Robust joint torque prediction method and system based on self-supervised reconstruction

Abstract

The invention provides a robust joint torque prediction method and system based on self-supervised reconstruction in the technical field of biomechanical signal processing and machine learning crossing. The method comprises the steps that S1, a large number of historical surface electromyogram signals are obtained, preprocessed and labeled, and then a data set is constructed; 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 torque prediction model based on a self-supervised reconstruction feature extraction module, a time-dependent modeling module and a regression prediction module; s4, through the training set, the verification set and the test set, training, verifying and testing the joint torque prediction model, and deploying the joint torque prediction model passing the test; and S5, joint torque prediction is carried out through the deployed joint torque prediction model. The joint torque prediction method has the advantages that the joint torque prediction precision, robustness and generalization ability are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of biomechanical signal processing and machine learning, and in particular to a robust joint torque prediction method and system based on self-supervised reconstruction. Background Technology

[0002] Joint torque is a core biomechanical parameter used to quantify human motor function, assess the health of the musculoskeletal system, and achieve precise control of intelligent rehabilitation equipment such as exoskeletons and robot-assisted systems. Accurate and continuous prediction of joint torque has significant research value and application prospects in fields such as sports medicine, rehabilitation engineering, and sports science.

[0003] Currently, obtaining joint torque mainly relies on the following two technical approaches: The first category is inverse dynamics analysis methods based on laboratory environments. This method typically requires expensive equipment such as high-speed optical motion capture systems and force tables to simultaneously collect human kinematic data and ground reaction forces, then combine this with multibody dynamics models to calculate joint torques. While this method offers high accuracy, it has significant limitations: high equipment costs, stringent experimental environment requirements, complex setup procedures, and difficulty in applying it to continuous, long-term monitoring in daily activities or real-world scenarios, thus greatly limiting its widespread adoption and practical application.

[0004] The second category comprises data-driven prediction methods. With the development of wearable sensing technology, predicting joint torque using surface electromyography (sEMG) signals has become a research hotspot. sEMG signals directly reflect the neural activation state of muscles and are physiologically related to torque generation. Early methods often employed linear models or simple machine learning models, extracting time-domain and frequency-domain features of sEMG signals to establish a mapping relationship between them and joint torque. In recent years, deep learning models (such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have demonstrated superior performance in this field due to their powerful automatic feature extraction and sequence modeling capabilities.

[0005] However, existing deep learning-based prediction methods still face several serious challenges: (1) Vulnerability to noise interference: In practical applications, the sEMG signals collected by wearable devices are susceptible to various noises such as power frequency interference, motion artifacts, and poor electrode contact, and the signal quality may fluctuate drastically. Traditional supervised learning models are mostly trained on relatively "clean" laboratory data. In real-world scenarios with low signal-to-noise ratio and complex noise, their prediction performance often drops significantly, showing poor robustness and generalization ability.

[0006] (2) Insufficient robustness of feature representation: Existing methods usually use joint torque prediction as the single objective for end-to-end training. The learned feature representation is prone to overfitting to specific patterns (including noise patterns) in the training data. It fails to fully learn the essential features of muscle activity that are less affected by noise and more universal, resulting in unstable prediction results when the signal quality changes.

[0007] (3) Lack of integration and application of self-supervised learning: Self-supervised learning has been proven to learn robust feature representations from unlabeled data in fields such as computer vision and natural language processing. However, in the specific technical problem of joint moment prediction, how to effectively combine its advantages (such as learning noise-resistant features through reconstruction tasks) with downstream supervised prediction tasks and build a unified framework that can reduce dependence on labeled data and significantly improve prediction stability in noisy environments is still a technical problem that needs to be solved.

[0008] Therefore, how to provide a robust joint torque prediction method and system based on self-supervised reconstruction to improve the accuracy, robustness and generalization ability of joint torque prediction has become an urgent technical problem to be solved. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide a robust joint torque prediction method and system based on self-supervised reconstruction, so as to improve the accuracy, robustness and generalization ability of joint torque prediction.

[0010] In a first aspect, the present invention provides a robust joint torque prediction method based on self-supervised reconstruction, comprising the following steps: Step S1: Obtain a large number of historical surface electromyography (EMG) signals from different channels, preprocess and label each historical surface EMG signal, and then construct a dataset. Step S2: Divide the dataset into a training set, a validation set, and a test set based on a preset ratio; Step S3: Create a joint torque prediction model based on the self-supervised reconstruction feature extraction module, time-dependent modeling module, and regression prediction module; Step S4: Train, validate, and test the joint torque prediction model using the training set, validation set, and test set, respectively, and deploy the joint torque prediction model that passes the test. Step S5: Predict joint torque using the deployed joint torque prediction model.

