the current motion mode, where the reduction ratio range covers the entire operating range of the equipment, and the rotor inertia parameter is dynamically compensated based on temperature sensor readings. Torque conversion calculations establish a two-mass spring dynamic equation, decoupling displacement-related and velocity-related features in the enhanced feature matrix to generate two independent state vectors. These vectors are input to the state-space converter, which, combined with the reduction ratio parameter, calculates the equivalent displacement at the load end. The dynamic converter then solves the system's differential algebraic equations of motion, where the biomechanical feature terms in the feature matrix are imported as external excitation functions, and the transmission chain stiffness coefficient is used as a coupling term gain parameter.

[0035] The numerical solution of the differential equation employs the fourth-order Runge-Kutta method, iteratively outputting the instantaneous solution of the load torque along the time series with a fixed step size. After rotational coordinate transformation, the solution results are projected along the axial direction of the actuator to generate the main driving torque component. During the output dimension alignment stage, linear interpolation is used to synchronize the torque signal sampling rate with the original feature matrix, ultimately forming a three-dimensional dynamic feature data matrix with a dual correspondence between the time series and the feature dimensions. The third dimension contains three fundamental physical quantities: torque amplitude, phase, and rate of change.

[0036] The preset weight allocation strategy includes a feature priority coefficient table and a spatial mapping rule base. The priority coefficient table is indexed according to the type of movement: the weight of electromyography features is set to 0.6 in resistance training mode, and the weight of movement trajectory smoothness features is increased to 0.7 in isokinetic training mode. The fusion matching process adopts a hierarchical weight loading mechanism. The primary weight module parses the torque phase parameters of dynamic feature data, calculates the direction cosine value of each feature channel relative to the target motion plane, and generates feature direction adaptation factors.

[0037] The joint range of motion threshold is transformed into a spatial boundary condition, and Gaussian attenuation weighting is applied to feature data exceeding the boundary. The reconstruction phase initiates a dual-processing pipeline: the temporal reconstruction channel uses a dynamic time warping algorithm to align the phase shifts of different feature sequences, eliminating time differences caused by sensor position variations; the frequency domain reconstruction channel applies wavelet packet decomposition and reconstruction technology, performing frequency band energy normalization within preset feature sub-bands (0-4Hz basic force generation band, 4-8Hz coordination control band). The weighted fusion unit receives the time-frequency processed corrected feature matrix, performs channel weighted summation according to priority coefficients, and compresses the output dimension. The reconstruction verification module verifies the consistency of the physical dimensions of the output features, ensuring that torque unit features are not mixed with non-dimensional features. The generated motion feature data includes a timestamp, a feature value array, and a confidence level identifier—three attribute fields.

[0038] This embodiment extracts the coupling features between muscle activation intensity and joint motion stability by performing time-series alignment and frequency domain reconstruction on electromyographic signals, joint motion data, and three-dimensional acceleration information. This forms a high-information-density hybrid feature vector characterizing the biomechanical state, effectively eliminating the perception blind spot of traditional single-modal signal processing. Combining the reduction ratio and torsional stiffness parameters of the transmission system, a two-mass spring model is used to transform kinematic features into equivalent torque components at the load end, achieving precise conversion of physical dimensions from biometric features to driving commands and avoiding control deviations caused by mismatched dynamic parameters.

[0039] In one embodiment, historical motion performance data of the target device is acquired, and intent recognition is performed with motion feature data to obtain the target resistance value and associated confidence parameters, including: The target device's historical database stores historical motion performance data from past training cycles, including torque curves, power output spectra, and motion completion records. The periodic segmentation recognition engine calls a motion feature period detection algorithm to identify the start and end points of the motion cycle based on the zero-crossing events of the device's actuator angular displacement signal, and marks the boundary between the complete force exertion phase and the relaxation phase on the torque-time curve.

[0040] The feature segmentation module segments the continuous torque data stream based on marked points, generating independent action units containing the force rise segment, peak maintenance segment, and decay segment. After time-domain normalization to a unified time base, each unit extracts seven key mechanical features: maximum torque value, peak time, peak maintenance duration, decay rate, impulse value, average force rate, and power integral value. The feature aggregator maps these features into dimensionality-reduced feature vectors and uses principal component analysis to retain core components with a cumulative variance contribution rate exceeding a threshold, forming time-aligned standard historical force value feature segments.

[0041] The movement type tag library defines three basic movement templates: resistance training, isokinetic training, and explosive power training. Each template includes a standard force curve envelope and biomechanical constraints.

[0042] The conventional pattern matching processor performs third-order operations: calculates the Euclidean distance matrix between historical feature segments and each action template, corrects the nonlinear scaling effect of the time axis through a dynamic time warping algorithm; analyzes the phase-sensitive parameters of the feature segments, performs spectral clustering analysis on the slope of the rising edge of the force and the stability of the peak value; and finally establishes a state transition probability matrix to describe the transition characteristics between different force stages.

[0043] The pattern classifier output module assigns action category labels based on probability thresholds, clustering feature fragments into three preset action type clusters. For each action cluster, it performs feature distribution statistical analysis, generating a motion pattern data set including the mean torque envelope, power standard deviation range, and dynamic fluctuation coefficient, forming a standard reference system characterizing the user's typical force application characteristics.

[0044] The dynamic warping matcher receives a real-time generated motion feature data stream, whose feature vector includes fused parameters of electromyographic activation timing, joint angular velocity, and three-dimensional acceleration. The pattern alignment stage first performs spatiotemporal registration: using the force application start point of the conventional pattern data as a reference, a sliding window correlation analysis method is used to calculate the time offset compensation of the real-time data. The similarity measurement engine performs a dual-path evaluation: one path applies the dynamic time warping algorithm to calculate the minimum bending path cost of the feature trajectory and generates a trajectory similarity coefficient; the other path compares the co-occurrence probability of the real-time action state sequence with the conventional pattern state transition chain using transition probability analysis.

