2.1 Evaluation method of muscle fatigue in running
The measurement method of bioelectrical signal sEMG (surface myoelectricity) is a comprehensive superposition of the electrical signals emitted by the relevant motor units during muscle contraction at the electrode in time and space, which can objectively and quantitatively reflect neuromuscular activities to a certain extent. sEMG’s evaluation of muscle fatigue is mainly divided into two methods: time domain analysis and frequency domain analysis [9, 10].
(1) Time domain evaluation method. Time domain analysis refers to the characteristics of the amplitude changes of the EMG signal in the time dimension. Its indicators include RMS (root mean square amplitude), iEMG (integrated EMG), and other indicators such as MA (average amplitude) [11, 12]. RMS can reflect the average change characteristic of EMG signal in a period of time. A large number of studies believe that it is related to the recruitment of motor units and the synchronization of excitable rhythms [13, 14]. Its definition is:
$$ RMS=\sqrt{\frac{1}{N}\sum \limits_{i=1}^N{x_i}^2} $$
(1)
This represents the time series of EMG signal of length N. The change of EMG signal amplitude is affected by many factors, such as the number of sports potential recruitment, sports potential release frequency, action potential transmission speed, and the distance between muscle fiber and sensor; the strength of the muscle will also affect the RMS; the greater the muscle strength, the larger the RMS; these factors will affect the accuracy of the RMS [15, 16]. iEMG is the sum of the area of the muscle electrical signal in the unit time after rectification and filtering [17]. It can characterize the strength and weakness of the myoelectric signal over time. It is important for fatigue evaluation. It is one of the means [18, 19]. Its definition is:
$$ {X}_{iEGM}=\frac{1}{N}\sum \limits_{i=1}^N\left|{x}_i\right| $$
(2)
Within a certain period of time, the more the unit’s motor potential and the discharge amount of each sports unit, the larger the iEMG value, and the response to muscle fatigue is very specific. MA is commonly used to characterize the discharge level of muscles within a certain period of time, reflecting that the strength of EMG signal is related to the number of motor units involved and the degree of synchronization of discharge frequency [20, 21].
(2) Frequency domain evaluation method. Frequency analysis refers to the analysis of EMG signals from the frequency of signal changes. It has better stability than the time domain indicators of EMG signals. The frequency domain indicators are obtained by FFT (fast Fourier transform) [22]. The indexes of frequency domain analysis mainly include MF (median frequency) and MPF (average power frequency) [23].
MF refers to the intermediate value of the muscle contraction discharge frequency, which is mainly affected by the composition ratio of the fast muscle and the slow muscle of the muscle tissue. When the fast muscle is excited, the signal spectrum moves to a high frequency, and when the slow muscle is excited, the signal spectrum moves at a low frequency. Its definition is:
$$ MF=\frac{1}{2M}\sum \limits_{j=1}^M{p}_j $$
(3)
The changes in the EMG signal spectrum during fatigue are related to factors such as muscle fiber conduction velocity and the duration of the motor unit potential waveform. As the degree of fatigue increases, the EMG signal spectrum will shift and the MF value will gradually decrease. MF is stable, has strong anti-interference ability, and has a wide range of load levels. It is widely used in the monitoring and analysis of surface EMG signals. MPF represents the frequency of the center of gravity of the power spectrum curve and is also a commonly used method in frequency domain analysis, but the response to time series changes is less consistent with MF. Therefore, only MF is selected as an indicator of frequency domain analysis.
2.2 BP neural network training and testing
A successful BP neural network model is not only reflected in the rationality of the neural network layer setting, but also in the practicality of solving the problem. Only the trained and tested network can meet the actual application requirements.
The structure of the BP neural network determines the nature and performance of the BP neural network. Its structure refers to the number of layers and the number of neurons in the BP neural network. The number of input layer and output layer neurons in the neural network is determined according to the actual problem itself, and the number of hidden layer neurons is variable. The determination of the structure of the BP neural network must first determine the input layer of the BP neural network according to the actual problem and the number of neurons in the output layer, and then determine the number of hidden layers and the number of hidden layer neurons. As long as the BP neural network contains enough hidden layers and hidden layer neurons, the BP neural network can approximate almost any function, but it may affect the performance of the BP neural network. With the increase of the number of hidden layers and neurons, the ability of the neural network to process information will be enhanced, but the transmission speed of the input signal from the input layer to the output layer will be slower, and the correction of the weights and thresholds will become more because of complexity; the speed of backpropagation becomes slower. It can be explained using the Occam’s razor principle, that is, as long as there is a simpler network that works, do not use a more complicated network.
By collecting data from similar projects in the early stage for training, the weights are constantly adjusted to meet the error requirements of the BP neural network. Combined with the approximate degree of the test data results to verify the feasibility of the network, a more mature BP neural network is finally obtained. The flow of BP network work is shown in Fig. 1.
2.3 Fatigue feature value extraction
As the two main biological electrical signals of the human body, surface myoelectric signals and electrocardiographic signals contain a large amount of useful physiological information of the human body. The important task of fatigue detection technology is to extract effective and reliable characteristic indexes from the collected physiological signals for fatigue detection. The quality of feature extraction directly determines the accuracy of multi-class recognition. Therefore, it is necessary to select a feature index that represents the essence of fatigue and changes in feature differences. ECG signals are electrical signals that reflect heart activity, and can objectively reflect different physiological states of the human body. The human body is an organism with a complex structure and function. To evaluate the physiological state of the human body, a comprehensive analysis of multiple information parameters is required to realize the feature extraction evaluation method of multi-signal acquisition.