[0011] Furthermore, step S1 specifically includes: A large number of historical surface electromyography (EMG) signals from different subjects and different channels were acquired. Each historical surface EMG signal was preprocessed, including filtering, normalization and sliding window segmentation. After labeling the joint torque values ​​of each preprocessed historical surface EMG signal, a dataset was constructed.

[0012] Furthermore, step S2 specifically includes: Based on chronological order, the dataset is divided into training, validation, and test sets in a ratio of 7:2:1.

[0013] Furthermore, in step S3, the self-supervised reconstruction feature extraction module is constructed based on an encoder and a decoder; the encoder is used to extract local temporal features representing instantaneous muscle activation changes from the surface electromyography (EMG) signal through a one-dimensional convolutional layer, learn the long-term dependencies of the EMG signal through a multilayer perceptron to extract global features, and fuse the local temporal features and global features into a multi-scale spatial feature representation; the decoder is used to reconstruct the input EMG signal based on the multi-scale spatial feature representation, and use the reconstruction error as a self-supervised learning objective to optimize the multi-scale spatial feature representation output by the encoder. The time-dependent modeling module is used to capture the dynamic time dependency between muscle activity and joint torque from the multi-scale spatial feature representation by combining a long short-term memory network with an attention mechanism, and generate a dynamic temporal feature representation. The regression prediction module is used to flatten the dynamic temporal feature representation and realize the nonlinear regression mapping between the flattened dynamic temporal feature representation and the target joint torque through a fully connected network, and output the joint torque prediction value.

[0014] Furthermore, step S4 specifically includes: The joint torque prediction model is trained using the training set until a preset early stopping condition is met. At the end of each training batch, the joint torque prediction model is validated using the validation set to optimize the hyperparameters of the joint torque prediction model. Then, RMSE, MAE, R², and VAF are calculated using the test set to test the trained joint torque prediction model. The joint torque prediction model that passes the test is then deployed.

[0015] Secondly, the present invention provides a robust joint torque prediction system based on self-supervised reconstruction, comprising the following modules: The dataset construction module is used to acquire a large number of historical surface electromyography (EMG) signals from different channels, and to construct the dataset after preprocessing and labeling each of the historical surface EMG signals. The dataset partitioning module is used to divide the dataset into a training set, a validation set, and a test set based on a preset ratio. The joint torque prediction model creation module is used to create a joint torque prediction model based on the self-supervised reconstruction feature extraction module, the time-dependent modeling module, and the regression prediction module. The joint torque prediction model training module is used to train, validate, and test the joint torque prediction model using the training set, validation set, and test set, respectively, and to deploy the joint torque prediction model that passes the test. A joint torque prediction module is used to predict joint torques using the deployed joint torque prediction model.

[0016] Furthermore, the dataset construction module is specifically used for: A large number of historical surface electromyography (EMG) signals from different subjects and different channels were acquired. Each historical surface EMG signal was preprocessed, including filtering, normalization and sliding window segmentation. After labeling the joint torque values ​​of each preprocessed historical surface EMG signal, a dataset was constructed.

[0017] Furthermore, the dataset partitioning module is specifically used for: Based on chronological order, the dataset is divided into training, validation, and test sets in a ratio of 7:2:1.

[0018] Furthermore, in the joint torque prediction model creation module, the self-supervised reconstruction feature extraction module is built based on an encoder and a decoder. The encoder is used to extract local temporal features representing instantaneous muscle activation changes from surface electromyography (EMG) signals through a one-dimensional convolutional layer, learn the long-term dependencies of EMG signals through a multilayer perceptron to extract global features, and fuse the local temporal features and global features into a multi-scale spatial feature representation. The decoder is used to reconstruct the input EMG signals based on the multi-scale spatial feature representation, and uses the reconstruction error as a self-supervised learning objective to optimize the multi-scale spatial feature representation output by the encoder. The time-dependent modeling module is used to capture the dynamic time dependency between muscle activity and joint torque from the multi-scale spatial feature representation by combining a long short-term memory network with an attention mechanism, and generate a dynamic temporal feature representation. The regression prediction module is used to flatten the dynamic temporal feature representation and realize the nonlinear regression mapping between the flattened dynamic temporal feature representation and the target joint torque through a fully connected network, and output the joint torque prediction value.

