CN-121996062-A - Gesture recognition method based on friction nano generator array

CN121996062ACN 121996062 ACN121996062 ACN 121996062ACN-121996062-A

Abstract

The invention discloses a gesture recognition method based on a friction nano generator array, which comprises the following steps of S1 designing a drum-shaped single-electrode friction nano generator DS-TENG into a 2X 2 array, fixing the array on a wrist, synchronously collecting four-channel time sequence voltage signals, S2 carrying out sliding window sampling of 50% overlapping area data enhancement on one-dimensional time sequence voltage signals, mapping the sliding window sampling into a 32X 32 multichannel structured gray matrix according to row priority, carrying out convolution and pooling operation on the gray matrix in sequence, inputting the obtained one-dimensional feature vector into a random forest classifier, and outputting gesture types, wherein super parameters of the classifier are determined through Bayesian optimization. The scheme of the invention realizes a classification and inference integrated process of self-powered signal acquisition, sliding window mapping and convolution-pooling feature extraction and Bayesian optimization, has the technical effects of low power consumption, low computational complexity and overfitting inhibition, and is suitable for a wearable gesture recognition terminal.

Inventors

YANG FENG
MA JIAQING
Huo Yiwen
Gao Chushan
SONG ZIWEN
WANG XINYUAN
ZHANG YIFAN
ZHU SHU
MIN FUHONG

Assignees

南京师范大学

Dates

Publication Date: 20260508
Application Date: 20251022

Claims (7)

1. The gesture recognition method based on the friction nano generator array is characterized by comprising the following steps of: S1, forming a2 multiplied by 2 array by a drum-shaped single-electrode friction nano generator DS-TENG, fixing the array at a wrist, circumferentially arranging four positions covering the wrist, and synchronously sampling superficial tendon vibration signals containing gesture information at the four positions by utilizing pressure sensing characteristics of the array to obtain four-channel time sequence voltage signals; S2, sampling the four-channel time sequence voltage signal by adopting a time sequence convolution characteristic compression TCF method, mapping each section of sample into a 32×32 structured gray matrix according to row priority, sequentially performing convolution and pooling operation on the gray matrix to obtain a one-dimensional characteristic vector, and inputting the one-dimensional characteristic vector into a classifier to output gesture types.
2. The gesture recognition method based on a friction nano generator array according to claim 1, wherein in step S1, the wearable pressure sensor comprises polytetrafluoroethylene sensitive layer PTFE, rectangular ring polyimide insulating layer PI, copper foil layer Cu, and polyvinyl chloride substrate PVC.
3. The method for gesture recognition based on a friction nano-generator array according to claim 1, wherein the four-channel time-series voltage signal obtained in the step S1 comprises a pulse vibration component for reflecting wrist blood flow pulsation and a muscle vibration component related to muscle group position, wherein the muscle vibration component is used as a main information source of gesture classification.
4. The gesture recognition method based on the friction nano generator array according to claim 1, wherein in order to accurately recognize microampere-level current and normal voltage from DS-TENG, a signal conditioning and sampling module is adopted in step S1, and the gesture recognition method based on the friction nano generator array comprises a low-bias operational amplifier with input bias current not more than 1 pA and an analog-to-digital converter with resolution of 24 bits, and further comprises a low-power wireless communication unit for data transmission, and is suitable for wearable and low-power application scenes.
5. The gesture recognition method based on a friction nano-generator array according to claim 1, wherein step S2 comprises: s21, preprocessing data, namely sampling by adopting a sliding window overlapped by 50%, and carrying out low-pass filtering on each channel signal, wherein the cut-off frequency of the low-pass filtering is 8 Hz; S22, sequential convolution feature compression, namely mapping each section of four-channel signal into a 32×32 structured gray matrix according to row priority, and sequentially executing two-stage convolution and pooling operation to obtain a one-dimensional feature vector; s23, classifying and identifying, namely inputting the feature vectors into a random forest classifier, determining the super parameters of the tree number, the maximum depth and the minimum leaf node sample number of the classifier through Bayesian optimization, and improving the accuracy and the robustness.
6. The method for gesture recognition based on a friction nano-generator array according to claim 5, wherein the structured gray scale matrix of step S2 comprises window-length-wise applying the four-channel time domain voltage signal, respectively Step size Segmenting, record the first Sample vector for each sliding window is Through 50% overlapping sampling, ensure that the middle sample point will be collected twice; ; Wherein, the Represents the first time sequence signal A number of sampling points are used to sample the sample, For the index of the starting point of the sample, The value of (2) is in the range of 1 to ; Will be As the first Line sample data is mapped to a two-dimensional space in a line-first manner to obtain × Is a matrix of samples of (a) The mapping process is shown in formula (2); ; Then, normalize each sample point from the voltage value sample space Mapping to gray pixel value sample space Thus obtaining × The specific process is shown in formula (3); ; Wherein, the Represent the first Gray pixel values mapped by the sample points, And (3) with The upper limit and the lower limit of the measuring range of the sampling circuit are respectively, Representing a rounding operation.
7. The method of claim 6, wherein to extract vibration features related to gestures from the gray matrix and complete classification, a combination of TCF and bayesian-optimized random forest classification BO-RF is used to compress TCF-BO-RF, comprising: For each section of four channels × The gray matrix is respectively input into a convolution network and respective space-time characteristics are extracted, and the maximum pooling operation is introduced after each layer of convolution operation to reduce the dimension of the characteristic matrix and improve the robustness of the system, wherein the pooling window has the size of × The step length is 2, and the pooling process is shown in formula (5); ; ; Wherein, the Is the first A convolution kernel of the layer of size , A characteristic diagram after the convolution is shown, And Is the position index of the feature map; convolving and pooling the four-way eigenvalue Flattening to obtain four channels 1× Vector quantity , wherein, The characteristic dimension after flattening; After four-way feature vectors are obtained, the four-way feature vectors are input into a random forest RF classifier which comprises Decision trees, each tree Training based on different feature subsets and data subsets, and for each tree Its classification rules are represented by the following recursive decision functions: ; Wherein is One region of the decision tree is identified, Is a classification label, and the final result is Majority vote results for all trees: ; In order to avoid the problem of overfitting and sinking into a local optimal solution caused by too high sample dimension, a Bayesian optimization method is adopted to adjust and optimize the RF super-parameters; By constructing a gaussian process model Evaluating results by means of previous hyper-parameters To predict the next most likely hyper-parametric combination that improves model performance: ; ; Wherein, the Is a predictive function based on a Gaussian process, and is expressed in a given evaluation result Under the condition of (a) super parameter Expected contribution to model performance; The super parameters at least comprise the number of trees And the Bayesian optimization selects the next set of candidate super parameters based on the Gaussian process agent model and the acquisition function until the stopping condition is met.

