Lightweight wrist joint continuous angle decoding system and method based on multi-channel sEMG
By combining multi-channel sEMG signals with a lightweight CNN-Transformer model, the real-time and accuracy issues of wrist joint motion angle decoding on portable devices are solved, achieving low-latency continuous angle decoding, which is suitable for real-time embedded devices.
Patent Information
- Application Number
- CN202510877881.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-27
- Publication Date
- 2025-10-31
AI Technical Summary
Existing technologies struggle to achieve high-precision, real-time, and continuous wrist joint motion angle decoding, especially on portable devices, and there is a lack of research on deep learning models for continuous wrist joint motion angles.
By employing multi-channel sEMG signals and combining them with a lightweight CNN-Transformer model, the model is made lightweight through preprocessing, quantization, and pruning, and can be used for real-time prediction on mobile or embedded devices with hardware acceleration.
It achieves low-latency, high-precision continuous angle decoding of the wrist joint, is suitable for portable devices, and supports the development of closed-loop systems from signal acquisition to prediction.
Smart Images
Figure CN120873584A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wrist joint motion angle decoding technology, and in particular to a lightweight wrist joint continuous angle decoding system and method based on multi-channel sEMG. Background Technology
[0002] As a crucial movement part of the upper limb, the wrist joint's precise decoding of movement angles has significant application value in fields such as prosthetic control, rehabilitation training, and robot interaction. Surface electromyography (sEMG) signals, due to their non-invasiveness and portability, have become the mainstream signal source for decoding human movement intentions. However, the non-stationarity, low signal-to-noise ratio, and individual variability of sEMG signals make it difficult for traditional signal processing and shallow learning models to achieve high-precision, real-time, and continuous angle decoding.
[0003] In recent years, deep learning methods, especially convolutional neural networks (CNNs) and Transformer architectures, have demonstrated good performance in sEMG gesture recognition. However, existing models mainly focus on discrete gesture classification and are bulky, making it difficult to meet the real-time continuous decoding requirements of portable devices. Furthermore, there is a lack of research on deep learning models for accurate decoding of continuous wrist joint movement angles. Summary of the Invention
[0004] To achieve the above-mentioned objectives and other advantages of the present invention, the first objective of the present invention is to provide a lightweight continuous angle decoding method for wrist joints based on multi-channel sEMG, comprising the following steps:
[0005] Acquire surface electromyographic signals of wrist muscle groups and preprocess the surface electromyographic signals;
[0006] Train a lightweight CNN-Transformer model using preprocessed data;
[0007] By reducing the model size through quantization and pruning, it is converted into a lightweight inference format to enable low-latency real-time predictions on mobile or embedded devices with hardware acceleration.
[0008] Furthermore, the surface electromyographic signals of the wrist muscle groups are configured as surface electromyographic signals of the radial flexor carpi radialis, ulnar flexor carpi radialis, extensor carpi radialis longus, extensor carpi ulnaris, superficial flexor digitorum group, and pronator teres.
[0009] Furthermore, the preprocessing of the surface electromyography signal includes:
[0010] Continuous surface electromyography signals are divided into input samples according to a fixed length;
[0011] The surface electromyography signals were preprocessed using a normalization method;
[0012] The normalized signal is downsampled over time.
[0013] Furthermore, the lightweight CNN-Transformer model is based on a two-stage CNN-Transformer hybrid structure, with each stage including multiple sub-modules to achieve feature extraction and global sequence modeling.
[0014] Furthermore, in the first stage, the data first passes through a convolutional downsampling block, which uses a 1D convolutional kernel to downsample the signal in the time dimension to extract primary temporal features;
[0015] Subsequently, the Switch activation function, LayerNorm, and Dropout were used to prevent overfitting.
[0016] Next, extended convolution is performed through inverse residual blocks, and Swish activation, LayerNorm and depthwise separable convolution are applied. After activation and layer normalization, dimensionality is reduced by 1D convolution and the original input residual connection is added to enhance the fusion of features at different scales.
