A Deep Learning-Based Method and System for Risk Assessment of Military Physical Training

Abstract

This invention provides a deep learning-based method and system for assessing the risk of military physical training, belonging to the field of military physical training. The method includes acquiring and preprocessing the surface electromyography (EMG) signals of trainees; performing time-domain, frequency-domain, and CNN feature extraction on the preprocessed EMG signals; and then filtering and combining the extracted features to obtain a feature sequence. Based on this feature sequence, the system inputs it into a muscle fatigue assessment model to obtain the current muscle fatigue level of the trainee. The system also acquires and processes vital sign data; after normalizing the muscle fatigue level and vital sign data (excluding outliers), it inputs them into a risk assessment model to obtain the training risk assessment result for the current trainee. This method solves the problem of inaccurate risk assessment for military physical training participants, providing guidance and decision support for military physical training, and improving the accuracy and operability of training risk assessment.

Description

Technical Field

[0001] This invention relates to the field of military physical training technology, and in particular to a method and system for risk assessment of military physical training based on deep learning. Background Technology

[0002] Military physical training plays a crucial role in maintaining the combat effectiveness and health of trainees. However, in traditional military physical training, risk assessments can identify the degree of physical discomfort among trainees. If a high-risk situation arises during training, trainees should be advised to adjust the intensity or even stop training. However, most current risk assessments still rely on trainees' subjective feelings, subjective judgment, or simple physiological parameter monitoring to determine their current physical condition. This approach struggles to accurately capture changes in trainees' condition and assess risks, lacking objectivity and precision, and is prone to overtraining and injury. Therefore, research on training risk assessment for trainees must be given high priority.

[0003] Traditional military training risk assessments are highly subjective and have limited data processing capabilities. The complexity and diversity of training risks limit the accuracy and precision of traditional methods. Furthermore, risk assessments for different trainees require consideration of multiple vital signs and potential cross-influences, further complicating the process. Summary of the Invention

[0004] Based on the above analysis, this invention aims to provide a deep learning-based method for assessing the risk of military physical training. It utilizes technologies such as Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTMs), and Fuzzy Neural Networks (FNNs) in this risk assessment. This method combines surface electromyography (EMG) signals and vital sign data from multiple sources to comprehensively analyze the trainees' status from different perspectives, accurately assessing training risks. This addresses the technical problem of the inaccurate assessment of training risks in existing military physical training methods and provides real-time guidance and decision support for military physical training. Through the automatic learning and feature extraction capabilities of deep learning, the accuracy and operability of risk assessment for military physical training are improved.

[0005] This manual provides a deep learning-based method for assessing the risks of military physical training, including the following steps:

[0006] The surface electromyography (EMG) signals of the current trainees are acquired and preprocessed. Time-domain, frequency-domain, and CNN feature extraction are performed on the preprocessed EMG signals. The extracted features are then filtered and combined to obtain the feature sequence after feature filtering and combination.

[0007] Based on the feature sequence obtained by feature selection and combination, input it into the muscle fatigue assessment model to obtain the muscle fatigue level of the current trainee at the current moment.

[0008] The vital signs data are acquired and processed. The muscle fatigue level and vital signs data after removing outliers are normalized and then input into the risk assessment model to assess the risk level, thereby obtaining the training risk assessment result of the current trainee.

[0009] Furthermore, the acquisition and preprocessing of surface electromyography signals includes:

[0010] First, the surface electromyography signal is subjected to SG filtering;

[0011] Wavelet thresholding is performed on the surface electromyography signal after SG filtering.

[0012] Furthermore, the wavelet threshold denoising includes:

[0013] The Symlets-2 wavelet basis function was used to perform a four-level orthogonal decomposition of the SG-filtered surface electromyography (EMG) signal. The wavelet coefficients on each decomposition level were calculated using the SG-filtered EMG signal as the function input.

[0014] Wavelet coefficients of each decomposition layer are filtered to retain wavelet coefficients with lower amplitudes.

[0015] The wavelet coefficients are selected using the following formula:

[0016]

[0017]

[0018] in, For wavelet coefficient selection, cd j,k This represents the wavelet coefficients, where k is the position at scale j, and λ is the threshold at scale j; λ is the threshold value, j represents the layer number, and cD j This represents the wavelet coefficient array of this layer, where N is the data length of the surface electromyography signal after SG filtering;

[0019] Based on the retained wavelet coefficients, the surface electromyography (EMG) signal is reconstructed by inverse wavelet transform to obtain the surface EMG signal after wavelet threshold denoising.

[0020] Furthermore, the step of performing time-domain, frequency-domain, and CNN feature extraction on the preprocessed surface electromyography signal includes:

[0021] Based on the surface electromyography signal after wavelet threshold denoising, the data segmentation of the surface muscle signal is obtained by using a sliding window.

[0022] Based on the surface electromyography signals after data segmentation, a time-domain feature set is extracted;

[0023] Perform a Fast Fourier Transform on the segmented surface electromyography (EMG) signals to extract the frequency domain feature set;

[0024] The surface electromyography signal after data segmentation was decomposed into four-level wavelet coefficients using the db4 basis function to extract the time-frequency domain feature set.

[0025] The surface electromyography signal after wavelet threshold denoising is used as the input of the CNN model. The data output of the penultimate Dense layer with a size of 16 of the CNN model is extracted to form the CNN feature set.

[0026] Furthermore, the step of performing feature filtering and combination on the extracted features to obtain the feature sequence after feature filtering and combination includes:

[0027] Based on the time domain, frequency domain, and time-frequency domain feature sets and each feature sequence in the CNN feature set, each feature sequence in the feature set is decomposed into two parts by the moving average method, including the stationary information of the feature and the random part of the feature;

[0028] The stationary information and random part of each feature sequence in the feature set are normalized.

[0029] Based on the stationary information and random components of the normalized features, correlation, monotonicity, and robustness indices are used for evaluation, resulting in the following comprehensive evaluation index:

[0030] J = w1Corr + w2Mon + w3Rob

[0031]

[0032] Where J is the comprehensive evaluation index, w1, w2, and w3 are the weights of correlation, monotonicity, and robustness, respectively, and Corr, Mon, and Bob are the correlation, monotonicity, and robustness indices, respectively.

[0033] Select the n features with the highest comprehensive evaluation index to obtain the feature sequence of the feature selection combination.

[0034] Furthermore, the fatigue prediction model is trained based on an LSTM model, and the training process includes:

[0035] Based on the feature sequences obtained during the training phase, the fatigue sample labels are labeled for each feature sequence according to the Borg scale.

[0036] The feature sequence and corresponding sample labels of the feature selection combination are input into the LSTM model for training to obtain the fatigue prediction model.

