32-channel surface myoelectricity real-time gesture man-machine interaction system and method for cognitive rehabilitation

Abstract

The invention provides a cognitive rehabilitation-oriented 32-channel surface myoelectricity real-time gesture man-machine interaction system and method.The system comprises a flexible array electrode, a hardware circuit, an upper computer interface and an interaction module, and the flexible array electrode is attached to a wrist flexor muscle group and an extensor muscle group of the wrist through a 2 * 16-channel snakelike wiring structure; the sensor is used for collecting surface electromyogram signals; the hardware circuit integrates an electrophysiology amplification and acquisition module, a main control module, a power management module and a wireless communication module, and amplification, filtering, analog-to-digital conversion and wireless transmission of electromyographic signals are achieved. The upper computer analyzes, processes and displays the received multi-channel myoelectricity data, completes gesture recognition based on activation detection and a multi-scale residual attention network model, and maps a recognition result into an interaction control event to drive a virtual rehabilitation scene; according to the invention, real-time identification and man-machine interaction control of micro-actions of wrists and fingers can be realized under a natural wearing condition, and the system is suitable for cognitive rehabilitation training and related interaction application scenes.

Description

Technical Field

[0001] This invention relates to the field of human-computer interaction and rehabilitation engineering technology, and more specifically, to a 32-channel surface electromyography real-time gesture human-computer interaction system and method for cognitive rehabilitation. Background Technology

[0002] Gesture recognition, as the most intuitive and flexible natural interface in the field of human-computer interaction, can effectively bridge human intentions and machine responses, and has unlimited potential in expanding the boundaries of human capabilities and assisting rehabilitation training. However, existing gesture interaction mainly relies on traditional large-amplitude movements of the five fingers. This method not only lacks social discretion, but also easily causes muscle fatigue after prolonged operation, making it difficult to meet the needs of next-generation AR / VR and metaverse applications for immersive, low-load experiences. In contrast, micro-movements centered on the thumb (joint rotation <5°) have extremely high subtlety and fatigue resistance, which aligns with the concept of intuitive control. However, current thumb-based interactions are still limited to simple directional commands, making it difficult to accurately recognize progressive, fine-grained movements such as amplitude and speed, thus limiting the diversity and depth of interaction.

[0003] In signal capture and analysis methods, traditional visual imaging is easily affected by light occlusion and spatial constraints, while mechanical sensing lacks accuracy in capturing subtle movements. Surface electromyography (SEMG) technology, due to its advantages of being non-invasive, environment-independent, and directly recording neuromuscular electrical activity, has become the preferred solution in this field. Nevertheless, current gesture interaction systems based on SEMG signals still face three major bottlenecks: First, limited recognition ability, especially in single-handed mode, where highly similar electromyographic patterns lead to fewer interaction categories and a high misclassification rate; second, insufficient precision, with existing algorithms mostly handling end-to-end full-range movements, making it difficult to effectively decode micro-movements containing gradual changes; and finally, a single interaction paradigm, lacking a multi-variable synchronous control mechanism, making it difficult to extend to multi-user collaboration or complex multi-tasking environments.

[0004] In summary, to overcome the shortcomings of existing technologies in terms of micro-motion capture accuracy, interaction diversity, and collaborative modes, the development of a fully integrated, wireless, wrist-worn, high-density surface electromyography (EMG) signaling system is particularly urgent. Achieving high-fidelity acquisition of process-oriented thumb micro-movements by employing a rigid-flexible hybrid architecture, integrating low-noise amplification chips (such as the RHD2132), and high-density flexible dry electrode arrays (such as 32 channels) to construct novel human-computer interaction modes with multiple scenarios, multiple commands, and dual-user interaction has become a key direction for breakthroughs in current human-computer interaction technology.