[0017] Finally, using a Transformer encoding block with a single-head self-attention mechanism, query, key, and value mappings are computed to capture global temporal dependencies.
[0018] Furthermore, in the second stage, the time step is first reduced by using convolutional downsampling blocks;
[0019] Then, after passing through the inverse residual block and the Transformer encoding block with single-head self-attention, 1D global average pooling is performed on the second-stage output in the global output layer to obtain the feature vector;
[0020] Then, dimensionality reduction is achieved through a fully connected layer with ReLU activation;
[0021] Finally, continuous wrist joint motion angle values are output through a fully connected layer.
[0022] Furthermore, the step of training a lightweight CNN-Transformer model using the preprocessed data also includes:
[0023] During model training, a regression loss function is defined, and the Adam optimizer is used to iteratively train the model parameters.
[0024] Furthermore, the step of reducing the model size through quantization and pruning, and converting it into a lightweight inference format to achieve low-latency real-time prediction on mobile or embedded devices with hardware acceleration includes:
[0025] The trained model is first exported as a SavedModel format;
[0026] Then, TensorFlow Lite Converter is used for conversion and dynamic range quantization, converting the weights and activations from 32-bit floating-point to 8-bit integers, thereby compressing the model size;
[0027] Finally, the compressed model is integrated into the mobile application, and the interpreter interface is called to complete the real-time surface electromyography signal angle prediction. The inference performance is optimized in conjunction with the hardware acceleration interface of mobile or embedded devices.
[0028] The second objective of this invention is to provide a lightweight wrist joint continuous angle decoding system based on multi-channel sEMG, employing the aforementioned method, including multi-channel distributed sEMG electrodes, a data preprocessing module, a model training module, and a model compression and deployment optimization module; wherein,
[0029] The multi-channel distributed sEMG electrode is used to collect surface electromyographic signals of the wrist muscle group.
[0030] The data preprocessing module is used to preprocess the surface electromyography signal;
[0031] The model training module is used to train a lightweight CNN-Transformer model using the preprocessed data;
[0032] The model compression and deployment optimization module is used to reduce the model size through quantization and pruning, and convert it into a lightweight inference format to achieve low-latency real-time prediction on mobile or embedded devices with hardware acceleration.
[0033] A third object of the present invention is to provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.
[0034] Compared with the prior art, the beneficial effects of the present invention are:
[0035] This invention proposes a lightweight continuous angle decoding system and method for wrist joints based on multi-channel sEMG. By fusing local convolution and global Transformer attention mechanisms, the model can effectively extract and utilize the spatiotemporal features of sEMG signals to achieve continuous angle regression. Furthermore, the model has a lightweight structure, consumes few computational resources, and is suitable for real-time embedded and portable devices. A complete end-to-end data processing and training system supports the development of a closed-loop system from signal acquisition to prediction. Specific model compression and hardware acceleration solutions ensure inference efficiency and adapt to practical application scenarios.
[0036] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it according to the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Specific embodiments of the present invention are given in detail below with reference to the accompanying drawings. Attached Figure Description
[0037] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:
[0038] Figure 1 Flowchart of a lightweight continuous angle decoding method for wrist joints based on multi-channel sEMG Figure 1 ;
[0039] Figure 2 Flowchart of a lightweight continuous angle decoding method for wrist joints based on multi-channel sEMG Figure 2 ;
[0040] Figure 3 Flowchart for preprocessing surface electromyography signals;
[0041] Figure 4 This is a flowchart of the first stage of the CNN-Transformer process;
[0042] Figure 5 The flowchart for the second stage of CNN-Transformer;
[0043] Figure 6 The flowchart illustrates the process of reducing model size through quantization and pruning, converting it into a lightweight inference format, and combining hardware acceleration to achieve low-latency real-time prediction on mobile or embedded devices.