[0037] The fatigue prediction model outputs a muscle fatigue value.

[0038] Furthermore, the acquisition and processing of vital sign data includes:

[0039] Zero values ​​in the vital signs data are excluded to obtain the filtered vital signs data;

[0040] For the filtered vital sign data, outlier detection is performed to obtain vital sign data after removing outliers, including:

[0041] Based on the filtered vital sign data, n sample data are selected as a sample subset, and an isolation tree is recursively constructed starting from the root node.

[0042] Randomly select a vital sign data point as a threshold, and divide the isolated tree into two branches. Data points below the threshold are placed on the left branch, and the rest are placed on the right branch.

[0043] When an isolated tree is growing, if a leaf node of a certain branch contains only one symptom data, that symptom data is an outlier.

[0044] When the isolated tree reaches the set height, the tree stops growing. Outliers in the vital signs data are then removed, resulting in vital signs data with outliers removed.

[0045] Furthermore, the risk assessment model is trained based on a fuzzy neural network model (FNN):

[0046] The muscle fatigue value and the vital sign data with outliers removed are labeled with risk values, and after normalization, they are used as input nodes of the input layer of the fuzzy neural network model FNN.

[0047] Fuzzy conditions are defined in the fuzzification layer, and a Gaussian function is used as the membership function to calculate the fuzzy conditional membership degree of each input node.

[0048] The number of nodes in the fuzzy rule layer is equal to the number of fuzzy rules; this layer has 2. 6 There are 64 fuzzy rules;

[0049] The normalization layer normalizes the output data of the fuzzy rule layer;

[0050] The deblurring layer performs a weighted summation of the output data from the normalization layer to obtain the training risk assessment result for the current trainees.

[0051] Furthermore, the vital signs data include heart rate, blood oxygen, body temperature, respiration, and blood pressure.

[0052] This manual also provides a deep learning-based risk assessment system for military physical training, including:

[0053] The data acquisition and preprocessing module is used to acquire surface electromyography (EMG) signals and perform preprocessing. It performs time-domain, frequency-domain, and time-frequency-domain feature extraction and CNN feature extraction on the preprocessed surface EMG signals. It then performs feature filtering and combination on the extracted features to obtain a feature sequence after feature filtering and combination.

[0054] The muscle fatigue assessment module is used to assess the muscle fatigue level of the current trainee at the current moment based on the feature sequence obtained by feature screening and combination.

[0055] The risk assessment module is used to acquire and process vital sign data. Based on muscle fatigue and normalized vital sign data after removing outliers, the module assesses the training risk level.

[0056] Compared with the prior art, the beneficial effects of the present invention are:

[0057] 1. This solution improves the assessment of training risks for trainees in military physical training from the traditional subjective judgment-based method to a deep learning-based assessment method, which is more objective, scientific, and real-time.

[0058] 2. This scheme combines vital signs parameters during military physical training with surface electromyography signals from the training sites, increasing the reliability of training risk assessment results.

[0059] 3. This solution introduces a convolutional neural network (CNN) when extracting surface electromyography (EMG) signal features, which enriches the feature extraction dimensions and improves the accuracy of the results. This solves the technical problem that existing military physical training participants cannot accurately assess training risks, and provides real-time guidance and decision support for military physical training.

[0060] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0061] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0062] Figure 1 This is a diagram illustrating the overall architecture of a deep learning-based risk assessment system for military physical training.

[0063] Figure 2 This is a schematic diagram illustrating the specific implementation of a deep learning-based risk assessment method for military physical training.

[0064] Figure 3 This is a diagram illustrating the decomposition process of surface electromyography signals after wavelet filtering.

[0065] Figure 4 To analyze surface electromyography signal feature maps using convolutional neural networks;

[0066] Figure 5 Numerical structure diagram of muscle fatigue during training of the Long Short-Term Memory (LSTM) network;

[0067] Figure 6 Flowchart for outlier detection in isolated forest models;

[0068] Figure 7 This is a training graph for a fuzzy neural network.

[0069] Figure 8 This is a schematic diagram of a deep learning-based risk assessment system for military physical training. Detailed Implementation

[0070] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0071] This solution primarily addresses the problem of timely and accurate assessment of training risks for trainees during military physical training. To resolve this technical issue, this solution provides a deep learning-based method and system for assessing military physical training risks.

[0072] Specifically, the biological signals generated by trainees during military physical training are taken into account, including electrical and non-electrical signals. The electrical signals are surface electromyography (sEMG) signals of the muscles being trained, while the non-electrical signals are vital signs data, including the trainee's current heart rate, blood oxygen, body temperature, respiration, and blood pressure.

[0073] Example 1:

[0074] This scheme performs CNN feature extraction, time-domain feature extraction, frequency-domain feature extraction, and time-frequency domain feature extraction on the preprocessed surface electromyography (sEMG) signals. The CNN features are combined with the time-domain, frequency-domain, and time-frequency domain features through feature selection and selection, and then used as the input to the fatigue assessment model to evaluate and calculate the muscle fatigue of the trainees.

[0075] The processed vital signs data and the calculated muscle fatigue level are then normalized and input into the risk assessment model for evaluation and analysis to obtain the training risk level of the current trainees.

[0076] The muscle fatigue assessment model was trained using a Long Short-Term Memory (LSTM) network, while the risk assessment model was trained using a fuzzy neural network model.

[0077] like Figure 1 and Figure 2 As shown, the method includes steps S1-S3.

[0078] Step S1: Obtain the surface electromyography (EMG) signal of the current trainee and preprocess it. Perform time-domain, frequency-domain, and time-frequency domain feature extraction and CNN feature extraction on the preprocessed surface EMG signal. Perform feature selection and combination on the extracted features to obtain the feature sequence after feature selection and combination.

[0079] Step S2: Input the feature sequence obtained by feature filtering and combination into the muscle fatigue assessment model to obtain the muscle fatigue level of the current trainee at the current moment.

[0080] Step S3: Acquire and process vital sign data. After normalizing the muscle fatigue level and vital sign data after removing outliers, input the data into the risk assessment model to conduct a risk assessment and obtain the training risk assessment result of the current trainee.

[0081] Specifically, step S1 includes steps S11-S13.

[0082] Step S11: Acquire surface electromyography signals and perform preprocessing.

[0083] Electrode patches are attached to the surface of the active muscles of trainees to acquire surface electromyography (sEMG) signals. Savitzky-Golay smoothing filtering (hereinafter referred to as 'SG filtering') and wavelet thresholding are then applied to the surface EMG signals to reduce noise and interference in the signals, making them more suitable for subsequent evaluation and analysis.

