Hand motion intention recognition and control method and device based on AI neural decoding

Abstract

The application belongs to the field of artificial intelligence, and relates to a hand motion intention recognition and control method and device based on AI neural decoding, the method comprising: collecting multi-modal signals facing the peripheral nerves of the wrist to obtain original neural data streams; performing neural data compression and wireless transmission based on feature reservation according to the original neural data streams; pre-processing the received and decompressed original neural data streams and extracting neural features; decoding real-time motion intention based on a hybrid deep learning architecture according to the extracted neural features to obtain abstract motion intention; programmably and contextually mapping and generating control commands from the abstract motion intention; and performing incremental model training and calibration based on semi-supervised learning. The application realizes high-precision and low-invasive motion intention decoding, has good practicability and adaptability, and provides a solution for intelligent prosthetics, rehabilitation training and natural human-computer interaction.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for recognizing and controlling hand movement intentions based on AI neural decoding. Background Technology

[0002] Human-machine interfaces enable users to interact with digital systems using physical or physiological input. Most traditional interfaces, including game controllers, joysticks, keyboards, mice, and touchscreens, rely on direct mechanical interaction, which can limit the speed, precision, or accessibility for some users. Various alternative interfaces—such as electromyography-based control, motion tracking, eye tracking, and camera-based gesture recognition—have been developed to provide more natural control modes. While each system has its advantages in specific scenarios, they also have limitations, including sensitivity to environmental conditions, dependence on gaze, and reduced robustness in dynamic environments.

[0003] Existing invasive neural interfaces, such as implanted electrodes placed on or inside the brain or other neural tissue, can provide high-fidelity signals, but require surgery, pose medical risks, and are not suitable for widespread consumer use. Non-invasive brain-computer interface systems, such as EEG-based systems, avoid surgical implantation, but typically exhibit low spatial resolution, low signal-to-noise ratio, and limited responsiveness to fine or real-time control tasks such as gaming.

[0004] Peripheral neuro-computer interfaces offer an alternative, working by extracting neural signals from the nerves of the limbs, where movement-related activity is more concentrated and easier to interpret. Existing wearable electromyography (EMG) devices attempt to infer user intent from muscle activity rather than neural activity and are often plagued by muscle crosstalk, motion artifacts, and limited signal specificity. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a hand movement intention recognition and control method based on AI neural decoding, employing the following technical solution, including the following steps:

[0006] Multimodal signals from peripheral nerves in the wrist were acquired to obtain raw neural data streams;

[0007] Based on the original neural data stream, neural data compression and wireless transmission based on feature preservation are performed;

[0008] The received and decompressed raw neural data stream is preprocessed and neural features are extracted.

[0009] Based on the extracted neural features, real-time motion intent decoding is performed using a hybrid deep learning architecture to obtain abstract motion intent;

[0010] The abstract motion intent is programmable and context-aware control command mapping and generation is performed.

[0011] Incremental model training and calibration are performed based on semi-supervised learning.

[0012] Preferably, the step of acquiring multimodal signals from the peripheral nerves of the wrist and obtaining the raw neural data stream specifically includes:

[0013] The recording electrodes were positioned in the surface projection areas of the radial, median, and ulnar nerves, and a multi-electrode array was configured to record the amplitude of multimodal signals and sense the propagation direction of nerve impulses.

[0014] At the forefront of signal acquisition, the microvolt-level neural signal is amplified, environmental noise, electromyographic interference, and motion artifacts are suppressed to obtain the original neural data stream;

[0015] Hardware-level suppression of common-mode interference and motion artifacts.

[0016] Preferably, the step of performing feature-preserving neural data compression and wireless transmission based on the original neural data stream specifically includes:

[0017] Based on the original neural data stream, neural data compression is performed using sparse representation and perceptual coding.

[0018] The wireless transmission parameters are dynamically adjusted based on signal characteristics, decoding performance requirements, and device battery status.

[0019] The original neural data stream is wirelessly transmitted based on forward error correction and packet loss concealment.

[0020] Preferably, the step of preprocessing the received and decompressed raw neural data stream and extracting neural features specifically includes:

[0021] The received and decompressed raw neural data stream is subjected to adaptive filtering and power frequency interference elimination.

[0022] The signal-to-noise ratio (SNR) of each channel is dynamically evaluated, and a subset of channels whose SNR exceeds a preset optimal SNR are selected as channels for neural feature extraction.

[0023] From the selected subset of channels, a set of features that characterize the changes in neural activity patterns in different dimensions are extracted from the original neural data stream after adaptive filtering and power frequency interference elimination, forming a feature vector.

[0024] Preferably, the step of decoding the abstract motion intent in real time based on the extracted neural features using a hybrid deep learning architecture specifically includes:

[0025] Construct a convolutional neural network branch for spatial feature extraction;

[0026] Construct a recurrent neural network branch for time-dynamic modeling;

[0027] Multi-task learning and lightweight model optimization are performed to enable a single model to simultaneously output discrete gesture classification and continuous joint angle regression, and optimization is performed for the limited computing resources of embedded decoding devices.

[0028] Preferably, the step of mapping and generating programmable, context-aware control commands for the abstract motion intent specifically includes:

[0029] Based on the application scenario and user preferences, the abstract motion intent is mapped into discrete trigger commands or continuous proportional control;

[0030] Perform two-handed coordination and generate advanced compound commands;

[0031] Automatically switch application configurations based on the running application.

[0032] Preferably, the step of incremental model training and calibration based on semi-supervised learning specifically includes:

[0033] Guided gestures are used to induce neural signal data, which is collected simultaneously for training purposes.

[0034] Automatically generate interval labels and gesture category labels for the collected neural signal data;

[0035] Using newly collected user data, update or retrain the model in the cloud and deploy the optimized model back to the user's device.

[0036] To address the aforementioned technical problems, this invention also provides a hand movement intention recognition and control device based on AI neural decoding, employing the following technical solution, including:

[0037] The acquisition module is used to collect multimodal signals from the peripheral nerves of the wrist and acquire the raw neural data stream.

[0038] A compression module is used to perform feature-preserving neural data compression and wireless transmission based on the original neural data stream;

[0039] The extraction module is used to preprocess the received and decompressed raw neural data stream and extract neural features.

[0040] The decoding module is used to perform real-time motion intent decoding based on the extracted neural features and a hybrid deep learning architecture to obtain abstract motion intent.

[0041] The mapping module is used to map and generate programmable, context-aware control commands for the abstract motion intent;

[0042] The training module is used for incremental model training and calibration based on semi-supervised learning.

[0043] To address the aforementioned technical problems, the present invention also provides a computer device that employs the technical solution described below, comprising a memory and a processor. The memory stores computer-readable instructions, and the processor, when executing the computer-readable instructions, implements the steps of the aforementioned AI neural decoding-based hand movement intention recognition and control method.

[0044] To address the aforementioned technical problems, the present invention also provides a computer-readable storage medium, which employs the technical solution described below. The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the aforementioned AI neural decoding-based hand movement intention recognition and control method.

[0045] Compared with the prior art, the present invention has the following main advantages:

[0046] (1) By collecting multimodal signals through the peripheral nerve of the wrist, the high invasive risk of traditional brain-computer interfaces is avoided, and rich neural information can be obtained. By combining feature-preserving compression technology and hybrid deep learning architecture, stable features can be extracted from high-dimensional data, and real-time and accurate decoding of fine movements such as fingers and wrists can be achieved. The control latency is low and the reliability is strong, realizing high-precision and low-invasive motion intention decoding.

[0047] (2) It adopts wireless transmission and lightweight processing, making it easy to wear and use on the go; through programmable, context-aware command mapping, it can generate corresponding instructions according to different tasks (such as gripping, typing, and operating the device), which significantly enhances the naturalness and flexibility of human-computer interaction and has good practicality and adaptability.

