CN-122020153-A - Robust joint moment prediction method and system based on self-supervision reconstruction

CN122020153ACN 122020153 ACN122020153 ACN 122020153ACN-122020153-A

Abstract

The invention provides a robust joint moment prediction method and a robust joint moment prediction system based on self-supervision reconstruction, which belong to the technical field of biomechanical signal processing and machine learning intersection, wherein the method comprises the following steps of S1, acquiring a large number of historical surface electromyographic signals for preprocessing and labeling, and constructing a data set; the method comprises the steps of S2, dividing a data set into a training set, a verification set and a test set based on a preset proportion, S3, creating a joint moment prediction model based on a self-supervision reconstruction feature extraction module, a time-dependent modeling module and a regression prediction module, S4, training, verifying and testing the joint moment prediction model through the training set, the verification set and the test set respectively, deploying the joint moment prediction model which passes the test, and S5, predicting the joint moment through the deployed joint moment prediction model. The method has the advantages that the precision, the robustness and the generalization capability of joint moment prediction are greatly improved.

Inventors

XIONG BAOPING
GUO YINGHUI
ZHANG JILIN

Assignees

福建理工大学

Dates

Publication Date: 20260512
Application Date: 20251218

Claims (10)

1. A robust joint moment prediction method based on self-supervision reconstruction is characterized by comprising the following steps: step S1, acquiring a large number of historical surface electromyographic signals of different channels, preprocessing each historical surface electromyographic signal, and constructing a data set after marking; step S2, dividing the data set into a training set, a verification set and a test set based on a preset proportion; s3, creating a joint moment prediction model based on the self-supervision reconstruction feature extraction module, the time-dependent modeling module and the regression prediction module; Step S4, training, verifying and testing the joint moment prediction model respectively through the training set, the verification set and the test set, and deploying the joint moment prediction model which passes the test; And S5, predicting the joint moment through the deployed joint moment prediction model.
2. The robust joint moment predicting method based on self-supervision reconstruction of claim 1, wherein the step S1 specifically comprises the following steps: Acquiring a large number of historical surface electromyographic signals of different subjects in different channels, preprocessing at least comprising filtering, normalizing and sliding window segmentation on each historical surface electromyographic signal, and labeling joint moment values on each preprocessed historical surface electromyographic signal to construct a data set.
3. The robust joint moment predicting method based on self-supervision reconstruction of claim 1, wherein the step S2 specifically comprises: the data set is divided into a training set, a validation set, and a test set in a ratio of 7:2:1 based on a time sequence.
4. The robust joint moment prediction method based on self-supervision reconstruction of claim 1, wherein in the step S3, the self-supervision reconstruction feature extraction module is constructed based on an encoder and a decoder, the encoder is used for extracting local time domain features representing instantaneous muscle activation changes from surface electromyographic signals through a one-dimensional convolution layer, extracting global features through long-term dependence of surface electromyographic signals learned by a multi-layer perceptron, fusing the local time domain features and the global features into multi-scale spatial feature representation, and the decoder is used for reconstructing the input surface electromyographic signals according to the multi-scale spatial feature representation and optimizing the multi-scale spatial feature representation output by the encoder by taking reconstruction errors as self-supervision learning targets; the time-dependent modeling module is used for capturing a dynamic time-dependent relation between muscle activity and joint moment from the multi-scale space feature representation through a long-period memory network and a concentration mechanism to generate a dynamic time sequence feature representation; The regression prediction module is used for flattening the dynamic time sequence characteristic representation, realizing nonlinear regression mapping between the flattened dynamic time sequence characteristic representation and the target joint moment through a fully connected network, and outputting a joint moment predicted value.
5. The robust joint moment predicting method based on self-supervision reconstruction of claim 1, wherein the step S4 specifically comprises the following steps: and training the joint moment prediction model through the training set until a preset early-stop condition is met, verifying the joint moment prediction model through the verification set at the end of each training batch to optimize the hyper-parameters of the joint moment prediction model, then calculating RMSE, MAE, R and VAF through the testing set to test the trained joint moment prediction model, and deploying the joint moment prediction model passing the test.
6. A robust joint moment prediction system based on self-supervision reconstruction is characterized by comprising the following modules: The data set construction module is used for acquiring a large number of historical surface electromyographic signals of different channels, preprocessing each historical surface electromyographic signal and constructing a data set after marking; the data set dividing module is used for dividing the data set into a training set, a verification set and a test set based on a preset proportion; The joint moment prediction model creation module is used for creating a joint moment prediction model based on the self-supervision reconstruction feature extraction module, the time-dependent modeling module and the regression prediction module; The joint moment prediction model training module is used for respectively training, verifying and testing the joint moment prediction model through the training set, the verification set and the test set, and deploying the joint moment prediction model which passes the test; and the joint moment prediction module is used for predicting the joint moment through the deployed joint moment prediction model.
7. The robust joint moment prediction system based on self-supervised reconstruction as recited in claim 6, wherein the dataset construction module is configured to: Acquiring a large number of historical surface electromyographic signals of different subjects in different channels, preprocessing at least comprising filtering, normalizing and sliding window segmentation on each historical surface electromyographic signal, and labeling joint moment values on each preprocessed historical surface electromyographic signal to construct a data set.
8. The robust joint moment prediction system based on self-supervised reconstruction as recited in claim 6, wherein the dataset partitioning module is configured to: the data set is divided into a training set, a validation set, and a test set in a ratio of 7:2:1 based on a time sequence.
9. The robust joint moment prediction system based on self-supervised reconstruction of claim 6, wherein in the joint moment prediction model creation module, the self-supervised reconstruction feature extraction module is constructed based on an encoder and a decoder, the encoder is used for extracting local time domain features representing transient muscle activation changes from surface electromyographic signals through a one-dimensional convolution layer, learning long-term dependence of the surface electromyographic signals through a multi-layer perceptron to extract global features, fusing the local time domain features and the global features into a multi-scale spatial feature representation, and the decoder is used for reconstructing the input surface electromyographic signals according to the multi-scale spatial feature representation and optimizing the multi-scale spatial feature representation output by the encoder by taking reconstruction errors as self-supervised learning targets; the time-dependent modeling module is used for capturing a dynamic time-dependent relation between muscle activity and joint moment from the multi-scale space feature representation through a long-period memory network and a concentration mechanism to generate a dynamic time sequence feature representation; The regression prediction module is used for flattening the dynamic time sequence characteristic representation, realizing nonlinear regression mapping between the flattened dynamic time sequence characteristic representation and the target joint moment through a fully connected network, and outputting a joint moment predicted value.
10. The robust joint moment prediction system based on self-supervised reconstruction as recited in claim 6, wherein the joint moment prediction model training module is configured to: and training the joint moment prediction model through the training set until a preset early-stop condition is met, verifying the joint moment prediction model through the verification set at the end of each training batch to optimize the hyper-parameters of the joint moment prediction model, then calculating RMSE, MAE, R and VAF through the testing set to test the trained joint moment prediction model, and deploying the joint moment prediction model passing the test.