[0084] Surface electromyography (EMG) signals are bioelectrical currents generated by the contraction of muscles on the human body; they are electrical signals. Electrode patches are attached to the surface of the skin to monitor and record these electrical signals of muscle activity. Electrode patches are typically attached near specific muscle groups to capture the electrical activity generated during muscle contraction and relaxation. In military physical training risk assessments, the placement of electrode patches varies depending on the application requirements, but typically includes:

[0085] 1) Quadriceps: This is a muscle group located at the front of the thigh, covering the front of the thigh. During some training activities, the quadriceps may bear a heavy load; monitoring its surface electromyography (EMG) signals can help assess muscle fatigue.

[0086] 2) The hamstrings are located at the back of the thigh, opposite the quadriceps. Monitoring the surface electromyographic signals of the hamstrings can provide information on the activity of the muscles in the back of the thigh.

[0087] 3) Abdominal muscles, located in the abdomen, are part of the core muscle group. By monitoring the surface electromyographic signals of the abdominal muscles, the activity of the core muscle group of the trainees can be understood.

[0088] 4) Back muscles: Monitoring the electrical activity of back muscles can provide information on back stability and support capacity.

[0089] 5) Biceps brachii, located in the front of the upper arm, can be monitored by measuring its electromyographic signals to understand the activity of the upper arm;

[0090] 6) The triceps brachii is located in the back of the upper arm. Monitoring its electromyographic signals can help understand the activity of the upper arm.

[0091] 7) Other key muscle groups: Select other important muscle groups, such as calf muscles and chest muscles, depending on the specific training goals.

[0092] The specific placement of the electrode patches can be adjusted and selected according to the actual application and training goals. Typically, the electrode patches are applied to the central area of ​​the muscle to obtain the most accurate electromyographic signals. During application, ensure the electrode patches adhere well to the skin to obtain accurate surface electromyographic signals.

[0093] As the degree of local muscle fatigue increases, the randomness of surface electromyography (sEMG) signals decreases and the periodicity increases. Therefore, the degree of local muscle fatigue in trainees can be assessed by analyzing these surface EMG signals.

[0094] The surface electromyography (EMG) signal is preprocessed, including SG filtering and wavelet thresholding to obtain the denoised EMG signal.

[0095] First, the surface electromyography signal is subjected to SG filtering.

[0096] While effectively preserving the features in the signal, it reduces and eliminates high-frequency noise in the signal, reduces noise interference in surface electromyography (sEMG) signals, and improves the smoothness of surface electromyography (sEMG) signals, thereby making the signal smoother and more continuous.

[0097] Next, wavelet thresholding denoising is performed on the surface electromyography signal after SG filtering.

[0098] Wavelet thresholding is used to further reduce noise levels and enhance the effective signal in the signal.

[0099] In wavelet thresholding denoising, it is necessary to select appropriate thresholding strategies and wavelet basis functions. Symlets wavelet functions possess good symmetry and regularity, and also perform well in terms of continuity, compact support, and orthogonality.

[0100] Therefore, this scheme selects Symlets-2 as the wavelet basis function to perform a 4-level orthogonal decomposition on the SG-filtered surface electromyography signal, and uses the SG-filtered surface electromyography signal as the input of the wavelet basis function to calculate the wavelet coefficients on each decomposition level.

[0101] Regarding the threshold function, this paper employs a trade-off between soft and hard thresholds to reconstruct the surface electromyography signal after SG filtering, and selects wavelet coefficients. The calculation formula is shown in formula (1).

[0102]

[0103] Among them, cd j,k Let λ represent the wavelet coefficient, k be the position at scale j, and λ be the threshold at scale j. The calculation formula is shown in equation (2).

[0104]

[0105] Where j represents the layer number, cD j This represents the wavelet coefficient array of this layer, and N represents the data length of the filtered surface electromyography (sEMG) signal.

[0106] In wavelet transform, wavelet coefficients are obtained by decomposing the signal at different scales (frequency) and locations (time). Scale represents the frequency characteristics of the surface electromyography (EMG) signal, location represents the temporal position of the EMG signal, and wavelet coefficients represent the changes in the signal at different scales and locations.

[0107] The basic idea of ​​wavelet thresholding denoising is to convert the time-domain signal to the wavelet domain, and then decompose the effective signal and noise signal in the surface electromyography signal after SG filtering by wavelet transform. Taking advantage of the fact that the wavelet coefficient of the noise signal is significantly smaller than the wavelet coefficient of the effective signal, the wavelet transform coefficient of the noise signal is eliminated.

[0108] The main steps of wavelet thresholding denoising are as follows:

[0109] ① The first step is to use the Symlets-2 wavelet basis functions to perform a 4-level orthogonal decomposition on the SG-filtered surface electromyography (sEMG) signal. The decomposition process is as follows: Figure 3 As shown, the surface electromyography signal after SG filtering is used as the input of the wavelet basis function. The wavelet coefficients of each decomposition layer are screened and processed, and wavelet coefficients with lower amplitudes are retained.

[0110] The wavelet coefficients cd1, cd2, cd3, and cd4 with lower amplitudes in each decomposition layer are retained. Among them, ca1, ca2, ca3, and ca4 represent the higher amplitude components obtained from the first, second, third, and fourth wavelet decomposition layers, respectively.

[0111] ② In the second step, for the layers of wavelet coefficients with higher amplitude, a threshold is defined to select useful wavelet coefficients. The threshold is calculated as shown in equation (2). This scheme adopts a trade-off between soft and hard thresholds, and performs wavelet coefficient screening on the wavelet coefficients with higher amplitude in each decomposition layer according to formula (1).

[0112] By setting wavelet coefficients less than or equal to the threshold λ to zero, the influence of noise is reduced. By processing wavelet coefficients at different scales and locations, signal characteristics are preserved while noise is removed, thereby improving the quality and accuracy of surface electromyography signals.

[0113] The surface electromyography (sEMG) signal is reconstructed by inverse wavelet transform based on the retained wavelet coefficients, and the sEMG signal after wavelet threshold denoising is obtained.

[0114] By combining SG filtering and wavelet thresholding for denoising, the quality of surface electromyography signals can be effectively improved, thus enabling more accurate muscle fatigue analysis in the future.

[0115] Step S12: Perform time-domain, frequency-domain, and CNN feature extraction on the preprocessed surface electromyography signal.

[0116] The preprocessed sEMG signal was used to extract time-domain, frequency-domain, and time-frequency-domain features. Simultaneously, the preprocessed sEMG signal was input into a pre-trained convolutional neural network (CNN) model to extract CNN data features, which can be used to describe muscle fatigue.

