A flight student simulation flight training psychological stress data acquisition method combined with virtual reality

Abstract

The application discloses a kind of flight student simulation flight training psychological stress data acquisition methods combined with virtual reality, it is related to machine learning technical field.The technical points of the present application include: collecting the skin electric signal of flight student in virtual reality environment carries out simulation flight training, and carries out feature extraction, as physiological signal feature dataset;Flight information of flight student in high dynamic simulation platform is collected and is used as flight feature dataset;Flight student carries out flight simulation training in virtual reality environment using hand throttle and joystick, and outputs hand throttle and joystick instruction as control signal dataset, while the key information of virtual reality scene is output as, constitute virtual reality scene information dataset;Psychological stress data acquisition model based on deep convolutional neural network is trained using physiological signal feature dataset, flight feature dataset, control signal dataset and virtual reality scene information dataset.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and specifically to a method for acquiring psychological stress data of flight trainees in simulated flight training using virtual reality. Background Technology

[0002] Through a high-dynamic simulation platform combined with virtual reality technology, flight cadets undergo repeated multi-dimensional flight mission simulation training. This hones their ability to respond to emergencies, adapt to unfamiliar environments, enhance their psychological resilience, raise their stress threshold, and strengthen their flying skills. Consequently, their survivability and combat effectiveness in modern warfare will be significantly improved. Utilizing virtual reality technology, a virtual world with diverse visual and auditory information, real-time interaction, and multi-channel perception capabilities creates a novel and modern flight training simulation environment.

[0003] Traditional flight simulation training is mainly conducted outdoors. Due to limitations in conditions and training costs, flight simulators are expensive, difficult to maintain, and their safety is hard to guarantee. Secondly, the scenarios in flight simulation training are limited and do not take into account various special situations and emergencies in flight missions. Furthermore, the psychological state of flight trainees in flight simulation training is not given sufficient attention, resulting in training effectiveness not meeting expectations.

[0004] Early studies monitored changes in skin conductivity by recording the electrodermal response (EDS) on areas such as the forehead and palms. This provided a direct visual representation of the sympathetic nervous system's moderating effect on skin conductivity during changes in psychological states, such as tension and happiness. As a fundamental physiological signal, EDS reflects the activity of the sympathetic nervous system, and its intensity varies with fluctuations in psychological state, thus serving as an indicator of psychological state fluctuations. The skin resistance on the palm surface provides the most direct feedback on psychological state and, due to its high density of sweat glands, is the most sensitive location, making it a common site in psychophysiological studies.

[0005] However, existing methods do not fully extract features from electrodermal signals, thus failing to assess the psychological stress of flight trainees during simulated flight training based on deep convolutional neural networks. Summary of the Invention

[0006] Therefore, this invention proposes a method for acquiring psychological stress data of flight trainees in simulated flight training using virtual reality, in an attempt to solve or at least alleviate at least one of the problems mentioned above.

[0007] A method for acquiring psychological stress data of flight trainees during simulated flight training using virtual reality includes the following steps:

[0008] S1: Collect skin conductance signals from flight trainees during simulated flight training in a virtual reality environment, and extract features to create a physiological signal feature dataset.

[0009] S2: Collect flight information from flight trainees during simulated flight training on a high-dynamic simulation platform, and use it as a flight feature dataset;

[0010] S3: Collect data on flight trainees using manual throttle and joystick for flight simulation training in a virtual reality environment, output manual throttle and joystick commands as a control signal dataset, and output key information of the virtual reality scene to form a virtual reality scene information dataset.

[0011] S4: Train a psychological stress data acquisition model based on a deep convolutional neural network using physiological signal feature datasets, flight feature datasets, control signal datasets, and virtual reality scene information datasets;

[0012] S5: Input the skin conductance signals, flight information and control signals of the new flight trainees during flight simulation training into the trained psychological stress data acquisition model to obtain the psychological stress data of the new flight trainees.

[0013] Furthermore, the skin electrical signals acquired in S1 specifically include:

[0014] Electrodermal signals were collected from flight trainees before, during, and after flight simulation training.