[0019] Furthermore, the joint torque prediction model training module is specifically used for: The joint torque prediction model is trained using the training set until a preset early stopping condition is met. At the end of each training batch, the joint torque prediction model is validated using the validation set to optimize the hyperparameters of the joint torque prediction model. Then, RMSE, MAE, R², and VAF are calculated using the test set to test the trained joint torque prediction model. The joint torque prediction model that passes the test is then deployed.

[0020] The advantages of this invention are: 1. By acquiring a large number of historical surface electromyography (EMG) signals from different channels, a dataset is constructed after preprocessing and labeling of each historical surface EMG signal. The dataset is divided into training, validation, and test sets. Then, a joint torque prediction model is created based on a self-supervised reconstruction feature extraction module, a time-dependent modeling module, and a regression prediction module. The joint torque prediction model is trained, validated, and tested using the training, validation, and test sets, respectively. The joint torque prediction model that passes the test is deployed. Finally, joint torque prediction is performed using the deployed joint torque prediction model. That is, by introducing an encoder-decoder architecture... The self-supervised reconstruction feature extraction module reconstructs the original electromyographic signals to learn their essential features. This reconstruction task forces the encoder to ignore random noise and extract robust multi-scale spatial feature representations. These optimized multi-scale spatial feature representations are then input into a time-dependent modeling module that incorporates an attention mechanism to accurately capture the dynamic temporal feature representations between muscle activity and joint torque. Finally, a nonlinear regression module achieves accurate mapping. This process reduces the model's sensitivity to noise from the signal source, making the learned features more universal and ultimately greatly improving the accuracy, robustness, and generalization ability of joint torque prediction.

[0021] 2. By introducing a self-supervised reconstruction module (encoder-decoder structure), the encoder extracts local temporal features and global long-term dependencies of electromyography (EMG) signals through one-dimensional convolution and multilayer perceptron, respectively, and performs multi-scale fusion; the decoder drives the encoder to learn more discriminative and robust feature representations by reconstructing the original signal. This pre-training method does not rely on a large amount of labeled data and can learn the intrinsic structure and patterns of EMG signals using unlabeled data. It effectively overcomes the problem that surface electromyography (sEMG) signals are susceptible to noise from individual differences, electrode displacement, skin impedance, and cross-interference, thereby enhancing the stability and adaptability of the model in complex real-world environments.

[0022] 3. The self-supervised module extracts information-rich "multi-scale spatial feature representations"; the temporal dependency modeling module (combining LSTM and attention mechanisms) specifically handles the temporal evolution of these features to capture the dynamic causal and dependency relationship between muscle activity and torque generation; finally, the regression module completes the mapping; this cascaded design of "spatial feature extraction → temporal relationship modeling → regression prediction" realizes hierarchical analysis from static signal features to dynamic physiological processes, with clear physical meaning; in particular, the introduction of the attention mechanism enables the model to adaptively focus on the key electromyographic activity periods most relevant to torque output, improving the model's interpretability and prediction accuracy, constituting an important technical feature that distinguishes it from simple end-to-end black box models.

[0023] 4. In the feature extraction module, one-dimensional convolutional layers efficiently capture local temporal patterns, while multilayer perceptrons (MLPs) learn global interactions. This combination achieves a balance between computational efficiency and representational capability. The entire model is based on deep learning and is highly modular, making it easy to adjust and transfer to different joints (such as knees and elbows), different torque ranges, or different acquisition devices. For example, the objective function of the self-supervised reconstruction task can be optimized relatively independently, or LSTM can be replaced with other temporal networks (such as Transformers) without affecting the overall framework.

[0024] 5. By introducing a self-supervised reconstruction feature extraction module, the encoder-decoder structure is used to learn multi-scale spatial features of electromyography signals on unlabeled data, effectively improving the model's robustness to noise and individual differences. Furthermore, by combining long short-term memory networks and attention mechanisms, the model accurately models temporal dynamic dependencies, realizing the complex mapping analysis from muscle activation to joint torque. The entire process includes rigorous data preprocessing, time-partitioned dataset processing, and multi-index evaluation, ensuring the reliability and reproducibility of the method. Its modular design balances computational efficiency and scalability, providing an innovative solution for high-precision and high-stability joint torque prediction. Attached Figure Description

[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0026] Figure 1 This is a flowchart of a robust joint torque prediction method based on self-supervised reconstruction according to the present invention.

[0027] Figure 2 This is a schematic diagram of the structure of a robust joint torque prediction system based on self-supervised reconstruction according to the present invention.

[0028] Figure 3 This is a flowchart illustrating the joint torque prediction model of the present invention.

