CN-117064410-B - Military physical training risk assessment method and system based on deep learning

CN117064410BCN 117064410 BCN117064410 BCN 117064410BCN-117064410-B

Abstract

The invention provides a military physical training risk assessment method and system based on deep learning, which belong to the field of military physical training and comprise the steps of obtaining surface electromyographic signals of a training person, preprocessing, carrying out time domain, frequency domain and time domain feature extraction and CNN feature extraction on the preprocessed surface electromyographic signals, carrying out feature screening combination to obtain a feature sequence after feature screening combination, inputting a muscle fatigue assessment model based on the feature sequence to obtain the muscle fatigue of the current training person at the current moment, obtaining and processing physical sign data, carrying out normalization processing on the muscle fatigue and the physical sign data with abnormal values removed, and inputting a risk assessment model to carry out risk assessment to obtain the training risk assessment result of the current training person. The problem that training risk assessment of military physical training parametrics cannot be accurately assessed is solved, guidance and decision support are provided for the military physical training, and training risk assessment accuracy and operability are improved.

Inventors

LI HAO
ZHANG DONGXU
ZHANG CHI
ZHANG TONG
Xue Zhuxin
YAO SHUAI
BAI YANG
LIU XIONGJUN
LI MIAO
LIU LIYUAN

Assignees

北京京航计算通讯研究所

Dates

Publication Date: 20260505
Application Date: 20230828

Claims (7)