[0015] Furthermore, in S1, the Fast Fourier Transform (FFT) method is used for feature extraction, including frequency domain features and time domain features. The frequency domain features include the mean, median, standard deviation, root mean square deviation, maximum value, minimum value, maximum value ratio, minimum value ratio, and range. The time domain features include the mean, median, standard deviation, maximum value, minimum value, maximum value ratio, minimum value ratio, difference, mean of the first-order difference, median of the first-order difference, standard deviation of the first-order difference, maximum value of the first-order difference, minimum value of the first-order difference, maximum value ratio of the first-order difference, and minimum value ratio of the first-order difference; and the mean of the second-order difference, median of the second-order difference, standard deviation of the second-order difference, maximum value of the second-order difference, minimum value of the second-order difference, maximum value ratio of the second-order difference, and minimum value ratio of the second-order difference.

[0016] Furthermore, in S1, after acquiring the frequency domain features and time domain features, an improved frequency conversion demodulation method is used to construct a time-spectrum feature set of the electrodermal signal. The frequency domain features, time domain features, and the time-spectrum feature set of the electrodermal signal together constitute a physiological signal feature dataset. Specific steps include:

[0017] S11. Downsampling of skin electrical signals;

[0018] S12. Filter the downsampled electrodermal signal using a high-pass filter;

[0019] S13. Decompose the filtered skin electrodermal signal using a low-pass filter, initially dividing the skin electrodermal signal into two frequency band groups; use bandwidth F ω and length N ω The low-pass filters are used for filtering, with the center frequency band being:

[0020]

[0021] The bandwidth of the adjacent center segment is 2F. ω f max The highest frequency in the band;

[0022] S14. Use the complex demodulation method to extract the main frequency of each bandwidth, and increase f in each of the two frequency band groups respectively. 0i Repeated demodulation is performed, and the demodulated electrodermal signal is sinusoidally modulated; the instantaneous frequency of the electrodermal signal is calculated according to the following formula:

[0023]

[0024] Where f0 is the center frequency; φ(t) represents the phase; T is the sampling interval of the processing cycle of the electrodermal signal; and L is the signal length of each processing cycle. ε is the instantaneous phase of the i-th processing cycle; i The weighting coefficient for the i-th processing cycle;

[0025] but

[0026]

[0027] Where x(t) represents the electrodermal signal before demodulation, dc(t) represents the DC component of the electrical signal, and A(t) represents the instantaneous amplitude of the electrical signal;

[0028] S15. For a given center frequency band, calculate the amplitude and phase of the electrodermal signal, and use the Hilbert transform to calculate the phase φ(t) and instantaneous amplitude A(t) of each sinusoidally modulated electrodermal signal component:

[0029]

[0030] A(t) = 2|z 1p (t)|

[0031]

[0032] Among them, z 1p(t) represents the demodulated electrodermal signal, Im represents the imaginary part of the instantaneous electrodermal signal, and Re represents the real part of the instantaneous electrodermal signal;

[0033] S16. Using the estimated phase φ(t) and instantaneous amplitude A(t), plot the high-resolution time spectrum of the electrodermal signal and correct the two initially divided frequency band groups;

[0034] S17. Divide the electrodermal signal into multiple frequency band groups; repeat steps S14-S16 until the time-spectrum characteristics of the electrodermal signal at a 10-fold scale frequency band are obtained.

[0035] Furthermore, the flight information described in S2 includes flight time, aircraft longitude position, aircraft latitude position, aircraft altitude, indicated airspeed, aircraft absolute speed, aircraft yaw angle, roll angle, pitch angle, and overload.

[0036] Furthermore, the virtual reality scenarios described in S3 include routine flight training scenarios, special situation training scenarios, multi-aircraft combat training scenarios, and target practice training scenarios.

[0037] Furthermore, the deep convolutional neural network structure in the psychological stress data acquisition model based on deep convolutional neural networks in S4 includes an input layer, three convolutional layers, three pooling layers, a fully connected layer, a batch normalization layer, and an output softmax layer. The input layer consists of a 4D feature matrix constructed from physiological signal feature datasets, flight feature datasets, control signal datasets, and virtual reality scene information datasets, serving as the input to the deep convolutional neural network. Convolutional layers and pooling layers are connected to form a group, and their training results are output to the next group of convolutional and pooling layers. The fully connected layer is a hidden layer, its function being to reduce the size of the feature matrix and the feature dimension using the backpropagation algorithm. The purpose of the batch normalization layer is to stabilize the training results of the deep convolutional neural network after the fully connected layer and to execute an "exit" instruction in the backpropagation algorithm between fully connected layers, reducing the possibility of overfitting. The output softmax layer is a one-dimensional matrix of the trained feature set.