[0005] Therefore, there is an urgent need for a 32-channel surface electromyography real-time gesture human-computer interaction system and method for cognitive rehabilitation to solve the above problems. Summary of the Invention

[0006] The purpose of this invention is to solve the technical problems mentioned in the background section and to provide a 32-channel surface electromyography real-time gesture human-computer interaction system for cognitive rehabilitation, comprising: A wireless wrist-worn 32-channel high-density surface electromyography (EMG) acquisition device for acquiring surface EMG signals from the wrist. The host computer interface is used to establish a communication connection with the acquisition device and to parse, display and store electromyographic data, while deploying a gesture recognition model to output gesture recognition results; An interaction module is used to receive the gesture recognition results and map them into virtual scene control events to drive cognitive rehabilitation training tasks; The wireless wrist-worn 32-channel high-density surface electromyography (EMG) acquisition device consists of a flexible array electrode and hardware circuitry. The flexible array electrode is a 2×16-channel flexible array structure adapted to the curved surface of the wrist. The hardware circuitry includes a main control module, a power management module, an FPC connector, an electrophysiological amplification and acquisition chip, and a Wi-Fi module.

[0007] As a preferred technical solution of the present invention, the flexible array electrode adopts a serpentine wiring structure to realize electrical connection and form a 2×16 channel electrode array to ensure the spatial coverage and adhesion flexibility of signal acquisition.

[0008] As a preferred technical solution of the present invention, the flexible array electrode covers the wrist flexor and extensor muscles in the wrist area, and uses medical pressure-sensitive tape to achieve stable contact and low impedance matching between the electrode and the skin surface.

[0009] As a preferred technical solution of the present invention, the hardware circuit has an integrated protective shell, which is a three-layer structure consisting of an upper shell, a middle shell, and a lower shell; the middle shell isolates the supporting circuit from the power supply lithium battery in upper and lower layers.

[0010] As a preferred technical solution of the present invention, the data transmission link of the acquisition device is as follows: the electromyographic signal acquired by the flexible array electrode is amplified by electrophysiological amplification and acquisition chip, and then amplified, bandpass filtered and analog-to-digital converted to obtain digital sampling data. Subsequently, it is transmitted to the main control module through the serial peripheral interface SPI2 and written into the ring buffer for temporary storage. Then, the main control module frames and encapsulates the data and transmits it to the Wi-Fi module through SPI1. The Wi-Fi module wirelessly transmits the encapsulated electromyographic data to the host computer interface based on the transmission control protocol TCP.

[0011] This invention proposes a 32-channel surface electromyography real-time gesture human-computer interaction method for cognitive rehabilitation, comprising: Acquisition steps: Acquire 32-channel surface electromyography signals from the wrist using a flexible array electrode; Transmission steps: The electromyographic signal is amplified, filtered, converted from analog to digital, and framed and encapsulated by hardware circuitry, and then wirelessly transmitted to the host computer interface via Wi-Fi. Processing steps: The host computer interface performs protocol parsing and filtering on the received data and forms input segments for identification; Recognition steps: Call the multi-scale residual attention network model deployed on the host computer to classify the gestures of the input segment and output gesture labels; Interaction steps: Map the gesture tags to control events and send them to the interaction module to drive the character behavior response in the virtual rehabilitation scenario.

[0012] As a preferred technical solution of the present invention, the host computer interface is laid out using Qt Designer and implemented using Python+PyQt5, and includes at least a TCP connection module, a data parsing and filtering module, an arbitrary channel waveform display module, a data storage module, a single channel spectrum graph module, and a model deployment and real-time inference module.

[0013] As a preferred technical solution of the present invention, a time sliding window mechanism and an absolute average-root mean square dual threshold activation segment screening method are used for real-time activation extraction: the time window length is 100ms and the overlap rate is 90%; four core channels with the best signal-to-noise ratio distribution are selected from 32 channels, namely channels 27-30, and the activation segment is marked when the envelope intensity first exceeds the fluctuation threshold α≥0.4. The activation start point is determined when the intensity threshold β≥1.0 is exceeded twice in the subsequent observation window, and the electromyographic segment containing all 32 channels is extracted for subsequent identification.

[0014] As a preferred technical solution of the present invention, the multi-scale residual attention network model is MSE-Net, which employs three sequentially stacked multi-scale convolutional blocks. Each convolutional block contains four parallel convolutional branches. The branch outputs are concatenated in the channel dimension, fused by a 1×1 convolution, and then connected to the residual attention mechanism. The branch kernel length and dilation rate parameters of the three convolutional blocks are as follows: The first layer consists of (3,1), (5,2), (7,4), and (9,4). The second layer consists of (3,1), (7,2), (11,4), and (15,8). The third layer consists of (3,1), (9,2), (15,4), and (21,8).

