A joint angle prediction method and system based on rule dimension reduction and a medium

Abstract

A joint angle prediction method and system based on rule dimension reduction and a medium are related to the technical field of human-computer interaction. The method comprises the following steps: acquiring signals related to joint movement, extracting time domain features, frequency domain features and nonlinear features from the signals to form a multi-dimensional input feature vector; clustering the multi-dimensional input feature vector to generate a plurality of cluster centers, configuring a Gaussian membership function for the cluster centers, and constructing an initial fuzzy system; calculating the membership degrees of the current input multi-dimensional feature vector to each original fuzzy rule to form an original membership degree vector; mapping the original membership degree vector to a low-dimensional membership degree vector by using a learnable dimension reduction matrix; performing fuzzy reasoning by using the low-dimensional membership degree vector and corresponding consequent function parameters to output a joint angle prediction value; and determining the dimension reduction matrix and the consequent function parameters through joint optimization. The application can realize accurate prediction of joint angle.

Description

Technical Field

[0001] This invention relates to the field of human-computer interaction technology, and in particular to a method, system and medium for predicting joint angles based on rule-based dimensionality reduction. Background Technology

[0002] In modern rehabilitation medicine and assistive robotics, the performance of devices such as exoskeleton robots, intelligent prostheses, and remote rehabilitation systems is highly dependent on the real-time and accurate prediction of joint angles in hemiplegic patients in order to achieve precise motor assistance and functional assessment.

[0003] Currently, joint angle prediction mainly employs time series analysis methods based on sensor data, including statistical models, artificial neural networks, support vector machines, and fuzzy systems. These methods perform well in specific scenarios, but when processing multidimensional sensor data, fuzzy systems are prone to a sharp increase in the number of rules, leading to high computational complexity and difficulty in meeting the low-latency requirements of real-time control.

[0004] More importantly, existing methods typically build a complete rule base first, and then reduce dimensionality through heuristic rule merging or post-processing pruning. This not only introduces redundant computation, but also easily loses motion details because the dimensionality reduction is disconnected from the prediction target, making it difficult to adapt to the dynamic motion patterns of hemiplegic patients at different stages of rehabilitation.

[0005] For example, Chinese patent document CN111476204A discloses a method and device for predicting the ankle joint angle of a prosthesis, which uses deep learning to perform end-to-end dimensionality reduction and prediction of surface electromyography signals; however, this method is still a black box model, which makes it difficult to establish an interpretable mapping relationship between motion rules and joint angles, and the model structure is fixed, making it impossible to dynamically adjust the complexity according to the patient's rehabilitation process, and thus difficult to meet the needs of personalized rehabilitation training.

[0006] The fundamental reason is that existing technologies, whether using traditional fuzzy systems or deep learning methods, have not achieved end-to-end joint optimization among rule representation, dimensionality reduction mechanism and prediction target, thus making it difficult to balance the interpretability, rehabilitation adaptability and practical application performance of the model. Summary of the Invention

[0007] The purpose of this invention is to provide a joint angle prediction method based on rule-based dimensionality reduction to achieve accurate prediction of joint angles.

[0008] This invention provides the following technical solution:

[0009] In a first aspect, the present invention provides a joint angle prediction method based on rule-based dimensionality reduction, comprising the following steps:

[0010] Acquire signals related to joint movement;

[0011] Time-domain features, frequency-domain features, and nonlinear features are extracted from the signal to form a multidimensional input feature vector;

[0012] Clustering is performed based on the multidimensional input feature vector to generate multiple cluster centers, and a Gaussian membership function is configured for each cluster center, thereby constructing an initial fuzzy system containing multiple original fuzzy rules;

[0013] Based on the current input multidimensional feature vector, calculate its membership degree to each original fuzzy rule to form the original membership degree vector;

[0014] The original membership vector is mapped to a low-dimensional membership vector using a learnable dimensionality reduction matrix.

[0015] Fuzzy inference is performed using the low-dimensional membership vector and the corresponding consequent function parameters to output the predicted joint angle value.

[0016] The dimensionality reduction matrix and the consequent function parameters are jointly determined through optimization.

[0017] Furthermore, the dimension of the low-dimensional membership vector is lower than the number of the original fuzzy rules.

[0018] Furthermore, the joint motion-related signals include at least one of inertial measurement unit (IMU) signals and surface electromyography (sEMG) signals.