[0045] The matching result fusion module performs a weighted synthesis of the two outputs, with the trajectory similarity weight set as a dynamically calculated value that changes linearly with the stage of motion. A pattern matching similarity score within a given range is generated, which reflects in real time the degree of deviation between the current action and the historical best pattern, providing a quantitative basis for resistance prediction.

[0046] The multi-mode resistance prediction engine establishes a dual-channel fusion architecture. The first channel receives mode matching similarity parameters and maps them to a preset resistance baseline range to generate a baseline resistance projection. The second channel analyzes electromyographic activation levels, joint range of motion, and movement speed parameters from motion feature data, and uses a feature cross-correlation calculator to solve for the correlation coefficients between each parameter and historical average force exertion efficiency.

[0047] The dual-channel output is fed into a cascaded neural network: the first hidden layer performs feature dimensionality reduction and nonlinear activation operations, aligning the baseline resistance projection with the correlation coefficient vector space; the second hidden layer uses gated recurrent units to handle time-series dependencies and capture the transition characteristics of the action phase; the output layer generates initial resistance prediction values ​​through linear regression. A prediction range constraint module uses historical maximum load data to limit the range, ensuring that the output value always stays within the safe operating boundary of the equipment. During the training phase, the cascaded network uses backpropagation to optimize weights, and the prediction process preserves the activation state of neurons to achieve incremental learning capabilities.

[0048] A dynamic adaptation system with three compensation paths is constructed: the real-time fatigue compensation path analyzes the median frequency change rate of electromyography in motion feature data and generates a fatigue decay factor to implement exponential decay of the initial predicted value; the amplitude compensation path calculates the deviation of the joint angle from the preset optimal range of motion and modulates the resistance level using an inflection function; the power maintenance compensation path compares the ratio of the current output power to the historical peak power and fine-tunes the resistance increment through a proportional-derivative controller.

[0049] The compensation calculation employs a priority arbitration logic: the main compensation path is activated when the action amplitude deviates from the threshold; the secondary compensation path is activated when the fatigue coefficient fluctuates abnormally; and power maintenance requirements are fused through a third channel. The compensation result is filtered by a limiting filter to eliminate high-frequency abrupt changes, outputting a target resistance value with a smooth transition. This value is bound to a timestamp and compensation operation record for subsequent module tracking and decision-making.

[0050] The confidence calculation module establishes an inference framework, with core calculations including three dimensions: historical matching stability dimension, which calculates the standard deviation of the current pattern matching similarity and the recent mean similarity, triggering a confidence penalty when the dispersion exceeds the range; resistance deviation assessment dimension, which compares the absolute error rate between the target resistance value and the real-time measured resistance, and applies an S-shaped function to transform it into a basic confidence factor; and sensor redundancy verification dimension, which analyzes the data conflict rate of electromyography, inertial and force sensing, and calculates a multi-source consistency score.

[0051] The parameter input probability fusion engine establishes a confidence network topology through a conditional probability table and solves for the posterior confidence probability. The final output association confidence parameter adopts a five-level quantization method: 1.0 (high confidence), 0.8 (good confidence), 0.6 (medium confidence), 0.4 (low confidence), and 0.2 (poor confidence).

[0052] This embodiment generates a standardized force feature fragment library by deconstructing the action stages of historical torque curves and extracting key mechanical features. This overcomes the shortcomings of traditional schemes in utilizing insufficient historical data and provides a high-precision time alignment benchmark for motion pattern matching. Combining a spatiotemporal registration algorithm with a dual-path similarity evaluation mechanism, the trajectory matching degree between the current motion and the historical optimal pattern is quantified in real time, effectively solving the problem of motion trajectory deviation caused by individual state fluctuations. Based on multi-source feedback of median electromyographic frequency, joint amplitude, and power output, fatigue decay, amplitude modulation, and power maintenance compensation paths are constructed respectively, significantly improving the physiological adaptability of resistance output and the matching degree of motion targets.

[0053] In one embodiment, a resistance output command is obtained by performing feedforward compensation control and closed-loop error correction on the target resistance value and associated confidence parameters based on a preset dynamic resistance prediction model, including: The dynamic resistance prediction model is trained based on a composite neural network architecture using a three-stage transfer learning strategy: First, a dual-stream temporal convolutional network is pre-trained using a pre-set biomechanical dataset (including synchronously acquired electromyographic signals, six-dimensional torque, and motion trajectory), where physiological dynamic feature streams and equipment mechanical feature streams are fused through a cross-modal attention mechanism. Second, domain-adaptive training is performed using an equipment dynamics dataset (covering the torque response of servo mechanisms under different temperatures, wear conditions, and load conditions) to optimize the transmission system characteristic modeling capability. Third, online fine-tuning is performed using incremental Gaussian processes, continuously adjusting network weights by loading motion feature data from the target user. After training, the output is a dynamic resistance prediction model containing a feedforward compensation layer, a closed-loop correction layer, and a command output layer.

[0054] During the initialization phase of the dynamic resistance prediction model, the pre-trained convolutional kernel parameter set and feature space mapping table are loaded. The feature segmentation module performs parallel processing on the input dual-channel data: the target resistance value sequence is standardized by a sliding window, and the mean, variance, and gradient within the window are extracted to form a mechanical feature vector; the associated confidence parameter is discretized into five control levels by a gradation quantizer, and each level activates the corresponding gain correction coefficient.

[0055] The feature coupling engine receives the aforementioned dual-path features, and the core coupling operation is implemented in three levels: Level 1 coupling calculates the Hadamard product of the mechanical feature vector and the confidence gain coefficient to generate an initial fusion quantity; Level 2 coupling applies inertial delay compensation to the fusion quantity through a temporal convolution kernel, with the kernel size adaptively adjusted according to the motion acceleration; Level 3 coupling applies a feature cross-gating mechanism, selectively injecting ambient temperature compensation parameters using the confidence level as the switching condition. The final output contains the coupling resistance features of the feature channels.