[0048] (3) Relying on semi-supervised learning for incremental training and calibration, it can continuously adapt to the changes in the user's neural signal characteristics and habits during use, gradually improve decoding accuracy, reduce dependence on a large amount of labeled data, and is more suitable for long-term practical applications, supporting continuous optimization and personalized adaptation; it has achieved a balance in three aspects: technical feasibility, user experience and system sustainability, and provides efficient solutions for fields such as intelligent prostheses, rehabilitation training and natural human-computer interaction. Attached Figure Description

[0049] To more clearly illustrate the solutions in this invention, the accompanying drawings used in the description of the embodiments of this invention will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0050] Figure 1 This is a flowchart of an embodiment of the hand movement intention recognition and control method based on AI neural decoding of the present invention;

[0051] Figure 2 This is a schematic diagram illustrating an example use case of the AI ​​neural decoding-based hand movement intention recognition and control method of the present invention applied to games;

[0052] Figure 3 This is a schematic diagram of the system architecture for implementing the AI ​​neural decoding-based hand movement intention recognition and control method of the present invention;

[0053] Figure 4 This is a schematic diagram of a neural interface device worn on the wrist used in the hand movement intention recognition and control method based on AI neural decoding of the present invention.

[0054] Figure 5 This is a detailed schematic diagram of the neural interface device used in the AI ​​neural decoding-based hand movement intention recognition and control method of the present invention;

[0055] Figure 6 This is a schematic diagram of the neural recording configuration of each electrode group used in the AI ​​neural decoding-based hand movement intention recognition and control method of the present invention;

[0056] Figure 7 This is a schematic diagram illustrating the mapping between the decoded motion intent and the output command used in the AI ​​neural decoding-based hand motion intent recognition and control method of the present invention.

[0057] Figure 8 This is a schematic diagram of the deep learning-based AI model used for neural decoding in the hand movement intention recognition and control method based on AI neural decoding of the present invention.

[0058] Figure 9 This is a flowchart of the AI ​​model training data collection process used in the AI ​​neural decoding-based hand movement intention recognition and control method of the present invention.

[0059] Figure 10 This is a schematic diagram of the data collected by the training AI model used in the AI-based neural decoding hand movement intention recognition and control method of the present invention;

[0060] Figure 11This is a schematic diagram of a structure of an embodiment of the hand movement intention recognition and control device based on AI neural decoding of the present invention;

[0061] Figure 12 This is a schematic diagram of the structure of an embodiment of the computer device of the present invention.

[0062] Figure labels: 1-Game console or computer, 2-Neural decoding device, 3-Neural interface device, 4-Ulnar nerve, 5-Median nerve, 6-Radial nerve, 7-Surface electrode, 8-Ulnar nerve electrode, 9-Main body of device, 10-Median nerve electrode, 11-Radial nerve electrode, 12-Ground / reference electrode, 13-Wristband, 14-Electrode, 15-Peripheral nerve, 16-Active gesture, 17-Rest posture. Detailed Implementation

[0063] 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 specification is for the purpose of describing particular embodiments only and is not intended to limit the invention; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings are used to distinguish different objects and not to describe a particular order.

[0064] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0065] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0066] It should be noted that the hand movement intention recognition and control method based on AI neural decoding provided in the embodiments of the present invention is generally executed by a server / terminal device, and correspondingly, the hand movement intention recognition and control device based on AI neural decoding is generally set in the server / terminal device.

[0067] It should be understood that the number of terminal devices, networks, and servers is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be used.

[0068] Example 1

[0069] Please refer to Figure 1 The diagram illustrates a flowchart of an embodiment of the AI ​​neural decoding-based hand movement intention recognition and control method of the present invention. The AI ​​neural decoding-based hand movement intention recognition and control method includes the following steps:

[0070] Step S1: Acquire multimodal signals from the peripheral nerves of the wrist to obtain the raw neural data stream.

[0071] In this embodiment, the electronic device (e.g., a server / terminal device) on which the AI ​​neural decoding-based hand motion intention recognition and control method runs can receive the AI ​​neural decoding-based hand motion intention recognition and control request via a wired or wireless connection. It should be noted that the aforementioned wireless connection methods may include, but are not limited to, 3G / 4G / 5G connections, WiFi connections, Bluetooth connections, WiMAX connections, Zigbee connections, UWB (ultra wideband) connections, and other currently known or future known wireless connection methods.

[0072] In this embodiment, step S1, acquiring multimodal signals from the peripheral nerves of the wrist to obtain the raw neural data stream, may specifically include the following steps:

[0073] S11, the recording electrodes are positioned in the surface projection areas of the radial nerve, median nerve and ulnar nerve, and a multi-electrode array is configured to record the amplitude of multimodal signals and sense the propagation direction of nerve impulses.

[0074] The wristband is designed with three independent electrode clusters, corresponding to common surface locations of the radial nerve (near the radial styloid process on the dorsal side of the wrist), median nerve (below the transverse carpal ligament on the palmar side of the wrist), and ulnar nerve (near the pisiform bone on the ulnar side of the wrist). The position of each electrode cluster can be finely adjusted in millimeters via a micro-slide rail or elastic adjustment mechanism to accommodate anatomical differences among users and ensure that the electrodes are located at the position of strongest nerve signal.

[0075] A longitudinal three-electrode differential configuration is possible, providing each target nerve with three electrodes arranged longitudinally along the nerve's course: a proximal electrode, a central electrode, and a distal electrode. The central electrode serves as a common reference point. This configuration creates two differential recording pairs: proximal electrode-central electrode and distal electrode-central electrode.

[0076] A directional information extraction circuit was constructed, connecting two differential pairs to two high-performance, low-noise bioelectric amplifiers (such as the IA in a custom neural recording chip). By comparing the time delay (Δt) and phase relationship between the output signals of the two amplifiers, the propagation direction of the action potential along the nerve fiber can be inferred. For example, if the signal peak of the proximal-central pair is earlier than that of the distal-central pair, it indicates that the nerve impulse is propagating distally (to the hand), possibly corresponding to a motor (efferent) command; conversely, it may correspond to sensory (infferent) feedback. This information provides an additional dimension for subsequent decoding.

[0077] The function of step S11 is to sense the direction of nerve impulse propagation, such as afferent / efferent, thereby reflecting the state of neural activity more comprehensively.

[0078] S12, amplify the microvolt-level neural signal at the forefront of signal acquisition, suppress environmental noise, electromyographic interference and motion artifacts, and acquire the original neural data stream.

[0079] A custom ASIC (Application-Specific Integrated Circuit) is used, with its input stage employing a special layout of complementary metal-oxide-semiconductor transistors to achieve an input impedance higher than 1 GΩ, minimizing signal attenuation caused by electrode-skin contact impedance variations. The amplifier's equivalent input noise density must be lower than [missing value]. @ 1kHz.

[0080] The ASIC chip integrates a programmable gain amplifier, with a gain range typically adjustable between 100 and 5000 times to accommodate differences in signal intensity among different users and neurons. Simultaneously, it integrates an analog bandpass filter, set to (100 Hz - 2 kHz), to preserve the effective spectral components of peripheral nerve signals (primarily concentrated in the hundreds to thousands of hertz), and to initially filter out low-frequency motion artifacts (<100 Hz) and high-frequency noise (>2 kHz).

[0081] Synchronous sampling and analog-to-digital conversion are performed. All channels (3 neurons × 2 differential pairs = 6 channels in this embodiment) use a synchronous sampling ADC (analog-to-digital converter) with a sampling rate set to at least 4 kHz (satisfying the Nyquist sampling theorem and allowing for margin). A high-precision (e.g., 16-bit) ADC ensures that minute changes in neural signals can be digitally captured.