[0019] Furthermore, the time-domain features include at least one of the mean, variance, and root mean square (RMS) values;

[0020] The frequency domain features include at least one of the dominant frequency and power spectral density;

[0021] The nonlinear characteristics include at least one of sample entropy and complexity.

[0022] Furthermore, the clustering adopts the fuzzy C-means clustering algorithm.

[0023] Furthermore, the process of jointly determining the joint optimization includes minimizing an objective function, which includes a prediction error term for measuring the joint angle prediction error, a dimensionality reduction fidelity term for maintaining consistency between the original membership degree and the low-dimensional membership degree, and a regularization term for controlling the complexity of the dimensionality reduction matrix.

[0024] Furthermore, the objective function is:

[0025]

[0026] in, Original membership degree; It is a dimension-reduced matrix; y represents the membership degree after dimensionality reduction; Y represents the joint angle. The input matrix after mapping; and For consequent parameters; For regularization parameters; For dimensional reduction and fidelity preservation; This is the prediction error term; For the regularization term of the dimension reduction matrix; This is a regularization term for the consequent parameter.

[0027] Furthermore, by adjusting the weight coefficients of the regularization term used to control the complexity of the dimensionality reduction matrix, the sparsity of the dimensionality reduction matrix is ​​dynamically adjusted, thereby changing the number of effective fuzzy rules to adapt to the complexity of the subject's movement patterns at different rehabilitation stages.

[0028] Furthermore, the fuzzy inference model is a Takagi-Sugeno-Kang (TSK) type fuzzy model (i.e., the TS fuzzy model), and its consequent function is a linear function of the input features.

[0029] Secondly, the present invention provides a joint angle prediction system based on rule-based dimensionality reduction, comprising:

[0030] The signal acquisition module is used to acquire signals related to joint movement.

[0031] The feature extraction module is used to extract time-domain features, frequency-domain features, and nonlinear features from the signal to form a multi-dimensional input feature vector;

[0032] The fuzzy rule construction module is used to perform clustering based on the multidimensional input feature vector to generate multiple cluster centers, and to configure Gaussian membership functions for each cluster center, thereby constructing an initial fuzzy system containing multiple original fuzzy rules;

[0033] The membership calculation module is used to calculate the membership degree of the current input multidimensional feature vector to each original fuzzy rule, forming the original membership degree vector;

[0034] The rule-based dimensionality reduction module is configured with a learnable dimensionality reduction matrix, which is used to map the original membership vector into a low-dimensional membership vector.

[0035] The fuzzy inference module is used to perform fuzzy inference using the low-dimensional membership vector and the corresponding consequent function parameters, and output the predicted value of the joint angle.

[0036] The parameter optimization module is used to jointly optimize the dimensionality reduction matrix and the consequent function parameters to jointly determine their values.

[0037] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the aforementioned rule-based dimensionality reduction-based joint angle prediction method.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. The number of rules has been significantly reduced while maintaining high prediction accuracy.

[0040] This invention maps multiple original fuzzy rules into low-dimensional membership vectors using a learnable dimensionality reduction matrix, and performs fuzzy inference based on this low-dimensional representation. Since the dimensionality reduction matrix and consequent function parameters are jointly determined through optimization, the optimization process directly targets the joint angle prediction error, ensuring that prediction accuracy is effectively maintained while significantly reducing computational complexity. This avoids the performance loss caused by task disconnect in the traditional two-stage "modeling first, then pruning" method.

[0041] 2. Preserve clinical interpretability and enhance physician trust.

[0042] Each original fuzzy rule is defined by a cluster center and a Gaussian membership function, corresponding to a specific motion pattern. The dimensionality reduction process fuses the rules through linear mapping, and the relationship is explicitly represented by the dimensionality reduction matrix. Experts can analyze this matrix to trace the original motion features (such as key phases in gait like heel strike and push-off) contained in the dimensionality-reduced rules, verifying whether key clinical patterns have been preserved. Compared to traditional heuristic rule merging based on preset thresholds, this solution provides transparent and interpretable fusion criteria, improving credibility in medical scenarios.

[0043] 3. Adapt to individual differences and support personalized rehabilitation.