[0117] Feature extraction is performed on the surface electromyography signals after data preprocessing.

[0118] Since the surface electromyography (EMG) signal after data preprocessing (SG filtering and wavelet thresholding denoising) is a one-dimensional data segment, feature extraction from it can only yield a single feature value, making it unsuitable for network training. Furthermore, the longer the signal acquisition time and the larger the data volume, the longer the data processing time.

[0119] Therefore, this scheme uses the sEMG signal after data preprocessing with a sliding window for data segmentation, with a window length of 300 sampling points and a step size of 10 sampling points.

[0120] Since some characteristic components in sEMG signals can characterize the fatigue state of muscles, feature extraction from sEMG signals is a fundamental step in assessing muscle fatigue.

[0121] The most commonly used features for assessing muscle fatigue using sEMG signals are time-domain features, frequency-domain features, and time-frequency-domain features. However, these cannot comprehensively and accurately characterize the state of muscle fatigue. Therefore, this solution introduces a CNN convolutional neural network model. By leveraging the ability of CNNs to deeply mine data features, deep-level features in sEMG signals are extracted and combined with time-domain, frequency-domain, and time-frequency-domain features to jointly describe the state of muscle fatigue.

[0122] Based on the preprocessed surface electromyography (sEMG) signal, i.e. the surface electromyography signal after wavelet threshold denoising, a sliding window is used to segment the data to obtain the segmented surface electromyography signal.

[0123] (1) Extract time-domain feature set based on surface electromyography signal after data segmentation.

[0124] The time-domain feature set includes: integrated electromyography value, root mean square value, mean absolute value, zero-crossing rate, variance, maximum fractal length, and slope change rate;

[0125] (2) Perform a fast Fourier transform on the surface electromyography signal after data segmentation to extract the frequency domain feature set.

[0126] First, a fast Fourier transform is performed on the surface electromyography signal after data segmentation. Then, the average power frequency, median frequency, cepstral coefficients, and power spectral ratio of the Fourier transformed signal are extracted as a frequency domain feature set.

[0127] (3) Use the db4 basis function to perform four-level wavelet decomposition on the surface electromyography signal after data segmentation to generate wavelet coefficients of each level and extract time-frequency domain feature set;

[0128] This scheme uses a four-level wavelet decomposition to generate wavelet coefficients c of the segmented surface electromyography (EMG) signal.i (c i ∈[cd1,cd2,cd3,cd4,ca4]), where wavelet coefficients cd1, cd2, cd3, and cd4 represent the low-frequency components (coarsedetail) obtained from the first, second, third, and fourth levels of wavelet decomposition, respectively, and ca4 represents the high-frequency components (approximation) obtained from the fourth level of wavelet decomposition. It contains most of the high-frequency components of the original signal and can be considered an approximate signal. The average absolute value, average energy, standard deviation, and the ratio of the mean to the mean of the wavelet coefficients are extracted as the time-frequency domain feature set.

[0129] (4) The surface electromyography signal after wavelet threshold denoising is used as the input of the CNN model. The data output of the second to last Dense layer of the CNN model with a size of 16 is extracted to form the CNN feature set.

[0130] Because Convolutional Neural Networks (CNNs) have excellent local feature sensing capabilities for signal data, they can mine deeper-level feature signals and possess weight-sharing characteristics, greatly reducing the number of model parameters. Therefore, this scheme chooses to use CNNs for deep feature mining of sEMG signals. Since the signals to be processed in this scheme are one-dimensional data, a one-dimensional CNN is used to extract signal features. The CNN model is as follows: Figure 4 As shown.

[0131] The convolutional neural network (CNN) model designed in this scheme consists of four modules, each composed of two one-dimensional convolutional layers and one one-dimensional pooling layer, and two fully connected layers. During training, the preprocessed surface electromyography (sEMG) signal is used as input, and the corresponding muscle fatigue level is used as output. The CNN model is trained using mean squared error as the loss function, and the Adam optimization algorithm is used to optimize the model parameters. The learning rate is set to 0.001.

[0132] The preprocessed surface electromyography (sEMG) signal is used as input, and the corresponding muscle fatigue level is output. The purpose is to extract the original features from the surface electromyography signal.

[0133] Furthermore, the data from the second-to-last Dense layer of the trained CNN model, with a size of 16, is extracted as features to form a 16-dimensional CNN feature set.

[0134] The function of the feature extraction step is to extract features from the surface electromyography (sEMG) signals after data preprocessing, converting the surface EMG signals into easily processed data, reducing data dimensionality, and simultaneously extracting important information from the surface EMG signals. This scheme utilizes the deep feature mining capabilities of the Convolutional Neural Network (CNN) model to uncover deep-level features of the preprocessed signals. The CNN feature set is then combined with the extracted time-domain, frequency-domain, and time-frequency-domain feature sets to jointly describe muscle fatigue states.

[0135] Step S13: Based on the time domain, frequency domain, and frequency domain feature sets and the CNN feature set, perform feature selection and combination to obtain the feature sequence after feature selection and combination.

[0136] The purpose of this step is to select the most meaningful features from the feature set, reduce the feature dimensions, and retain important information.

[0137] By comprehensively considering correlation, monotonicity, and robustness from the time-domain feature set, frequency-domain feature set, time-frequency domain feature set, and CNN feature set, the feature sequences with the highest comprehensive scores are selected, and the feature sequences most meaningful for fatigue assessment are selected, in order to reduce feature dimensionality and retain important information.

[0138] The extracted signal features described above can effectively reflect the muscle fatigue state of trainees, and these features should be retained. However, some features contain a lot of redundant information and are not sensitive enough to capturing muscle fatigue state, affecting the muscle fatigue assessment results. Therefore, before using feature sets (time domain, frequency domain, time-frequency domain feature sets, and CNN feature sets) for muscle fatigue assessment, feature selection must be performed on the existing feature sets to filter out features useful for muscle fatigue prediction and eliminate redundant and useless features as much as possible.

[0139] This scheme adopts a multi-indicator fusion method, which linearly integrates three evaluation indicators: correlation, monotonicity, and robustness, to select some features with higher comprehensive evaluation.

[0140] Suppose F = [f(1), f(2), ..., f(T)] represents each feature sequence passed through the time-domain feature set, frequency-domain feature set, time-frequency-domain feature set, and CNN feature set, and t = 1, 2...T represents the time series. Each feature sequence in the feature set is decomposed into two parts using the moving average method, as shown in equation (3).

[0141] f(t) = f S (t)+f R (t) (3)

[0142] In the formula, f S (t) represents stationary information of the feature, f R(t) represents the random part of the feature.