[0015] This invention proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements a 32-channel surface electromyography real-time gesture human-computer interaction method for cognitive rehabilitation.

[0016] Compared with existing technologies, this invention constructs a wireless, wrist-worn 32-channel high-density surface electromyography (EMG) acquisition system. Employing an integrated design of flexible array electrodes and lightweight hardware circuitry, it achieves stable acquisition of EMG signals from the wrist flexor and extensor muscles under natural wearing conditions. The flexible serpentine wiring electrode structure conforms well to the curved contour of the wrist, improving electrode contact stability and signal coverage without restricting hand movement. Combined with a zoned power supply hardware architecture and wireless transmission, it effectively reduces noise interference and user burden, making the system comfortable to wear, compact, lightweight, and suitable for long-term use, ideal for continuous training applications such as cognitive rehabilitation.

[0017] In terms of signal processing and recognition methods, this invention employs a two-stage recognition strategy combining a dual-threshold activation detection mechanism based on core channel selection and a multi-scale residual attention network. By selecting the core channel with the optimal signal-to-noise ratio from multi-channel electromyography (EMG) signals for real-time activation detection, the real-time computational complexity is significantly reduced while ensuring recognition accuracy. The multi-scale parallel convolution and residual attention mechanism can effectively extract EMG features at different time scales, improving the recognition ability of wrist and finger micro-movements, especially progressive fine-grained thumb movements, thereby achieving stable, low-latency real-time gesture recognition.

[0018] This invention maps electromyographic gesture recognition results into standardized interactive events, which can directly drive virtual scenes and rehabilitation training tasks, constructing a complete closed-loop control system for human-computer interaction. This system is compatible with both two-dimensional planar interaction and three-dimensional virtual reality interaction environments, supporting multiple interaction modes such as single-role control, dual-role collaboration, and continuous action triggering. It can cover diverse cognitive rehabilitation training needs within a unified system framework. Due to the use of a universal event mapping method, this system can be seamlessly integrated into existing rehabilitation equipment and virtual interaction platforms, possessing good scalability and engineering application value, and helping to lower the barriers to clinical application and system deployment. Attached image description: Figure 1 This is a schematic diagram of the overall system structure of the wireless wrist-worn 32-channel high-density surface electromyography real-time gesture human-computer interaction system (WFWES) for cognitive rehabilitation of the present invention, including flexible array electrodes, hardware circuits, host computer interface and interaction module.

[0019] Figure 2 This is a schematic diagram of the structure of the wireless wrist-worn 32-channel surface electromyography (EMG) acquisition device of the present invention, wherein: (i) is a schematic diagram of the flexible array electrode structure; (ii) is a schematic diagram of the structure of the hardware circuit and the integrated protective shell.

[0020] Figure 3This is a schematic diagram of the data acquisition and wireless transmission link for the 32-channel surface electromyography signal of the present invention, showing the complete process from flexible array electrode acquisition, analog-to-digital conversion, main control processing to Wi-Fi wireless transmission to the host computer.

[0021] Figure 4 This is a schematic diagram of the actual wearing and attachment position of the WFWES of the present invention on the human wrist, showing the flexible array electrodes covering the wrist flexor and extensor muscles and the fixation method of the hardware circuit.

[0022] Figure 5 This is a schematic diagram of 25 standard hand gestures used for experimental verification of the present invention.

[0023] Figure 6 This is a schematic diagram of the overall network structure of the multi-scale residual attention network model used in this invention.

[0024] Figure 7 Schematic diagrams of 12 progressive fine-grained thumb and finger micro-movement gestures designed for this invention.

[0025] Figure 8 This is a schematic diagram of the algorithm for detecting activated fragments based on absolute average-root mean square (RMS) according to the present invention.

[0026] Figure 9 This is a schematic diagram of the real-time human-computer interaction control process for two-person collaboration implemented on the Unity platform according to the present invention, showing the complete process from electromyography acquisition, gesture recognition to virtual character control.

[0027] Figure 10 This is a schematic diagram illustrating the effects of different progressive thumb movements on the independent and synchronous control of two virtual characters according to the present invention.