[0044] Thanks to the end-to-end joint optimization framework, the degree of dimensionality reduction can be flexibly controlled by adjusting the regularization intensity: lower regularization is used in the early stages of rehabilitation to retain more original rules to capture a wide range of compensatory movements; in the later stages of rehabilitation, regularization is increased to achieve a higher degree of rule reduction and focus on fine motor control. The system can dynamically adjust the model complexity as the patient's motor ability recovers, providing truly personalized real-time assistance for devices such as exoskeletons. Detailed Implementation

[0045] The present invention will be further described below with reference to specific embodiments, but the scope of protection of the present invention is not limited thereto.

[0046] This embodiment provides a joint angle prediction method and system based on rule-based dimensionality reduction, which can be used to achieve accurate prediction of joint angles. Compared with the prior art, this embodiment has made improvements in model structure, optimization mechanism and calculation process, and has been optimized in particular for the actual needs of exoskeleton robots and rehabilitation training for hemiplegic patients.

[0047] In one specific embodiment, the joint angle prediction system includes: a signal acquisition module, a feature extraction module, a fuzzy rule construction module, a membership degree calculation module, a rule dimensionality reduction module, a fuzzy inference module, and a parameter optimization module. Specifically, the signal acquisition module acquires IMU and / or sEMG signals related to joint motion; the feature extraction module extracts time-domain, frequency-domain, and nonlinear features from the original signals and constructs a multi-dimensional input feature vector; the fuzzy rule construction module performs clustering based on the multi-dimensional input feature vector, generates cluster centers, and configures Gaussian membership functions to construct an initial fuzzy system; the membership degree calculation module calculates the membership degree of the current input feature vector to each original fuzzy rule, forming an original membership degree vector; the rule dimensionality reduction module is configured with a learnable dimensionality reduction matrix to map the original membership degree vector to a lower-dimensional membership degree vector; the fuzzy inference module employs a Takagi-Sugeno-Kang (TSK) fuzzy inference mechanism, using the low-dimensional membership degree vector and corresponding consequent function parameters to output the predicted joint angle value; and the parameter optimization module jointly optimizes the dimensionality reduction matrix and consequent function parameters.

[0048] This embodiment replaces the traditional heuristic rule fusion mechanism by performing learnable dimensionality reduction directly in the membership space, effectively avoiding information loss and ensuring that the dimensionality reduction process is consistent with the prediction target.

[0049] In some embodiments, the feature extraction module can be used to perform bandpass filtering, denoising and normalization preprocessing on the raw sensor data, and extract time-domain features, frequency-domain features and nonlinear features from it to form a multi-dimensional input feature vector.

[0050] In some embodiments, the joint angle prediction system may further include an output control module for sending the joint angle prediction results to the actuator of the exoskeleton robot or rehabilitation device to achieve closed-loop control.

[0051] Specifically, the joint angle prediction method in this embodiment includes the following steps:

[0052] Step 1: Constructing the membership matrix

[0053] Let the training samples be ,in Let y be the n-dimensional feature vector extracted by the sliding window. i ∈[0,180] represents the corresponding target joint angle. Assume there are initially R fuzzy rules, each rule's antecedent consisting of a Gaussian membership function:

[0054]

[0055] in, and Let r be the center and width of the r-th rule on the j-th feature, respectively, and initialize them using the fuzzy C-means clustering algorithm. This is the value of the current input sample in the j-th feature dimension.

[0056] Membership matrix Each element is:

[0057]

[0058] Then, normalization was performed:

[0059]

[0060] This step retains the advantages of the traditional TSK fuzzy system, namely, capturing expert knowledge through predefined membership functions and avoiding the instability caused by training membership functions.

[0061] Step 2: Dimensionality Reduction Matrix Design and Optimization

[0062] This embodiment introduces a learnable dimensionality reduction matrix. This replaces traditional rule-based fusion, thereby projecting high-dimensional membership degrees onto a low-dimensional space:

[0063]

[0064] in, This is the membership matrix after dimensionality reduction. The dimensionality reduction matrix Q is not obtained through eigenvalue decomposition, but rather through an optimization process that satisfies orthogonality constraints. In rehabilitation training for hemiplegic patients, different stages of movement require rule representations of different granularities. Learnable dimensionality reduction matrices can adaptively adjust the rule granularity without redesigning the entire fuzzy system.