[0082] The purpose of step S12 is to obtain raw data with high signal-to-noise ratio (SNR).

[0083] S13 performs hardware-level suppression of common-mode interference and motion artifacts.

[0084] An integrated circuit drives the right leg. The overall common-mode voltage (primarily power frequency interference) is monitored via a separate reference electrode (such as a large-area electrode on the underside of the device body). Then, an inverting amplifier injects an out-of-phase signal into the user's body (usually via the same reference electrode), thereby actively canceling the common-mode interference sensed at the recording electrode, suppressing power frequency interference by 20-40 dB.

[0085] A hardware high-pass filter with a cutoff frequency of approximately 0.5-1 Hz (such as an RC circuit or a switched capacitor filter) is incorporated into the amplifier chain to block extremely low-frequency baseline drift caused by slow changes in skin-electrode interface potential or large limb movements, thus preventing saturation of subsequent circuits.

[0086] The neural recording chip and front-end circuitry are encapsulated within an electromagnetically shielded shell. Shielded electrode leads are used, and the shielding layer is connected to the output of the circuitry driving the right leg, forming active shielding that effectively suppresses spatial electromagnetic coupling noise. The device casing and wristband are made of flexible and conformal materials to reduce frictional electrical noise caused by relative movement.

[0087] The purpose of step S13 is to actively suppress 50 / 60 Hz power frequency interference and low-frequency artifacts generated by wrist movement through hardware circuit design before the signal is digitized, so as to prevent these interference signals from drowning out weak nerve signals or saturating the preamplifier.

[0088] The purpose of step S1 is to non-invasively and stably capture peripheral nerve electrical activity related to hand movement intentions from the complex anatomy of the wrist. This solves the problems of weak and easily interfered surface electrode signals, as well as the deep and parallel location of nerves in the wrist.

[0089] Step S2: Based on the original neural data stream, perform neural data compression and wireless transmission based on feature preservation.

[0090] In this embodiment, step S2 may specifically include the following steps:

[0091] S21, based on the original neural data stream, perform neural data compression based on sparse representation and perceptual coding.

[0092] Discrete wavelet transform is performed on the neural signal segment (e.g., a 256-millisecond window) for each channel, decomposing the signal into different frequency band subspaces. The characteristics of the neural signal (e.g., action potential waveform) exhibit sparsity in the wavelet domain (i.e., most coefficients are close to zero). The wavelet coefficient calculation formula is: ,in, These are primitive neural signals. It is the wavelet mother function. It is a scale parameter (corresponding to frequency). It is the translation parameter (corresponding to time). These are wavelet coefficients. The asterisk * indicates complex conjugate. This transformation maps the time-domain signal to the time-frequency domain, facilitating coefficient selection based on frequency importance.

[0093] Based on the prior knowledge of the decoding model, different perceptual weights are assigned to different wavelet subbands. Mid-to-high frequency subbands, which are highly correlated with action potential morphology, are given high weights and retain more bits; low-frequency noise and ultra-high frequency irrelevant components are given low weights and subjected to aggressive quantization or directly set to zero. The retained wavelet coefficients are then subjected to scalar or vector quantization.

[0094] The quantized non-zero wavelet coefficients and their position information (such as run-length encoding) are further compressed using Huffman coding or arithmetic coding to reduce the average code length by taking advantage of their statistical properties.

[0095] The purpose of step S21 is to compress the data at a rate much higher than that of traditional general compression algorithms, while retaining as many signal features as possible that are most important for decoding motion intent and discarding redundant information that is not perceptually important.

[0096] S22 dynamically adjusts wireless transmission parameters based on signal characteristics, decoding performance requirements, and device battery status.

[0097] The compressed data packets are categorized into high, medium, and low priorities. High-priority packets contain the highest-energy or most significant coefficients in the current window's sparse representation and must be transmitted. Low-priority packets contain detailed information and can be selectively discarded when the channel is congested or the battery is low. The receiver can then reconstruct the data at different quality levels based on the received packets.

[0098] Wireless modules (such as BLE 5.0+) monitor link quality (e.g., Received Signal Strength Indicator (RSSI) and Packet Error Rate (PER)). When link quality is good, higher-order modulation (e.g., 2Mbps PHY) and low error correction overhead are used to improve effective throughput; when quality is poor, it automatically switches to more robust lower-order modulation and stronger forward error correction coding to ensure data reliability.

[0099] The microcontroller works in conjunction with the wireless module. During the intervals between data acquisition and compression, the wireless module and processor are quickly put into deep sleep mode. The wake-up cycle is precisely calculated based on the update rate required for real-time control (e.g., 50-100 Hz) to minimize average power consumption.

[0100] The purpose of step S22 is to optimize power consumption while ensuring real-time performance in wireless streaming transmission with adaptive bandwidth and power management.

[0101] S23, based on forward error correction and packet loss concealment, the original neural data stream is wirelessly transmitted.

[0102] Concatenated codes are applied at the data link layer. For example, the inner code uses convolutional codes (constrained length K=7, code rate 1 / 2) for random error correction, while the outer code uses Reed-Solomon codes (such as RS(255, 223)) for burst error correction. This combination can effectively combat various types of interference in the wireless channel.

[0103] Selective retransmission with acknowledgment is used for high-priority data packets. For extremely critical control frames (such as system synchronization information), instantaneous redundant transmission (sending the same data in two consecutive time slots) can be used to sacrifice a small amount of bandwidth for extremely high reliability.

[0104] At the neural decoding device, if a data packet loss is detected, it is not directly set to zero, but rather hidden using an interpolation algorithm. For example, the wavelet coefficients of the corresponding channel successfully received in the previous time step are used to estimate the coefficients of the lost window through linear or spline interpolation, and then an inverse transform is performed to reconstruct the time-domain signal, ensuring that the data stream input to the decoding model is continuous in time.

[0105] The purpose of step S23 is to address the inherent instability of the wireless channel and ensure that even if data packets are lost or corrupted, the decoded control commands will not experience catastrophic jumps or interruptions.

[0106] The purpose of step S2 is to perform real-time, lossy but intelligent compression of the high sampling rate, multi-channel raw neural data stream (data rate up to 6 channels × 16 bits × 4000 samples / second = 384 kbps) on resource-constrained wearable devices to adapt to bandwidth-constrained wireless links such as Bluetooth Low Energy and ensure the reliability of transmission.

[0107] Step S3: Preprocess the received and decompressed raw neural data stream and extract neural features.

[0108] In this embodiment, step S3 may specifically include the following steps:

[0109] S31, adaptive filtering and power frequency interference elimination are performed on the received and decompressed raw neural data stream.

[0110] Multiple adaptive filters are connected in parallel to form a harmonic notch filter group. The reference signal is provided by locally generated sine / cosine pairs with the same frequency as the fundamental power frequency and major harmonics (e.g., 100 / 120 Hz, 150 / 180 Hz). The filter weights are updated using a normalized least mean square algorithm to estimate and subtract the power frequency interference component mixed into the neural signal in real time.

[0111] The NLMS algorithm formula is: ,in, It is the filter weight vector at time n. It is the step size factor. It is a small positive number used to prevent division by zero. It is the reference signal vector. It is the current error signal (the difference between the desired signal and the filter output). This algorithm can dynamically track minute changes in the phase and amplitude of power frequency interference.

[0112] To address the periodicity of power frequency interference, a recursive comb filter is designed to generate a deep notch at the interference frequency. Its transfer function in the Z-domain is expressed as: ,in, , It is the sampling rate. It's power frequency. It is a pole attenuation factor close to 1 (e.g., 0.995) used to control the notch bandwidth. This filter can effectively filter out the fundamental frequency and harmonics.