[0056] The core of the feedforward compensation layer is a dual-path processing architecture. The main path implements multi-level nonlinear transformations: the first level uses grouped normalization to process the coupled feature tensors, dividing the channels into kinematic and biological groups according to the physical properties of the features for separate processing; the second level uses a channel attention mechanism to calculate the feature channel weights and assign enhancement coefficients to high-confidence features; the final level performs an improved ReLU activation function transformation, maintaining linearity in the positive interval while introducing a weakly decaying response in the negative interval.

[0057] The auxiliary path establishes a confidence-guided mechanism: It analyzes the confidence level markers in the coupling features. When a low-confidence state is detected, the shunt controller is activated, importing a portion of the feature channels into the auxiliary compensation network to generate compensation increments. In a high-confidence state, the feature selector is activated to suppress noise channels with fluctuations exceeding the threshold. The dual-path outputs undergo weighted synthesis at the feature fusion node: the main path output weights are set to the sigmoid mapping value of the confidence level, and the auxiliary path weights are their complementary values. Finally, the output shaping unit eliminates spectral aliasing components, generating a bandwidth-controlled dynamic feedforward compensation. This compensation includes phase lead correction information and is strictly matched to the device's dynamic characteristics.

[0058] The closed-loop correction layer employs a dual-loop feedback architecture to perform error prediction processing. The real-time signal acquisition unit captures the actual resistance output value of the actuator and compares it with the target resistance value using a difference comparator to generate the original error signal. The original error undergoes third-order feature deconstruction: extracting the steady-state error component (0-4Hz), the transient fluctuation component (4-15Hz), and the high-frequency noise component above 15Hz. The error feature reconstruction module performs frequency domain compensation based on the transmission system's damping characteristics: the steady-state component uses an integral compensator to eliminate accumulated errors; the transient component uses a proportional-derivative predictor to generate a phase lead compensation; and the high-frequency component is selectively attenuated using an adaptive filter.

[0059] The error feature fusion unit receives the compensated components and constructs a feature state space using a weighted autoregressive model: the state transition matrix is ​​dynamically updated based on the device's motion acceleration, and the observation matrix is ​​loaded with preset torque-velocity relationship parameters. The error predictor performs rolling time-domain optimization based on the feature state space to solve for the optimal error compensation sequence within the next 50ms time window. Finally, the compensation amount at the current moment is selected as the closed-loop error correction output, which undergoes transient smoothing and is bound to error source diagnostic markers.

[0060] The instruction construction phase implements a four-order processing flow: the primary spatial fusion unit decomposes the feedforward compensation into orthogonal motion spatial components (propulsion direction and lateral stabilization direction), and projects the closed-loop correction onto this spatial coordinate system to generate a fusion basis vector. The secondary dynamic weighting unit calculates the temporal correlation coefficient of the two inputs: the propulsion direction adopts a feedforward-dominated strategy (weight 0.6-0.8), and the stabilization direction implements closed-loop correction enhancement (weight 0.5-0.7).

[0061] The dead zone compensation unit includes a triple protection mechanism: the dead zone position detector constructs a position-backlash relationship surface based on historical transmission backlash data; the pre-compensation calculator generates a hyperbolic tangent function compensation curve based on the current motion direction; and the overshoot suppressor limits the maximum compensation slope based on the mechanism acceleration.

[0062] The instruction optimization phase involves frequency-time domain co-processing: a frequency domain notch filter eliminates mechanical resonant frequencies of a specific number (including preset gearbox meshing frequencies); and an adaptive window moving average algorithm maintains the continuity of the second derivative. The optimized instruction is then converted to the device drive signal format, outputting a composite resistance instruction containing the instruction value, security check code, and feature traceability index.

[0063] Based on motion characteristic data, dynamic boundary verification and incremental parameter updates are performed on the composite resistance command to obtain the resistance output command.

[0064] This embodiment utilizes a feature segmentation module to perform Hadamard product operations and gated fusion on drag values ​​and confidence parameters, constructing a high-dimensional feature tensor that includes environmental compensation parameters. This significantly enhances the dynamic representation capability of the drag model and overcomes the decoupling defect between physical signals and physiological parameters in traditional schemes. The feedforward compensation layer employs grouping normalization and channel attention mechanisms to process the feature tensor. Combined with a confidence-activated auxiliary path compensation network, it implements an on-demand switching nonlinear mapping strategy, ensuring output accuracy under high confidence and maintaining stability under low confidence. The closed-loop correction layer uses third-order frequency band decomposition technology to implement differentiated compensation for steady-state, transient, and high-frequency errors, overcoming the phase lag limitation of traditional single integral compensation and eliminating response delay.

[0065] In one embodiment, the coupled resistance characteristics are nonlinearly activated and mapped through the feedforward compensation layer of the dynamic resistance prediction model to obtain the dynamic feedforward compensation amount, including: The feedforward extraction module of the feedforward compensation layer receives coupling resistance features containing 16 channels. The core of the feedforward extraction module performs three-level processing: the first-level channel filter classifies mechanical feature channels and biological feature channels according to the physical properties of the features. The mechanical channels (torque amplitude, gradient change, etc.) are assigned to the high-pass filter group to eliminate baseline drift, while the biological channels (electromyographic activation, joint stability, etc.) are guided to the band-pass filter group to extract the energy of the feature frequency bands.

[0066] The second-stage spatial projector constructs a motion plane reference coordinate system, decoupling the feature vectors in the propulsion and stability directions for projection, and calculating the feature contribution factors in each direction. The third-stage vector compensation operation injects compensation based on equipment dynamic characteristic parameters: the transmission mechanism hysteresis compensation term is generated through a parabolic function fitted from historical backlash data, and the temperature drift compensation term is corrected online based on thermal sensor readings. Finally, the processed feature segments are stitched together along the time axis to generate a compensated feature vector, which includes a timestamp, an array of feature values, and a compensation operation log triplet.