[0065] Step 3: Construction of the dimensionality-reduced TS fuzzy model (i.e., TSK fuzzy model)

[0066] The dimensionality-reduced TS fuzzy model is represented as follows:

[0067]

[0068] in, This represents the membership degree of the k-th dimension after dimensionality reduction. and These are the consequent parameters after dimensionality reduction. This is a mapping function that maps the original input to a high-dimensional feature space. These are the predicted joint angle values. The mapping function can transform low-dimensional nonlinear problems into high-dimensional linear problems, enabling the system to accurately model specific movement patterns in rehabilitation training.

[0069] Step 4: Design of Joint Optimization Objective Function

[0070] Design a joint optimization objective function that simultaneously optimizes the dimensionality reduction matrix and fuzzy system parameters:

[0071]

[0072] in, The input matrix after mapping, For regularization parameters. In rehabilitation training, the system needs to consider both model accuracy and complexity. The joint optimization framework described above balances these two requirements, ensuring optimal predictive performance with limited computational resources.

[0073] in, Original membership degree; It is a dimension-reduced matrix; y represents the membership degree after dimensionality reduction; Y represents the joint angle. The input matrix after mapping; and For consequent parameters; For regularization parameters; For dimensional reduction and fidelity preservation; This is the prediction error term; For the regularization term of the dimension reduction matrix; This is a regularization term for the consequent parameter.

[0074] Step 5: Implementation of the Alternating Optimization Algorithm

[0075] 1. An alternating optimization strategy is used to solve the objective function to ensure stable convergence of the algorithm;

[0076] 2. Fix Q and optimize consequent parameters and ;

[0077] 3. Fix the parameters of the successor and optimize Q;

[0078] 4. Update the membership degree of the dimension reduction ;

[0079] 5. Update mapping parameters ;

[0080] 6. Repeat the above steps until convergence.

[0081] Step Six: Joint Angle Prediction

[0082] After the model training is complete, perform the following operations for new sensor input data:

[0083] (1) Extract the same multidimensional feature vectors as those used in the training phase;

[0084] (2) Calculate its membership degree to the original fuzzy rule to form the original membership degree vector;

[0085] (3) Calculate the low-dimensional membership vector using the optimized dimensionality reduction matrix;

[0086] (4) Use the reduced-dimensional TSK fuzzy model to output the predicted joint angle values.

[0087] In one specific embodiment, the joint angle prediction method includes the following steps:

[0088] Step 1: Data Preprocessing and Feature Extraction

[0089] Bandpass filtering is applied to the raw sensor data (e.g., 0.5-20Hz for IMU signals and 20-150Hz for sEMG signals).

[0090] The filtered signal is segmented using a sliding window (window length 200ms, overlap rate 50%).

[0091] Extract multi-dimensional feature vectors for each window, such as 45-dimensional features, including time-domain features (e.g., mean, variance, RMS value), frequency-domain features (e.g., dominant frequency, power spectral density), and nonlinear features (e.g., sample entropy, complexity).

[0092] Step 2: Fuzzy rule initialization

[0093] Use the fuzzy C-means clustering algorithm to cluster the feature data of the training set, for example, to generate 100 cluster centers;

[0094] Assign a Gaussian membership function to each cluster center;

[0095] Initialize the center parameters of each membership function and width parameter .

[0096] Step 3: Joint optimization of dimensionality reduction matrix and model parameters

[0097] Calculate the original membership matrix based on the training samples. ;

[0098] Initialize the dimensionality reduction matrix For example, by using QR decomposition to make it satisfy orthogonal constraints;

[0099] Set regularization parameters ;

[0100] Adopt an alternating optimization strategy:

[0101] a. Fix Q, and update the TSK consequent parameters using a closed-form solution;

[0102] b. Fix the consequent parameters, update Q while maintaining orthogonal constraints;

[0103] c. Update kernel mapping parameters;

[0104] d. Check convergence and repeat the above steps until convergence is achieved.

[0105] Step 4: Model Simplification and Validation

[0106] Calculate the row norm of the dimensionality-reduced matrix Q to identify important original rules;

[0107] The model's predictive performance was evaluated using 10-fold cross-validation.

[0108] Adjust regularization parameters according to the patient's recovery stage To balance model complexity and accuracy;

[0109] Save the optimized model parameters, including the dimensionality reduction matrix. and consequent parameters .