1. The military physical training risk assessment method based on deep learning is characterized by comprising the following steps of: Acquiring surface electromyographic signals of current parametrics and preprocessing, respectively carrying out time domain, frequency domain and time domain feature extraction and CNN feature extraction on the preprocessed surface electromyographic signals, and carrying out feature screening combination on the extracted features to obtain feature sequences after feature screening combination; inputting a muscle fatigue evaluation model based on the feature sequences obtained by the feature screening combination to obtain the muscle fatigue of the current time of the current parameter training personnel; Acquiring and processing physical sign data, normalizing the muscle fatigue degree and the physical sign data with abnormal values removed, and inputting a risk degree assessment model to carry out risk degree assessment to obtain a training risk degree assessment result of the current training personnel; The step of acquiring and preprocessing the surface electromyographic signals comprises the following steps: firstly, SG filtering is carried out on the surface electromyographic signals; Performing wavelet threshold denoising on the surface electromyographic signals after SG filtering; The wavelet threshold denoising includes: 4 layers of orthogonal decomposition are carried out on the surface electromyographic signals after SG filtering by using a Symlets-2 wavelet basis function, the surface electromyographic signals after SG filtering are used as function input, and wavelet coefficients on all decomposition layers are calculated; Carrying out wavelet coefficient screening treatment on the wavelet coefficient of each decomposition layer, and reserving the wavelet coefficient with lower amplitude; wherein, the wavelet coefficient screening is performed using the following formula: Wherein, the The values are selected for the wavelet coefficients and, Representing wavelet coefficients, k being the position at scale j, Is a threshold value at a scale j, j representing the number of layers, Representing the wavelet coefficient array of the layer, wherein N is the data length of the surface electromyographic signals after SG filtering; Performing inverse wavelet transformation to reconstruct a surface electromyographic signal based on the retained wavelet coefficient to obtain the surface electromyographic signal after wavelet threshold denoising; the time domain, frequency domain and time domain feature extraction and CNN feature extraction of the preprocessed surface electromyographic signals comprise the following steps: based on the surface myoelectric signal after wavelet threshold denoising, carrying out data segmentation by adopting a sliding window to obtain a surface myoelectric signal after data segmentation; Extracting a time domain feature set based on the surface electromyographic signals after data segmentation; performing fast Fourier transform on the surface electromyographic signals after the data segmentation to extract a frequency domain feature set; performing four-layer wavelet decomposition on the surface electromyographic signals subjected to data segmentation by using a db4 basis function to generate wavelet coefficients of each layer, and extracting a time-frequency domain feature set; and taking the surface electromyographic signals after wavelet threshold denoising as the input of a CNN model, and extracting data output by a Dense layer with the size of 16 of the penultimate layer of the CNN model to form a CNN feature set.
2. The risk assessment method according to claim 1, wherein the feature screening and combining the extracted features to obtain a feature sequence after feature screening and combining includes: Based on the time domain, frequency domain and time domain feature set and each feature sequence in the CNN feature set, decomposing each feature sequence in the feature set into two parts by a moving average method, wherein the two parts comprise stationary information of features and random parts of the features; Normalizing the stationary information of each feature sequence and the random part of the feature in the feature set; and evaluating the stationary information of the features and the random part of the features based on the normalized features by using correlation, monotonicity and robustness indexes to obtain comprehensive evaluation indexes as follows: Wherein J is a comprehensive evaluation index, The weights of the correlation, the monotonicity and the robustness are respectively, and Corr, mon and Bob are respectively indexes of the correlation, the monotonicity and the robustness; And selecting n features with highest comprehensive evaluation indexes to obtain a feature sequence of the feature screening combination.
3. The risk assessment method according to claim 2, wherein the muscle fatigue assessment model is trained based on an LSTM model, the training process comprising: Feature sequences of the feature screening combination obtained in the training stage are based, and corresponding fatigue degree sample labels are marked for each feature sequence according to the Borg scale; inputting the feature sequences of the feature screening combination and the corresponding sample labels into an LSTM model for training to obtain the muscle fatigue evaluation model; the muscle fatigue evaluation model outputs a muscle fatigue value.
4. A risk assessment method according to claim 3, wherein said acquiring and processing of sign data is: excluding zero values of the sign data to obtain screened sign data; And detecting abnormal values of the screened sign data to obtain sign data with abnormal values removed, wherein the method comprises the following steps of: Selecting n sample data as a sample subset based on the screened sign data, and recursively constructing an isolated tree from a root node; randomly selecting one piece of sign data as a threshold value, dividing the isolated tree into two branches, wherein the left branch is smaller than the threshold value, and the rest branches are arranged on the right branch; when the isolated tree grows, the leaf node of a certain branch only contains one sign data, and the sign data is an abnormal value; When the height of the isolated tree reaches the set height, stopping the growth of the tree, and eliminating abnormal values of the physical sign data to obtain the physical sign data with the abnormal values eliminated.
5. The risk assessment method according to claim 4, wherein the risk assessment model assessment is based on fuzzy neural network model FNN training: Marking risk values by the muscle fatigue value and the sign data with the abnormal values removed, and taking the risk values as input nodes of the FNN input layer after normalization processing; Defining a fuzzy condition at a fuzzification layer, and calculating the membership degree of the fuzzy condition of each input node by adopting a Gaussian function as a membership degree function; The number of nodes of the fuzzy rule layer is the fuzzy rule number, and the layer has 2 6 fuzzy rules, namely 64 nodes; The normalization layer performs normalization processing on the output data of the fuzzy rule layer; and the anti-blurring layer performs weighted summation on the output data of the normalization layer to obtain a training risk assessment result of the current training personnel.
6. The method for risk assessment according to any one of claims 1 to 5, wherein, The sign data includes heart rate, blood oxygen, body temperature, respiration, and blood pressure.
7. A deep learning-based military physical training risk assessment system, comprising: The data acquisition and preprocessing module is used for acquiring and preprocessing the surface electromyographic signals, carrying out time domain, frequency domain and time domain feature extraction and CNN feature extraction on the preprocessed surface electromyographic signals, and carrying out feature screening combination on the extracted features to obtain feature sequences after feature screening combination; The muscle fatigue evaluation module is used for evaluating the characteristic sequences obtained by screening based on the characteristic screening combination to obtain the muscle fatigue of the current time of the current parameter training personnel; The risk assessment module is used for acquiring and processing the sign data, and obtaining a training risk assessment result based on muscle fatigue and the normalized sign data with abnormal values removed; The step of acquiring and preprocessing the surface electromyographic signals comprises the following steps: firstly, SG filtering is carried out on the surface electromyographic signals; Performing wavelet threshold denoising on the surface electromyographic signals after SG filtering; The wavelet threshold denoising includes: 4 layers of orthogonal decomposition are carried out on the surface electromyographic signals after SG filtering by using a Symlets-2 wavelet basis function, the surface electromyographic signals after SG filtering are used as function input, and wavelet coefficients on all decomposition layers are calculated; Carrying out wavelet coefficient screening treatment on the wavelet coefficient of each decomposition layer, and reserving the wavelet coefficient with lower amplitude; wherein, the wavelet coefficient screening is performed using the following formula: Wherein, the The values are selected for the wavelet coefficients and, Representing wavelet coefficients, k being the position at scale j, Is a threshold value at a scale j, j representing the number of layers, Representing the wavelet coefficient array of the layer, wherein N is the data length of the surface electromyographic signals after SG filtering; Performing inverse wavelet transformation to reconstruct a surface electromyographic signal based on the retained wavelet coefficient to obtain the surface electromyographic signal after wavelet threshold denoising; the time domain, frequency domain and time domain feature extraction and CNN feature extraction of the preprocessed surface electromyographic signals comprise the following steps: based on the surface myoelectric signal after wavelet threshold denoising, carrying out data segmentation by adopting a sliding window to obtain a surface myoelectric signal after data segmentation; Extracting a time domain feature set based on the surface electromyographic signals after data segmentation; performing fast Fourier transform on the surface electromyographic signals after the data segmentation to extract a frequency domain feature set; performing four-layer wavelet decomposition on the surface electromyographic signals subjected to data segmentation by using a db4 basis function to generate wavelet coefficients of each layer, and extracting a time-frequency domain feature set; and taking the surface electromyographic signals after wavelet threshold denoising as the input of a CNN model, and extracting data output by a Dense layer with the size of 16 of the penultimate layer of the CNN model to form a CNN feature set.