[0067] The piecewise linear mapping engine loads a predefined inflection point parameter table, which divides the equipment's operating range into four segments: no-load, light-load, rated, and overload. After the compensation feature vector is input, the segment controller performs real-time operating segment determination: it calculates the mean Euclidean distance of the feature vectors within 10 consecutive sampling windows, and generates a segment switching command when the mean crosses a preset segment threshold.

[0068] The linear mapping kernel performs differentiated operations within the active segment: a fixed gain ratio mapping is used in the unloaded segment, adaptive slope adjustment is implemented in the lightly loaded segment (the slope changes inversely proportionally to the feature variance), multidimensional hyperplane mapping is enabled in the rated segment (the summation of the Hadamard product of the eigenvector and the weight matrix), and saturation limiting protection is implemented in the overload segment. The mapping result is processed by the phase correction unit: the phase jump caused by segment switching is compensated by an all-pass filter, and the group delay hold algorithm is used to ensure the continuity of the output signal waveform. The final output preliminary compensation value is supplemented with segment markers and mapping parameter indices for subsequent traceability calculations.

[0069] The amplitude range adjustment module establishes a confidence-driven dynamic gain control system. The confidence level parameter is input to a graded quantizer and discretized into five control levels. Each level activates a preset gain curve: high confidence level enables an exponential enhancement curve to expand the compensation upper limit; medium confidence level adopts a linear sustain mode; and low confidence level triggers a logarithmic decay function to compress the output range.

[0070] The dynamic limiter monitors the gradient rate of change of the initial compensation value in real time. When the rate of change exceeds a threshold, an adaptive limiting strategy is activated: a hyperbolic tangent saturation constraint is applied when the positive gradient changes abruptly, and a minimum output hold mechanism is activated when the negative gradient drops sharply. The range adjustment process preserves the phase characteristics of the original waveform, and phase distortion introduced by gain adjustment is eliminated through a constant group delay filter. A confidence state flag and gain operation record are appended to the output value to generate an adjustment compensation value with a controlled amplitude range.

[0071] The deviation correction system constructs a triple compensation path: the mechanical backlash compensation path loads the historical backlash database of the transmission system and generates the backlash compensation amount that varies with displacement through a parabolic fitting function; the temperature drift compensation path analyzes real-time data from thermal sensors and calculates the temperature drift correction term based on the material expansion coefficient; the zero-position cumulative error compensation path estimates the zero-position compensation value based on the initial position offset of the motor recorded by the Hall sensor and applies the recursive least squares method.

[0072] The correction calculation employs priority arbitration logic: primary backlash compensation is activated when the motion direction reverses; temperature drift compensation is enabled after continuous operation exceeding the threshold time; and zero-position compensation is executed when the zero-position calibration command is triggered. The correction result, after being filtered by an anti-aliasing filter to eliminate high-frequency interference, is input to the output shaping unit for waveform envelope optimization, generating a dynamic feedforward compensation amount whose phase characteristics strictly match the device's dynamic response. This amount is bound to a timestamp and correction parameter fingerprint for the servo driver to parse and execute.

[0073] This embodiment constructs a compensation feature vector with high characterization accuracy by classifying and screening mechanical and biological features and decoupling them through spatial projection, combined with the dynamic injection of transmission backlash and temperature drift parameters, significantly improving the physical adaptability of feedforward compensation. Based on a four-segment division strategy of the equipment's working range, it implements differentiated processing of no-load fixed gain, light-load slope adaptive, rated multi-dimensional hyperplane mapping, and overload saturation protection, breaking through the limitations of traditional single compensation curves in terms of working condition coverage. According to the confidence parameter, it activates three gain modes: exponential enhancement, linear maintenance, and logarithmic decay, and combines them with a hyperbolic limiting mechanism triggered by the gradient rate of change to achieve coordinated control of performance optimization under high confidence conditions and safety protection under low confidence conditions.

[0074] In one embodiment, the composite resistance command is obtained by constructing commands and compensating dead zones for the dynamic feedforward compensation and closed-loop error correction through the command output layer of the dynamic resistance prediction model: The proportional overlay module of the command output layer establishes a spatial projection coordinate system, decomposes the dynamic feedforward compensation into a principal component in the propulsion direction and a lateral stabilization component, and projects the closed-loop error correction along this coordinate system to generate an error compensation component. The overlay operation employs a dynamic weight allocation strategy: the weight of the principal component in the propulsion direction is linearly adjusted based on the motion acceleration (increase in acceleration leads to an increase in feedforward weight), while the lateral component implements an error correction-dominated strategy. The spatial synthesizer performs vector synthesis operations: the weighted sum of the feedforward and error components is calculated in the main propulsion direction, while the feedforward component is limited and the error component is integrally fused in the lateral stabilization direction.

[0075] The instruction construction phase introduces motion state constraints: when the joint angular velocity exceeds a threshold, motion plane rotation compensation is activated, transforming the synthesized vector to the instantaneous motion tangent plane. The final output encapsulates a three-dimensional data structure containing timestamps, spatial component arrays, and synthesized parameters, generating the initial composite drag instruction. This instruction retains the original waveform characteristics, and the phase delay is controlled within the microsecond range.

[0076] The stability adjustment system implements triple anti-jitter processing: the frequency domain notch filter eliminates command components that coincide with the inherent resonant frequency of the equipment (the center frequency is preset to the torsional vibration fundamental frequency of the transmission system); the time domain smoother adopts an adaptive window length moving average algorithm, and the window length is adjusted inversely proportional to the rate of change of the command (the window shortens as the rate of change increases); the phase synchronizer compensates for the delay differences of each frequency component group through an all-pass filter.