[0110] Step 5: Real-time Reasoning

[0111] Acquire real-time sensor data within a 200ms window;

[0112] Calculate the same multidimensional feature vectors as those used in the training phase, such as 45-dimensional feature vectors;

[0113] Calculate the original membership vector ;

[0114] Calculate the reduced membership degree using the optimized reduced-dimensional matrix Q. ;

[0115] Calculate Gaussian kernel mapping ;

[0116] Calculate the predicted joint angle based on the dimensionality-reduced TSK fuzzy model;

[0117] Apply a low-pass filter to smooth the output;

[0118] The prediction results were used for exoskeleton control at an update rate of 100 Hz.

[0119] In summary, this embodiment achieves high-performance joint angle prediction through three key technical mechanisms:

[0120] First, a learnable dimensionality reduction matrix replaces rule fusion. This embodiment introduces an optimizable dimensionality reduction matrix Q, achieving effective rule reduction through orthogonal constraints and joint optimization. Unlike traditional PCA dimensionality reduction, Q is not obtained through eigenvalue decomposition, but rather learned through an optimization process, ensuring that the dimensionality-reduced representation is consistent with the prediction task. The dimensionality reduction matrix Q satisfies orthogonal constraints. Furthermore, it is jointly optimized with the TSK fuzzy system parameters, preserving the semantic information of the original rules.

[0121] Second, a clinically-oriented parameter adaptation mechanism. This embodiment establishes a dynamic correlation mechanism between the patient's rehabilitation status and algorithm parameters, enabling adaptive adjustment of model complexity. Unlike static models with fixed parameters in existing technologies, this embodiment monitors the patient's motor function recovery level and dynamically adjusts the sparsity of the dimensionality reduction matrix and model complexity, allowing the system to adapt to the needs of different rehabilitation stages.

[0122] Third, the synergistic optimization of low-dimensional representation and kernel mapping. This embodiment first reduces the number of rules by dimensionality reduction, and then enhances nonlinear modeling capabilities through kernel mapping. Dimensionality reduction dimension d and kernel parameters... Collaborative optimization and kernel mapping are applied in dimensionality reduction space to effectively solve the problem of decreased nonlinear expressive power after dimensionality reduction.

[0123] The above mechanisms together bring about the following significant effects:

[0124] First, the number of rules is significantly reduced while maintaining prediction accuracy. This embodiment reduces the original R rules to d effective rules using a dimensionality reduction matrix, significantly lowering the computational burden. The dimensionality reduction matrix Q is learned through an optimization process to ensure the membership degrees after dimensionality reduction. Able to retain the original membership degree to the greatest extent The relationship with the joint angle Y. The prediction error term in the joint optimization objective function ensures the consistency between the dimensionality reduction process and the prediction task, avoiding the task misalignment problem in traditional two-stage methods.

[0125] Secondly, it preserves clinical interpretability and enhances physician trust. The elements Q of the dimensionality-reduced matrix Q. i,j The value represents the contribution of the original i-th rule to the j-th rule after dimensionality reduction, allowing for a direct analysis of which movement patterns were merged. Experts can verify whether the dimensionality reduction process preserved key clinical features, such as crucial phases in gait like heel strike and push-off. Traditional rule fusion mechanisms merge rules using predefined similarity thresholds, making it difficult to explain the basis for merging decisions. While the dimensionality reduction matrix in this embodiment is obtained through optimization learning, its elements still directly reflect the mapping relationship between the original rules and the dimensionality-reduced rules. By analyzing the row norm of Q, the original rules that contribute the most to the dimensionality-reduced representation can be identified; these rules often correspond to key movement patterns in clinical significance.

[0126] Finally, it adapts to individual differences and supports personalized rehabilitation. This is achieved by adjusting the regularization parameters. It can control the sparsity of the dimensionality reduction matrix Q, adapting to the motor complexity at different rehabilitation stages. For example, patients in the early stages of rehabilitation use smaller... It retains more of the original rules and captures a wide range of compensatory movements; it is more suitable for patients in the later stages of rehabilitation. This achieves a higher degree of rule simplification, focusing on fine motor control. The complexity of joint movements changes dynamically with the rehabilitation process, and the parameters... By controlling the Frobenius norm of the dimensionality reduction matrix Q, the degree of dimensionality reduction is indirectly adjusted, enabling the exoskeleton system to dynamically adjust along with the patient's rehabilitation process and provide personalized motion assistance.

