CN-121979384-A - Non-contact medical picture operation terminal gesture decoding algorithm based on electromyographic signal recognition

Abstract

The application relates to a gesture decoding algorithm of a non-contact medical painting operation terminal based on electromyographic signal identification. The system comprises a multichannel electromyographic signal acquisition module, an anti-interference preprocessing module, a multi-dimensional feature fusion extraction module, a CNN-LSTM hybrid network decoding module and a scene self-adaptive optimization module, wherein a multichannel flexible electromyographic sensor array is used for acquiring electromyographic signals from key muscle groups, high-quality signal preprocessing is realized, an initial feature set is constructed through the cooperation features of time domain, frequency domain, time frequency and muscle groups, a genetic algorithm and LASSO regression are adopted to jointly optimize and screen core features, a time-frequency spectrogram extracts spatial features through a CNN network, LSTM network extracts time sequence features, accurate decoding of degree-of-freedom gesture actions is realized, and a dynamic optimization is realized by combining a signal quality monitoring. The application improves the myoelectric recognition precision and robustness under complex environment.

Inventors

• HUANG XUELAN
• HUANG JINCHENG
• FANG YUEYUE
• WANG NAN
• ZHOU JUNRUI

Assignees

• 绍兴职业技术学院

Dates

Publication Date

20260505

Application Date

20251225

Claims (10)

1. 1. A gesture decoding algorithm of a non-contact medical picture operation terminal based on electromyographic signal recognition is characterized by comprising the following steps, Collecting multichannel electromyographic signals, namely collecting the electromyographic signals through an 8-channel flexible electromyographic sensor array attached to the surface of muscle groups such as biceps brachii, flexor carpi radialis, supinator and the like, and dynamically adjusting the sampling frequency and the gain coefficient of multiplied by 1000 to multiplied by 5000 within the range of 500-2000 Hz according to scene requirements by an FPGA control module after the electromyographic signals are digitized by an ADS1298 bioelectric chip; Performing anti-interference preprocessing, namely performing band-pass filtering (20-500 Hz), 50Hz notch filtering, self-adaptive sliding window smoothing processing and Z-score normalization on the digital electromyographic signals, and comprehensively detecting start and stop points of an action segment based on an energy threshold, a length constraint, a waveform characteristic and a slope threshold to obtain an effective action signal segment; Extracting time domain features, frequency domain features, time-frequency features and muscle group cooperative features from the effective action signal segment, and obtaining 8 core features including IEMG, RMS, MPF, MF, WPE 3 、CR、IPD 1 、IPD 2 from 32 initial features by a combined screening mode of genetic algorithm and LASSO regression; The fourth step of CNN-LSTM mixed network decoding, which is to construct the 8-channel core feature into a time-frequency spectrogram with the size of 128 multiplied by 8, extract the space feature through a CNN network comprising three convolution layers and two pooling layers, extract the time sequence feature through two bidirectional LSTM networks, and finally output a 6-degree-of-freedom control instruction corresponding to the medical picture terminal; And fifthly, scene self-adaptive optimization, namely updating model parameters according to new characteristic data in the use process of a user by adopting an online incremental learning mode, respectively adjusting dynamic response speed, a signal-to-noise ratio threshold value and control sensitivity in a sterile operating room mode or a postoperative rehabilitation mode, and triggering control suspension and prompt when the myooxygen saturation is lower than a preset threshold value or the signal quality does not reach the standard.
2. 2. The algorithm of claim 1, wherein the slope threshold is set to be greater than 0.5 mV/s for rising or falling signal slope in the start-stop point detection of the action segment.
3. 3. The algorithm of claim 1, wherein the sliding window smoothing process uses a moving average method with a window size of 0.2 s.
4. 4. The algorithm of claim 1, wherein the genetic algorithm has a population size of 50, a number of iterations of 30, a crossover probability of 0.6, and a mutation probability of 0.05.
5. 5. The algorithm of claim 1, wherein the time dimension and the frequency dimension of constructing the time-frequency spectrogram each employ 128-point resolution.
6. 6. The algorithm of claim 1, wherein the convolution kernel size of the CNN is 3×3, the step size is 1, the padding is 1, and the pooling layer uses a2×2 pooling kernel.
7. 7. The algorithm of claim 1, wherein the LSTM network is a bi-directional structure and comprises 128 hidden units per layer.
8. 8. The algorithm of claim 1, wherein the online incremental learning performs a model parameter update after every 10 active operations, the update employing 1/10 of a pre-trained learning rate.
9. 9. The algorithm of claim 1, wherein in the sterile operating room mode, the gain factor is automatically increased to enhance signal quality when the signal-to-noise ratio is below 30 dB.
10. 10. The algorithm of claim 1, wherein the sensitivity adjustment is performed according to MRC muscle strength level in a postoperative rehabilitation mode, wherein MRC 1-2 level is set to be x 2.0, MRC 3 level is set to be x 1.5, and MRC 4 level is set to be x 1.0; When the signal-to-noise ratio of the continuous 5s is lower than 25dB or the failure rate of the action segment recognition exceeds 30%, the system triggers a hardware self-check and prompts a user to reattach the sensor or adjust the position.

