Physiological feedback-based psychological adaptation capacity intelligent regulation system and method

Abstract

The application discloses a kind of psychological adaptive capacity intelligent regulation and control system and method based on physiological feedback, it is related to psychological adaptive capacity regulation and control technical field;By acquiring multi-modal physiological signals, ship physical environment parameters and post task load parameters and constructing space-time synchronous state matrix;Adaptive filtering is used to strip ship sway interference to generate purified physiological feature vector;Through space-time attention state evaluation model, real-time psychological load index is output;When the index exceeds dynamic early warning threshold, based on reinforcement learning, match initial adaptive training scheme;During execution, calculate actual breathing instantaneous phase and physiological feedback deviation rate using hilbert transform, when exceeding limit, through variable step optimization gradient dynamic update guide parameter to establish physiological resonance and smooth traction;According to the recovery rate of communication operation efficiency, update action value matrix;The application eliminates environmental interference, realizes closed-loop dynamic regulation and control of physiological phase and strategy self-evolution.

Description

Technical Field

[0001] This invention belongs to the field of psychological adaptability regulation technology, and specifically relates to an intelligent regulation system and method for psychological adaptability based on physiological feedback. Background Technology

[0002] With the development of ocean-going navigation technology, the task of ensuring communication for ships during long voyages and in complex sea conditions is becoming increasingly demanding. Ship communication personnel are often in confined spaces, exposed to high-decibel engine noise, and subjected to violent rocking, while simultaneously handling high-frequency, high-security message transmission, reception, and decryption. This combination of multiple physical stressors and high cognitive load can easily lead to psychological overload, excessive activation of the sympathetic nervous system, and accumulation of negative emotions among the personnel.

[0003] Existing psychological adjustment and stress relief techniques are mostly applied in static indoor environments, primarily through playing relaxing music or white noise with fixed rhythms, or providing standardized breathing guidance videos for training. In practice, the system typically triggers a simple threshold based on a single physiological indicator (such as heart rate), and then executes a preset static training program, requiring the trainee to passively follow the given rhythm to adjust.

[0004] This reveals the following problems with existing technologies: 1. They do not consider the direct interference of physical environments such as ship roll and pitch on the human vestibular nerve and physiological signals, leading to inaccurate psychological state assessments; 2. Adaptive training programs are fixed and lack a micro-closed-loop feedback mechanism between training guidance rhythms and real-time physiological phases, which can easily lead to rhythmic discontinuity and psychological resistance when personnel are under high stress; 3. They lack quantitative assessment of the recovery of actual job performance after training and the ability for strategy self-evolution. Summary of the Invention

[0005] (a) Technical problems to be solved In view of the problems in related technologies, the present invention provides an intelligent regulation system and method for psychological adaptability based on physiological feedback, so as to overcome the above-mentioned technical problems existing in the existing related technologies.

[0006] (II) Technical Solution To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: On the one hand, this invention provides an intelligent regulation method for psychological adaptability based on physiological feedback, comprising the following steps: S1. Obtain the multimodal physiological signal sequence, job task load parameter sequence, and ship physical environment parameter sequence of the trainees; S2. Based on the ship's physical environment parameter sequence, perform motion artifact removal and vestibular compensation heart rate subtraction calculation on the multimodal physiological signal sequence to generate purified physiological feature vectors. The purified physiological feature vector and the job task load parameter sequence are input into the spatiotemporal attention state assessment model, and the real-time psychological load index and job stress state classification label are output. S3. Determine whether the real-time psychological load index is greater than the dynamic baseline warning threshold. If so, input the real-time psychological load index and the occupational stress state classification label into the reinforcement learning-based regulation strategy generation model to obtain the initial adaptive training scheme and the initial training guidance parameter set; Otherwise, if the trainee's mental state is deemed normal, proceed to step S1. S4. Execute the initial adaptive training program and continuously determine whether the real-time physiological feedback deviation rate of the trainees is greater than the phase lock threshold. If so, based on the gradient direction of the real-time physiological feedback deviation rate, the variable step size optimization algorithm is used to update the guidance frequency parameter and stimulus intensity parameter in the initial training guidance parameter set in real time, and generate a dynamic control instruction sequence to be sent to the execution terminal; otherwise, the initial training guidance parameter set is kept unchanged, and execution continues until the training time ends. S5. Extract the job task load parameter sequence of trainees after the training period ends, and calculate the communication operation efficiency recovery rate; based on the communication operation efficiency recovery rate and the decline of physiological indicators during the training period, update the action value matrix of the generation model through the reward function and the regulation strategy. This invention acquires multimodal physiological signal sequences, job task load parameter sequences, and ship physical environment parameter sequences from trainees. It utilizes adaptive filtering technology to remove motion artifacts and seasickness compensation responses, and combines this with a spatiotemporal attention state assessment model to accurately output a real-time psychological load index. When psychological load exceeds limits, it matches an initial adaptive training scheme based on reinforcement learning. During execution, it calculates the real-time physiological feedback deviation rate and dynamically updates the training guidance parameters using a variable step-size optimization algorithm, forming a microsecond-level closed-loop control. The action value matrix is ​​updated based on the communication operation efficiency recovery rate. This eliminates interference from complex sea conditions, achieves personalized matching and dynamic adaptive adjustment of training strategies, and improves the accuracy and effectiveness of psychological regulation.