[0029] Figure 4 This is a flowchart illustrating the encoder of the present invention.

[0030] Figure 5 This is a flowchart illustrating the decoder of the present invention. Detailed Implementation

[0031] The overall approach of the technical solution in this application is as follows: By introducing a self-supervised reconstruction feature extraction module based on an encoder-decoder architecture, the original electromyographic signal is reconstructed to learn its essential features. This reconstruction task forces the encoder to ignore random noise and extract robust multi-scale spatial feature representations. These optimized multi-scale spatial feature representations are then input into a time-dependent modeling module that incorporates an attention mechanism to accurately capture the dynamic temporal feature representations between muscle activity and joint torque. Finally, a nonlinear regression module is used to achieve accurate mapping. This process reduces the model's sensitivity to noise from the signal source, making the learned features more universal, thereby improving the accuracy, robustness, and generalization ability of joint torque prediction.

[0032] Please refer to Figures 1 to 5 As shown, a preferred embodiment of the robust joint torque prediction method based on self-supervised reconstruction of the present invention includes the following steps: Step S1: Obtain a large number of historical surface electromyography (EMG) signals from different channels, preprocess and label each historical surface EMG signal, and then construct a dataset. Step S2: Divide the dataset into a training set, a validation set, and a test set based on a preset ratio; Step S3: Create a joint torque prediction model based on the self-supervised reconstruction feature extraction module, time-dependent modeling module, and regression prediction module; Step S4: Train, validate, and test the joint torque prediction model using the training set, validation set, and test set, respectively, and deploy the joint torque prediction model that passes the test. Step S5: Perform joint torque prediction using the deployed joint torque prediction model to achieve time-by-time prediction of hip flexion / extension (Hip FE) and hip adduction / abduction (Hip AA) torques.

[0033] Step S1 specifically involves: A large number of historical surface electromyography (EMG) signals from different subjects and different channels were acquired. Each historical EMG signal was preprocessed by at least the following methods: filtering (using filtering methods to suppress interference such as electrical noise and motion artifacts to improve the usability of the original signal, such as using 2Hz high-pass filtering), normalization, and sliding window segmentation (using sliding window technology to perform time-series segmentation to meet the needs of continuous state monitoring and real-time torque prediction). After labeling the joint torque values ​​of each preprocessed historical EMG signal, a dataset was constructed.

[0034] A channel represents an independent signal acquisition unit, which is a single independent muscle electrical signal recorded from a specific location on the skin surface.

[0035] In practice, data acquisition should encompass multiple subjects (e.g., 10-20 healthy adults) and multi-channel sEMG signals (e.g., 8-16 channels) to ensure data diversity. The preprocessing stage needs to be specific: filtering can use a 2Hz high-pass filter to eliminate baseline drift; normalization uses the z-score method to scale the signal to zero mean and unit variance; for sliding window segmentation, a window length of 200ms and an overlap rate of 50% are recommended to balance temporal resolution and computational efficiency. When labeling joint torque values, inverse dynamics analysis equipment (such as force tables and motion capture systems) should be used simultaneously to obtain the true torque values ​​to ensure labeling accuracy. This step lays the data foundation for model training, improving the model's generalization ability through diversified data.

[0036] Step S2 specifically involves: Based on chronological order, the dataset is divided into training, validation, and test sets in a ratio of 7:2:1.

[0037] The partitioning should be based on chronological order to avoid data leakage. For example, for continuously acquired sEMG signals, first sort them by time series, then partition them proportionally: the training set covers the first 70% of the data for model learning; the validation set covers the middle 20% for hyperparameter tuning; and the test set covers the last 10% for final evaluation. This temporal partitioning simulates the data flow in real-world scenarios, ensuring model performance on unknown data. The partition ratio can be adjusted according to the amount of data; for example, when the data volume is large, the test set ratio can be appropriately increased to 20% to improve evaluation reliability.