[0038] Furthermore, in S4, the improved backpropagation algorithm is used to train the connection weights w and biases b of the deep convolutional neural network during model training; according to the Leaky ReLU activation function, the model's loss function is defined as f(w,b):

[0039]

[0040] Where n represents the length of the data to be processed; x k ,y k The input consists of two sets of feature data; γ and υ represent the decay coefficients of the connection weights. This represents the connection weight between the j-th neuron in layer p and the i-th neuron in layer (p+1); n p s represents the data length of the p-th layer; p This represents the number of neurons in the p-th layer; the first term on the right side of the equals sign is the sum of the squared errors between the output of the model's Softmax layer and the target value, and the second term is the regularization part;

[0041] According to the above formula, the gradient term of the p-th layer is expressed as:

[0042]

[0043] Where × represents the inner product operation;

[0044] Update the connection weights w and biases b using a stochastic optimization method; iterate using the following equations:

[0045]

[0046]

[0047] in, and Δw in the k-th iteration p and △b p Value; △w p Δb represents the connection weights of the p-th layer of the neural network. p Ψ represents the bias of the p-th layer of the neural network; p+1 Let g( represent the gradient term of the (p+1)th layer. p) Represents the LeakyReLU function;

[0048] The connection weights and biases are continuously updated before the convergence rate approaches zero.

[0049]

[0050]

[0051] In the formula, ρ represents the learning rate. and These represent w in the (k+1)th iteration. p and b p The value of w, where α is the impulse term. p and b p These represent the connection weights and biases of the p-th layer of the neural network, respectively. These represent w in the k-th iteration. p and b p The value of .

[0052] The beneficial technical effects of this invention are:

[0053] This invention utilizes virtual reality technology to create a realistic and continuous flight environment. Through continuous training in this virtual reality environment, the skin conductance signals of flight trainees are monitored in real time during simulated flight training. An improved frequency conversion demodulation method is used to construct a time-spectrum feature set of the skin conductance signals, enabling accurate and comprehensive feature extraction. Furthermore, an improved deep convolutional neural network training model is used. During training, an improved backpropagation algorithm is used to train the connection weights and biases of the deep convolutional neural network, making the acquired psychological stress data more effective and accurate. This invention can further enhance the flight training effect for flight trainees, improve their flying ability, and enhance their flight skills and flight environment analysis capabilities. Attached Figure Description

[0054] The present invention can be better understood by referring to the description given below in conjunction with the accompanying drawings, which together with the following detailed description are included in and form part of this specification, and are used to further illustrate preferred embodiments of the invention and explain the principles and advantages of the invention.

[0055] Figure 1 This is a flowchart of a method for acquiring psychological stress data of flight trainees in simulated flight training using virtual reality, according to an embodiment of the present invention.

[0056] Figure 2 This is a flowchart illustrating the training process of a deep convolutional neural network in an embodiment of the present invention. Detailed Implementation

[0057] To enable those skilled in the art to better understand the present invention, exemplary embodiments or examples of the present invention will be described below in conjunction with the accompanying drawings. Obviously, the described embodiments or examples are merely some, not all, of the embodiments or examples of the present invention. All other embodiments or examples obtained by those skilled in the art based on the embodiments or examples of the present invention without inventive effort should fall within the scope of protection of the present invention.

[0058] This invention proposes a method for acquiring psychological stress data of flight trainees in simulated flight training using virtual reality, such as... Figure 1 As shown, the method includes the following steps:

[0059] S1: Collect skin conductance signals from flight trainees during simulated flight training in a virtual reality environment, and extract features to create a physiological signal feature dataset.

[0060] S2: Collect flight information from flight trainees during simulated flight training on a high-dynamic simulation platform, and use it as a flight feature dataset;

[0061] S3: Collect data on flight trainees using manual throttle and joystick for flight simulation training in a virtual reality environment, output manual throttle and joystick commands as a control signal dataset, and output key information of the virtual reality scene to form a virtual reality scene information dataset.

[0062] S4: Train a psychological stress data acquisition model based on a deep convolutional neural network using physiological signal feature datasets, flight feature datasets, control signal datasets, and virtual reality scene information datasets;

[0063] S5: Input the skin conductance signals, flight information and control signals of the new flight trainees during flight simulation training into the trained psychological stress data acquisition model to obtain the psychological stress data of the new flight trainees.

[0064] The present invention will now be described in detail.