Description

Military physical training risk assessment method and system based on deep learning Technical Field The invention relates to the technical field of military physical training, in particular to a risk assessment method and a risk assessment system for military physical training based on deep learning. Background Military physical training plays an important role in maintaining combat power and health status of parametrics. However, in the conventional physical training process, the training risk assessment of the training personnel can find out the uncomfortable degree of the training personnel, and if the training risk is high in the training process, the training personnel should pay attention to the training strength, and even stop training. However, most of risk degree assessment still judges the current physical state of the participant through subjective feeling, manual subjective judgment or simple physiological parameter monitoring of the participant, and the state change and risk condition of the participant are difficult to accurately capture, so that the participant cannot be accurately assessed, excessive training is easy to occur, and training damage is caused. Therefore, a study of training risk assessment by a participant must be highly appreciated. The traditional military training risk assessment has strong subjectivity and limited data processing capacity, and the complexity and diversity of the training risk make the traditional method have limitations in terms of accurate and precise risk assessment. And for risk assessment of different parametrics, multiple sign data and possible cross-influences need to be considered, increasing the complexity of the assessment. Disclosure of Invention In view of the above analysis, the embodiment of the invention aims to provide a military physical training risk assessment method based on deep learning, which uses technologies such as a convolutional neural network CNN (Convolutional Neural Network), a Long Short-Term Memory network LSTM (Long Short-Term Memory), a fuzzy neural network FNN (Fuzzy Neural Network) and the like in the military physical training risk assessment, combines surface electromyographic signals and physical sign data multisource data of physical training, comprehensively analyzes the states of the training personnel from different angles, accurately assesses the training risk, is used for solving the technical problem that the training risk assessment of the existing military physical training personnel cannot accurately assess, and provides real-time guidance and decision support for the military physical training. Through the automatic learning and feature extraction capability of deep learning, the accuracy and operability of risk assessment of military physical training are improved. The specification provides a military physical training risk assessment method based on deep learning, which comprises the following steps: Acquiring surface electromyographic signals of current parametrics and preprocessing, respectively carrying out time domain, frequency domain and time domain feature extraction and CNN feature extraction on the preprocessed surface electromyographic signals, and carrying out feature screening combination on the extracted features to obtain feature sequences after feature screening combination; inputting a muscle fatigue evaluation model based on the feature sequences obtained by the feature screening combination to obtain the muscle fatigue of the current time of the current parameter training personnel; and acquiring and processing physical sign data, normalizing the muscle fatigue degree and the physical sign data with the abnormal value removed, inputting a risk degree assessment model for risk degree assessment, and obtaining a training risk degree assessment result of the current training personnel. Further, the acquiring and preprocessing the surface electromyographic signals includes: firstly, SG filtering is carried out on the surface electromyographic signals; and carrying out wavelet threshold denoising on the surface electromyographic signals after SG filtering. Further, the wavelet threshold denoising includes: 4 layers of orthogonal decomposition are carried out on the surface electromyographic signals after SG filtering by using a Symlets-2 wavelet basis function, the surface electromyographic signals after SG filtering are used as function input, and wavelet coefficients on all decomposition layers are calculated; Carrying out wavelet coefficient screening treatment on the wavelet coefficient of each decomposition layer, and reserving the wavelet coefficient with lower amplitude; wherein, the wavelet coefficient screening is performed using the following formula: Wherein, the For the wavelet coefficient screening value, cD j,k represents the wavelet coefficient, k is the position under the scale j, lambda is the threshold, j represents the layer number, cD j represents the wavelet coefficient