[0028] Figure 11 This is a schematic diagram of a single-person three-dimensional immersive virtual reality rehabilitation interaction scene after the present invention is integrated into a commercial cognitive rehabilitation device.

[0029] Figure 12 This is a schematic diagram of typical interactive actions in a single-person 3D virtual reality rehabilitation scenario of the present invention, wherein: (i) is a schematic diagram of view rotation based on thumb movements; (ii) is a schematic diagram of a virtual character performing various interactive behaviors by controlling micro-motions with one hand.

[0030] In the diagram: 1: Flexible array electrode; 2: Hardware circuit; 3: Host computer interface; 4: Interaction module; 1-1: Serpentine wiring structure; 1-2: 2×16 channel flexible electrode array; 2-1: Upper shell; 2-2: Power management module; 2-3: FPC connector; 2-4: Lower shell; 2-5: Power supply lithium battery; 2-6: Wi-Fi module; 2-7: Middle shell; 2-8: Main control module; 2-9: Electrophysiological amplification and acquisition chip; 3-1: TCP communication module; 3-2: Data parsing and filtering module; 3-3: Arbitrary channel waveform display module; 3-4: Data storage module; 3-5: Single channel spectrum analysis module; 3-6: Neural network model deployment and real-time inference module. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with embodiments and appendices. Figure 1-12 The present invention will be further described in detail below. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0032] The present invention will be further described below with reference to specific embodiments, but the scope of protection of the present invention is not limited to the following embodiments. Equivalent substitutions or modifications made by those skilled in the art without departing from the technical concept of the present invention should all fall within the scope of protection of the present invention.

[0033] Example 1: This example provides a wireless, flexible, wrist-worn 32-channel high-density surface electromyography (SEMG) real-time gesture human-computer interaction system (WFWES) for cognitive rehabilitation. It is used to achieve stable acquisition, real-time recognition, and human-computer interaction control of wrist and finger micro-movements under natural wearing conditions. The system mainly consists of a flexible array electrode 1, hardware circuitry 2, a host computer interface 3, and an interaction module 4.

[0034] like Figure 1 , Figure 2 As shown, the flexible array electrode 1 is a 2×16 channel flexible array structure. The electrodes are electrically connected through a serpentine wiring structure 1-1, thus ensuring both channel density and flexibility and bending resistance. The flexible array electrode 1 is attached to the skin surface of the human wrist, covering the wrist flexor and extensor muscle groups. It is stably fixed using medical pressure-sensitive tape to ensure low-impedance contact between the electrode and the skin surface and reduce the impact of motion artifacts on signal acquisition.

[0035] Hardware circuit 2 is connected to flexible array electrode 1 via FPC connector 2-3. It adopts an integrated protective housing structure, including an upper housing 2-1, a middle housing 2-7, and a lower housing 2-4. The middle housing 2-7 isolates the supporting circuitry from the power supply lithium battery 2-5 in layers, improving physical protection while optimizing battery heat dissipation efficiency. Internally, hardware circuit 2 includes a power management module 2-2, an electrophysiological amplification and acquisition chip 2-9, a main control module 2-8, and a Wi-Fi module 2-6. The power management module 2-2 employs a zoned power supply design, providing a stable 3.3V operating power to both analog and digital circuits, effectively reducing cross-domain interference.

[0036] During signal acquisition and transmission, the weak surface electromyography (EMG) signals (μV level) acquired by the flexible array electrodes 1 are first fed into the electrophysiological amplification and acquisition chip 2-9. Inside the chip, high-fidelity amplification, noise suppression, bandpass filtering, and analog-to-digital conversion are performed to obtain digital sampling data. Subsequently, the digital sampling data is transmitted to the main control module 2-8 (STM32U5) via the serial peripheral interface SPI2 and temporarily stored in a circular buffer. The main control module 2-8 sequentially reads the sample data from the buffer, frames and encapsulates it according to a custom communication protocol, and sends the encapsulated data to the Wi-Fi module 2-6 via the SPI1 interface. The Wi-Fi module 2-6 wirelessly transmits the EMG data to the host computer interface 3 based on the transmission control protocol TCP, achieving real-time transmission of 32-channel surface EMG signals. The overall dimensions of this hardware module are 48×47×16.2mm, and it weighs approximately 27.8g, featuring lightweight design, miniaturization, and good wearing comfort.