[0038] In step S3, the self-supervised reconstruction feature extraction module is constructed based on an encoder and a decoder. The encoder is used to extract local temporal features representing instantaneous muscle activation changes from surface electromyography (sEMG) signals through a one-dimensional convolutional layer, and to learn the long-term dependencies of sEMG signals through a multilayer perceptron to extract global features. The local temporal features and global features are then fused into a multi-scale spatial feature representation. The decoder is used to reconstruct the input sEMG signal based on the multi-scale spatial feature representation, and uses the reconstruction error as a self-supervised learning objective to optimize the multi-scale spatial feature representation output by the encoder. This enables the model to automatically learn feature representations with strong noise resistance, thereby enhancing its stability in real-world noisy environments. The self-supervised reconstruction feature extraction module is used to learn robust latent feature representations from the original sEMG signal. This module does not rely on manually labeled data, but rather guides the model to extract low-dimensional feature representations with good generalization through the pre-task of reconstructing the original signal. Assuming the input surface electromyography signal is ,in, B represents a real number, C represents the batch size, T represents the number of channels, and T represents the time series length. The local temporal features extracted by the one-dimensional convolutional layer from the surface electromyography signal input are represented as follows: ; in, Indicates the first Local temporal feature representation of the layer output; Represents the ReLU activation function; Indicates the convolution kernel weights; Indicates the kernel bias; Indicates the first Local temporal feature representation of the layer output; The time-dependent modeling module is used to capture the dynamic time dependency between muscle activity and joint torque from the multi-scale spatial feature representation by combining a long short-term memory network with an attention mechanism, generating a dynamic temporal feature representation. Since the generation of joint torque has a significant temporal correlation, a long short-term memory network (LSTM) is used to perform sequence modeling on the encoded multi-scale spatial feature representation to capture the dynamic relationship between muscle activity and torque changes. In addition, an attention mechanism is introduced to adaptively weight the importance of different time slices, enabling the network to focus on key muscle activation segments and further improve the prediction ability. The regression prediction module flattens the dynamic temporal feature representation and uses a fully connected network to achieve a nonlinear regression mapping between the flattened dynamic temporal feature representation and the target joint torque, outputting the predicted joint torque value. The fully connected network achieves an accurate correspondence between high-dimensional features and the target torque output through nonlinear transformation, thereby completing the joint torque prediction process.

[0039] Step S4 specifically involves: The joint torque prediction model is trained using the training set until a preset early stopping condition is met. At the end of each training batch, the joint torque prediction model is validated using the validation set to optimize the hyperparameters of the joint torque prediction model. Then, RMSE, MAE, R², and VAF are calculated using the test set to test the trained joint torque prediction model. The joint torque prediction model that passes the test is then deployed.

[0040] Training uses the Adam optimizer with a learning rate of 0.001. Early stopping can be set to ensure the validation set loss does not decrease for 5 consecutive epochs. Validation is performed after each batch, and hyperparameters such as dropout can be adjusted to 0.3 to prevent overfitting.

[0041] A preferred embodiment of the robust joint torque prediction system based on self-supervised reconstruction of the present invention includes the following modules: The dataset construction module is used to acquire a large number of historical surface electromyography (EMG) signals from different channels, and to construct the dataset after preprocessing and labeling each of the historical surface EMG signals. The dataset partitioning module is used to divide the dataset into a training set, a validation set, and a test set based on a preset ratio. The joint torque prediction model creation module is used to create a joint torque prediction model based on the self-supervised reconstruction feature extraction module, the time-dependent modeling module, and the regression prediction module. The joint torque prediction model training module is used to train, validate, and test the joint torque prediction model using the training set, validation set, and test set, respectively, and to deploy the joint torque prediction model that passes the test. The joint torque prediction module is used to predict joint torques through the deployed joint torque prediction model, thereby achieving time-by-time prediction of hip flexion / extension (Hip FE) and hip adduction / abduction (Hip AA) torques.

[0042] The dataset construction module is specifically used for: A large number of historical surface electromyography (EMG) signals from different subjects and different channels were acquired. Each historical EMG signal was preprocessed by at least the following methods: filtering (using filtering methods to suppress interference such as electrical noise and motion artifacts to improve the usability of the original signal, such as using 2Hz high-pass filtering), normalization, and sliding window segmentation (using sliding window technology to perform time-series segmentation to meet the needs of continuous state monitoring and real-time torque prediction). After labeling the joint torque values ​​of each preprocessed historical EMG signal, a dataset was constructed.

[0043] A channel represents an independent signal acquisition unit, which is a single independent muscle electrical signal recorded from a specific location on the skin surface.

[0044] In practice, data acquisition should encompass multiple subjects (e.g., 10-20 healthy adults) and multi-channel sEMG signals (e.g., 8-16 channels) to ensure data diversity. The preprocessing stage needs to be specific: filtering can use a 2Hz high-pass filter to eliminate baseline drift; normalization uses the z-score method to scale the signal to zero mean and unit variance; for sliding window segmentation, a window length of 200ms and an overlap rate of 50% are recommended to balance temporal resolution and computational efficiency. When labeling joint torque values, inverse dynamics analysis equipment (such as force tables and motion capture systems) should be used simultaneously to obtain the true torque values ​​to ensure labeling accuracy. This step lays the data foundation for model training, improving the model's generalization ability through diversified data.