[0007] Preferably, the specific steps for constructing the multidimensional spatiotemporal synchronization state matrix in S1 include: S11. Collect electrocardiogram signals, skin conductance level signals, and respiratory rhythm signals from trainees to obtain multimodal physiological signal sequences; S12. Collect the ship's pitch acceleration sequence, ship's roll angular velocity sequence, and cabin background noise sound pressure level sequence to obtain the ship's physical environment parameter sequence. S13. Statistically analyze the number of effective keystrokes, message retransmission frequency, and error correction backspace times of trainees within a unit of time to obtain the job task load parameter sequence. S14. Using the absolute time of the network time protocol server as a reference, the cubic spline interpolation algorithm is used to resample the multimodal physiological signal sequences, ship physical environment parameter sequences and job task load parameter sequences with different sampling rates to a unified reference frequency, so as to obtain a multidimensional spatiotemporal synchronization state matrix. This invention acquires physiological signals through dry electrodes, skin conductance, and strain sensors, obtains environmental parameters through an inertial measurement unit and a decibel meter, and combines the underlying driver interface to statistically analyze the task load. Relying on the absolute timestamp of the network time protocol server and the cubic spline interpolation algorithm, it resamples heterogeneous multi-source data with different sampling rates to a unified reference frequency. It eliminates the time phase difference of multimodal data and constructs a high-dimensional, spatiotemporally aligned synchronization state matrix, providing a reliable data foundation for subsequent high-precision fusion calculations.

[0008] Preferably, the specific steps in S2 for generating the purified physiological feature vector include: S211. Extract the ship pitch acceleration sequence from the aligned ship physical environment parameter sequence as the reference input signal, and extract the electrocardiogram signal from the aligned multimodal physiological signal sequence as the main input signal. S212. Construct a normalized minimum mean square adaptive filter and calculate the error signal between the main input signal and the reference input signal after processing by the filter weight vector. S213. Iteratively update the filter weight vector based on the error signal. When the filter weight vector converges, extract the continuous error signal as the pure electrocardiogram signal stripped of seasickness compensation response. S214. Detect the peak value of the R wave in the pure electrocardiogram signal, calculate the time interval sequence of adjacent R waves, and use an autoregressive model to estimate the power spectral density of the time interval sequence to obtain the sympathetic vagus nerve balance index. Calculate the peak value of the phase awakening wave of the skin conductance level signal in the multimodal physiological signal sequence, and combine the sympathetic vagal nerve balance index with the peak value of the phase awakening wave to form a purified physiological feature vector; This invention uses the ship's pitch acceleration as a reference input to construct a normalized least mean square adaptive filter to process the electrocardiogram (ECG) signal. It iterative convergence is used to remove compensatory heart rate fluctuations caused by ship rolling. An autoregressive model is used to estimate the power spectral density of the RR interval of the pure ECG signal to calculate the sympathetic-vagal balance index, and the phase arousal peak value of skin conductance is combined. This effectively filters out physical noise interference caused by vestibular dysfunction and extracts physiological feature vectors that reflect the pure psychological stress state.