Description

Non-contact medical picture operation terminal gesture decoding algorithm based on electromyographic signal recognition Technical Field The application relates to the field of bioelectric signal processing and intelligent control, in particular to a gesture decoding algorithm of a non-contact medical picture operation terminal based on electromyographic signal recognition. Background In recent years, with the development of intelligent medical equipment and a contactless operation terminal, gesture recognition and motion decoding technology based on surface electromyographic signals (sEMG) is gradually applied to scenes such as clinical operation control, robot-assisted surgery and nerve rehabilitation training. The electromyographic signals have the characteristics of non-invasiveness, quick response, easy acquisition and the like, can send operation instructions to the medical terminal under the condition of not directly touching equipment, are beneficial to reducing the risk of cross contamination in the operation process, and improve the continuous operation efficiency. However, the existing electromyographic signal decoding technology still has various defects in a complex clinical environment, and cannot completely meet the requirement of high-precision medical control. Current state of the art in industry The current contactless myoelectric control technology mainly comprises key components such as a myoelectric sensor array, a signal preprocessing algorithm, a feature extraction model, a gesture decoding network and the like. The sensor and hardware integration technology is that a multichannel surface myoelectric sensor array with a flexible substrate is commonly adopted in the prior art, PDMS materials are used as supporting structures, electrodes are constructed through conductive silver adhesives, and the attached acquisition of key muscle groups such as biceps brachii, flexor carpi radialis, supinator and supinator is realized. Typical products are configured as an 8-channel array, typically 80mm by 40mm by 0.3mm in size, capable of covering six muscle groups that are primarily involved in hand fine motion and forearm rotation. The hardware acquisition end is generally integrated with an ADS1298 bioelectric acquisition chip, and the chip can provide a sampling frequency range of 1000-2000Hz and support the gain adjustment of multiplied by 1000 multiplied by 5000 so as to adapt to myoelectric signals with different intensities. Under the control of the FPGA, the sampling rate can be switched according to the specific operation action requirement, and meanwhile, a magnetic shielding structure (magnetic permeability mu > 5000) is added and is inhibited against 50Hz power frequency interference, so that the signal quality and the stability of the system are improved. The current state of the art of motion recognition and scene adaptation is that in a sterile operating room scene, the prior art generally determines the effective motion segment in the electromyographic signal based on an improved motion segment detection algorithm by combining an energy threshold, a length constraint and a waveform characteristic. For example, by setting the resting energy threshold to 3 times the resting time average, the duration of the rotational motion is greater than or equal to 0.8s, the scaling motion is greater than or equal to 0.5s, and determining that the motion occurs in combination with zero crossing rate ZCR > 0.8. Rotation control is based on cooperative activation of supinator and supinator, and when the MPF (average power frequency) difference exceeds a certain threshold (such as 30 Hz), a rotation instruction can be triggered, and scaling action is triggered by the integral myoelectricity value change rate (> 50%/s) between biceps brachii and triceps brachii. Part of the system supports multi-mode action combination, such as fist making and rotation, so that complex control actions like 360-degree visual angle operation and the like are realized, the switching times of instruments are reduced, and the operation efficiency is improved. In a postoperative rehabilitation scenario, existing systems typically employ antagonistic muscle ratio (CR) to monitor muscle coordination, and when CR >1.8, prompt abnormal compensation of muscle groups to help patients adjust training intensity. Meanwhile, some systems introduce near infrared spectroscopy (NIRS) technology to monitor myooxygen saturation (StO 2), and automatically terminate training when StO 2 is below 60% to avoid excessive muscle fatigue. In addition, based on the patient muscle strength score (e.g., MRC grade), the control system adjusts the sensitivity ratio to achieve adaptive control for different rehabilitation stages, e.g., setting the sensitivity at MRC grade 3 to be 1.5 times that of the conventional one to compensate for the low amplitude signal generated by weak muscles. The defects of the prior art are that although the prior art has a