[0044] Figure 7 A schematic diagram of a lightweight wrist joint continuous angle decoding system based on multi-channel sEMG;
[0045] Figure 8 A schematic diagram of computer equipment;
[0046] Figure 9 This is a schematic diagram of a computer-readable storage medium. Detailed Implementation
[0047] The present invention will now be further described with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.
[0048] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.
[0049] The drawing numbers in this application are only used to distinguish the steps in the scheme and are not used to limit the execution order of the steps. The specific execution order is as described in the specification.
[0050] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.
[0051] Decoding wrist joint motion angles is crucial for the control accuracy and response speed of intelligent prostheses and human-computer interaction devices. Traditional methods mainly rely on shallow machine learning approaches based on feature extraction (such as support vector machines and linear discriminant analysis). These methods depend on manually designed electromyographic signal features, which are difficult to capture the complex spatiotemporal patterns of sEMG signals, and their performance degrades significantly with individual differences and environmental changes.
[0052] In recent years, deep learning techniques (Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Transformers) have been applied to discrete gesture recognition, achieving remarkable results. However, they have the following main shortcomings:
[0053] Large model complexity and size: Deep models have a large number of parameters, high computational resource requirements, and are difficult to embed into low-power mobile devices.
[0054] Focusing on classification rather than continuous regression: Most current research focuses on discrete gesture recognition and lacks end-to-end regression decoding schemes for continuous motion angles.
[0055] High inference latency: In embedded scenarios, complex models lead to long inference times, which affects the real-time control experience.
[0056] Lack of system-level hardware and software co-optimization: The lack of tight end-to-end process support between model training, compression, and deployment limits the practical application value of the technology.
[0057] This invention focuses on solving the above problems and proposes a lightweight and efficient CNN-Transformer continuous angle decoding model, supplemented by a complete end-to-end data processing and deployment optimization process, to achieve the goals of low latency, high accuracy and portability.
[0058] Example 1
[0059] A lightweight method for continuous wrist joint angle decoding based on multi-channel sEMG, such as Figures 1-2As shown, it includes the following steps:
[0060] S100: Acquire the surface electromyography (EMG) signals of the wrist muscle group and preprocess the EMG signals.
[0061] Furthermore, the surface electromyographic signals of the wrist muscle groups are configured as surface electromyographic signals of the radial flexor carpi radialis, ulnar flexor carpi radialis, extensor carpi radialis longus, extensor carpi ulnaris, superficial flexor digitorum group, and pronator teres.
[0062] In this embodiment, high-sampling-rate raw sEMG data is first collected. This embodiment uses a 6-channel distributed sEMG electrode to collect electromyographic signals of the main muscle groups of the wrist at a sampling frequency of 2000Hz, including the flexor carpi radialis (Ch1), flexor carpi ulnaris (Ch2), extensor carpi radialis longus (Ch3), extensor carpi ulnaris (Ch4), flexor digitorum superficialis (Ch5), and pronator teres (Ch6).
[0063] In some embodiments, such as Figure 3 As shown, the preprocessing of the surface electromyography signal includes:
[0064] S110. Divide the continuous surface electromyography signal into input samples according to a fixed length;
[0065] Optionally, the continuous sEMG signal is divided into a single input sample by a fixed length of 150 sampling points (equivalent to a 75ms time window), with no overlap between windows to ensure time synchronization.
[0066] In order to achieve signal balance and stability and suppress signal amplitude differences between individuals, S120 uses a normalization method to preprocess the surface electromyography signal;
[0067] Meanwhile, to reduce data dimensionality and alleviate the computational load on the model, S130 performs time downsampling on the normalized signal. Specifically, the step sizes are 3 and 4, implemented in conjunction with a convolutional downsampling block within the model.
[0068] It is worth noting that this embodiment does not use high-pass or low-pass filtering to avoid introducing time delay, but instead relies on the learning ability of the model structure itself to suppress noise. After final processing, the input tensor size of a single sample is (150, 6), that is, 150 time steps, with 6 channels of signal per step.