[0077] The dynamic gain scheduling module monitors the second derivative characteristics of the command: when a sudden acceleration change is detected, a nonlinear attenuator is activated, and a hyperbolic tangent function is used to softly limit the command amplitude; during continuous stable segments, a gain booster is activated to enhance command resolution exponentially. The output verification unit verifies the consistency of the command's physical dimensions: the unit of the drag command is forcibly converted to standard torque units, and mixed-dimensional data packets are eliminated. The processed adjusted drag parameters are then labeled with a stability rating.

[0078] A nonlinear dead zone identification system constructs a displacement-torque relationship analysis model. A historical transmission clearance database is loaded with records of backlash changes throughout the equipment's entire lifecycle, and a displacement-backlash probability distribution surface is generated through Gaussian process regression. A real-time displacement monitoring unit captures the actuator's position feedback signal. When a motion direction reversal event is detected, the dead zone scanner is activated: a probe torque sequence is applied in micro-steps within the neighborhood of the reversal point, and the inflection point of the torque-displacement response curve is recorded.

[0079] The second derivative extrema of the dead zone boundary locator's analytical response curve are used, combined with the confidence interval of the probability distribution surface, to calculate the displacement coordinates of the dead zone's start and end points. The compensation range generator outputs a three-dimensional compensation range parameter set containing the dead zone width, center position, and directional characteristics, with added confidence probability labels indicating positioning reliability. This parameter set is then filtered by a Kalman filter to eliminate measurement noise interference, generating a steady-state dead zone compensation range descriptor.

[0080] The segmented compensation engine establishes a direction-sensitive compensation strategy. A dead-zone crossing detector monitors the real-time trend of the change in the sign of the resistance parameter, triggering the segmented controller when the command value crosses the dead-zone center. The forward crossing segment employs an S-shaped compensation curve: an exponentially increasing compensation amount is applied at the dead-zone initiation point, a constant compensation value is maintained in the center, and hyperbolic tangent attenuation is enabled at the termination point. The reverse crossing segment uses asymmetric compensation: step compensation is used at the initiation point to eliminate idle time, and a linear ramp transition is implemented at the termination point. The compensation amount dynamic calibration module adaptively adjusts the compensation intensity based on the dead-zone width: high-frequency jitter compensation is used to suppress viscosity in narrow dead zones, and multi-level ramp compensation is performed in wide dead zones to avoid overshoot. The compensation phase synchronizer maintains consistent waveform group delay before and after compensation through an all-pass filter, outputting a compensation resistance command with a direction marker.

[0081] The boundary constraint system implements a two-level safety protection. The physiological safety boundary module analyzes the joint torque-angle relationship curves in the motion characteristic data to generate a dynamic torque threshold that changes in real time with body position. The mechanical protection module calls upon the equipment's maximum load parameters and transmission chain stiffness coefficients to calculate the mechanical structure resonance avoidance threshold. The dual-threshold coordinating controller limits the compensation resistance command: when the command value exceeds the physiological threshold, a hyperbolic tangent soft limit is activated; when it approaches the mechanical threshold, a gradient suppressor is activated. The command synchronization coordinator uses a clock synchronization protocol to align multi-source control commands: by calibrating the phase difference of the control cycle, a sliding window prediction algorithm is applied to compensate for network transmission jitter. The final output composite resistance command includes three fields: command value, safety status flag, and synchronization quality index.

[0082] This embodiment decouples the feedforward compensation and closed-loop correction in the propulsion and lateral stabilization directions, implementing a dynamic weight allocation strategy that differentiates between feedforward dominance and error correction. This effectively solves the directional conflict problem in traditional vector synthesis and improves the trajectory tracking accuracy of composite commands. Frequency-domain notch filtering eliminates mechanical resonance interference, time-domain adaptive smoothing suppresses command abrupt changes, and phase synchronization compensation ensures consistent group delay across the entire frequency band. This triple protection mechanism completely eliminates command output instability. A segmented strategy of forward S-curve and reverse asymmetric compensation is designed to address the reversal characteristics of motion direction. The compensation intensity is adaptively adjusted by dead-zone width, systematically solving the idle and viscosity problems in bidirectional transmission.

[0083] In one embodiment, dynamic boundary verification and incremental parameter updates are performed on the composite resistance command based on motion characteristic data to obtain the resistance output command, including: The safety boundary verification system performs physiological-mechanical dual constraint processing. The joint torque safety threshold from the motion characteristic data is input into the dynamic envelope calculation unit. Based on the real-time joint angles, the torque-angle dynamic relationship function is solved, and the dynamic physiological safety boundary, which changes continuously with body position, is output. The mechanical protection limiting module calls upon the equipment transmission chain stiffness parameter library and the maximum load threshold, and solves for the critical value for mechanical structure failure using the strain energy density calculation formula.

[0084] The dual-constraint coprocessor performs real-time conflict detection: when a composite resistance command simultaneously exceeds both the physiological safety boundary and the mechanical critical value, a minimum safety margin strategy is activated to select the lower constraint limit; when a single boundary exceeds the limit, a boundary approximation algorithm is used to generate an asymptotic limiting curve. The multi-level limiting engine classifies response levels based on the command's rate of change: when the rate of change is below a preset threshold A, a first-level smooth limiting (linear decay) is activated; when the rate of change is between thresholds A and B, a second-level gradient limiting (exponential decay) is activated; and when the rate of change exceeds threshold B, a third-level emergency truncation (step suppression) is triggered. The output is a mechanical limiting value containing the limiting value, response level, and constraint type.

[0085] The incremental resolver performs orthogonal projection operations. The mechanical limiting value input feature decoupling space is constructed from the motion plane normal vector and the equipment drive axis, forming an orthogonal basis. The torque adjustment component extraction module performs projection decomposition along the motion direction: calculating the dot product of the limiting value vector and the motion tangent vector, corrected by the inertia compensation coefficient to generate the instantaneous torque adjustment. The power adjustment component resolver performs frequency domain decomposition in the power dimension: extracting the fundamental frequency power spectral density of the limiting value signal by transformation, and calculating the effective power component by combining it with the equipment efficiency curve.