[0113] The signal spectrum is monitored in real time using short-time Fourier transform. When a persistent narrowband spike unrelated to the user's current action is detected (such as resonance at 5-15 Hz caused by equipment looseness), a second-order infinite impulse response bandstop filter is dynamically inserted, with its center frequency aligned with the spike, to filter it out.

[0114] The purpose of step S31 is to completely eliminate 50 / 60 Hz power frequency interference and its harmonics, and adaptively filter out specific frequency artifacts caused by non-ideal movements (such as slight wrist tremors).

[0115] S32, dynamically evaluate the signal-to-noise ratio of each channel, and select a subset of channels whose signal-to-noise ratio exceeds the preset optimal signal-to-noise ratio as channels for neural feature extraction.

[0116] The signal frequency band is divided into a signal band (e.g., 150-1000 Hz) and a noise band (e.g., 2000-3000 Hz, assuming that the neural activity energy in this range is negligible).

[0117] The power spectral density ratio of the two frequency bands within a short time window of each channel is calculated as an approximate estimate of the signal-to-noise ratio. The calculation formula is as follows: .

[0118] Calculate the low-frequency component (0.5-10 Hz) energy of each channel signal. If the energy rises sharply within a short period of time and exceeds an empirical threshold, the channel is considered to be severely contaminated by motion artifacts, and it is temporarily removed from the feature extraction channel list until the energy returns to normal.

[0119] Based on real-time estimated SNR and artifact status, for example, K channels with the highest SNR and no artifacts (e.g., K=4) are selected from 6 available channels. Feature extraction is performed only on these K channels. This simulates an attention mechanism, allowing the system to focus on the highest quality information source.

[0120] The purpose of step S32 is to address the fact that not all recorded channels consistently have good signal quality. This step dynamically evaluates the signal-to-noise ratio of each channel, selects the optimal subset of channels for subsequent feature extraction, and improves the robustness and adaptability of the system.

[0121] S33. From the selected channel subset, extract a set of features that characterize the changes in neural activity patterns in different dimensions from the original neural data stream after adaptive filtering and power frequency interference elimination, and form a feature vector.

[0122] The raw neural data was first preprocessed to remove unwanted components. This may include downsampling after applying an anti-aliasing filter, suppressing 50 / 60 Hz power line interference and / or its harmonics using a notch filter, isolating the frequency band of interest using a bandpass filter, and removing motion and / or stimulus artifacts. After preprocessing, a series of signal processing techniques were applied to extract descriptive features from the data. Feature extraction focused on key time-domain and frequency-domain features while significantly reducing the data dimensionality. The neural signals were segmented into non-overlapping 20 ms windows. For each window, multiple features were computed for each channel. A non-exhaustive list of these neural features—along with their formulas, descriptions, and principles—is shown in Table 1. These features are designed to capture fundamental characteristics of the signal, including amplitude, frequency components, energy, and spike potential activity. All features were normalized to the [0,1] range using the 0.5% and 99.5% percentiles of the dataset to reduce sensitivity to outliers. This process produced a feature vector time series with an effective sampling rate of 50 Hz, where each feature vector represents the neural dynamics within a 20 ms window.

[0123] Neural data features include, but are not limited to: mean square value (MSQ), waveform length (WL), mean absolute value (MAB), logarithmic detector (LD), third-order V quantity (V3), nonlinear energy operator (NEO), absolute standard deviation of difference (DABS), root mean square (RMS), maximum fractal length (MFL), mean absolute slope (MAVS), etc.

[0124] Table 1: List of Neural Data Features

[0125]

[0126] Differential channel energy ratio: For each nerve, the ratio of the signal energy of the two differential output pairs (proximal vs. distal) is calculated. This ratio may be related to the strength of muscle contraction or the topology of nerve activation.

[0127] Finally, the time-domain, frequency-domain, and spatiotemporal features of all channels are concatenated into a high-dimensional feature vector. , as input to the AI ​​decoding model.

[0128] The purpose of step S33 is to construct a time-frequency-space multidimensional feature engineering.

[0129] The purpose of step S3 is to purify and refine the received and decompressed raw neural signals, remove residual noise and artifacts, and convert the high-dimensional time-domain waveform into a low-dimensional and discriminative feature vector that better characterizes the essence of neural activity patterns, providing optimal input for the subsequent AI decoding model.

[0130] Step S4: Based on the extracted neural features, perform real-time motion intent decoding using a hybrid deep learning architecture to obtain abstract motion intent.

[0131] In this embodiment, step S4 may specifically include the following steps:

[0132] S41, Construct a convolutional neural network branch for spatial feature extraction.

[0133] One-dimensional feature vector The images are rearranged according to channel order and feature type to form a pseudo-image. For example, rows correspond to different channels (or neurons), and columns correspond to different features (MAV, WL, ZC, etc.). This allows CNNs to learn local spatial filters on the channel-feature plane using their convolutional kernels.

[0134] Parallel convolutional layers with varying kernel sizes (e.g., 1x3, 1x5, 1x7) are used to capture spatial patterns within different receptive fields. Smaller kernels focus on subtle relationships between neighboring channels / features, while larger kernels capture broader synergistic relationships. The outputs of each parallel branch are concatenated along the channel dimension.

[0135] A compression-activation module is introduced. First, global average pooling is performed on the convolutional feature map to obtain a global descriptor for each channel. Then, the importance weights of each channel are learned through two fully connected layers (with dimensionality reduction in between). Finally, these weights are multiplied by the original feature map to achieve channel-level feature recalibration, allowing the network to focus more on information-rich channels.

[0136] The purpose of step S41 is to effectively capture the spatial correlation between multichannel neural signals (such as the synergistic / antagonistic patterns between different recording sites within the same nerve or between different nerves), which are crucial for distinguishing different gestures.

[0137] S42, construct the recurrent neural network branch for time-dynamic modeling.

[0138] Use GRU instead of a basic RNN or LSTM because GRU has fewer parameters, is more computationally efficient, and is better suited for embedded deployments while maintaining similar performance. Stack 2-3 layers of GRU; higher-level GRUs learn dependencies over longer time scales. Serialize the feature vectors. Input the GRU network.

[0139] A bidirectional GRU is employed. The forward GRU processes the sequence from the past to the present, while the backward GRU processes the sequence from the future (within a sliding window) to the present. The final hidden states from both directions are concatenated, so that decoding is based not only on historical information but also incorporates the "context" information of the future short window, improving the accuracy of classification / regression at the current moment.

[0140] An attention layer is introduced on the top-level output sequence of BiGRU. This layer learns to assign different attention weights to the hidden states at different time steps in the sequence, allowing the model to focus on the few historical moments that are most critical to the current intention judgment, rather than treating all historical moments equally.

[0141] The purpose of step S42 is to demonstrate that neural activity exhibits strong temporal dependence. The intent at the current moment depends not only on the current neural features but also on the state of neural activity over the previous hundreds of milliseconds. RNN branches are specifically designed to model this temporal dynamic.

[0142] S43 performs multi-task learning and lightweight model optimization, enabling a single model to simultaneously output discrete gesture classification and continuous joint angle regression, and optimizes for the limited computing resources of embedded decoding devices.

[0143] Features extracted by CNN and BiGRU branches are fused after a fusion layer (e.g., concatenation or addition) to form a shared deep feature representation. Then, two parallel task heads are derived from this shared feature representation:

[0144] The classification head is a fully connected layer followed by a Softmax activation function, outputting the probability distribution of N discrete gesture categories. .

[0145] The regression head is a fully connected layer followed by a sigmoid or linear activation function (if the data has been normalized), outputting continuous values ​​of M joint angles (e.g., 1-2 major joints of each finger). .