[0069] S200. Train a lightweight CNN-Transformer model using the preprocessed data;
[0070] Furthermore, the lightweight CNN-Transformer model is based on a two-stage CNN-Transformer hybrid structure, with each stage including multiple sub-modules to achieve feature extraction and global sequence modeling.
[0071] In some embodiments, such as Figure 4 As shown, in the first stage, S210 and the data first pass through the convolution downsampling block, and the signal is downsampled in the time dimension using a 1D convolution kernel (stride DS=3, number of output channels is 24) to extract primary temporal features;
[0072] S220. Subsequently, the Switch activation function, LayerNorm, and Dropout (ratio 0.5) are used to prevent overfitting.
[0073] S230 Next, extended convolution is performed through inverted residual blocks (the number of output channels is expanded to 24×EF=72, EF=3), Swish activation, LayerNorm and depthwise separable convolution (the number of depthwise convolutional filters is 6), reactivation and layer normalization are performed, and finally 1D convolution is used to reduce the dimension back to 24 channels, and the original input residual connection is added to enhance the fusion of features at different scales;
[0074] S240. Finally, using a Transformer encoding block with a single-head self-attention mechanism, query (Q), key (K), and value (V) mappings are computed to capture global temporal dependencies. The Transformer includes layer normalization, residual connections, and two MLP layers (size reduced from 48 to 24).
[0075] The second phase is similar to the first phase. In some embodiments, such as Figure 5 As shown, in the second stage, S250 first reduces the time step by using a convolution downsampling block (1D convolution stride DS=4, number of output channels is 48);
[0076] S260. Then, after passing through the inverse residual block (expansion factor EF = 5, expanding the channels to 48 × 5 = 240) and the Transformer coding block with single-head self-attention, 1D global average pooling is performed on the second stage output in the global output layer to obtain a feature vector with a length and width of 48.
[0077] S270, Then, the dimension is reduced to 32 through a fully connected layer with ReLU activation;
[0078] S280 Finally, continuous wrist joint motion angle values are output through the fully connected layer.
[0079] Furthermore, the step of training a lightweight CNN-Transformer model using the preprocessed data also includes:
[0080] During model training, a regression loss function is defined, and the Adam optimizer is used to iteratively train the model parameters.
[0081] S300 reduces model size through quantization and pruning, converting it to a lightweight inference format (such as TensorFlow Lite) to enable low-latency real-time predictions on mobile or embedded devices by combining hardware acceleration (such as Apple Neural Engine and GPU).
[0082] Furthermore, such as Figure 6 As shown, the step of reducing the model size through quantization and pruning, and converting it into a lightweight inference format to achieve low-latency real-time prediction on mobile or embedded devices with hardware acceleration includes:
[0083] S310. The trained model is first exported as a SavedModel format;
[0084] S320. Then, TensorFlow Lite Converter is used for conversion and dynamic range quantization, converting the weights and activations from 32-bit floating-point to 8-bit integer, thereby compressing the model size.
[0085] S330. Finally, the compressed model is integrated into the mobile application, and the interpreter interface is called to complete real-time surface electromyography signal angle prediction. Inference performance is optimized with the hardware acceleration interface of mobile or embedded devices. For example, inference performance is optimized with ARM embedded hardware or mobile hardware acceleration interface.
[0086] This embodiment integrates local convolution and global Transformer attention mechanisms, enabling the model to effectively extract and utilize the spatiotemporal features of sEMG signals to achieve continuous angle regression. Furthermore, the model has a lightweight structure, consumes minimal computational resources, and is suitable for real-time embedded and portable devices. A complete end-to-end data processing and training system supports the development of a closed-loop system from signal acquisition to prediction. Specific model compression and hardware acceleration solutions ensure inference efficiency and adapt to real-world application scenarios.