[0086] The decomposition process introduces a fatigue accumulation factor: when the historical power integral value exceeds a set threshold, the power attenuation coefficient is activated, reducing the power component amplitude in real time. A dual-component recombination verification unit verifies the decomposition completeness: it calculates the vector norm product of the torque component and the power component, ensuring that the energy conservation error between the torque component and the original amplitude limit is below the tolerance threshold. The output is a pair of torque adjustment component and power adjustment component data with time synchronization markers.

[0087] The dynamic range matching unit loads peak torque distribution curves and fatigue decay characteristics from the historical motion performance database. The torque adjustment component is input to the amplitude-phase analyzer to extract three parameters: instantaneous torque amplitude, gradient of change, and phase offset angle. The historical matching engine performs a dual-track alignment operation: the amplitude track calls a Gaussian process regressor to calculate the probability matching degree between the current torque amplitude and the historical peak curve; the phase track uses a dynamic time warping algorithm to align the real-time phase with the historical optimal force application sequence.

[0088] The range optimizer generates correction coefficients based on the matching results: a strengthening mode is activated in high-probability matching intervals to increase the upper limit of torque amplitude; a conservative mode is activated in low-probability matching intervals to apply logarithmic decay constraints. The output verification module verifies the continuity of torque changes: acceleration abrupt changes are identified through a third-order derivative detector, and spline smooth transition segments are inserted in the neighborhood of the abrupt changes to generate optimized torque components that conform to biomechanical laws.

[0089] A fatigue state perception network is constructed using an attenuation factor correction system. A power adjustment component is input to a power-time integrator to calculate the cumulative power value within a continuous time window. A fatigue coefficient resolver simultaneously analyzes the median electromyographic frequency shift and joint range of motion attenuation in the motion characteristic data, generating a real-time fatigue index through a sigmoid function mapping.

[0090] The attenuation factor calculation unit establishes a power-fatigue coupling equation: when the fatigue index exceeds the threshold, the exponential attenuation function is activated (attenuation base = cumulative power × fatigue index); a linear maintenance strategy is adopted in the sub-fatigue state. The correction actuator implements phase compensation processing: an all-pass filter is inserted in the power attenuation interval to maintain a constant waveform group delay; pre-emphasis technology is used to compensate for high-frequency losses during the power boost phase. Fatigue state markers and correction operation logs are added to the output components to generate fatigue adaptive components.

[0091] The parameter incremental updater establishes the feature weight matrix. The matrix row vectors are associated with the stress moment component features (amplitude / phase / gradient), and the column vectors are associated with the power component parameters (effective power / power factor / harmonic distortion rate). The incremental learning engine performs a three-step update: first, it calculates the Euclidean distance between the current feature vector and the historical mean vector; second, it solves for the feature weight increment using recursive least squares; and third, it applies a weight update law with a forgetting factor.

[0092] The command construction phase implements spatiotemporal coordination: the spatial projector decomposes the torque components into the main direction of equipment motion and the stable direction; the power-torque converter converts the power components into equivalent compensation torque based on the transmission efficiency curve. The dual outputs are recombined through a singular value decomposition and fusion algorithm to generate basic command values. Finally, a time delay compensation observer compensates for the mechanism's response lag, outputting a resistance output command that is phase-synchronized with the user's actions.

[0093] This embodiment establishes a minimum safety margin selection mechanism and a three-level response strategy through real-time collaborative calculation of the dynamic envelope of physiological joint torque-angle and the critical value of mechanical strain energy. This effectively solves the problem that traditional static safety thresholds cannot adapt to dynamic changes in body position, significantly improving the operational safety of the equipment. Based on the dual-path decoupling technology of motion tangential projection and power frequency domain decomposition, combined with real-time feedback adjustment of the fatigue accumulation factor, the physical dimensions of torque and power components are separated in a consistent manner, ensuring optimal energy transfer efficiency during parameter adjustment. A collaborative strategy of Gaussian probability amplitude matching and dynamic time warping phase alignment is adopted, generating enhanced / conservative dual-mode correction coefficients based on historical peak curves and force exertion timing characteristics, greatly improving the biomechanical adaptability of torque output.

[0094] In one implementation, the composite resistance command is checked for safety boundaries and mechanically protected based on motion characteristic data to obtain mechanical limit values, including: The joint torque-angle relationship curve from the motion characteristic data is input into the dynamic envelope calculation unit, which calculates the maximum safe torque value under the current body position based on the real-time acquired joint angle data. The angular velocity threshold segmenter receives this torque value and the device transmission ratio parameters, and calculates the instantaneous motion capacity limit of the joint using the kinetic energy conservation equation. The segmentation algorithm performs three-stage processing: the first stage converts the capacity limit into an angular acceleration threshold; the second stage performs integral calculations to generate a safe angular velocity envelope; and the third stage introduces an electromyographic activation intensity correction factor to increase the angular velocity threshold in a highly activated muscle state. The real-time motion state monitoring unit captures joint angular velocity data, and activates a boundary reinforcement mechanism when the angular velocity approaches the safe envelope: abnormal fluctuation patterns are identified through sliding window variance analysis, and the safe boundary range is dynamically narrowed. Finally, an angular velocity safe boundary with a confidence rating is output.