[0146] The total loss function is a weighted sum of the classification loss and the regression loss:

[0147] ,in, It is cross-entropy loss. It is the mean squared error loss. It is a hyperparameter that balances the importance of the two tasks.

[0148] Post-training pruning is used to remove connections with weights whose absolute values ​​are close to zero, thus sparsifying the model. Then, the 32-bit floating-point weights and activation values ​​are quantized into 8-bit integers (INT8). Finally, a dedicated compiler for edge AI chips (such as the MAX78000) is used to compile the optimized model graph into code that can run efficiently on hardware accelerators, significantly reducing power consumption and latency.

[0149] The purpose of step S43 is to perform multi-task learning and lightweight model optimization.

[0150] The purpose of step S4 is: This is the core intelligent step in this embodiment, which involves processing the neural feature vectors extracted in the previous step. Mapping to the user's hand movement intentions requires processing both spatial (multi-channel) and temporal (sequence) information simultaneously, while balancing decoding accuracy, latency, and computational overhead.

[0151] Step S5: The abstract motion intent is mapped and generated as a programmable, context-aware control command.

[0152] In this embodiment, step S5 may specifically include the following steps:

[0153] S51, based on the application scenario and user preferences, the abstract motion intent is mapped into discrete trigger commands or continuous proportional control.

[0154] Discrete event mapping (virtual buttons): Assign one or more virtual button events (such as keyboard key codes, game controller buttons) to each decoded gesture category (e.g., "clench fist", "extend index finger", "abduct thumb"). When the probability of this category exceeds a set threshold (e.g., 0.8) and is debounced for a certain period of time (e.g., 50ms), a "button press" event is triggered; when the probability is below the threshold, a "button release" event is triggered. Support for key combination mapping (e.g., mapping "clench fist" + "pronation" to "Ctrl+C").

[0155] Continuous analog mapping (virtual joystick / mouse): This decodes continuous joint angles (such as the angle of the proximal interphalangeal joint of the index finger). It is mapped to an analog control quantity through a configurable transfer function.

[0156] For example: ,in, It is the gain coefficient. It's the angle of the resting position. This is the offset. The output is normalized to a standard analog joystick range (e.g., -32768 to 32767). It can also be used to control the speed of the mouse cursor.

[0157] Hybrid Mode and Mode Switching: Allows users to define multiple mapping modes in the same configuration file and cycle between different modes using a specific "mode switching gesture" (such as quickly bending the pinky twice). For example, Mode A is used for first-person shooter games (the gesture is mapped to move, fire, and reload), and Mode B is used for real-time strategy games (mapped to formation, attack, and move).

[0158] Operating modes can be selected via a physical button or a multi-position slider on the neural decoder. The following are some representative modes:

[0159] Training mode: The wristband device streams raw neural data to the neural decoder and can record this data for offline model training. In this mode, newly trained AI models can also be uploaded to the neural decoder for deployment.

[0160] PC Mode: The neural decoder connects to a PC and simulates standard human-computer interface devices such as a mouse or keyboard. This mode supports plug-and-play functionality, requiring no dedicated drivers. Neural signals are processed by an onboard AI agent and translated into predefined cursor movements or key presses, allowing users to directly control the computer via neural signals.

[0161] Console Mode: The neural decoder connects to a game console (such as Xbox or PlayStation) and emulates a standard game controller. Similar to PC mode, this mode supports plug-and-play and requires no additional drivers. The neural signals are decoded and mapped to controller inputs, allowing users to play games via neural controls.

[0162] User Configuration (PC Mode and Game Console Mode): Multiple user-defined configuration files can be configured for different application scenarios or game types. Each configuration file maps the decoded neural activity to a specific input pattern. For example:

[0163] First-person shooter: continuous cursor control for aiming, and discrete trigger controls for firing.

[0164] Real-time strategy game: a combination of cursor navigation and multi-key command shortcuts.

[0165] Prosthetic Mode: In this mode, the neural decoder interfaces with the robotic prosthetic hand. The decoded neural signals are converted into control commands for each finger and joint, enabling amputees to perform intuitive and proportionally controlled prosthetic operations.

[0166] Other device control modes: Other modes can also be defined for interfacing with various external systems, such as virtual reality platforms, drones, or musical instruments. In each mode, the decoded neural activity is mapped to device-specific control commands.

[0167] The purpose of step S51 is to perform dynamic mapping between discrete and continuous control modes.

[0168] S52 performs two-handed coordination and high-level compound command generation.

[0169] The neural decoding device maintains separate decoding instances and mapping tables for the wireless data streams from the left and right hands, respectively. The decoding results from the left and right hands are logically treated as two independent controller inputs.

[0170] Define compound commands based on the decoding states of both hands. For example, a special macro command (such as "unleash the ultimate skill" in a game) is triggered only when the left hand decodes as "open palm" and the right hand decodes as "fist". This is implemented through a simple logical AND gate.

[0171] Advanced interaction can be achieved by utilizing continuous parameters decoded from both hands. For example, the pronation / supination angle decoded from the left wrist can be used as a "proportional coefficient" to modulate the analog output controlled by the flexion of the right fingers (e.g., when controlling the throttle of an aircraft, the left hand adjusts the sensitivity, while the right hand controls the actual thrust). This is achieved by performing mathematical operations (such as multiplication) on the two continuous variables.

[0172] The purpose of step S52 is to enable richer and more complex interactive logic beyond single-handed control when the user wears the device with both hands.

[0173] S53 automatically switches application configurations based on the running application.

[0174] Develop a companion smartphone / computer application. Users can drag gesture icons onto virtual keyboard / gamepad buttons or draw continuous parameter mapping curves within the application. All configurations are saved as configuration files (JSON format).

[0175] The neural decoding device has built-in flash memory that can store multiple profiles. Users can switch between profiles by sending commands via physical buttons or specific gestures on the device, without needing to reconnect.

[0176] The decoding device runs a lightweight process that monitors the currently active window title or process name on the host. When it detects that the user has switched to "Game A", it automatically loads the default configuration file "Profile_GameA.json" for "Game A"; when switching to drawing software, it automatically loads "Profile_Drawing.json", achieving seamless context-adaptive control.

[0177] The purpose of step S53 is to provide a user-friendly interface that allows users to personalize the control system and to automatically switch configurations based on the running application.

[0178] Step S5 transforms the abstract motion intent (gesture category or joint angle) output by the AI ​​decoder into specific control commands (such as button presses, mouse movements, and joystick offsets) that the game console, computer, or other devices can understand. This step emphasizes flexibility and user customizability.

[0179] Step S6: Based on semi-supervised learning, perform incremental model training and calibration.

[0180] In this embodiment, step S6 may specifically include the following steps:

[0181] S61 performs guided gesture induction and simultaneously collects neural signal data for training.

[0182] The accompanying app guides users through its interface, using animations or highlighted prompts to repeat the target gesture at a specific rhythm (e.g., "prepare-action-hold-relax," each phase lasting 1-2 seconds). The screen simultaneously displays a timer and the icon of the gesture to be performed, ensuring that the action and data are synchronized.

[0183] While recording neural signals, the device's inertial measurement unit records the wrist's motion acceleration and angular velocity, or records hand video via a camera (with user authorization). This multimodal data can serve as supplementary information for subsequent, more precise automatic annotation or data augmentation.

[0184] The system guides users to perform the same gesture under different conditions: varying force (gentle, medium, forceful), forearm posture (pronation, neutral, supination), and arm position (placed on a table, suspended). This increases the diversity of training data and makes the model more robust.

[0185] The purpose of step S61 is to efficiently and accurately collect weakly labeled neural signal data while the user performs actions naturally, for use in training.

[0186] S62 automatically generates interval labels and gesture category labels for the collected neural signal data.