[0087] Example 2
[0088] A lightweight wrist joint continuous angle decoding system based on multi-channel sEMG is provided. The above-described method is applied. For a detailed description of the lightweight wrist joint continuous angle decoding method based on multi-channel sEMG, please refer to the corresponding description in the above-described embodiments of the lightweight wrist joint continuous angle decoding method based on multi-channel sEMG, which will not be repeated here.
[0089] This embodiment aims to construct a complete end-to-end system for decoding continuous angles from surface electromyography (sEMG) signals of the wrist joint. The system flow consists of three stages: data preprocessing, model training, and model compression and deployment optimization, aiming to achieve a complete closed loop from raw sEMG signals to continuous angle prediction of the wrist joint.
[0090] like Figure 7 As shown, the system 400 includes a multi-channel distributed sEMG electrode 410, a data preprocessing module 420, a model training module 430, and a model compression and deployment optimization module 440; wherein,
[0091] The multi-channel distributed sEMG electrode is used to collect surface electromyographic signals of the wrist muscle group.
[0092] The data preprocessing module is used to preprocess the surface electromyography signal;
[0093] The model training module is used to train a lightweight CNN-Transformer model using the preprocessed data;
[0094] The model compression and deployment optimization module is used to reduce the model size through quantization and pruning, and convert it into a lightweight inference format to achieve low-latency real-time prediction on mobile or embedded devices with hardware acceleration.
[0095] Based on the technical solutions of the above embodiments, optionally, the surface electromyography (EMG) signals of the wrist muscle group are configured as surface EMG signals of the radial flexor carpi radialis, ulnar flexor carpi radialis, extensor carpi radialis longus, extensor carpi ulnaris, flexor digitorum superficialis, and pronator teres.
[0096] Based on the technical solutions of the above embodiments, optionally, the preprocessing of the surface electromyography signal includes:
[0097] Continuous surface electromyography signals are divided into input samples according to a fixed length;
[0098] The surface electromyography signals were preprocessed using a normalization method;
[0099] The normalized signal is downsampled over time.
[0100] Based on the technical solutions of the above embodiments, optionally, the lightweight CNN-Transformer model takes a two-stage CNN-Transformer hybrid structure as its core, with each stage including multiple sub-modules to achieve feature extraction and global sequence modeling.
[0101] Based on the technical solution of the above embodiments, optionally, in the first stage, the data first passes through a convolution downsampling block, and the signal is downsampled in the time dimension using a 1D convolution kernel to extract primary temporal features;
[0102] Subsequently, the Switch activation function, LayerNorm, and Dropout were used to prevent overfitting.
[0103] Next, extended convolution is performed through inverse residual blocks, and Swish activation, LayerNorm and depthwise separable convolution are applied. After activation and layer normalization, dimensionality is reduced by 1D convolution and the original input residual connection is added to enhance the fusion of features at different scales.
[0104] Finally, using a Transformer encoding block with a single-head self-attention mechanism, query, key, and value mappings are computed to capture global temporal dependencies.
[0105] Based on the technical solutions of the above embodiments, optionally, in the second stage, the time step is first reduced by a convolutional downsampling block;
[0106] Then, after passing through the inverse residual block and the Transformer encoding block with single-head self-attention, 1D global average pooling is performed on the second-stage output in the global output layer to obtain the feature vector;
[0107] Then, dimensionality reduction is achieved through a fully connected layer with ReLU activation;
[0108] Finally, continuous wrist joint motion angle values are output through a fully connected layer.
[0109] Based on the technical solutions of the above embodiments, optionally, the step of training a lightweight CNN-Transformer model using preprocessed data further includes:
[0110] During model training, a regression loss function is defined, and the Adam optimizer is used to iteratively train the model parameters.