[0065] In S1, after entering the virtual reality environment, the testers use the high dynamic range simulation platform to select flight training scenarios, collect skin conductance signals, and extract features to form a physiological signal feature dataset.

[0066] According to embodiments of the present invention, the implemented hardware includes virtual reality glasses, a manual throttle and joystick, a skin conductance signal acquisition device, and an algorithm processor. The manual throttle and joystick can be used in conjunction with a high-dynamic flight platform for high-dynamic flight simulation training, or they can be placed on a desktop and used in conjunction with a virtual reality device for daily flight simulation training.

[0067] The electrodermal (EDS) signals were collected from the tips of the index and middle fingers of the flight trainee when operating the manual throttle and joystick, and from the base of the thumb. The EDS signal collection method included: collecting EDS signals from the flight trainee 3 minutes before the simulated flight training; collecting EDS signals from the flight trainee during the simulated flight training; and collecting EDS signals from the flight trainee 3 minutes after the simulated flight training ended. The sampling frequency was 200Hz. Through the above steps, EDS signals were obtained at three different stages.

[0068] Then, the EDS signal was downsampled at a sampling frequency of 8Hz, and motion artifact removal and filtering were performed on the downsampled EDS signal to obtain a clean EDS signal.

[0069] Then, the frequency domain and time domain features of the electrodermal signal are extracted using the Fast Fourier Transform (FFT) method. The frequency domain features include statistical characteristics of the electrodermal signal such as mean, median, standard deviation, root mean square, maximum, minimum, maximum ratio, minimum ratio, and range. The time domain features include mean, median, standard deviation, maximum, minimum, maximum ratio, minimum ratio, and difference; mean, median, standard deviation, maximum, minimum, maximum ratio, minimum ratio, and difference of the first-order difference; mean, median, standard deviation, maximum, minimum, maximum ratio, and minimum ratio of the first-order difference; mean, median, standard deviation, maximum, minimum, maximum ratio, and minimum ratio of the second-order difference; and mean, median, standard deviation, maximum, minimum, maximum ratio, and minimum ratio of the second-order difference.

[0070] Then, using an improved frequency conversion demodulation method, a time-spectral feature set of the electrodermal signal is constructed; the specific steps are as follows:

[0071] S11. Further downsample the skin electrodermal signal and lower the frequency to 4Hz.

[0072] S12. A high-pass filter is used to filter the downsampled electrodermal signal. The purpose is to filter out electrodermal signals above 2Hz and eliminate redundant noise in the electrodermal signal.

[0073] S13. Initially, the electrodermal signal is divided into two frequency bands: [0, 0.40] Hz and [0.4, 1.0] Hz. For these two frequency bands, bandwidth F is used... ω and length N ω The low-pass filters are used for filtering, with the center frequency band being:

[0074]

[0075] The bandwidth of the adjacent center segment is f max It is the highest frequency in the band, and its purpose is to avoid signal interference above 1.0Hz.

[0076] S14. Use the complex demodulation method to extract the main frequency of each bandwidth, and increase f in each of the two frequency band groups respectively. 0i Repeated demodulation is performed, and the demodulated electrodermal signal is sinusoidally modulated; the instantaneous frequency of the electrodermal signal is calculated according to the following formula:

[0077]

[0078] Where f0 is the center frequency; φ(t) represents the phase; T is the sampling interval of the processing cycle of the electrodermal signal; and L is the signal length of each processing cycle. ε is the instantaneous phase of the i-th processing cycle; i The weighting coefficient for the i-th processing cycle;

[0079] but

[0080]

[0081] Where x(t) represents the electrodermal signal before demodulation, dc(t) represents the DC component of the electrical signal, and A(t) represents the instantaneous amplitude of the electrical signal.

[0082] S15. For a given center frequency band of the electrodermal signal, its amplitude and phase can be calculated:

[0083]

[0084] A(t) = 2|z 1p (t)| (5)

[0085]

[0086] Among them, z 1p (t) represents the demodulated electrodermal signal, Im represents the imaginary part of the instantaneous electrodermal signal, and Re represents the real part of the instantaneous electrodermal signal;

[0087] Furthermore, using the Hilbert transform, the phase and instantaneous amplitude of each sinusoidally modulated electrodermal signal component are calculated by equations (5) and (6).