[0187] Calculate the Hilbert envelope of the neural signal for each channel. ,in yes Hilbert transform, It approximates the instantaneous amplitude of the signal. The envelope of the channel (with the largest variance) that performed the most actively during the target gesture training was selected as the primary annotation basis.

[0188] For the selected envelope signal Calculate its median over the entire recording session. and standard deviation Set a high threshold. and low threshold .when When, mark as "determine activity period"; when When the threshold is set, it is marked as the "determined quiescent period"; the region in between is smoothed by considering the context. This dual-threshold method is more noise-resistant than the single-threshold method.

[0189] Since each training session focuses on only one gesture, the active-period binary labels generated for that session are all associated with the category label of that gesture. The inactive-period labels are associated with the "rest" category. This results in weakly labeled data pairs with both temporal and category information. .

[0190] The purpose of step S62 is to automatically generate reliable activity / rest interval labels and gesture category labels for the recorded long-term data stream by utilizing the characteristics of neural signals themselves, in the absence of precise motion capture equipment.

[0191] S63 uses newly collected user data to update or retrain models in the cloud and deploys the optimized models back to user devices.

[0192] An initial general model is maintained in the cloud (pre-trained on a large amount of anonymized user data). When a new user begins training, this general model is downloaded to their mobile app as a starting point.

[0193] On the user's mobile app, the model is incrementally trained (fine-tuned) several times using newly collected local data. Before uploading the model update (weight difference) to the cloud, calibrated Gaussian noise is added to achieve differential privacy protection and prevent the original data from being inferred from the model update.

[0194] The cloud server aggregates encrypted or noisy model updates from multiple users, performs secure averaging, and updates the global model. Simultaneously, it can generate a more personalized model version for each user. After testing and verification, the updated model is pushed via the app and wirelessly updated to the user's neural decoding device, completing one learning cycle.

[0195] The purpose of step S63 is to use the newly collected user data to efficiently update or retrain the model in the cloud while protecting privacy, and to securely deploy the optimized model back to the user's device, enabling cloud-based collaborative incremental learning and model updates.

[0196] The purpose of step S6 is to provide a systematic process for initial training and continuous optimization of the AI ​​decoding model, enabling it to adapt to changes in the user's unique neural signal fingerprint and usage habits, achieving personalization and maintaining long-term performance.

[0197] Figure 2 This is a schematic diagram illustrating an example use case of the AI ​​neural decoding-based hand movement intention recognition and control method of the present invention applied to games. For example... Figure 2 As shown, the system includes: a game console or computer 1, a neural interface device 3, and a neural decoding device 2. The neural interface device 3 has a smartwatch-like form factor and is worn on the user's wrist. Two units can be worn simultaneously, one on each hand, with each device operating independently. The neural interface device 3 records peripheral nerve signals around the wrist and wirelessly transmits the data to the neural decoding device 2. The neural decoding device 2 has a compact USB dongle form factor and plugs into the game console, computer, or other host device via a standard USB port. The neural decoding device 2 is directly powered via the USB connection. For host devices, the neural decoding device 2 is enumerated as a standard human-machine interface device (similar to a game controller, keyboard, or joystick), enabling broad compatibility across many existing platforms. Upon receiving neural data, the neural decoding device performs real-time signal processing and applies artificial intelligence decoding algorithms to infer the user's intended movement or gesture. The decoded intent is then mapped to predefined control commands (e.g., "move left," "move right," "jump") and transmitted to the connected game console or computer. Users can customize command mappings for specific games or applications.

[0198] Figure 3 This is a schematic diagram of the system architecture for implementing the AI ​​neural decoding-based hand movement intention recognition and control method of the present invention. Figure 3 As shown, neural interface device 3 integrates neural recording chips. These chips acquire peripheral nerve signals from the radial nerve 6, median nerve 5, and ulnar nerve 4 using non-invasive surface electrodes placed on the wrist. The device then applies a neural data compression algorithm to reduce the data rate. This enables the streaming of neural signals over short-range wireless links with limited bandwidth, such as Bluetooth Low Energy. The compression algorithm can be implemented as a software module on a microcontroller or in dedicated hardware, such as a custom ASIC. The compressed neural data stream is then wirelessly transmitted in real time to neural decoding device 2.

[0199] Upon receiving the data, neural decoding device 2 executes complementary decompression algorithms to reconstruct the original neural data stream. Then, neural decoding device 2 applies a series of signal conditioning filters, followed by neural feature extraction algorithms, to extract basic information from the recorded neural activity. The generated feature data is passed to an AI inference engine, which uses the neural decoding model to predict the user's motion intentions in real time. The AI ​​inference engine can run on dedicated hardware (such as MAX78000 / MAX78002) or a custom ASIC. The decoded motion intentions can represent discrete gestures (e.g., bending a single finger, clenching a fist, wrist pronation or supination, etc.) or continuous regressions of single finger joint trajectories. The decoded intentions are mapped to predefined output commands and transmitted in real time to a connected arcade machine or computer.

[0200] Furthermore, the neural decoding device 2 can store incoming neural data in onboard flash memory for offline processing. This data is used to train or fine-tune the AI ​​model when the user first sets up the system or when more accurate predictions are needed. During training, the neural decoding device 2 connects to a smartphone application via a wireless link. The stored neural data is transferred to the application, which then uploads the data to a cloud server for AI model training. Once training is complete, the updated AI model is downloaded to the application and transmitted back to the neural decoding device 2 to update the onboard inference engine. The smartphone application also allows users to customize the mapping between decoded motion intentions and output commands. Custom configurations are stored on the neural decoding device 2 for use during real-time operation.

[0201] Figure 4 This is a schematic diagram of the neural interface device worn on the wrist used in the hand movement intention recognition and control method based on AI neural decoding of the present invention. Figure 4 As shown, it has a form factor similar to a smartwatch and is worn around the wrist. The figure also illustrates the approximate placement of the non-invasive surface electrodes 7 used for nerve signal acquisition, which are located above the approximate anatomical locations of the radial nerve 6, median nerve 5, and ulnar nerve 4.

[0202] Figure 5 This is a detailed schematic diagram of the neural interface device used in the AI ​​neural decoding-based hand movement intention recognition and control method of the present invention. Figure 5As shown, the neural interface device 3 comprises a device body 9 and a battery. It includes basic user interface elements such as buttons, LED indicators, a buzzer, and a display screen. A ground / reference electrode 12 is integrated on the underside of the body. This electrode is structurally similar to the recording electrode but has a larger surface area. The device also includes three sets of recording electrodes mounted on a wristband 13: a radial nerve electrode 11, a median nerve electrode 10, and an ulnar nerve electrode 8. The positions of these electrode sets are adjustable to ensure proper alignment with the radial nerve 6, median nerve 5, and ulnar nerve 4, respectively, during use.

[0203] Figure 6 This is a schematic diagram of the neural recording configuration of each electrode group used in the AI ​​neural decoding-based hand movement intention recognition and control method of the present invention. Figure 6 As shown, each group consists of three electrodes 14 arranged longitudinally along the peripheral nerve. The electrodes 14 are surrounded by peripheral nerves 15. This configuration allows the device to capture not only the amplitude of the nerve signal but also the direction of signal propagation (efferent or afferent). Each electrode group is connected to two differential amplifiers. The central electrode serves as the shared negative input to both amplifiers, while the distal and proximal electrodes serve as the positive inputs to their respective amplifiers. The outputs of the two amplifiers provide measurements of the nerve signal amplitude, and the time offset between these outputs may indicate the directionality of nerve signal propagation.