First, the time domain analysis method is used to extract the characteristics of the ECG signal and the EMG signal, and the ECG signal extracts the time domain indicators such as heart rate and heart rate variability. EMG signals are extracted from time-domain indicators such as root mean square (RMS), reshaping average (ARV), and integrated EMG (IEMG). The frequency domain analysis method is to convert the information in the time domain to the information in the frequency dimension through Fourier transform. The ECG signal extracts high-frequency power (HF, 0.15-0.4 Hz), low-frequency power (HF, 0.04-0.15 Hz), Frequency domain indicators such as the ratio of low frequency power to high frequency power (LF/HF), EMG signals are extracted from frequency domain indicators such as mean power frequency (Mean Power Frequency: MPF) and median frequency (Media Frequency: MF). Frequency domain analysis can reflect the frequency spectrum, power spectrum and other information of the signal in the frequency dimension.
This paper separately extracts features from the time domain and frequency domain, analyzes the changing trends of different characteristics of ECG and EMG signals under different fatigue states of the human body, and selects through experiments to change the features with large differences, good classification effect, and can reflect the movement state Substantial features are input as multiple classifiers. Using the multi-level structure and learning ability of Long-Short-Term Memory Neural Network (LSTM), these features are learned to find the inherent relationship between these features and the state of movement to find indicators with large differences for exercise fatigue judgment.
2.4 BP neural network principle algorithm
BP neural network (backpropagation neural network) is an artificial neural network based on BP algorithm, which uses BP algorithm to adjust the weight and threshold. In the 1980s, several different scholars developed backpropagation algorithms for training multilayer perceptrons. The backpropagation algorithms proposed by David Rumelhart and James McClelland are the most influential. It contains the two main processes of BP, namely the forward propagation of the working signal and the backward propagation of the error signal. The backward propagation of the error signal is to modify the weights and thresholds layer by layer from back to front, in order to make the actual output closer to the expected output.
Input p learning samples, which are represented by x1, x2, …xp, …. After inputting the pth sample to the network, the output yjp(J = 1, 2, …m) is obtained. Using the squared error function, the error Ep of the pth sample is obtained.
$$ {E}_p=\frac{1}{2}\sum \limits_{j-1}^m{\left({t}_j^p-{y}_j^p\right)}^2 $$
(4)
Use the cumulative error BP algorithm to adjust wjk to make the global error E smaller, that is
$$ \Delta {w}_{jk}=-n\frac{\partial E}{\partial {w}_{jk}}=-n\frac{\partial }{\partial {w}_{jk}}\left(\sum \limits_{p-1}^p{E}_p\right)=\sum \limits_{p-1}^p\left(-n\frac{\partial {E}_p}{\partial {w}_{jk}}\right) $$
(5)
In the test, n represents the learning rate. Define the error signal as:
$$ {\delta}_{yj}=-\frac{\partial {E}_p}{\partial {S}_j}=-\frac{\partial {E}_p}{\partial {y}_j}.\frac{\partial {y}_j}{\partial {S}_j} $$
(6)
The first of them:
$$ \frac{\partial {E}_p}{\partial {y}_j}=\frac{\partial }{\partial {y}_j}\left[\frac{1}{2}\sum \limits_{j-1}^m{\left({t}_j^p-{y}_j^p\right)}^2\right]=-\sum \limits_{j-1}^m\left({t}_j^p-{y}_j^p\right) $$
(7)
Second section:
$$ \frac{\partial {y}_j}{\partial {S}_j}={f_2}^{\prime}\left({S}_j\right) $$
(8)
The partial differential of the transfer function of the output layer, so:
$$ {\delta}_{yj}=\sum \limits_{j-1}^m\left({t}_j^p-{y}_j^p\right){f_2}^{\prime}\left({S}_j\right) $$
(9)
From the chain theorem;
$$ \frac{\partial {E}_p}{\partial {w}_{jk}}=\frac{\partial {E}_p}{\partial {S}_j}.\frac{\partial {S}_j}{\partial {w}_{jk}}=-{\delta}_{yj}{z}_k $$
(10)
$$ -{\delta}_{yj}{z}_k=-\sum \limits_{j-1}^m\left({t}_j^p-{y}_j^p\right){f_2}^{\prime}\left({S}_j\right).{z}_k $$
(11)
Therefore, the weight adjustment formula of each neuron in the output layer is:
$$ \Delta {w}_{jk}=\sum \limits_{p-1}^p\sum \limits_{j-1}^mn\left({t}_j^p-{y}_j^p\right){f_2}^{\prime}\left({S}_j\right){z}_k $$
(12)
-
(1)
The working signal is propagated forward. The input signal enters from the input layer, enters the hidden layer neurons through synapses, and is transferred to the output layer after the transfer function calculation, and the output signal is calculated and output at the output layer. When the working signal is propagated in the forward direction, the weights and thresholds are fixed, and the state of each layer in the neural network is only related to the net output, weights, and thresholds of the previous layer. If the forward propagation obtains the desired output at the output layer, the learning ends, and the current weights and thresholds are retained.
-
(2)
The error signal propagates in reverse. If the expected output is not obtained after the working signal is propagated in the forward direction, the error signal is calculated for backward propagation, and the difference between the actual output of the BP neural network and the expected output is calculated as the error signal. The output layer propagates layer by layer to the input layer. In this process, the weights and thresholds of this layer are modified every time a layer is propagated forward, thereby propagating forward to the input layer. This process is to make the result of the neural network closer to the expected result. After the forward propagation and the backward propagation, if the error still cannot meet the requirements, the process continues until the error meets the accuracy or meets other set end conditions such as the number of iterations.