[0088] S16. Using the estimated phase φ(t) and instantaneous amplitude A(t), plot the high-resolution time spectrum of the electrodermal signal and correct the two initially divided frequency band groups;

[0089] S17. Divide the electrodermal signal into multiple frequency band groups; repeat steps S14-S16 until the time-spectrum characteristics of the electrodermal signal at a 10-fold scale frequency band are obtained.

[0090] As an example, the electrodermal signal is divided into four frequency bands: [0, 0.04], [0.04, 0.25] Hz, [0.25, 0.4] Hz, and [0.4, 1.0] Hz. This process is repeated until a sufficient number of frequency bands (e.g., 10 times the scale) of the electrodermal signal's time-spectral characteristics are obtained. Then, using the frequency conversion demodulation method, the characteristics reflecting psychological state fluctuations in the time-spectrum of the electrodermal signal are obtained.

[0091] In S2, a high dynamic simulation platform is used to conduct simulated flight training and output flight information as a flight feature dataset.

[0092] According to an embodiment of the present invention, flight information data includes flight time, aircraft longitude position, aircraft latitude position, aircraft altitude, etc.; flight indication airspeed, aircraft absolute speed, aircraft yaw angle, roll angle, pitch angle and overload. Flight data is output twice per second to form flight datasets for different stages.

[0093] By utilizing real-time flight information data, further extract flight data features, such as the rate of climb during the climb phase, the glide slope and localizer deviation angles during the approach phase, the rate of change of course, the turn rate, the rate of climb, the loop and half-loop, etc., and integrate them into the flight data feature set.

[0094] In S3, flight simulation training is conducted using manual throttle and joystick in a virtual reality context. The manual throttle and joystick commands are output as a control signal dataset, and key information of the virtual reality scene is output as a virtual reality scene information dataset.

[0095] According to embodiments of the present invention, simulated flight training scenarios created using virtual reality technology include routine flight training, emergency training, multi-aircraft combat training, and target practice training. Key information is extracted from these virtual reality scenarios to form a virtual reality scenario information dataset, which describes the characteristics of the virtual reality environment and enhances the robustness of the constructed psychological stress data acquisition model.

[0096] In S4, a psychological stress data acquisition model based on a deep convolutional neural network is trained using physiological signal feature datasets, flight feature datasets, control command datasets, and virtual reality scene information datasets.

[0097] According to an embodiment of the present invention, a deep convolutional neural network is used to identify and evaluate the psychological fluctuations of flight trainees during simulated flight training, thereby obtaining psychological stress data of flight trainees during simulated flight training.

[0098] Deep convolutional neural networks (DCNNs) automatically extract high-dimensional features from four different categories of raw signals and low-level features. Through multi-layer fusion optimization, they extract feature combinations to accurately represent the psychological fluctuations of flight trainees during simulated flight training. The DCNN consists of an input layer, three convolutional layers, three pooling layers, a fully connected layer, a batch normalization layer, and a softmax layer. The input layer comprises a 4D feature matrix constructed from physiological signal feature datasets, flight feature datasets, control signal datasets, and virtual reality scene information datasets, serving as the input to the DCNN. Convolutional and pooling layers are connected as a group, and their training results are output to the next group of convolutional and pooling layers. The fully connected layer is a hidden layer that reduces the size of the feature matrix and the feature dimensionality using backpropagation. The batch normalization layer stabilizes the training results of the DCNN after the fully connected layers and executes an "exit" instruction in the backpropagation algorithm between fully connected layers, reducing the possibility of overfitting. The output layer is a one-dimensional matrix of the trained feature set.

[0099] The training process of a deep convolutional neural network is as follows: Figure 2 As shown, the 4D feature matrices constructed from physiological signal feature datasets, flight feature datasets, control signal datasets, and virtual reality scene information datasets are estimated and merged into a four-dimensional feature matrix, serving as the input to the DCNN. Then, the DCNN is constructed layer by layer, performing convolution calculations and maximum-based subsampling operations to achieve prominent features. This step involves local neural network convolution calculations and weight iterations to reduce the number of training parameters and improve the computational efficiency of the cascaded layers. Next, an improved backpropagation algorithm is used to slightly adjust the connection weights and biases by minimizing the error between the predicted value and the model output. The DCNN training ends when the error approaches a preset threshold. Finally, the trained model is used to evaluate the fluctuations in the psychological state of flight trainees during simulated flight training, i.e., psychological stress data.

[0100] The activation function of a deep convolutional neural network is the Leaky ReLU function, and according to the steepest gradient descent method, its expression is as follows:

[0101]

[0102] λ is defined as an adaptive function.