[0204] Figure 7 This is a schematic diagram illustrating the mapping between the decoded motion intent and the output command used in the AI ​​neural decoding-based hand motion intent recognition and control method of the present invention. Figure 7 The diagram illustrates an example of how decoded hand gesture intentions are translated into control commands for a game console or computer. In discrete mode, the AI ​​model categorizes specific gestures. Each gesture is mapped to a predefined button or keystroke. In this example, a thumb bend gesture might be mapped to a "move left" command, an index finger bend gesture might be mapped to "move right," and a fist clench gesture might trigger an "up / jump" action. In continuous mode, the AI ​​model performs regression to estimate the joint angles of individual fingers. These continuous joint angle values ​​can be mapped to analog joystick movements or proportional control inputs. The system supports independent mapping for each hand, allowing users to control individual functions using the neural interface devices 3 of both their left and right hands. Users can also define combined dual-hand mappings for more complex multi-button or multi-axis control. All mappings are fully customizable via a companion smartphone application. Users can create multiple configuration profiles tailored for different games or applications. Multiple profiles can be stored on the neural decoding device 2 and selected via onboard buttons.

[0205] Figure 8This is a schematic diagram of the deep learning-based AI model used for neural decoding in the hand movement intention recognition and control method based on AI neural decoding of the present invention. Figure 8 Example AI models that can be used to implement the neural decoder in the proposed system are illustrated. The neural decoder can employ any suitable deep learning architecture capable of mapping neural input signals to user motion intentions. Such architectures can include, but are not limited to, convolutional neural networks, residual networks, recurrent neural networks based on long short-term memory or gated recurrent units, Transformer-based sequence models, or hybrid models thereof. Typically, the deep learning model may contain multiple layers for feature extraction and temporal modeling. These layers may include convolutional layers, fully connected layers, recurrent layers (such as LSTM or GRU), attention layers, pooling layers, or other neural network components. The architecture may also include support layers, such as linear rectified units, sigmoid, tanh, softmax, batch normalization, dropout, and residual connections, to improve the model's stability, generalization ability, and prediction accuracy. The number of layers, the dimension per layer, and the order of the layers can be selected or optimized to strike a balance between prediction performance, computational efficiency, and robustness against overfitting to the available dataset.

[0206] also, Figure 8 Examples of architectures designed for classification or regression are provided, depending on the desired control mode. In the classification model, the network outputs probabilities over a discrete set of gestures; this corresponds to the discrete mapping control mode described above. In the regression model, the network outputs one or more continuous variables, such as predicted joint angles or other kinematic parameters, supporting either continuous or proportional control modes. Furthermore, Figure 8 Two types of examples are provided: non-temporally dependent models (such as CNN-based models) and temporally dependent models (such as RNN-based models). CNN models process neural data within a fixed window and do not have explicit temporal recursion. RNN models contain temporal components (e.g., LSTM or GRU layers) that explicitly capture the temporal dynamics of neural activity.

[0207] Figure 9 This is a flowchart illustrating the process of collecting labeled training data for AI model training in the hand motion intention recognition and control method based on AI neural decoding of the present invention. This process is performed when the user first sets up the system or when periodic fine-tuning is needed to improve decoding accuracy. The user first selects an active gesture to be trained (e.g., bending a single finger, clenching a fist, pointing, pinching, etc.) from a predefined set of discrete gestures. Prompted, the user repeatedly performs the selected gesture, alternating between a static state and active gestures. The gesture can be performed at different speeds, force levels, and hand or arm positions to capture diverse neural patterns. Neural signals are recorded throughout the process.

[0208] After recording, feature extraction is performed to generate neural feature data that will serve as input to the AI ​​model. Simultaneously, the signal power over time is calculated using methods such as the Hilbert envelope. Since each recording session contains only two states—rest and a selected active gesture—these states can be reliably distinguished using signal power thresholding. Active gestures consistently exhibit significantly higher power compared to the resting state. To improve robustness, thresholding can be performed using the channel or channel exhibiting the highest signal power, as neural activity may be stronger on different neurons depending on the gesture. The generated binary segmentation results are then combined with user-selected gesture labels to generate ground truth labels corresponding to the neural feature data. This process is repeated for each gesture the user wants the AI ​​model to learn. By aggregating all extracted feature data and their corresponding labels across all gestures, a complete training dataset for training the AI ​​model is constructed.

[0209] Figure 10 This is a schematic diagram of the data collected during the training of the AI ​​model used in the AI-based neural decoding-based hand movement intention recognition and control method of the present invention, showing an example of neural data obtained during the aforementioned training data collection process. The raw neural signals clearly show the state cycles of multiple alternating resting postures 17 and active gestures 16. The signal power is then calculated using the Hilbert envelope to emphasize the changes in neural activity between resting and active periods. By selecting an intermediate value to threshold the obtained power signal, a binary segment is generated to distinguish between resting intervals and active gesture intervals that vary over time. This binary segment is then combined with a user-selected gesture label to generate corresponding ground truth labels for AI model training or fine-tuning.

[0210] The beneficial effects of implementing this embodiment are:

[0211] (1) By collecting multimodal signals through the peripheral nerve of the wrist, the high invasive risk of traditional brain-computer interfaces is avoided, and rich neural information can be obtained. By combining feature-preserving compression technology and hybrid deep learning architecture, stable features can be extracted from high-dimensional data, and real-time and accurate decoding of fine movements such as fingers and wrists can be achieved. The control latency is low and the reliability is strong, realizing high-precision and low-invasive motion intention decoding.

[0212] (2) It adopts wireless transmission and lightweight processing, making it easy to wear and use on the go; through programmable, context-aware command mapping, it can generate corresponding instructions according to different tasks (such as gripping, typing, and operating the device), which significantly enhances the naturalness and flexibility of human-computer interaction and has good practicality and adaptability.

[0213] (3) Relying on semi-supervised learning for incremental training and calibration, it can continuously adapt to the changes in the user's neural signal characteristics and habits during use, gradually improve decoding accuracy, reduce dependence on a large amount of labeled data, and is more suitable for long-term practical applications, supporting continuous optimization and personalized adaptation; it has achieved a balance in three aspects: technical feasibility, user experience and system sustainability, and provides efficient solutions for fields such as intelligent prostheses, rehabilitation training and natural human-computer interaction.

[0214] This invention can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This invention can 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 modules can reside in local and remote computer storage media, including storage devices.

[0215] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing related hardware through computer-readable instructions. These computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, optical disk, or read-only memory (ROM), or random access memory (RAM).

[0216] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0217] Example 2

[0218] Further reference Figure 11 As a response to the above Figure 1 The present invention provides an embodiment of a hand movement intention recognition and control device based on AI neural decoding, which implements the method shown. Figure 1 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0219] like Figure 11 As shown, the hand movement intention recognition and control device 70 based on AI neural decoding described in this embodiment includes: an acquisition module 71, a compression module 72, an extraction module 73, a decoding module 74, a mapping module 75, and a training module 76. Wherein:

[0220] Acquisition module 71 is used to acquire multimodal signals of peripheral nerves of the wrist and obtain raw neural data streams;

[0221] Compression module 72 is used to perform feature-preserving neural data compression and wireless transmission based on the original neural data stream;

[0222] Extraction module 73 is used to preprocess the received and decompressed raw neural data stream and extract neural features;

[0223] The decoding module 74 is used to perform real-time motion intent decoding based on the extracted neural features and a hybrid deep learning architecture to obtain abstract motion intent;

[0224] Mapping module 75 is used to map and generate programmable, context-aware control commands for the abstract motion intent;

[0225] Training module 76 is used for incremental model training and calibration based on semi-supervised learning.

[0226] The beneficial effects of implementing this embodiment are: it achieves high-precision, low-intrusion motion intent decoding; it has good practicality and adaptability; it supports continuous optimization and personalized adaptation; and it achieves a balance in terms of technical feasibility, user experience, and system sustainability, providing an efficient solution for fields such as intelligent prostheses, rehabilitation training, and natural human-computer interaction.