[0143] It helps to analyze and separate trend and random information in the features.

[0144] Before calculating the index, the stationary information and random components of the feature sequences need to be normalized to ensure that the values ​​of all variables involved in the calculation, except for the time series T, are within the range of [0,1]. It is also important to ensure that all feature sequences are calculated on the same scale to avoid the influence of values ​​from different ranges on the index calculation.

[0145] Based on the stationary information of the normalized features and the random part of the features, the correlation, monotonicity and robustness indices are used for evaluation.

[0146] Correlation, monotonicity, and robustness play a role based on the stationary information of features and the random part of features.

[0147] Correlation (Corr): Indicates the degree of association between different features. If there is a high correlation between features, the weights need to be adjusted appropriately when selecting and combining features to avoid repeating similar information.

[0148] Monotonicity (Mon): Indicates the monotonic trend of a feature under different conditions;

[0149] Robustness (Rob): Indicates the sensitivity of a feature to outliers. Features with higher robustness are more likely to remain stable in the presence of noise or outliers, thereby improving the reliability of the evaluation.

[0150] The formulas for Correlation (Corr), Monotonicity (Mon), and Robustness (Rob) are shown in equations (3) to (5).

[0151]

[0152]

[0153]

[0154] Where M = [m(1), m(2), ..., m(T)] is the time series of muscle fatigue, and δ(·) is the step function. Each represents the average value of the F and M data sequences.

[0155] In order to integrate the three feature evaluation indicators, this scheme constructs a linear weighted index to integrate the correlation, monotonicity and robustness indicators, and uses the final weighted value as the evaluation standard for a certain feature. By combining them in a weighted manner, a comprehensive evaluation index is obtained.

[0156] Specifically, a weight is assigned to each evaluation indicator—relevance, monotonicity, and robustness—to represent their importance in the overall evaluation. The evaluation of each indicator is multiplied by its corresponding weight, and the weighted evaluations are summed to obtain the final overall evaluation index.

[0157] The formula for the comprehensive evaluation index is shown in equation (7).

[0158]

[0159] In the formula, w i The weights of each evaluation indicator relative to the comprehensive evaluation indicator are given by J, where J is the comprehensive evaluation indicator, and w1, w2, and w3 are the weights of correlation, monotonicity, and robustness, respectively.

[0160] w i The settings can be customized to the specific circumstances, with each individual indicator having a fixed value range to prevent masking the effects of other indicators.

[0161] The value of each weight can be determined based on experience, domain knowledge, or through quantitative methods such as statistical methods or machine learning models.

[0162] The choice of weights varies depending on the actual business scenario. For example, some prioritize relevance, while others focus more on monotonicity or robustness. In practice, the weight selection is adjusted based on the specific problem to obtain evaluation results that reflect the actual situation.

[0163] For different feature sequences, the final calculated comprehensive evaluation index J will be different. The key to distinguishing the quality of a feature sequence lies in the magnitude of the comprehensive evaluation index J. If J is larger, the corresponding feature will have a better effect on muscle fatigue assessment. Therefore, in this scheme, the same calculation is performed on each feature sequence in each feature set in step S12 to calculate the comprehensive evaluation index, which will not be elaborated here.

[0164] The comprehensive evaluation indicators are then sorted from largest to smallest, and the n (n≥1) features with the highest comprehensive evaluation indicators are selected.

[0165] The specific features to be selected will be determined based on the specific application.

[0166] Based on the time domain, frequency domain, and CNN feature sets, this step performs feature selection and combination to obtain the feature sequence after feature selection and combination.

[0167] Step S2, specifically.

[0168] The feature sequences obtained through feature filtering and combination are input into a pre-trained fatigue prediction model, which is trained based on an LSTM model.

[0169] The fatigue prediction model is used to model and analyze the muscle fatigue level of trainees at the current moment. Based on the LSTM model, the temporal relationships in time series data can be captured, further improving the accuracy of muscle fatigue assessment.

[0170] Based on the feature sequences obtained during the training phase, the fatigue sample labels are labeled for each feature sequence according to the Borg scale.

[0171] This program uses the Borg scale. The core idea of ​​the Borg scale is to have trainees select a number to represent their perceived exercise intensity or effort during training. This number represents the individual's subjective perception of the difficulty or intensity of the exercise, rather than being based on objective physiological indicators. The higher the number on the Borg scale, the greater the exercise intensity.

[0172] The purpose of the Borg scale is to help trainees subjectively assess their perceived level of physical effort, thereby enabling them to regulate and control their bodies in different activities.

[0173] This program classifies muscle fatigue into 11 levels, using numbers from 1 to 11. Training samples and sample labels are obtained through participants' self-perception of muscle activity. This number reflects the participants' subjective feeling about their level of effort.

[0174] Table 1 Borg Scale for Muscle Fatigue

[0175]

[0176]

[0177] The time-domain, frequency-domain, time-frequency-domain, and CNN feature sequences, along with their corresponding sample labels, after feature filtering are input into the Long Short-Term Memory (LSTM) network model for training.

[0178] The LSTM network structure used in this scheme has a total of three hidden layers, including two LSTM layers and one fully connected layer. Each hidden layer has 128 nodes. The loss function, optimization algorithm, and learning rate are consistent with the CNN model. The LSTM model structure is as follows: Figure 5 As shown.

[0179] Muscle fatigue values ​​were trained using a Long Short-Term Memory (LSTM) network.

[0180] Depend on Figure 5 As shown:

[0181] (1) Input layer: The feature sequence k = [k1, k2, ..., k2] after feature selection and combination. n As input to the LSTM network.

[0182] (2) Hidden Layer: The hidden layer consists of two LSTM layers and one fully connected layer, each with 128 nodes. The LSTM layers process the feature sequence data to obtain the fully connected layer data h. k =[h k1 ,h k2 ,…,h kn ].

[0183] (3) Output layer: The output layer contains one node, where f is the final muscle fatigue value, derived from the data h of the fully connected layer. k The calculation yielded the result.

[0184] (4) Training the LSTM model: Using backpropagation and optimization algorithms (such as gradient descent or Adam optimization), with mean squared error as the loss function, the parameters in the LSTM network are iteratively optimized to make the muscle fatigue value f output by the model as close as possible to the actual observed value. The training dataset will include known feature sequences and corresponding muscle fatigue values ​​f.

[0185] (5) Calculate the muscle fatigue value: After training is completed, the new feature sequence is input into the trained LSTM model, and the corresponding muscle fatigue value f is calculated through the forward propagation process. This value represents the current level of muscle fatigue of the trainee according to the model's prediction.

[0186] The fatigue prediction model outputs muscle fatigue values.