[0037] like Figure 3 , Figure 4As shown, the host computer interface 3 is used to receive, parse, display, and store the electromyographic data, and deploy a gesture recognition model to output real-time recognition results. The host computer interface 3 uses QtDesigner 6.9.0 for interface layout design and Python combined with PyQt5 for functional development. It mainly includes a TCP communication module 3-1, a data parsing and filtering module 3-2, an arbitrary channel data waveform display module 3-3, a data storage module 3-4, a single-channel spectrum analysis module 3-5, and a neural network model deployment and real-time inference module 3-6. The TCP communication module 3-1 is used to establish a wireless communication connection with the lower-level hardware; the data parsing and filtering module 3-2 is used to parse the received electromyographic data according to the lower-level hardware data transmission protocol and perform bandpass filtering and power frequency suppression processing; the waveform display module 3-3 allows users to select any channel for high refresh rate waveform display as needed; the data storage module 3-4 is used to store the gesture data collected during the experiment; the spectrum analysis module 3-5 is used to characterize the frequency domain characteristics of single-channel electromyographic signals; the neural network model deployment and real-time inference module 3-6 is used to load and run the gesture recognition model, output real-time gesture recognition results, and generate corresponding control events.

[0038] Regarding the gesture recognition method, this embodiment employs a time-sliding window-based absolute average-root mean square (MAV-RMS) dual-threshold activation detection algorithm to extract activation segments from continuously acquired 32-channel surface electromyography (EMG) signals in real time. With a time window length of 100ms and an overlap rate of 90%, the algorithm first selects four core channels (channels 27–30) with the optimal SNR distribution from the 32 channels based on the signal-to-noise ratio distribution during the static gesture phase, thus reducing the computational complexity of real-time processing. Subsequently, the core channel signals are z-score standardized using the global mean and standard deviation of the training set to improve the consistency of signals across subjects. The absolute average of the selected core channels is calculated window by window, and the root mean square is further calculated as the envelope strength index. When the envelope strength first exceeds the fluctuation threshold α ≥ 0.4, it is marked as an activated segment. Within a subsequent 120ms observation window, it is monitored whether the strength exceeds the intensity threshold β ≥ 1.0 a second time. If the condition is met, the activation starting point is determined, and an EMG segment containing all 32 channels is automatically extracted for subsequent recognition.

[0039] In the model recognition stage, such as Figure 6As shown, this embodiment uses the multi-scale residual attention network model MSE-Net to classify and identify the electromyographic segments. This model consists of three sequentially stacked multi-scale convolutional blocks, each containing four parallel one-dimensional convolutional branches to cover the receptive fields of features at different time scales. The branch outputs are concatenated in the channel dimension and then fused through 1×1 convolutions, with a residual attention mechanism incorporated to enhance the modeling ability of key signal channels. The model's final layer sequentially employs batch normalization, ReLU activation, adaptive global pooling, Dropout, and a fully connected layer to output the probability distribution corresponding to the gesture category. This model is implemented using the PyTorch framework and deployed on the host computer interface 3 for real-time inference.

[0040] The recognized gesture tags are converted into keyboard or mouse control events through an event mapping mechanism and sent to the interaction module 4. The interaction module 4 is used to drive virtual scenes or cognitive rehabilitation training tasks to achieve real-time human-computer interaction control.

[0041] Example 2: Based on the system and method described in Example 1, this example further provides a real-time human-computer interaction application example based on progressive fine-grained thumb movements, used to verify the practicality and stability of the system in complex cognitive rehabilitation tasks and immersive interaction scenarios.

[0042] like Figure 5 As shown, this embodiment first collects surface electromyography (EMG) signals from 6 healthy adult male subjects (aged 25.0 ± 2.19 years) based on 25 preset standard gestures. Each subject performed each gesture 5 times. The raw EMG signals were bandpass filtered from 20 to 450 Hz to remove DC drift and high-frequency noise, and a comb notch filter was used to suppress 50 Hz power frequency and its integer multiples of harmonic interference to obtain continuous motion EMG signals. Subsequently, the start and end times of the gestures were located by EMG activation detection. A sliding time window with a 320 ms window and a 160 ms step size was used to extract signal segments, and a standardized dataset was constructed for model training and validation.