[0045] The dataset partitioning module is specifically used for: Based on chronological order, the dataset is divided into training, validation, and test sets in a ratio of 7:2:1.

[0046] The partitioning should be based on chronological order to avoid data leakage. For example, for continuously acquired sEMG signals, first sort them by time series, then partition them proportionally: the training set covers the first 70% of the data for model learning; the validation set covers the middle 20% for hyperparameter tuning; and the test set covers the last 10% for final evaluation. This temporal partitioning simulates the data flow in real-world scenarios, ensuring model performance on unknown data. The partition ratio can be adjusted according to the amount of data; for example, when the data volume is large, the test set ratio can be appropriately increased to 20% to improve evaluation reliability.

[0047] In the joint torque prediction model creation module, the self-supervised reconstruction feature extraction module is built based on an encoder and a decoder. The encoder is used to extract local temporal features representing instantaneous muscle activation changes from surface electromyography (sEMG) signals through a one-dimensional convolutional layer, and to learn the long-term dependencies of sEMG signals through a multilayer perceptron to extract global features. The local temporal features and global features are then fused into a multi-scale spatial feature representation. The decoder is used to reconstruct the input sEMG signals based on the multi-scale spatial feature representation, and uses the reconstruction error as a self-supervised learning objective to optimize the multi-scale spatial feature representation output by the encoder. This enables the model to automatically learn feature representations with strong noise resistance, thereby enhancing its stability in real-world noisy environments. The self-supervised reconstruction feature extraction module is used to learn robust latent feature representations from the original sEMG signals. This module does not rely on manually labeled data, but rather guides the model to extract low-dimensional feature representations with good generalization through the pre-task of reconstructing the original signal. Assuming the input surface electromyography signal is ,in, B represents a real number, C represents the batch size, T represents the number of channels, and T represents the time series length. The local temporal features extracted by the one-dimensional convolutional layer from the surface electromyography signal input are represented as follows: ; in, Indicates the first Local temporal feature representation of the layer output; Represents the ReLU activation function; Indicates the convolution kernel weights; Indicates the kernel bias; Indicates the first Local temporal feature representation of the layer output; The time-dependent modeling module is used to capture the dynamic time dependency between muscle activity and joint torque from the multi-scale spatial feature representation by combining a long short-term memory network with an attention mechanism, generating a dynamic temporal feature representation. Since the generation of joint torque has a significant temporal correlation, a long short-term memory network (LSTM) is used to perform sequence modeling on the encoded multi-scale spatial feature representation to capture the dynamic relationship between muscle activity and torque changes. In addition, an attention mechanism is introduced to adaptively weight the importance of different time slices, enabling the network to focus on key muscle activation segments and further improve the prediction ability. The regression prediction module flattens the dynamic temporal feature representation and uses a fully connected network to achieve a nonlinear regression mapping between the flattened dynamic temporal feature representation and the target joint torque, outputting the predicted joint torque value. The fully connected network achieves an accurate correspondence between high-dimensional features and the target torque output through nonlinear transformation, thereby completing the joint torque prediction process.

[0048] The joint torque prediction model training module is specifically used for: The joint torque prediction model is trained using the training set until a preset early stopping condition is met. At the end of each training batch, the joint torque prediction model is validated using the validation set to optimize the hyperparameters of the joint torque prediction model. Then, RMSE, MAE, R², and VAF are calculated using the test set to test the trained joint torque prediction model. The joint torque prediction model that passes the test is then deployed.

[0049] Training uses the Adam optimizer with a learning rate of 0.001. Early stopping can be set to ensure the validation set loss does not decrease for 5 consecutive epochs. Validation is performed after each batch, and hyperparameters such as dropout can be adjusted to 0.3 to prevent overfitting.

[0050] In summary, the advantages of this invention are as follows: 1. By acquiring a large number of historical surface electromyography (EMG) signals from different channels, a dataset is constructed after preprocessing and labeling of each historical surface EMG signal. The dataset is divided into training, validation, and test sets. Then, a joint torque prediction model is created based on a self-supervised reconstruction feature extraction module, a time-dependent modeling module, and a regression prediction module. The joint torque prediction model is trained, validated, and tested using the training, validation, and test sets, respectively. The joint torque prediction model that passes the test is deployed. Finally, joint torque prediction is performed using the deployed joint torque prediction model. That is, by introducing an encoder-decoder architecture... The self-supervised reconstruction feature extraction module reconstructs the original electromyographic signals to learn their essential features. This reconstruction task forces the encoder to ignore random noise and extract robust multi-scale spatial feature representations. These optimized multi-scale spatial feature representations are then input into a time-dependent modeling module that incorporates an attention mechanism to accurately capture the dynamic temporal feature representations between muscle activity and joint torque. Finally, a nonlinear regression module achieves accurate mapping. This process reduces the model's sensitivity to noise from the signal source, making the learned features more universal and ultimately greatly improving the accuracy, robustness, and generalization ability of joint torque prediction.