[0187] In summary, by modeling feature sequences using an LSTM model, the temporal information and feature relationships within the sequence data can be learned, thereby predicting muscle fatigue values. The training process of the LSTM model involves parameter optimization using known feature sequences so that the model can make accurate predictions on unknown data.

[0188] Step S3 is detailed below.

[0189] Step S31: Acquire and process vital sign data. Acquire vital sign data of trainees, including heart rate, blood oxygen, body temperature, respiration and blood pressure. Use the isolated forest algorithm to detect outliers and remove them.

[0190] (1) Obtain vital signs data.

[0191] The system acquires vital signs data of trainees, including heart rate, blood oxygen saturation, body temperature, respiration, and blood pressure. These vital signs data are non-electrical signals and are collected in real time by sensors worn on the human body.

[0192] Sensors can typically be worn on different parts of the body to collect vital sign data; the specific location depends on the metric being monitored and the application scenario. For example:

[0193] Wrist / Arm: Sensors on the wrist can measure heart rate, blood oxygen saturation, etc.;

[0194] Chest: The chest position can be used for electrocardiogram (ECG) measurements to obtain more detailed heart rate activity data;

[0195] Fingertip: A fingertip sensor can measure blood oxygen saturation;

[0196] Ears: The ear can be used to measure heart rate and body temperature.

[0197] The sensors can also be worn on the waist / abdomen, ankle / foot, and head, with the placement varying depending on the application and objectives. When collecting vital sign data, ensure the sensors are correctly positioned and detect and process any outliers to obtain accurate and reliable data.

[0198] (2) Perform data preprocessing on vital signs data.

[0199] Vital signs data are non-electrical signals collected in real time by sensors worn on the human body. Abnormalities are inevitable, such as heart rate or blood oxygen levels showing 0 or abnormally high values. Therefore, outlier detection is necessary. The Isolation Forest algorithm is used for outlier detection and removal to ensure the accuracy of the vital signs data.

[0200] This solution uses the Isolation Forest algorithm for outlier detection and removes abnormal physical characteristic data samples. The construction process of the Isolation Forest model used in this solution is as follows: Figure 6 As shown.

[0201] like Figure 6 As shown, the Isolation Forest algorithm for outlier detection in vital sign data follows these steps:

[0202] Step 1: Exclude zero-value data to obtain filtered vital sign data.

[0203] Exclude vital sign data samples with a value of 0 (including heart rate, blood oxygen, body temperature, respiration, and blood pressure).

[0204] Step 2: Construct an isolation tree based on the filtered vital sign data.

[0205] Randomly select n data samples from the filtered vital signs data as a subset of samples; create an isolated tree, starting with the root node and placing it into the root node of the isolated tree.

[0206] Step 3: Recursively construct the isolated tree.

[0207] A feature is randomly selected from the data sample.

[0208] Within the data value range of the feature at the current node, randomly select a vital sign data point as a threshold, and divide all vital sign data samples into two branches based on the threshold:

[0209] Items below the threshold are placed on the left branch, and the rest are placed on the right branch.

[0210] Step 4: Remove outliers from vital sign data.

[0211] Determining whether an isolated tree leaf node has only one vital sign data sample or whether the tree has grown to the set height are two stopping conditions in the growth process of an isolated tree. If so, the process ends; otherwise, continue to step three.

[0212] When an isolated tree is growing, if a leaf node of a branch contains only one sample data, it means that the sample data is relatively isolated in the branch of the tree and there are no other sample data nearby. This usually means that the sample data is an outlier.

[0213] When an isolated tree reaches a set height, it stops growing to prevent overfitting.

[0214] Outlier values ​​in the vital signs data were removed.

[0215] Vital health data collected by sensors may be subject to various interferences, resulting in outliers such as zero heart rate or abnormally high blood oxygen levels. The Isolation Forest algorithm can effectively monitor and eliminate these outliers, ensuring the quality of the vital health data.

[0216] In military training, the safety of trainees is of paramount importance. By eliminating outliers in vital signs data, misjudgments or incorrect assessments caused by abnormal data can be reduced, thereby improving the safety of training.

[0217] Step S32: Based on the muscle fatigue value and the physical sign data after removing outliers, normalization processing is performed to obtain the input data of the fuzzy neural network (FNN).

[0218] The risk assessment model is trained based on the fuzzy neural network model (FNN).

[0219] The muscle fatigue values ​​obtained from the muscle fatigue assessment model are normalized together with the physical signs data after outlier removal. The normalized data is then used as the input data for the fuzzy neural network model (FNN). The normalization process ensures that data with different features are on the same scale. The output data is the corresponding training risk assessment result.

[0220] The results of the training risk assessment are shown in Table 2.

[0221] Step S33: Based on the normalized data, the risk assessment model evaluates the training risk assessment results of the current trainees.

[0222] The risk assessment model is trained based on the fuzzy neural network model (FNN).

[0223] The normalized data is input into a pre-trained fuzzy neural network model to calculate the training risk assessment result for the current trainee. The fuzzy neural network model (FNN) can handle uncertainty and fuzziness, helping to derive the training risk assessment result for the current trainee.

[0224] Construction and training of the fuzzy neural network model FNN.

[0225] This scheme uses a fuzzy neural network (FNN) model to analyze vital sign data (heart rate, blood oxygen, body temperature, respiration, and blood pressure) after outlier removal and the calculated muscle fatigue values. For example... Figure 7 As shown, it includes:

[0226] Step 1: Preparing input data for the fuzzy neural network (FNN) model.

[0227] The vital signs data (heart rate, blood oxygen, body temperature, respiration, and blood pressure) collected by the sensors after removing outliers, as well as the calculated muscle fatigue value f, are normalized and used as input data for the fuzzy neural network model FNN.

[0228] Step 2: Setting up the Fuzzy Neural Network (FNN) network structure.

[0229] Design the network structure of the fuzzy neural network model FNN, adopting a five-layer structure, and determine the number of nodes and connection methods of the input layer, fuzzification layer, fuzzy rule layer, normalization layer, and defuzzification layer (output layer).

[0230] Five vital signs (heart rate, blood oxygen, body temperature, respiration, and blood pressure) with outliers and one calculated muscle fatigue value were removed. The training risk value corresponding to each training sample data was obtained through the trainee's verbal description or sensor alarm. The input sample data were marked with risk values. The output variables (training risk values) corresponding to different combinations of input variables are shown in Table 2.

[0231] Table 2 Training Risk and its Numerical Representation

[0232]

[0233] After labeling the muscle fatigue value and the vital sign data after removing outliers with risk values, normalization processing is performed.