[0111] Based on the technical solutions of the above embodiments, optionally, the step of reducing the model volume through quantization and pruning, converting it into a lightweight inference format, and combining hardware acceleration to achieve low-latency real-time prediction on mobile or embedded devices includes:
[0112] The trained model is first exported as a SavedModel format;
[0113] Then, TensorFlow Lite Converter is used for conversion and dynamic range quantization, converting the weights and activations from 32-bit floating-point to 8-bit integers, thereby compressing the model size;
[0114] Finally, the compressed model is integrated into the mobile application, and the interpreter interface is called to complete the real-time surface electromyography signal angle prediction. The inference performance is optimized in conjunction with the hardware acceleration interface of mobile or embedded devices.
[0115] This embodiment integrates local convolution and global Transformer attention mechanisms, enabling the model to effectively extract and utilize the spatiotemporal features of sEMG signals to achieve continuous angle regression. Furthermore, the model has a lightweight structure, consumes minimal computational resources, and is suitable for real-time embedded and portable devices. A complete end-to-end data processing and training system supports the development of a closed-loop system from signal acquisition to prediction. Specific model compression and hardware acceleration solutions ensure inference efficiency and adapt to real-world application scenarios.
[0116] Example 3
[0117] A computer device 500, such as Figure 8 As shown, the system includes a memory 510, a processor 520, and a computer program 530 stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of a lightweight wrist joint continuous angle decoding method based on multi-channel sEMG. For a detailed description of the method, please refer to the corresponding description in the above method embodiments; it will not be repeated here.
[0118] Example 4
[0119] A computer-readable storage medium, such as Figure 9 As shown, a computer program is stored thereon. When executed by a processor, the computer program implements the steps of a lightweight wrist joint continuous angle decoding method based on multi-channel sEMG. For a detailed description of the method, please refer to the corresponding description in the above method embodiments, which will not be repeated here.
[0120] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention. Applications, modifications, and variations of the invention will be readily apparent to those skilled in the art.
[0121] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.
[0122] The apparatus, computer device, and non-volatile computer storage medium and method provided in the embodiments of this specification are corresponding. Therefore, the apparatus, computer device, and non-volatile computer storage medium also have similar beneficial technical effects as the corresponding method. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the corresponding apparatus, computer device, and non-volatile computer storage medium will not be repeated here.
[0123] Those skilled in the art will also know that, besides implementing the controller in the form of purely computer-readable program code, the same functions can be achieved by logically programming the method steps, making the controller take the form of logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers (PLCs), and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the devices included within it for implementing various functions can also be considered structures within that hardware component. Alternatively, the devices for implementing various functions can be considered as both software units implementing the method and structures within a hardware component.
[0124] The systems, apparatuses, or units described in the above embodiments can be implemented by computer chips or physical entities, or by products with certain functions. For ease of description, the above apparatuses are described separately as various units based on their functions. Of course, when implementing one or more embodiments of this specification, the functions of each unit can be implemented in one or more software and / or hardware.
[0125] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, the embodiments of this specification can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the embodiments of this specification can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0126] This specification is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0127] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0128] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0129] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0130] This specification may be described in the general context of computer-executable instructions, such as program units, that are executed by a computer. Generally, program units include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification may also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program units may reside in local and remote computer storage media, including storage devices.
[0131] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.
[0132] The above description is merely an embodiment of this specification and is not intended to limit the scope of one or more embodiments of this specification. Various modifications and variations can be made to one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of one or more embodiments of this specification.
Claims
1. A lightweight continuous angle decoding method for wrist joints based on multi-channel sEMG, characterized in that, Includes the following steps: Acquire surface electromyographic signals of wrist muscle groups and preprocess the surface electromyographic signals; Train a lightweight CNN-Transformer model using preprocessed data; By reducing the model size through quantization and pruning, it is converted into a lightweight inference format to enable low-latency real-time predictions on mobile or embedded devices with hardware acceleration.