[0009] Preferably, the specific steps in S2 for outputting the real-time psychological load index and occupational stress state classification labels of trainees include: S221. Normalize the purified physiological feature vector and the job task load parameter sequence, and concatenate them into a fused feature tensor. S222. Input the fused feature tensor into the spatiotemporal attention state evaluation model, and extract spatial dimension features from the fused feature tensor through a one-dimensional convolutional layer to generate a local physiological response feature map. S223. The long short-term memory network layer of the spatiotemporal attention state assessment model receives local physiological response feature maps, extracts the cumulative evolution features of psychological stress response in the time dimension, and calculates the hidden layer state sequence. S224. The self-attention mechanism layer of the spatiotemporal attention state evaluation model assigns attention weights to the hidden layer state sequence to generate a weighted hidden layer state sequence. The weighted hidden layer state sequence is input into the fully connected layer; the regression branch of the fully connected layer outputs a continuous real-time psychological load index, and the classification branch of the fully connected layer outputs the job stress state classification label through the Softmax function; This invention fuses purified physiological feature vectors with job task load parameter sequences and inputs them into a spatiotemporal attention state assessment model. It extracts spatial nonlinear mapping features through a one-dimensional convolutional layer, extracts the temporal cumulative evolution features of psychological stress using a long short-term memory network layer, and assigns higher weights to key mutation nodes by combining a self-attention mechanism. This achieves spatiotemporal joint feature extraction of physiological reactions and task load, improving the accuracy of real-time psychological load index regression calculation and job stress state classification in complex environments.

[0010] Preferably, the specific steps in S3 for determining whether the real-time psychological load index is greater than the dynamic baseline warning threshold include: S311. Retrieve the historical real-time psychological load index of trainees during periods of normal navigation and without communication tasks, and calculate the mean and standard deviation of the historical real-time psychological load index. S312. Add a constant multiple of the standard deviation of the historical real-time psychological load index to the mean of the historical real-time psychological load index to obtain the dynamic baseline warning threshold for trainees. S313. Determine whether the real-time psychological load index at the current time point is greater than the dynamic baseline warning threshold; if so, determine that the trainee is in a high psychological load state; otherwise, determine that the trainee is in a normal psychological state. This invention retrieves the historical real-time psychological load index of trainees during rest periods without tasks, calculates its mean and standard deviation, and introduces a constant multiple bias to construct a personalized dynamic baseline warning threshold. It abandons the limitations of fixed thresholds, fully considers the differences in psychological stress resistance baselines among different individuals, and effectively avoids false alarms and false negatives in the judgment of abnormal psychological state.

[0011] Preferably, the specific steps in S3 for obtaining the initial adaptive training scheme and the initial training guidance parameter set include: S321. Discretize the real-time psychological load index into discrete values ​​for light load, moderate load, and heavy load states; combine the discrete values ​​of states with the work stress state classification labels to define the current state variable; S322. Read the action value matrix in the control strategy generation model. The row index of the action value matrix corresponds to the current state variable, and the column index of the action value matrix corresponds to all candidate adaptive training schemes in the training action library. S323. Use an ε-greedy strategy to select actions from the action value matrix; generate a random number between zero and one. Determine whether the random number is less than the exploration probability threshold; if so, randomly select a candidate adaptive training program from the training action library as the initial adaptive training program; the initial adaptive training program includes at least one of breathing rhythm traction training, binaural beat auditory training, and progressive muscle relaxation. Otherwise, find the column with the largest value in the row corresponding to the current state variable in the action value matrix, and use the candidate adaptive training scheme corresponding to the column with the largest value as the initial adaptive training scheme. S324. Based on the initial adaptive training program, extract the corresponding initial training guidance parameter set from the parameter mapping table; when the initial adaptive training program includes respiratory rhythm traction training, the initial training guidance parameter set includes the initial scaling period of the visual respiratory halo and the initial beat frequency of the auditory guidance sound. This invention discretizes the continuous real-time psychological load index and combines it with negative emotion labels to form the current state variable. It generates the action value matrix of the model based on the reinforcement learning regulation strategy and adopts an ε-greedy strategy to balance exploration and utilization. It matches the initial adaptive training scheme with the highest expected return from the action library. This achieves adaptive optimization matching of adaptive training schemes and ensures that the training strategies for different psychological states and negative emotions have high relevance and initial effectiveness.