[0127] This invention is not limited to the above embodiments. Those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this invention. The scope of this invention is defined by the claims and their equivalents.

Claims

1. A joint angle prediction method based on rule-based dimensionality reduction, characterized in that, Includes the following steps: Acquire signals related to joint movement; Time-domain features, frequency-domain features, and nonlinear features are extracted from the signal to form a multidimensional input feature vector; Clustering is performed based on the multidimensional input feature vector to generate multiple cluster centers, and a Gaussian membership function is configured for each cluster center, thereby constructing an initial fuzzy system containing multiple original fuzzy rules; Based on the current input multidimensional feature vector, calculate its membership degree to each original fuzzy rule to form the original membership degree vector; The original membership vector is mapped to a low-dimensional membership vector using a learnable dimensionality reduction matrix. Fuzzy inference is performed using the low-dimensional membership vector and the corresponding consequent function parameters to output the predicted joint angle value. The dimensionality reduction matrix and the consequent function parameters are jointly determined through optimization.

2. The joint angle prediction method based on rule-based dimensionality reduction according to claim 1, characterized in that, The joint motion-related signals include at least one of IMU signals and sEMG signals.

3. The joint angle prediction method based on rule-based dimensionality reduction according to claim 1, characterized in that, The time-domain features include at least one of mean, variance, and RMS value; The frequency domain features include at least one of the dominant frequency and power spectral density; The nonlinear characteristics include at least one of sample entropy and complexity.

4. The joint angle prediction method based on rule-based dimensionality reduction according to claim 1, characterized in that, The clustering algorithm used is fuzzy C-means clustering.

5. The joint angle prediction method based on rule-based dimensionality reduction according to claim 1, characterized in that, The joint optimization process includes minimizing an objective function, which includes a prediction error term to measure the joint angle prediction error, a dimensionality reduction fidelity term to maintain consistency between the original membership degree and the low-dimensional membership degree, and a regularization term to control the complexity of the dimensionality reduction matrix.

6. The joint angle prediction method based on rule-based dimensionality reduction according to claim 5, characterized in that, The objective function is: in, Original membership degree; It is a dimension-reduced matrix; The membership degree is for dimensionality reduction; Y is the joint angle. The input matrix after mapping; and For consequent parameters; For regularization parameters; For dimensional reduction and fidelity preservation; This is the prediction error term; For the regularization term of the dimension reduction matrix; This is a regularization term for the consequent parameter.

7. The joint angle prediction method based on rule-based dimensionality reduction according to claim 5 or 6, characterized in that, By adjusting the weight coefficients of the regularization term used to control the complexity of the dimensionality reduction matrix, the sparsity of the dimensionality reduction matrix is ​​dynamically adjusted, thereby changing the number of effective fuzzy rules to adapt to the complexity of the subject's movement patterns at different rehabilitation stages.

8. The joint angle prediction method based on rule-based dimensionality reduction according to claim 1, characterized in that, The fuzzy inference model is a TSK fuzzy model, and its consequent function is a linear function of the input features.

9. A joint angle prediction system based on rule-based dimensionality reduction, characterized in that, include: The signal acquisition module is used to acquire signals related to joint movement. The feature extraction module is used to extract time-domain features, frequency-domain features, and nonlinear features from the signal to form a multi-dimensional input feature vector; The fuzzy rule construction module is used to perform clustering based on the multidimensional input feature vector to generate multiple cluster centers, and to configure Gaussian membership functions for each cluster center, thereby constructing an initial fuzzy system containing multiple original fuzzy rules; The membership calculation module is used to calculate the membership degree of the current input multidimensional feature vector to each original fuzzy rule, forming the original membership degree vector; The rule-based dimensionality reduction module is configured with a learnable dimensionality reduction matrix, which is used to map the original membership vector into a low-dimensional membership vector. The fuzzy inference module is used to perform fuzzy inference using the low-dimensional membership vector and the corresponding consequent function parameters, and output the predicted value of the joint angle. The parameter optimization module is used to jointly optimize the dimensionality reduction matrix and the consequent function parameters to jointly determine their values.

10. A computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the joint angle prediction method based on rule-based dimensionality reduction as described in any one of claims 1 to 8.

Images