2. The lightweight wrist joint continuous angle decoding method based on multi-channel sEMG as described in claim 1, characterized in that: The surface electromyography (EMG) signals of the wrist muscle groups were configured as the surface EMG signals of the radial flexor carpi radialis, ulnar flexor carpi radialis, extensor carpi radialis longus, extensor carpi ulnaris, superficial flexor digitorum group, and pronator teres.
3. The lightweight wrist joint continuous angle decoding method based on multi-channel sEMG as described in claim 1, characterized in that, The preprocessing of the surface electromyography signal includes: Continuous surface electromyography signals are divided into input samples according to a fixed length; The surface electromyography signals were preprocessed using a normalization method; The normalized signal is downsampled over time.
4. The lightweight wrist joint continuous angle decoding method based on multi-channel sEMG as described in claim 1, characterized in that: The lightweight CNN-Transformer model is based on a two-stage CNN-Transformer hybrid structure, with each stage including multiple sub-modules to achieve feature extraction and global sequence modeling.
5. A lightweight wrist joint continuous angle decoding method based on multi-channel sEMG as described in claim 4, characterized in that: In the first stage, the data first passes through a convolutional downsampling block, which uses a 1D convolutional kernel to downsample the signal in the time dimension and extract primary temporal features. Subsequently, the Switch activation function, LayerNorm, and Dropout were used to prevent overfitting. Next, extended convolution is performed through inverse residual blocks, and Swish activation, LayerNorm and depthwise separable convolution are applied. After activation and layer normalization, dimensionality is reduced by 1D convolution and the original input residual connection is added to enhance the fusion of features at different scales. Finally, using a Transformer encoding block with a single-head self-attention mechanism, query, key, and value mappings are computed to capture global temporal dependencies.
6. The lightweight wrist joint continuous angle decoding method based on multi-channel sEMG as described in claim 5, characterized in that: In the second stage, the time step is first reduced by using convolutional downsampling blocks; Then, after passing through the inverse residual block and the Transformer encoding block with single-head self-attention, 1D global average pooling is performed on the second-stage output in the global output layer to obtain the feature vector; Then, dimensionality reduction is achieved through a fully connected layer with ReLU activation; Finally, continuous wrist joint motion angle values are output through a fully connected layer.
7. A lightweight wrist joint continuous angle decoding method based on multi-channel sEMG as described in claim 1, characterized in that, The step of training a lightweight CNN-Transformer model using preprocessed data also includes: During model training, a regression loss function is defined, and the Adam optimizer is used to iteratively train the model parameters.
8. A lightweight wrist joint continuous angle decoding method based on multi-channel sEMG as described in claim 1, characterized in that, The step of reducing the model size through quantization and pruning, and converting it into a lightweight inference format to achieve low-latency real-time prediction on mobile or embedded devices with hardware acceleration includes: The trained model is first exported as a SavedModel format; Then, TensorFlow Lite Converter is used to perform conversion and dynamic range quantization, converting the weights and activations from 32-bit floating-point to 8-bit integers, thereby compressing the model size; Finally, the compressed model is integrated into the mobile application, and the interpreter interface is called to complete the real-time surface electromyography signal angle prediction. The inference performance is optimized in conjunction with the hardware acceleration interface of mobile or embedded devices.
9. A lightweight wrist joint continuous angle decoding system based on multi-channel sEMG, using the method described in any one of claims 1 to 8, characterized in that: It includes a multi-channel distributed sEMG electrode, a data preprocessing module, a model training module, and a model compression and deployment optimization module; among which, The multi-channel distributed sEMG electrode is used to collect surface electromyographic signals of the wrist muscle group. The data preprocessing module is used to preprocess the surface electromyography signal; The model training module is used to train a lightweight CNN-Transformer model using the preprocessed data; The model compression and deployment optimization module is used to reduce the model size through quantization and pruning, and convert it into a lightweight inference format to achieve low-latency real-time prediction on mobile or embedded devices with hardware acceleration.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 8.