[0012] Preferably, the specific steps for calculating the real-time physiological feedback deviation rate of trainees in step S4 include: S411. Set a sliding time window; when the initial adaptive training program includes respiratory rhythm traction training, capture the real-time respiratory rhythm signal within the sliding time window. S412. Perform Hilbert transform on the real-time respiratory rhythm signal within the sliding time window, calculate the analytic signal of the real-time respiratory rhythm signal, and extract the imaginary part and the real part of the analytic signal. S413. Calculate the ratio of the imaginary part of the analytic signal to the real part of the analytic signal using the arctangent function to obtain the actual instantaneous breathing phase of the trainee; S414. Extract the instantaneous phase of the visual breathing halo sent to the execution terminal from the initial training guidance parameter set; S415. Calculate the absolute difference between the actual instantaneous breathing phase and the guided instantaneous phase, and divide the absolute difference by twice pi to obtain the real-time physiological feedback deviation rate. In respiratory rhythm traction training, this invention extracts real-time respiratory rhythm signals within a sliding time window, uses Hilbert transform to extract the real and imaginary parts of the analytical signal, and accurately calculates the actual instantaneous respiratory phase of the trainee through the arctangent function. The absolute difference between the actual phase and the guiding instantaneous phase of the visual respiratory halo is calculated and normalized. This achieves quantitative tracking of the human respiratory rhythm phase, providing a high-precision evaluation index for determining whether the trainee is effectively following the training guidance.

[0013] Preferably, the specific steps in S4 for generating the dynamic control instruction sequence and sending it to the execution terminal include: S421. Determine whether the real-time physiological feedback deviation rate is greater than the preset phase lock threshold. S422. When the real-time physiological feedback deviation rate is greater than the phase lock threshold, it is determined that the trainee has failed to follow the guided rhythm; calculate the derivative of the actual respiratory instantaneous phase with respect to time to obtain the trainee's actual respiratory rate. S423. Forcefully modify the guidance frequency parameter in the initial training guidance parameter set to the actual breathing frequency, generate a synchronous follow command and send it to the execution terminal, so that the scaling period of the visual breathing halo completely matches the actual breathing frequency of the trainee and establishes a physiological resonance state. S424. After maintaining the preset breathing cycle in the physiological resonance state, the guide frequency parameter is gradually reduced by optimizing the gradient with variable step size, generating a slow descent traction command and sending it to the execution terminal until the guide frequency parameter drops to the set deep relaxation target frequency. When the real-time physiological feedback deviation rate is not greater than the preset phase lock threshold, no synchronous follow command or descent traction command is generated. When the real-time physiological feedback deviation rate exceeds the limit, this invention obtains the actual respiratory rate by calculating the time derivative, forcibly modifies the guidance frequency parameters to generate a synchronous following instruction, and quickly establishes a human-machine physiological resonance state; it uses variable step size to optimize the gradient to generate a slow descent traction instruction, smoothly guiding the respiratory rate to the deep relaxation target value; it avoids the psychological resistance caused by reverse guidance, realizes a smooth transition from passive following to active traction, and significantly improves the compliance and training efficiency of respiratory rhythm regulation.

[0014] Preferably, in step S4, when the initial adaptive training program includes progressive muscle relaxation, the rules for calculating the real-time physiological feedback deviation rate and generating instructions are as follows: S431. Acquire electromyographic signals and calculate the current root mean square amplitude of the electromyographic signals within the sliding time window; S432. Extract the initial root mean square amplitude of electromyography (EMG) of the trainee before training, and subtract the current root mean square amplitude from the initial EMG amplitude to obtain the amplitude difference. The muscle relaxation rate is calculated by dividing the amplitude difference by the initial root mean square amplitude of electromyography; the real-time physiological feedback deviation rate is defined as 100% minus the muscle relaxation rate. S433. When the real-time physiological feedback deviation rate is greater than the set threshold, it is determined that the target muscle group has not reached the relaxation standard; a pause command and a strengthening stimulation command are generated and sent to the execution terminal. The execution terminal stops playing the relaxation instructions for the next muscle group according to the pause command, and repeats the instructions for the current muscle group. It also outputs binaural beat audio according to the reinforcement stimulation command. This invention calculates the root mean square amplitude of the surface electromyography (EMG) signal of the target muscle group, compares it with the initial amplitude before training to calculate the muscle relaxation rate and the real-time physiological feedback deviation rate; when the deviation rate exceeds the limit, it automatically triggers a pause command to repeat the guidance, and superimposes binaural beat audio stimulation; it constructs a closed-loop monitoring and correction mechanism for the muscle relaxation process, ensuring that the local muscle group reaches the standard of substantial relaxation.