Description

Robust joint moment prediction method and system based on self-supervision reconstruction Technical Field The invention relates to the technical field of biomechanical signal processing and machine learning intersection, in particular to a robust joint moment prediction method and system based on self-supervision reconstruction. Background The joint moment is a core biomechanical parameter for quantifying the motion function of a human body, evaluating the health condition of a musculoskeletal system, and realizing the accurate control of intelligent rehabilitation equipment such as an exoskeleton and a robot auxiliary system. In fields such as sports medicine, rehabilitation engineering, sports science and the like, the method has important scientific research value and application prospect for realizing accurate and continuous prediction of joint moment. At present, the acquisition of joint moment mainly depends on the following two technical routes: The first is a laboratory environment-based inverse kinetic analysis method. The method generally needs to synchronously acquire human body kinematic data and ground reaction force by means of expensive equipment such as a high-speed optical motion capturing system, a force measuring table and the like, and then calculates joint moment by combining a multi-body dynamics model. Although the method has higher precision, the method has obvious limitations of high equipment cost, strict requirements on experimental environment and complex construction flow, and is difficult to be suitable for continuous and long-term monitoring in daily activities or real scenes, thereby greatly limiting the popularization and practical application range. The second category is data-driven based prediction methods. With the development of wearable sensing technology, predicting joint moment using surface myoelectricity (sEMG) signals has become a research hotspot. The sEMG signal can directly reflect the nerve activation state of the muscle and has physiological relevance to moment generation. In the early method, a linear model or a simple machine learning model is mostly adopted, and the mapping relation between the sEMG signal and the joint moment is established by extracting the time domain and frequency domain characteristics of the sEMG signal. In recent years, deep learning models (such as convolutional neural network CNN and cyclic neural network RNN) have been shown to have superior performance in the field by virtue of strong feature automatic extraction and sequence modeling capabilities. However, existing deep learning-based prediction methods still face several serious challenges: (1) Vulnerability under noise interference in practical application, sEMG signals collected by the wearable equipment are easily affected by various noises such as power frequency interference of a power supply, motion artifacts, poor electrode contact and the like, and the signal quality may fluctuate severely. The traditional supervised learning model is mostly based on relatively clean laboratory data training, and in a real scene with low signal-to-noise ratio and complex noise, the prediction performance of the traditional supervised learning model is often obviously reduced, and poor robustness and generalization capability are shown. (2) The characteristic representation has insufficient robustness, the existing method usually uses joint moment prediction as a single target to perform end-to-end training, the learned characteristic representation is easy to overfit to a specific mode (including a noise mode) in training data, and the essential characteristic of the muscle activity which is less affected by noise and more universal cannot be sufficiently learned, so that the prediction result is unstable when the signal quality changes. (3) The self-supervision learning fusion application lacks that self-supervision learning has proved to learn robust feature representation from unlabeled data in the fields of computer vision, natural language processing and the like, but in the specific technical problem of joint moment prediction, how to effectively combine the advantages of the self-supervision learning fusion application (such as learning anti-noise features through a reconstruction task) with a downstream supervision prediction task to construct a unified framework capable of reducing the dependence on labeling data and remarkably improving the prediction stability in a noise environment, and still remains a technical problem to be broken through in urgent need at present. Therefore, how to provide a robust joint moment prediction method and system based on self-supervision reconstruction, so as to improve the precision, robustness and generalization capability of joint moment prediction, is a technical problem to be solved urgently. Disclosure of Invention The invention aims to solve the technical problem of providing a robust joint moment prediction method and a robust joint mo