[0234] The normalized muscle fatigue values ​​marked with risk values ​​and the vital signs data after removing outliers are used as the input sample data for the input layer, with a total of 6 input nodes x1-x6.

[0235] Step 3: Define fuzzy conditions.

[0236] In this scheme, two fuzzy conditions are defined for each input sample data: normal and abnormal. A Gaussian function is used as the membership function, and the fuzzy layer calculates the fuzzy condition membership degree of each input node.

[0237] Step 4: Define fuzzy rules.

[0238] There are 10 nodes in total; the number of nodes in the fuzzy rule layer is equal to the number of fuzzy rules, and this layer has 2. 6 There are 64 fuzzy rules.

[0239] Step 5: Normalization.

[0240] The normalization layer normalizes the output data of the fuzzy rule layer to ensure that different feature data are on the same scale.

[0241] Step 6: Deblurring. Finally, the deblurring layer performs a weighted summation of the output data from the normalization layer to obtain the training risk assessment result for the current trainees.

[0242] Step 7: Training the Fuzzy Neural Network (FNN) Model. The backpropagation algorithm is used to iteratively optimize the parameters of the FNN model, with mean squared error as the loss function, and the Adam optimization algorithm is used to further optimize the model parameters.

[0243] Step 8: Assess Training Risk: Input the normalized data into the pre-trained fuzzy neural network model (FNN) to calculate the training risk assessment result for the current trainee. The fuzzy neural network model (FNN) can handle uncertainty and fuzziness, thus deriving the final training risk level.

[0244] This process encompasses key steps such as data processing, model building, training, and evaluation to achieve a comprehensive assessment of the training risk level of participants. The use of fuzzy neural networks helps to handle uncertainty and ambiguity, providing more comprehensive risk assessment results.

[0245] Step S34: Application analysis of model results.

[0246] By acquiring surface electromyography (EMG) signals and vital signs data (heart rate, blood oxygen saturation, body temperature, respiration, and blood pressure) and inputting them into this deep learning-based military physical training risk assessment model, the training risk level of trainees can be directly obtained. This ensures the safety of trainees and reduces risks while providing real-time guidance and decision support for military physical training.

[0247] This solution can be widely applied in military physical training and health monitoring. Here are some application scenarios and examples:

[0248] (1) Military physical training risk assessment: This method can be applied to military training to assess the training risks of trainees in a timely and accurate manner. By monitoring the muscle fatigue and vital signs data of trainees in real time and combining them with deep learning models for comprehensive analysis, it can help instructors and military commanders better understand the training status of each trainee, thereby making appropriate adjustments and decisions and reducing potential risks in training;

[0249] (2) Health management and rehabilitation training: This method can also be used for individual health management and rehabilitation training. By continuously monitoring vital signs such as heart rate, blood oxygen, and body temperature, and assessing muscle fatigue, individuals can understand their own physical condition and adjust their lifestyle, exercise plan, or rehabilitation program in a timely manner to promote health and recovery;

[0250] (3) Optimization of sports training: In addition to the military field, this method can also be applied to sports training. Coaches can use this method to monitor athletes' vital signs and muscle fatigue, and develop more effective training plans and adjustment strategies to improve training effectiveness and competitive performance;

[0251] (4) Emergency situation monitoring: In special environments, such as high-altitude areas, polar environments or remote expeditions, this method can be used to monitor the physical condition of participants, predict potential physical adaptation problems or health risks, and provide timely early warning and support for emergency situations.

[0252] In summary, this deep learning-based risk assessment method for military physical training can be applied in multiple fields to help achieve individualized training and health management, provide real-time risk assessment and early warning, thereby improving the effectiveness and safety of physical training and health monitoring.

[0253] Example 2:

[0254] Another embodiment of the present invention discloses a military physical training risk assessment system based on deep learning, thereby realizing the military physical training risk assessment method based on deep learning in Embodiment 1.

[0255] The specific implementation methods of each module are described in the corresponding description in Example 1.

[0256] A deep learning-based risk assessment system for military physical training utilizes deep learning technology combined with surface electromyography (EMG) signals and vital sign data to accurately assess the training risks of trainees in real time, providing valuable decision support for instructors and military commanders. Figure 8 As shown, the system includes a data acquisition and preprocessing module M1, a muscle fatigue assessment module M2, and a risk assessment module M3.

[0257] They are respectively:

[0258] (1) Data acquisition and preprocessing module M1: used to acquire surface electromyography signals and perform preprocessing, extract time-domain, frequency-domain and CNN features from the preprocessed surface electromyography signals, and perform feature selection and combination on the extracted features to obtain the feature sequence after feature selection and combination.

[0259] (2) Muscle fatigue assessment module M2: used to assess the muscle fatigue of the current trainee at the current moment based on the feature sequence obtained by feature screening combination;

[0260] (3) Risk assessment module M3: It is used to acquire and process physical signs data. Based on muscle fatigue and physical signs data after removing outliers, it is normalized and then assessed to obtain the final training risk assessment result.

[0261] By integrating the above modules, this deep learning-based military physical training risk assessment system can achieve real-time and accurate risk assessment, providing scientific basis and decision support for military physical training.

[0262] Since the system in this embodiment and the method in Embodiment 1 are related and can be referenced from each other, this description is redundant and will not be repeated here. Because this system embodiment shares the same principle as the above method embodiment, it also possesses the corresponding technical effects of the above method embodiment.

[0263] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0264] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for risk assessment of military physical training based on deep learning, characterized in that, Includes the following steps: The surface electromyography (EMG) signals of the current trainees are acquired and preprocessed. Time-domain, frequency-domain, and CNN feature extraction are performed on the preprocessed EMG signals. The extracted features are then filtered and combined to obtain the feature sequence after feature filtering and combination. Based on the feature sequence obtained by feature selection and combination, input it into the muscle fatigue assessment model to obtain the muscle fatigue level of the current trainee at the current moment. Acquire and process vital sign data. After normalizing the muscle fatigue level and vital sign data after removing outliers, input the data into the risk assessment model to assess the risk level and obtain the training risk assessment result of the current trainee. Acquiring and preprocessing surface electromyography signals includes: First, the surface electromyography signal is subjected to SG filtering; Wavelet threshold denoising is performed on the surface electromyography signal after SG filtering; The wavelet threshold denoising includes: The Symlets-2 wavelet basis function was used to perform a four-level orthogonal decomposition of the SG-filtered surface electromyography (EMG) signal. The wavelet coefficients on each decomposition level were calculated using the SG-filtered EMG signal as the function input. Wavelet coefficients of each decomposition layer are filtered to retain wavelet coefficients with lower amplitudes. The wavelet coefficients are selected using the following formula: in, For wavelet coefficient selection values, This represents the wavelet coefficients, where k is the position at scale j. For the scale j The threshold below; j Indicates the number of floors. This represents the wavelet coefficient array of this layer. N The data length of the surface electromyography signal after SG filtering; Based on the retained wavelet coefficients, the surface electromyography signal is reconstructed by inverse wavelet transform to obtain the surface electromyography signal after wavelet threshold denoising. The process of extracting time-domain, frequency-domain, and CNN features from the preprocessed surface electromyography (EMG) signals includes: Based on the surface electromyography signal after wavelet threshold denoising, the data segmentation of the surface muscle signal is obtained by using a sliding window. Based on the surface electromyography signals after data segmentation, a time-domain feature set is extracted; Perform a Fast Fourier Transform on the segmented surface electromyography (EMG) signals to extract the frequency domain feature set; The surface electromyography signal after data segmentation was decomposed into four-level wavelet coefficients using the db4 basis function to extract the time-frequency domain feature set. The surface electromyography signal after wavelet threshold denoising is used as the input of the CNN model. The data output of the penultimate Dense layer with a size of 16 of the CNN model is extracted to form the CNN feature set.