Description

Gesture recognition method based on friction nano generator array Technical Field The invention belongs to the field of man-machine interaction, relates to a gesture recognition method based on a friction nano generator array, and particularly relates to a gesture recognition method based on a drum-shaped friction nano generator (DS-TENG) array, time sequence convolution feature compression (TCF) and Bayesian optimization random forest (BO-RF) classification model. Background With the rapid development of human-machine interaction (HMI) technology, intelligent control systems based on gesture recognition have wide application prospects in the fields of Augmented Reality (AR), artificial Intelligence (AI), internet of things (IoT), medical rehabilitation and the like. The existing gesture recognition is dependent on active detection devices such as piezoelectric sensors and infrared sensors, and the problems of high power consumption, high cost, poor portability and the like exist in the scheme, so that the gesture recognition method is not suitable for low-power consumption wearable equipment and daily life of a wearer. The friction nano generator (TENG) is used as an emerging self-powered sensing technology, has the advantages of low power consumption, high sensitivity, easiness in integration and the like, and has been widely studied in the field of flexible sensing. However, a single TENG sensor has the problem of insufficient capability in sensing complex signals, and meanwhile, along with the improvement of waveform data complexity, the traditional neural network model is easy to have bottlenecks of fitting, large calculation amount and the like, so that a new signal processing and recognition method is needed to improve recognition precision and robustness. Disclosure of Invention The invention aims to solve the problem of providing a gesture recognition method based on a friction nano generator array, in particular to a gesture recognition method based on a drum-shaped friction nano generator (DS-TENG) array, time sequence convolution feature compression (TCF) and Bayesian optimization random forest (BO-RF) classification model. The specific scheme is as follows: a gesture recognition method based on a friction nano generator array comprises the following steps: S1, aiming at the extraction scheme of superficial tendon vibration signals, a novel sensing array structure is adopted, a drum-shaped single-electrode friction nano generator DS-TENG is designed into a 2X 2 array, so that a wearable pressure sensor is formed and fixed at a wrist, and the high-sensitivity and self-powered characteristics of a nano friction motor are utilized. The four-channel time sequence voltage signal acquisition device can synchronously acquire superficial tendon vibration signals of four positions of the wrist part containing gesture information, so that four-channel time sequence voltage signals are obtained, the spatial resolution and sensitivity of signal acquisition are effectively improved, the detection effect is improved, and the use cost is reduced; S2, in order to extract harmonic component information from muscle group vibration contained in the time sequence voltage signals, a time sequence convolution feature compression TCF method is adopted, the four-channel time sequence voltage signals are sampled in a sliding window overlapped by 50%, each section of sample is mapped into a 32×32 structured gray matrix according to row priority, the gray matrix can effectively retain time sequence features and space correlation of the time sequence voltage signals, convolution and pooling operation are sequentially carried out on the gray matrix to obtain one-dimensional feature vectors, data dimension and redundancy are reduced, and the feature vectors are input into a classifier to output gesture types. Further, in step S1, the wearable pressure sensor prepared by using the high-sensitivity pressure sensor is circumferentially distributed around the wrist, so that vibration waveform signals corresponding to different gestures can be better monitored. The waveform characteristics molded by the harmonic components are researched and classified by utilizing an algorithm model, so that gesture gestures corresponding to different waveforms can be well confirmed; Further, in step S1, the wearable pressure sensor includes a polytetrafluoroethylene layer (PTFE), a rectangular ring polyimide insulating layer (PI), a copper foil layer (Cu), and a polyvinyl chloride substrate (PVC). When the pressure sensor is worn by volunteers, the copper foil layer and the PTFE layer are contacted and rubbed with each other under the action of pressure, and the surface of the copper foil layer and the PTFE layer can generate equal friction charges with opposite polarities due to the electrophilic property of the PTFE material. Most of the time, the PTFE material separates from the copper foil layer and maintains a certain negative charge due t