[0095] The mechanical rigidity database stores the elastic modulus, moment of inertia, and allowable strain energy density parameters of the drive shaft. The gradient limiting processor loads the time derivative of the composite resistance command and calculates the gradient value of the resistance change rate. The protection threshold generator performs real-time strain energy density calculation: inputting the resistance gradient value into the cantilever beam mechanics equations to solve for the maximum bending stress of the drive shaft; and combining this with the elastic modulus to calculate the real-time strain energy density value. The threshold setting employs a three-level strategy: when the strain energy density is below 50% of the material's yield strength, the command protection threshold is set to the rated load value; linear attenuation limiting is enabled in the 50%-80% range; and exponential attenuation protection is triggered above 80%. The dynamic calibration module corrects the threshold based on temperature sensor readings: for every 5°C increase in temperature, the yield strength reference value is lowered by a preset percentage. The output results are appended with a material fatigue accumulation warning flag and generate a command protection threshold data package.

[0096] The safety boundary verification process establishes a dual-parameter collaborative processing mechanism. Angular velocity safety boundary data is input into the motion plane analysis unit, decomposed into independent angular velocity threshold components of the flexion-extension plane and the rotation plane. Command protection thresholds are input into the load distribution calculation unit, allocated to key transmission nodes based on the equipment's transmission structure characteristics. The collaborative verification processor constructs multi-dimensional decision logic: dividing the angular velocity range into low-speed, medium-speed, and high-speed zones, and combining this with load conditions to classify light-load, medium-load, and heavy-load operating conditions.

[0097] The system monitors the position of composite resistance commands in the working condition matrix in real time. In high-risk areas, a geometric average strategy is used to generate temporary constraint values, while in low-risk areas, maximum value selection is implemented. Dynamic weight allocation is adjusted based on joint pose: angular velocity constraints are emphasized in neutral positions, and load limitations are strengthened in extreme positions. The output undergoes transient smoothing to eliminate decision jumps, generating primary mechanical constraint values ​​with working condition labels.

[0098] The mechanical fatigue correction system performs time-varying damage analysis. Accumulated equipment usage time data is input into the stress cycle statistics unit, recording the stress cycle history of key components under typical load spectra. The fatigue damage calculation unit calls upon the material durability property database and evaluates the real-time damage state using a linear accumulation criterion. The correction coefficient generation module dynamically adjusts the constraint values ​​based on the damage level: linear strengthening is enabled for low damage, the baseline value is maintained for medium damage, and progressive decay is triggered for high damage.

[0099] The environmental compensation unit integrates temperature and humidity sensor data, increasing the damage calculation weight when the temperature rises and activating the corrosion acceleration factor when the humidity exceeds the standard. Output verification ensures smooth changes in constraint values ​​and generates secondary mechanical constraint values ​​with material condition warnings.

[0100] The mechanical protection limiting process implements risk-based hierarchical control. Secondary mechanical constraint values ​​are input into the risk rating unit, which classifies risks into low, medium, and high levels based on the rate of change and absolute amplitude. The hierarchical limiting module performs differentiated processing: low-risk levels use smoothing filtering output; medium-risk levels enable exponential decay with limited rate of change; and high-risk levels trigger rapid step suppression.

[0101] The vibration monitoring unit analyzes the spectral characteristics in real time and activates selective frequency suppression when a resonant frequency is detected. The final output is a mechanical amplitude limit value that includes the amplitude limit, risk level, and vibration control markers.

[0102] This embodiment effectively solves the problem that traditional static thresholds cannot adapt to dynamic changes in body position by establishing a minimum safety margin selection mechanism and a three-level response strategy, significantly improving the operational safety of the equipment. Based on a dual-path decoupling technology of motion tangential projection and power frequency domain decomposition, combined with real-time feedback adjustment of the fatigue factor, it achieves consistent separation of the physical dimensions of torque and power components, ensuring optimal energy transfer efficiency during parameter adjustment. Employing a probability amplitude matching and dynamic phase alignment strategy, it generates enhanced / conservative dual-mode correction coefficients based on historical peak curves and force application timing characteristics, greatly improving the motion adaptability of torque output.

[0103] Reference Figure 2 As shown, the present invention also provides a machine learning-based resistance servo control system, applicable to any of the above-mentioned machine learning-based resistance servo control methods, comprising: The identification module is used to acquire multi-source signals and motion state data collected by the target device, and perform feature fusion to obtain motion feature data; The parsing module is used to acquire historical motion performance data of the target device, perform intent recognition with motion feature data, and obtain the target resistance value and associated confidence parameters. The processing module is used to perform feedforward compensation control and closed-loop error correction on the target resistance value and associated confidence parameters based on the preset dynamic resistance prediction model, and obtain the resistance output command.

[0104] This invention provides a machine learning-based resistance servo adjustment system. By fusing multi-source signals and motion state data collected from the target device, it can comprehensively capture the user's personalized motion characteristics, overcoming the problem of incomplete feature extraction caused by a single data source in traditional methods, and significantly improving the adaptability of resistance adjustment. By combining historical motion performance data with real-time motion feature data for intent recognition and outputting associated confidence parameters, the resistance value prediction is not only based on static patterns but can also dynamically adapt to changes in the user's real-time state, effectively solving the problem of lagging intent recognition in traditional methods. Based on motion feature data, dynamic boundary verification and incremental parameter updates are performed on composite resistance commands, ensuring that the resistance output can adapt to the user's real-time needs while automatically adjusting to a safe range, avoiding overtraining or insufficient protection problems caused by traditional fixed boundary methods. The generated resistance output commands fully consider the user's motion characteristics, real-time state, and safety boundaries, making resistance adjustment more precise and flexible, thereby improving the effectiveness of the training or rehabilitation process and the user experience.