2. The risk assessment method according to claim 1, characterized in that, The feature selection and combination of the extracted features to obtain the feature sequence after feature selection and combination includes: Based on the time domain, frequency domain, and time-frequency domain feature sets and each feature sequence in the CNN feature set, each feature sequence in the feature set is decomposed into two parts by the moving average method, including the stationary information of the feature and the random part of the feature; The stationary information and random part of each feature sequence in the feature set are normalized. Based on the stationary information and random components of the normalized features, correlation, monotonicity, and robustness indices are used for evaluation, resulting in the following comprehensive evaluation index: Where J is the comprehensive evaluation index. These represent the weights for correlation, monotonicity, and robustness, respectively, with Corr, Mon, and Bob being the correlation, monotonicity, and robustness indices, respectively. Select the n features with the highest comprehensive evaluation index to obtain the feature sequence of the feature selection combination.

3. The risk assessment method according to claim 2, characterized in that, The muscle fatigue assessment model is trained based on an LSTM model, and the training process includes: Based on the feature sequences obtained during the training phase, the fatigue sample labels are labeled for each feature sequence according to the Borg scale. The feature sequence and corresponding sample labels of the feature selection combination are input into the LSTM model for training to obtain the muscle fatigue assessment model. The muscle fatigue assessment model outputs a muscle fatigue value.

4. The risk assessment method according to claim 3, characterized in that, The acquisition and processing of vital sign data are as follows: Zero values ​​in the vital signs data are excluded to obtain the filtered vital signs data; For the filtered vital sign data, outlier detection is performed to obtain vital sign data after removing outliers, including: Based on the filtered vital sign data, n sample data are selected as a sample subset, and an isolation tree is recursively constructed starting from the root node. Randomly select a vital sign data point as a threshold, and divide the isolated tree into two branches. Data points below the threshold are placed on the left branch, and the rest are placed on the right branch. When an isolated tree is growing, if a leaf node of a certain branch contains only one symptom data, that symptom data is an outlier. When the isolated tree reaches the set height, the tree stops growing. Outliers in the vital signs data are then removed, resulting in vital signs data with outliers removed.

5. The risk assessment method according to claim 4, characterized in that, The risk assessment model is based on training of a fuzzy neural network (FNN) model. The muscle fatigue value and the vital sign data with outliers removed are labeled with risk values, and after normalization, they are used as input nodes of the input layer of the fuzzy neural network model FNN. Fuzzy conditions are defined in the fuzzification layer, and a Gaussian function is used as the membership function to calculate the fuzzy conditional membership degree of each input node. The number of nodes in the fuzzy rule layer is equal to the number of fuzzy rules; this layer has 2. 6 There are 64 fuzzy rules; The normalization layer normalizes the output data of the fuzzy rule layer; The deblurring layer performs a weighted summation of the output data from the normalization layer to obtain the training risk assessment result for the current trainees.

6. The risk assessment method according to any one of claims 1-5, characterized in that, The vital signs data include heart rate, blood oxygen, body temperature, respiration, and blood pressure.

7. A military physical training risk assessment system based on deep learning, characterized in that, include: The data acquisition and preprocessing module is used to acquire surface electromyography (EMG) signals and perform preprocessing. It performs time-domain, frequency-domain, and time-frequency-domain feature extraction and CNN feature extraction on the preprocessed surface EMG signals. It then performs feature filtering and combination on the extracted features to obtain a feature sequence after feature filtering and combination. The muscle fatigue assessment module is used to assess the muscle fatigue level of the current trainee at the current moment based on the feature sequence obtained by feature screening and combination. The risk assessment module is used to acquire and process vital sign data. Based on muscle fatigue and normalized vital sign data after removing outliers, the training risk assessment result is obtained. Acquiring and preprocessing surface electromyography signals includes: First, the surface electromyography signal is subjected to SG filtering; Wavelet threshold denoising is performed on the surface electromyography signal after SG filtering; The wavelet threshold denoising includes: The Symlets-2 wavelet basis function was used to perform a four-level orthogonal decomposition of the SG-filtered surface electromyography (EMG) signal. The wavelet coefficients on each decomposition level were calculated using the SG-filtered EMG signal as the function input. Wavelet coefficients of each decomposition layer are filtered to retain wavelet coefficients with lower amplitudes. The wavelet coefficients are selected using the following formula: in, For wavelet coefficient selection values, This represents the wavelet coefficients, where k is the position at scale j. For the scale j The threshold below; j Indicates the number of floors. This represents the wavelet coefficient array of this layer. N The data length of the surface electromyography signal after SG filtering; Based on the retained wavelet coefficients, the surface electromyography signal is reconstructed by inverse wavelet transform to obtain the surface electromyography signal after wavelet threshold denoising. The process of extracting time-domain, frequency-domain, and CNN features from the preprocessed surface electromyography (EMG) signals includes: Based on the surface electromyography signal after wavelet threshold denoising, the data segmentation of the surface muscle signal is obtained by using a sliding window. Based on the surface electromyography signals after data segmentation, a time-domain feature set is extracted; Perform a Fast Fourier Transform on the segmented surface electromyography (EMG) signals to extract the frequency domain feature set; The surface electromyography signal after data segmentation was decomposed into four-level wavelet coefficients using the db4 basis function to extract the time-frequency domain feature set. The surface electromyography signal after wavelet threshold denoising is used as the input of the CNN model. The data output of the penultimate Dense layer with a size of 16 of the CNN model is extracted to form the CNN feature set.

Images