[0043] Based on the above dataset, the MSE-Net model described in Example 1 was further designed and trained. The offline dataset was divided into training, validation, and test sets in a ratio of 0.6:0.2:0.2. The model was trained using a batch size of 128 and 100 training epochs, with AdamW as the optimizer and an initial learning rate of 4×10⁻⁻⁴. 4 It also combines cosine annealing with a warm-up strategy (8 rounds of warm-up, minimum learning rate 1×10⁻). 6 Meanwhile, an early stop mechanism is used to monitor and verify accuracy in order to avoid overfitting.

[0044] Based on this, such as Figure 7As shown, this embodiment further designs 12 progressive fine-grained hand gesture commands that include thumb half-range movements, covering 8 thumb progression movements such as thumb half-right swipe (HC-1, 20°), thumb right swipe (HC-2, 40°), thumb half-up swipe (HC-6, 30°), and thumb up swipe (HC-5, 60°), as well as 4 index finger micro-movement movements. Using a data acquisition protocol consistent with standard gestures, corresponding electromyographic signals were collected from 5 healthy adult subjects. Each subject contributed 1200 electromyographic segments, ultimately constructing a progressive fine-grained thumb movement dataset containing 7200 samples for progression gesture recognition tasks.

[0045] like Figure 8 , Figure 9 As shown, during real-time interaction, the 32-channel surface electromyography (EMG) signals collected by WFWES are used to detect activated segments in real time using the MAV-RMS dual-threshold algorithm described in Example 1. These signals are then input into the MSE-Net model deployed on the host computer for inference, outputting 12 types of gesture labels. These gesture labels are mapped to keyboard and mouse events via the pynput library and sent to the interaction module developed based on the Unity platform, thereby driving the virtual character's movements.

[0046] This embodiment constructs a two-person collaborative human-computer interaction scenario on the Unity platform, realizing the independent and synchronous control mapping of two virtual characters through a single-handed, process-based thumb action. The full-process action controls character 1, the partial-process action controls character 2, and the thumb click action (HC-10) enables simultaneous attacks by both characters. For example... Figure 10 As shown, different progressive thumb movements can achieve independent or collaborative control effects for two characters.

[0047] Furthermore, this embodiment seamlessly integrates the system into commercial cognitive rehabilitation training equipment, and develops a single-person interactive mode that supports 3D VR display, such as... Figure 10 Figure 11 and Figure 10 Figure 12 As shown. To enrich the expression of interactive behavior, a continuous action triggering mechanism is introduced, which binds logically related actions such as "pick up and use" and "get on the vehicle and drive". When the preceding action of a continuous action is triggered, the triggering switch of the subsequent action is automatically activated, thereby achieving the coverage of 20 complex interactive actions under a unified interaction framework.

[0048] Experimental results show that the WFWES system can achieve zero-configuration plug-and-play functionality in both two-dimensional planar interactive systems and three-dimensional VR rehabilitation systems. The action response is smooth and lag-free, and the recognition accuracy of core interactive actions can reach 100%. This verifies the practicality, stability, and cross-platform compatibility of the system in cognitive rehabilitation and immersive human-computer interaction scenarios.

[0049] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A 32-channel surface electromyography real-time gesture human-computer interaction system for cognitive rehabilitation, characterized in that, include: A wireless wrist-worn 32-channel high-density surface electromyography (EMG) acquisition device for acquiring surface EMG signals from the wrist. The host computer interface is used to establish a communication connection with the acquisition device and to parse, display and store electromyographic data, while deploying a gesture recognition model to output gesture recognition results; An interaction module is used to receive the gesture recognition results and map them into virtual scene control events to drive cognitive rehabilitation training tasks; The wireless wrist-worn 32-channel high-density surface electromyography acquisition device consists of a flexible array electrode (1) and a hardware circuit (2). The flexible array electrode (1) is a 2×16-channel flexible array structure to adapt to the curved surface of the wrist. The hardware circuit (2) includes a main control module (2-8), a power management module (2-2), an FPC connector (2-3), an electrophysiological amplification and acquisition chip (2-9), and a Wi-Fi module (2-6).