[0051] 2. By introducing a self-supervised reconstruction module (encoder-decoder structure), the encoder extracts local temporal features and global long-term dependencies of electromyography (EMG) signals through one-dimensional convolution and multilayer perceptron, respectively, and performs multi-scale fusion; the decoder drives the encoder to learn more discriminative and robust feature representations by reconstructing the original signal. This pre-training method does not rely on a large amount of labeled data and can learn the intrinsic structure and patterns of EMG signals using unlabeled data. It effectively overcomes the problem that surface electromyography (sEMG) signals are susceptible to noise from individual differences, electrode displacement, skin impedance, and cross-interference, thereby enhancing the stability and adaptability of the model in complex real-world environments.

[0052] 3. The self-supervised module extracts information-rich "multi-scale spatial feature representations"; the temporal dependency modeling module (combining LSTM and attention mechanisms) specifically handles the temporal evolution of these features to capture the dynamic causal and dependency relationship between muscle activity and torque generation; finally, the regression module completes the mapping; this cascaded design of "spatial feature extraction → temporal relationship modeling → regression prediction" realizes hierarchical analysis from static signal features to dynamic physiological processes, with clear physical meaning; in particular, the introduction of the attention mechanism enables the model to adaptively focus on the key electromyographic activity periods most relevant to torque output, improving the model's interpretability and prediction accuracy, constituting an important technical feature that distinguishes it from simple end-to-end black box models.

[0053] 4. In the feature extraction module, one-dimensional convolutional layers efficiently capture local temporal patterns, while multilayer perceptrons (MLPs) learn global interactions. This combination achieves a balance between computational efficiency and representational capability. The entire model is based on deep learning and is highly modular, making it easy to adjust and transfer to different joints (such as knees and elbows), different torque ranges, or different acquisition devices. For example, the objective function of the self-supervised reconstruction task can be optimized relatively independently, or LSTM can be replaced with other temporal networks (such as Transformers) without affecting the overall framework.

[0054] 5. By introducing a self-supervised reconstruction feature extraction module, the encoder-decoder structure is used to learn multi-scale spatial features of electromyography signals on unlabeled data, effectively improving the model's robustness to noise and individual differences. Furthermore, by combining long short-term memory networks and attention mechanisms, the model accurately models temporal dynamic dependencies, realizing the complex mapping analysis from muscle activation to joint torque. The entire process includes rigorous data preprocessing, time-partitioned dataset processing, and multi-index evaluation, ensuring the reliability and reproducibility of the method. Its modular design balances computational efficiency and scalability, providing an innovative solution for high-precision and high-stability joint torque prediction.

[0055] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A robust joint torque prediction method based on self-supervised reconstruction, characterized in that: Includes the following steps: Step S1: Obtain a large number of historical surface electromyography (EMG) signals from different channels, preprocess and label each historical surface EMG signal, and then construct a dataset. Step S2: Divide the dataset into a training set, a validation set, and a test set based on a preset ratio; Step S3: Create a joint torque prediction model based on the self-supervised reconstruction feature extraction module, time-dependent modeling module, and regression prediction module; Step S4: Train, validate, and test the joint torque prediction model using the training set, validation set, and test set, respectively, and deploy the joint torque prediction model that passes the test. Step S5: Predict joint torque using the deployed joint torque prediction model.

2. The robust joint torque prediction method based on self-supervised reconstruction as described in claim 1, characterized in that: Step S1 specifically involves: A large number of historical surface electromyography (EMG) signals from different subjects and different channels were acquired. Each historical surface EMG signal was preprocessed, including filtering, normalization and sliding window segmentation. After labeling the joint torque values ​​of each preprocessed historical surface EMG signal, a dataset was constructed.

3. The robust joint torque prediction method based on self-supervised reconstruction as described in claim 1, characterized in that: Step S2 specifically involves: Based on chronological order, the dataset is divided into training, validation, and test sets in a ratio of 7:2:

1.