[0227] Example 3

[0228] To address the aforementioned technical problems, embodiments of the present invention also provide a computer device. Please refer to [link / reference needed]. Figure 12 , Figure 12 This is a basic structural block diagram of the computer device in this embodiment.

[0229] The aforementioned computer device 8 includes a memory 81, a processor 82, and a network interface 83 that are interconnected via a system bus. It should be noted that only the computer device 8 with components 81, 82, and 83 is shown in the figure; however, it should be understood that it is not required to implement all the shown components, and more or fewer components can be implemented alternatively. Those skilled in the art will understand that the computer device described herein is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0230] The aforementioned computer devices can be desktop computers, laptops, handheld computers, and cloud servers, among other computing devices. These devices can facilitate human-computer interaction with users through keyboards, mice, remote controls, touchpads, or voice-activated devices.

[0231] The aforementioned memory 81 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the aforementioned memory 81 may be an internal storage unit of the aforementioned computer device 8, such as the hard disk or memory of the computer device 8. In other embodiments, the aforementioned memory 81 may also be an external storage device of the aforementioned computer device 8, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 8. Of course, the aforementioned memory 81 may also include both the internal storage unit and its external storage device of the aforementioned computer device 8. In this embodiment, the aforementioned memory 81 is typically used to store the operating system and various application software installed on the aforementioned computer device 8, such as computer-readable instructions for hand movement intention recognition and control methods based on AI neural decoding. In addition, the aforementioned memory 81 can also be used to temporarily store various types of data that have been output or will be output.

[0232] In some embodiments, the processor 82 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip. The processor 82 is typically used to control the overall operation of the computer device 8. In this embodiment, the processor 82 is used to execute computer-readable instructions stored in the memory 81 or to process data, such as executing the computer-readable instructions of the AI ​​neural decoding-based hand movement intention recognition and control method.

[0233] The network interface 83 may include a wireless network interface or a wired network interface, which is typically used to establish a communication connection between the computer device 8 and other electronic devices.

[0234] The beneficial effects of implementing this embodiment are: it achieves high-precision, low-intrusion motion intent decoding; it has good practicality and adaptability; it supports continuous optimization and personalized adaptation; and it achieves a balance in terms of technical feasibility, user experience, and system sustainability, providing an efficient solution for fields such as intelligent prostheses, rehabilitation training, and natural human-computer interaction.

[0235] Example 4

[0236] The present invention also provides another embodiment, namely, providing a computer-readable storage medium storing computer-readable instructions that can be executed by at least one processor to cause the at least one processor to perform the steps of the AI ​​neural decoding-based hand movement intention recognition and control method described above.

[0237] The beneficial effects of implementing this embodiment are: it achieves high-precision, low-intrusion motion intent decoding; it has good practicality and adaptability; it supports continuous optimization and personalized adaptation; and it achieves a balance in terms of technical feasibility, user experience, and system sustainability, providing an efficient solution for fields such as intelligent prostheses, rehabilitation training, and natural human-computer interaction.

[0238] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0239] Obviously, the embodiments described above are merely some embodiments of the present invention, not all embodiments. The accompanying drawings show preferred embodiments of the present invention, but do not limit the patent scope of the present invention. The present invention can be implemented in many different forms; rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the patent protection scope of this invention.

Claims

1. A method for hand movement intention recognition and control based on AI neural decoding, characterized in that, Includes the following steps: Multimodal signals from peripheral nerves in the wrist were acquired to obtain raw neural data streams; Based on the original neural data stream, neural data compression and wireless transmission based on feature preservation are performed; The received and decompressed raw neural data stream is preprocessed and neural features are extracted. Based on the extracted neural features, real-time motion intent decoding is performed using a hybrid deep learning architecture to obtain abstract motion intent; The abstract motion intent is programmable and context-aware control command mapping and generation is performed. Incremental model training and calibration are performed based on semi-supervised learning.

2. The hand movement intention recognition and control method based on AI neural decoding according to claim 1, characterized in that, The steps of acquiring multimodal signals from the peripheral nerves of the wrist and obtaining the raw neural data stream specifically include: The recording electrodes were positioned in the surface projection areas of the radial, median, and ulnar nerves, and a multi-electrode array was configured to record the amplitude of multimodal signals and sense the propagation direction of nerve impulses. At the forefront of signal acquisition, the microvolt-level neural signal is amplified, environmental noise, electromyographic interference, and motion artifacts are suppressed to obtain the original neural data stream; Hardware-level suppression of common-mode interference and motion artifacts.

3. The hand movement intention recognition and control method based on AI neural decoding according to claim 1, characterized in that, The step of performing feature-preserving neural data compression and wireless transmission based on the original neural data stream specifically includes: Based on the original neural data stream, neural data compression is performed using sparse representation and perceptual coding. The wireless transmission parameters are dynamically adjusted based on signal characteristics, decoding performance requirements, and device battery status. The original neural data stream is wirelessly transmitted based on forward error correction and packet loss concealment.

4. The hand movement intention recognition and control method based on AI neural decoding according to claim 1, characterized in that, The steps of preprocessing the received and decompressed raw neural data stream and extracting neural features specifically include: The received and decompressed raw neural data stream is subjected to adaptive filtering and power frequency interference elimination. The signal-to-noise ratio (SNR) of each channel is dynamically evaluated, and a subset of channels whose SNR exceeds a preset optimal SNR are selected as channels for neural feature extraction. From the selected subset of channels, a set of features that characterize the changes in neural activity patterns in different dimensions are extracted from the original neural data stream after adaptive filtering and power frequency interference elimination, forming a feature vector.

5. The hand movement intention recognition and control method based on AI neural decoding according to claim 1, characterized in that, The step of decoding the abstract motion intent in real time based on the extracted neural features using a hybrid deep learning architecture specifically includes: Construct a convolutional neural network branch for spatial feature extraction; Construct a recurrent neural network branch for time-dynamic modeling; Multi-task learning and lightweight model optimization are performed to enable a single model to simultaneously output discrete gesture classification and continuous joint angle regression, and optimization is performed for the limited computing resources of embedded decoding devices.

6. The hand movement intention recognition and control method based on AI neural decoding according to claim 1, characterized in that, The step of mapping and generating programmable, context-aware control commands for the abstract motion intent specifically includes: Based on the application scenario and user preferences, the abstract motion intent is mapped into discrete trigger commands or continuous proportional control; Perform two-handed coordination and generate advanced compound commands; Automatically switch application configurations based on the running application.

7. The hand movement intention recognition and control method based on AI neural decoding according to claim 1, characterized in that, The steps for incremental model training and calibration based on semi-supervised learning specifically include: Guided gestures are used to induce neural signal data, which is collected simultaneously for training purposes. Automatically generate interval labels and gesture category labels for the collected neural signal data; Using newly collected user data, update or retrain the model in the cloud and deploy the optimized model back to the user's device.

8. A hand movement intention recognition and control device based on AI neural decoding, characterized in that, include: The acquisition module is used to collect multimodal signals from the peripheral nerves of the wrist and acquire the raw neural data stream. A compression module is used to perform feature-preserving neural data compression and wireless transmission based on the original neural data stream; The extraction module is used to preprocess the received and decompressed raw neural data stream and extract neural features. The decoding module is used to perform real-time motion intent decoding based on the extracted neural features and a hybrid deep learning architecture to obtain abstract motion intent. The mapping module is used to map and generate programmable, context-aware control commands for the abstract motion intent; The training module is used for incremental model training and calibration based on semi-supervised learning.

9. A computer device, characterized in that, The method includes a memory and a processor, wherein the memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the steps of the hand movement intention recognition and control method based on AI neural decoding as described in claim 1.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the hand movement intention recognition and control method based on AI neural decoding as described in claim 1.

Images