2. The system according to claim 1, characterized in that, The flexible array electrode (1) adopts a serpentine wiring structure (1-1) to achieve electrical connection and form a 2×16 channel electrode array (1-2) to ensure the spatial coverage and adhesion flexibility of signal acquisition.

3. The system according to claim 1, characterized in that, The flexible array electrode (1) covers the wrist flexor and extensor muscles in the wrist area, and uses medical pressure-sensitive tape to achieve stable contact and low impedance matching between the electrode and the skin surface.

4. The 32-channel surface electromyography real-time gesture human-computer interaction system for cognitive rehabilitation according to claim 1, characterized in that, The hardware circuit (2) has an integrated protective shell, which is a three-layer structure consisting of an upper shell (2-1), a middle shell (2-7), and a lower shell (2-4); the middle shell (2-7) isolates the supporting circuit from the power supply lithium battery (2-5) in upper and lower layers.

5. A 32-channel surface electromyography real-time gesture human-computer interaction system for cognitive rehabilitation according to claim 1, characterized in that, The data transmission link of the acquisition device is as follows: the electromyographic signal acquired by the flexible array electrode (1) is amplified, bandpass filtered and analog-to-digital converted by the electrophysiological amplification and acquisition chip (2-9) to obtain digital sampling data. Then, it is transmitted to the main control module (2-8) through the serial peripheral interface SPI2 and written into the ring buffer for temporary storage. The main control module (2-8) then frames and encapsulates the data and transmits it to the Wi-Fi module (2-6) through SPI1. The Wi-Fi module (2-6) wirelessly transmits the encapsulated electromyographic data to the host computer interface based on the transmission control protocol TCP.

6. A gesture recognition and interactive control method based on the system according to any one of claims 1 to 5, characterized in that, include: Acquisition steps: Acquire 32-channel surface electromyography signals from the wrist using a flexible array electrode (1); Transmission steps: After the electromyographic signal is amplified, filtered, converted from analog to digital and framed by the hardware circuit (2), it is wirelessly transmitted to the host computer interface via Wi-Fi. Processing steps: The host computer interface performs protocol parsing and filtering on the received data and forms input segments for identification; Recognition steps: Call the multi-scale residual attention network model deployed on the host computer to classify the gestures of the input segment and output gesture labels; Interaction steps: Map the gesture tags to control events and send them to the interaction module to drive the character behavior response in the virtual rehabilitation scenario.

7. The method according to claim 6, characterized in that, The host computer interface is laid out using Qt Designer and implemented using Python + PyQt5. It includes at least a TCP connection module (3-1), a data parsing and filtering module (3-2), an arbitrary channel waveform display module (3-3), a data storage module (3-4), a single-channel spectrum graph module (3-5), and a model deployment and real-time inference module (3-6).

8. The method according to claim 6, characterized in that, Real-time activation extraction was performed using a time-sliding window mechanism and an absolute average-root mean square dual-threshold activation fragment selection method: the time window length was 100ms and the overlap rate was 90%; four core channels with the best signal-to-noise ratio distribution were selected from 32 channels, namely channels 27-30, and the activation fragment was marked when the envelope intensity first exceeded the fluctuation threshold α≥0.

4. The activation start point was determined when the intensity exceeded the intensity threshold β=1.0 twice in the subsequent observation window, and the electromyographic fragment containing all 32 channels was extracted for subsequent identification.

9. The method according to claim 6, characterized in that, The multi-scale residual attention network model is MSE-Net, which uses three sequentially stacked multi-scale convolutional blocks. Each convolutional block contains four parallel convolutional branches. The branch outputs are concatenated in the channel dimension, fused by a 1×1 convolution, and then connected to the residual attention mechanism. The branch kernel length and dilation rate parameters of the three convolutional blocks are as follows: The first layer consists of (3,1), (5,2), (7,4), and (9,4). The second layer consists of (3,1), (7,2), (11,4), and (15,8). The third layer consists of (3,1), (9,2), (15,4), and (21,8).

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 method as described in any one of claims 6 to 9.

Images