[0105] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the system and each module described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0106] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A resistance servo control method based on machine learning, characterized in that, include: Multi-source signals are acquired by the target device, and feature fusion is performed to obtain motion feature data; Acquire historical motion performance data of the target device, perform intent recognition with the motion feature data, and obtain the target resistance value and associated confidence parameters; Based on a preset dynamic resistance prediction model, feedforward compensation control and closed-loop error correction are performed on the target resistance value and the associated confidence parameter to obtain a resistance output command; specifically, this includes: inputting the target resistance value and the associated confidence parameter into the dynamic resistance prediction model, and performing feature segmentation and resistance feature coupling to obtain coupled resistance features; The dynamic resistance prediction model uses a feedforward compensation layer to perform nonlinear activation mapping on the coupled resistance features to obtain a dynamic feedforward compensation amount. The model's closed-loop correction layer then performs error prediction on the target resistance value to obtain a closed-loop error correction amount. Finally, the dynamic resistance prediction model's command output layer constructs and compensates for dead zones by combining the dynamic feedforward compensation amount and the closed-loop error correction amount to obtain a composite resistance command. Based on the motion feature data, the composite resistance command undergoes dynamic boundary verification and incremental parameter updates to obtain a resistance output command.

2. The machine learning-based resistance servo adjustment method according to claim 1, characterized in that, The process of acquiring multi-source signals and motion state data collected by the target device and performing feature fusion to obtain motion feature data includes: The multi-source signal is segmented and transformed to obtain a hybrid feature vector; Based on the motion state data, motion coupling calculation is performed on the hybrid feature vector to obtain the enhanced feature matrix; The mechanical parameters of the target device are obtained, and the enhanced feature matrix is ​​subjected to torque conversion to obtain dynamic feature data; The dynamic feature data is fused, matched, and recombined according to a preset weight allocation strategy to obtain the motion feature data.

3. The machine learning-based resistance servo adjustment method according to claim 1, characterized in that, The process of acquiring historical motion performance data of the target device and performing intent recognition with the motion feature data to obtain the target resistance value and associated confidence parameters includes: The historical motion performance data is periodically segmented and identified to obtain historical force value feature segments; Based on preset action type labels, the historical force value feature segments are matched using conventional patterns to obtain conventional motion pattern data. The motion feature data is dynamically normalized and matched with the conventional motion pattern data to obtain the pattern matching similarity. The pattern matching similarity and the motion feature data are used to perform multi-mode drag prediction to obtain the initial drag prediction value; The initial resistance prediction value is compensated based on the motion characteristic data to obtain the target resistance value; The association confidence parameter is obtained by performing an association confidence calculation based on the pattern matching similarity and the target resistance value.

4. The machine learning-based resistance servo adjustment method according to claim 1, characterized in that, The step of performing nonlinear activation mapping on the coupled resistance characteristics through the feedforward compensation layer of the dynamic resistance prediction model to obtain the dynamic feedforward compensation amount includes: The feedforward compensation layer performs feature extraction and vector compensation on the coupling resistance features to obtain a compensation feature vector; The compensation feature vector is subjected to piecewise linear mapping calculation to obtain preliminary compensation values; The amplitude range of the initial compensation value is adjusted based on the correlation confidence parameter to obtain the adjusted compensation value. The deviation offset correction is applied to the adjusted compensation value to obtain the dynamic feedforward compensation amount.

5. The machine learning-based resistance servo adjustment method according to claim 1, characterized in that, The composite resistance command is obtained by constructing commands and compensating dead zones for the dynamic feedforward compensation and the closed-loop error correction through the command output layer of the dynamic resistance prediction model. The initial composite resistance command is obtained by proportionally superimposing the dynamic feedforward compensation amount and the closed-loop error correction amount and constructing the command output layer. The initial composite resistance command is adjusted for command stability to obtain the adjusted resistance parameters; Nonlinear dead zone identification is performed on the adjusted resistance parameters to obtain the dead zone compensation range; Based on the dead zone compensation range and the adjusted resistance parameters, the initial composite resistance command is segmented and compensated to obtain a compensated resistance command. The composite resistance command is obtained by constraining the command boundary and coordinating command synchronization on the compensation resistance command.

6. The machine learning-based resistance servo adjustment method according to claim 1, characterized in that, The process of dynamically boundary-checking and incrementally updating the composite resistance command based on the motion characteristic data to obtain the resistance output command includes: Based on the motion characteristic data, the composite resistance command is subjected to safety boundary verification and mechanical protection limiting to obtain the mechanical limiting value; Incremental parameter decomposition is performed on the mechanical limit value to obtain torque adjustment component and power adjustment component; Based on the historical motion performance data, the torque adjustment component is dynamically range-matched to obtain the optimized torque component. The power adjustment component is corrected by an attenuation factor to obtain the fatigue adaptive component; The optimized torque component and the fatigue adaptive component are incrementally updated with parameters and commands are constructed to obtain the resistance output command.

7. The machine learning-based resistance servo adjustment method according to claim 6, characterized in that, The step of performing safety boundary verification and mechanical protection limiting on the composite resistance command based on the motion characteristic data to obtain the mechanical limit value includes: Based on the motion characteristic data, the composite resistance command is segmented by angular velocity threshold to obtain the angular velocity safety boundary; The composite resistance command is gradient-limited based on preset mechanical rigidity data to obtain the command protection threshold. The safety boundary of the angular velocity and the command protection threshold are checked to obtain the primary mechanical constraint value; Based on the motion characteristic data, the primary mechanical constraint value is corrected for mechanical fatigue accumulation to obtain the secondary mechanical constraint value; The secondary mechanical constraint value is mechanically protected and limited to obtain the mechanical limit value.

8. A resistance servo adjustment system based on machine learning, characterized in that, The method for adjusting resistance based on machine learning, applied to any one of claims 1-7, comprises: The identification module is used to acquire multi-source signals and motion state data collected by the target device, and perform feature fusion to obtain motion feature data; The parsing module is used to acquire the historical motion performance data of the target device, perform intent recognition with the motion feature data, and obtain the target resistance value and associated confidence parameters. The processing module is used to perform feedforward compensation control and closed-loop error correction on the target resistance value and the associated confidence parameter according to a preset dynamic resistance prediction model, so as to obtain a resistance output command.

Images