[0103] Deep convolutional neural networks (CNNs) are trained using an improved backpropagation algorithm to determine the connection weights and biases. Based on the Leaky ReLU activation function, the model's loss function is defined as f(w,b):

[0104]

[0105] Where n represents the length of the data to be processed; x k ,y k The input consists of two sets of feature data; γ and υ represent the decay coefficients of the connection weights. This represents the connection weight between the j-th neuron in layer p and the i-th neuron in layer (p+1); n p s represents the data length of the p-th layer; p This represents the number of neurons in the p-th layer; the first term on the right side of the equals sign is the sum of the squared errors between the output of the model's Softmax layer and the target value, and the second term is the regularization part;

[0106] According to the above formula, the gradient term of the p-th layer is expressed as:

[0107]

[0108] Where × represents the inner product operation. The connection weights w and biases b can be updated using stochastic optimization methods.

[0109] The connection weights and biases are continuously updated using iterative calculations using the following equations:

[0110]

[0111] in, and Δw in the k-th iteration p and △b p Value; △w p Δb represents the connection weights of the p-th layer of the neural network. p Ψ represents the bias of the p-th layer of the neural network; p+1 Let g( represent the gradient term of the (p+1)th layer. p) Represents the LeakyReLU function;

[0112] in and Δw in the k-th iteration p and △b p value.

[0113]

[0114] In the formula, ρ represents the learning rate. and These represent w in the (k+1)th iteration. p and b p The value of , where α is the impulse term, whose main function is to increase the backward search step size and accelerate the convergence speed of the neural network; w p and b p These represent the connection weights and biases of the p-th layer of the neural network, respectively. These represent w in the k-th iteration. p and b p The value of .

[0115] Although the invention has been described with respect to a limited number of embodiments, those skilled in the art will understand from the foregoing description that other embodiments are conceivable within the scope of the invention described herein. The disclosure of the invention is illustrative and not restrictive, and the scope of the invention is defined by the appended claims.

Claims

1. A method for acquiring psychological stress data of flight trainees in simulated flight training combined with virtual reality, characterized in that, Comprising the following steps: S1: Collecting the skin electrical signal of the flight trainee in the virtual reality environment for simulated flight training, and performing feature extraction as a physiological signal feature dataset; wherein the feature extraction is performed by using the fast Fourier transform method, including frequency domain features and time domain features; after obtaining the frequency domain features and time domain features, an improved variable frequency demodulation method is used to construct a skin electrical signal time-frequency spectrum feature set, and the frequency domain features, time domain features and skin electrical signal time-frequency spectrum feature set jointly constitute the physiological signal feature dataset; the specific steps include: S11, downsampling the skin electrical signal; S12, filtering the downsampled skin electrical signal by using a high-pass filter; S13, decompose the filtered skin electricity signal by using low-pass filter, initially divide the skin electricity signal into two frequency band groups; filter by using low-pass filter with bandwidth and length , wherein the center frequency band is: The bandwidth of the adjacent center segment is 2 , is the highest frequency of the frequency band. S14, using complex demodulation method to extract the main frequency of each bandwidth, and through increasing the The skin conductance signal is repeatedly demodulated and sinusoidally modulated. The instantaneous frequency of the skin conductance signal is calculated according to the following formula: wherein, is a center frequency; denotes a phase; T is a sampling interval of a processing period of the skin electrical signal; L is a signal length per processing period; is an instantaneous phase of the i-th processing period; is a weighting coefficient of the i-th processing period; S17, dividing the skin electrical signal into multiple frequency band groups; repeating steps S14-S16 until the skin electrical signal time-frequency spectrum feature of 10 times the scale frequency band is obtained; where x(t) represents the electrodermal signal before demodulation, represents the direct current component in the electrical signal, represents the instantaneous amplitude of the electrical signal; S15, for a given central frequency band of the electrodermal signal, compute its amplitude and phase, and use the Hilbert transform to compute the phase of each sinusoidal modulated electrodermal signal component and instantaneous amplitude : wherein, denotes the demodulated skin conductance signal, Im denotes the imaginary part of the instantaneous skin conductance signal, Re denotes the real part of the instantaneous skin conductance signal; S16, using the estimated phase and instantaneous amplitude Plotting the high-resolution electrodermal signal's frequency spectrum, revising the initially divided two frequency bands groups; S2: Collecting the flight information of the flight trainee in the high dynamic simulation platform for simulated flight training as a flight feature dataset; S3: Collecting the flight simulation training of the flight trainee in the virtual reality environment by using the manual throttle and joystick, outputting the manual throttle and joystick instructions as a control signal dataset, and simultaneously outputting the key information of the virtual reality scene as a virtual reality scene information dataset; S4: Training a psychological stress data acquisition model based on a deep convolutional neural network by using the physiological signal feature dataset, the flight feature dataset, the control signal dataset and the virtual reality scene information dataset; S5: Inputting the skin electrical signal, flight information and control signal of the new flight trainee during flight simulation training into the trained psychological stress data acquisition model to obtain the psychological stress data of the new flight trainee. The skin electrical signal collected in S1 specifically includes: collecting the skin electrical signal of the flight trainee before, during and after the simulated flight training, respectively.