4. The robust joint torque prediction method based on self-supervised reconstruction as described in claim 1, characterized in that: In step S3, the self-supervised reconstruction feature extraction module is constructed based on an encoder and a decoder. The encoder is used to extract local temporal features representing instantaneous muscle activation changes from the surface electromyography (EMG) signal through a one-dimensional convolutional layer, learn the long-term dependencies of the EMG signal through a multilayer perceptron to extract global features, and fuse the local temporal features and global features into a multi-scale spatial feature representation. The decoder is used to reconstruct the input EMG signal based on the multi-scale spatial feature representation, and uses the reconstruction error as a self-supervised learning objective to optimize the multi-scale spatial feature representation output by the encoder. The time-dependent modeling module is used to capture the dynamic time dependency between muscle activity and joint torque from the multi-scale spatial feature representation by combining a long short-term memory network with an attention mechanism, and generate a dynamic temporal feature representation. The regression prediction module is used to flatten the dynamic temporal feature representation and realize the nonlinear regression mapping between the flattened dynamic temporal feature representation and the target joint torque through a fully connected network, and output the joint torque prediction value.

5. The robust joint torque prediction method based on self-supervised reconstruction as described in claim 1, characterized in that: Step S4 specifically involves: The joint torque prediction model is trained using the training set until a preset early stopping condition is met. At the end of each training batch, the joint torque prediction model is validated using the validation set to optimize the hyperparameters of the joint torque prediction model. Then, RMSE, MAE, R², and VAF are calculated using the test set to test the trained joint torque prediction model. The joint torque prediction model that passes the test is then deployed.

6. A robust joint torque prediction system based on self-supervised reconstruction, characterized in that: Includes the following modules: The dataset construction module is used to acquire a large number of historical surface electromyography (EMG) signals from different channels, and to construct the dataset after preprocessing and labeling each of the historical surface EMG signals. The dataset partitioning module is used to divide the dataset into a training set, a validation set, and a test set based on a preset ratio. The joint torque prediction model creation module is used to create a joint torque prediction model based on the self-supervised reconstruction feature extraction module, the time-dependent modeling module, and the regression prediction module. The joint torque prediction model training module is used to train, validate, and test the joint torque prediction model using the training set, validation set, and test set, respectively, and to deploy the joint torque prediction model that passes the test. A joint torque prediction module is used to predict joint torques using the deployed joint torque prediction model.

7. A robust joint torque prediction system based on self-supervised reconstruction as described in claim 6, characterized in that: The dataset construction module is specifically used for: A large number of historical surface electromyography (EMG) signals from different subjects and different channels were acquired. Each historical surface EMG signal was preprocessed, including filtering, normalization and sliding window segmentation. After labeling the joint torque values ​​of each preprocessed historical surface EMG signal, a dataset was constructed.

8. A robust joint torque prediction system based on self-supervised reconstruction as described in claim 6, characterized in that: The dataset partitioning module is specifically used for: Based on chronological order, the dataset is divided into training, validation, and test sets in a ratio of 7:2:

1.

9. A robust joint torque prediction system based on self-supervised reconstruction as described in claim 6, characterized in that: In the joint torque prediction model creation module, the self-supervised reconstruction feature extraction module is built based on an encoder and a decoder. The encoder is used to extract local temporal features representing instantaneous muscle activation changes from surface electromyography (EMG) signals through a one-dimensional convolutional layer, learn the long-term dependencies of surface EMG signals through a multilayer perceptron to extract global features, and fuse the local temporal features and global features into a multi-scale spatial feature representation. The decoder is used to reconstruct the input surface EMG signals based on the multi-scale spatial feature representation, and uses the reconstruction error as a self-supervised learning objective to optimize the multi-scale spatial feature representation output by the encoder. The time-dependent modeling module is used to capture the dynamic time dependency between muscle activity and joint torque from the multi-scale spatial feature representation by combining a long short-term memory network with an attention mechanism, and generate a dynamic temporal feature representation. The regression prediction module is used to flatten the dynamic temporal feature representation and realize the nonlinear regression mapping between the flattened dynamic temporal feature representation and the target joint torque through a fully connected network, and output the joint torque prediction value.

10. A robust joint torque prediction system based on self-supervised reconstruction as described in claim 6, characterized in that: The joint torque prediction model training module is specifically used for: The joint torque prediction model is trained using the training set until a preset early stopping condition is met. At the end of each training batch, the joint torque prediction model is validated using the validation set to optimize the hyperparameters of the joint torque prediction model. Then, RMSE, MAE, R², and VAF are calculated using the test set to test the trained joint torque prediction model. The joint torque prediction model that passes the test is then deployed.

Images