2. The method according to claim 1, wherein, The frequency domain features in S1 include the mean, median, standard deviation, mean square error, maximum value, minimum value, maximum value ratio, minimum value ratio, and range of the frequency domain; the time domain features include the mean, median, standard deviation, maximum value, minimum value, maximum value ratio, minimum value ratio, difference value, mean value of first-order difference, median of first-order difference, standard deviation of first-order difference, maximum value of first-order difference, minimum value of first-order difference, maximum value ratio of first-order difference, minimum value ratio of first-order difference; mean value of second-order difference, median of second-order difference, standard deviation of second-order difference, maximum value of second-order difference, minimum value of second-order difference, maximum value ratio of second-order difference, minimum value ratio of second-order difference.

3. The method of claim 1, wherein the method further comprises: The flight information in S2 includes flight time, aircraft longitude position, aircraft latitude position, aircraft altitude, flight indicated airspeed, aircraft absolute speed, aircraft yaw angle, roll angle, pitch angle and overload.

4. The method of claim 1, wherein the method further comprises: The virtual reality scene in S3 includes a normalized flight training scene, a special situation training scene, a multi-aircraft combat training scene and a shooting training scene.

5. The method of claim 1, wherein the method further comprises: ​ 6. The method of claim 1, wherein the method further comprises: The structure of the deep convolutional neural network in the psychological stress data acquisition model in S4 comprises an input layer, three convolutional layers, three pooling layers, a fully connected layer, a batch normalization layer and an output Softmax layer; wherein the input layer is composed of a 4D feature matrix constructed by a physiological signal feature dataset, a flight feature dataset, a control signal dataset and a virtual reality scene information dataset, serving as the input of the deep convolutional neural network; the convolutional layer and the pooling layer are connected to form a group, and the training result of the group is output to the next group of convolutional layer and pooling layer; the fully connected layer is a hidden layer, which is used to reduce the size of the feature matrix and the feature dimension by using the back propagation algorithm; the purpose of the batch normalization layer is to stabilize the training result of the deep convolutional neural network after the fully connected layer, and to execute the "exit" instruction in the back propagation algorithm between the fully connected layers to reduce the possibility of overfitting; and the output Softmax layer is a one-dimensional matrix of the feature set obtained by training.

7. The method of claim 6, wherein the method further comprises: In the model training process in S4, the improved back propagation algorithm is used to train the connection weights w and the bias b of the deep convolutional neural network; according to the activation function of Leaky ReLU, the loss function of the model is defined as : wherein n represents the data length to be processed; represents two sets of input feature data; and represents the decay coefficient of the connection weight; represents the connection weight between the jth neuron of the pth layer and the ith neuron of the (p+1)th layer; represents the data length of the pth layer; represents the number of neurons of the pth layer; the first term on the right side of the equal sign is the sum of squared errors between the model Softmax layer output and the target value, and the second term is the regularization part; According to the above formula, the gradient term of the pth layer is represented as: wherein denotes an inner product operation; The connection weight w and the bias b are updated using a random optimization method; and are iteratively calculated using the following equation: wherein, and are the values of and at the kth iteration, respectively; represents the connection weights of the pth layer neural network, represents the bias of the pth layer neural network; represents the gradient term of the (p+1)th layer, represents the LeakyReLU function; The connection weight w and the bias b are constantly updated before the convergence speed approaches 0: wherein, denotes the learning rate, and denote the values of and at the (k+1)th iteration, is the momentum term, and denote the connection weights and biases of the pth layer neural network, , denote the values of and at the kth iteration.

Images