[0015] On the other hand, the present invention also provides an intelligent regulation system for psychological adaptability based on physiological feedback, used to implement the aforementioned intelligent regulation method for psychological adaptability based on physiological feedback, the system comprising: The multidimensional spatiotemporal synchronization matrix construction module is used to acquire multimodal physiological signal sequences, ship physical environment parameter sequences, and job task load parameter sequences, and perform timestamp alignment to construct a multidimensional spatiotemporal synchronization state matrix. The anti-interference status assessment module is used to perform adaptive filtering calculations to generate a purified physiological feature vector by using the ship's physical environment parameter sequence as a reference input signal, and to output a real-time psychological load index using a spatiotemporal attention status assessment model. The reinforcement learning strategy matching module is used to output an initial adaptive training scheme and an initial training guidance parameter set by matching the action value matrix when the real-time psychological load index is greater than the dynamic baseline warning threshold. The physiological closed-loop training guidance module is used to calculate the actual instantaneous phase of respiration using Hilbert transform, calculate the real-time physiological feedback deviation rate, and generate a dynamic control command sequence using a variable step size optimization algorithm. The performance recovery reward update module is used to calculate the communication operation performance recovery rate and comprehensive reward value, and update the action value matrix.

[0016] (III) Beneficial Effects The present invention has the following beneficial effects: This invention effectively eliminates vestibular compensatory heart rate fluctuations and motion artifacts caused by severe ship rolling by constructing a multidimensional spatiotemporal synchronous state matrix and using ship physical environment parameters as reference inputs for adaptive filters. By combining spatiotemporal attention mechanisms to extract the time-cumulative evolution characteristics of physiological and task load, it eliminates the interference of complex sea state physical noise on psychological state assessment and improves the objectivity and accuracy of real-time psychological load index calculation.

[0017] This invention introduces a physiological closed-loop feedback mechanism during training regulation; it uses Hilbert transform to calculate the actual instantaneous respiratory phase of the trainee in real time, and establishes a physiological resonance state by forcibly modifying the guidance frequency when the physiological feedback deviation rate exceeds the limit, and then uses variable step size to optimize the gradient for smooth traction; it avoids the rhythm discontinuity and adverse resistance caused by traditional static guidance, and realizes dynamic adaptive regulation in accordance with the individual's real-time physiological state.

[0018] This invention constructs a regulation strategy generation model based on a reinforcement learning architecture, generates personalized dynamic baseline warning thresholds based on historical data, and calculates a comprehensive reward value by quantifying the recovery rate of communication operation efficiency and the decline in physiological indicators after training. The action value matrix is ​​iteratively updated using the reward value, enabling the system to have the ability to self-evolve based on individual feedback and task urgency, thus ensuring the continuous effectiveness of the training strategy in long-haul environments.

[0019] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, the drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating an intelligent regulation method for psychological adaptability based on physiological feedback according to the present invention. Figure 2 This is a comparison chart of the adaptive filtering vestibular interference stripping effect in this invention; Figure 3 This is a diagram illustrating the effect of real-time psychological load index and dynamic baseline early warning in this invention. Figure 4 This is a diagram illustrating the effect of the closed-loop phase modulation process of respiratory rhythm in this invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the invention, and not all embodiments. Based on the embodiments of the invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the invention.

[0023] When large ships are on long-distance ocean voyages, the communications and radio operators are located in a confined compartment below the waterline, constantly exposed to high-decibel engine noise. In sea states of force 5 or higher, the ship's roll angle can exceed 15 degrees. Communications and radio operators must perform high-intensity tasks such as sending and receiving encrypted telegrams, decryption, and issuing commands to the control system in this confined, swaying, and noisy environment. Due to prolonged exposure to multiple physical and cognitive stressors, communications and radio operators are prone to sympathetic overactivation, narrowed attention spans, and the accumulation of negative emotions such as anxiety, leading to an increased error rate in message typing and seriously threatening the accuracy and timeliness of ship communications command. Traditional relaxation methods such as playing fixed light music cannot counteract the interference of severe sea conditions and lack accurate assessment and dynamic feedback of the personnel's actual physiological state.

[0024] To resolve the above issues, please refer to [link / reference]. Figure 1 This invention discloses an intelligent regulation method for psychological adaptability based on physiological feedback, comprising the following steps: S1. Acquire the multimodal physiological signal sequence of the trainees through wearable sensing nodes, simultaneously acquire the ship's physical environment parameter sequence through ship environment sensing nodes, and acquire the job task load parameter sequence of the trainees through communication operation terminals; The multimodal physiological signal sequence, the ship's physical environment parameter sequence, and the job task load parameter sequence are timestamped to construct a multidimensional spatiotemporal synchronization state matrix; S1 includes the following steps: S11. Collect electrocardiogram (ECG) signals through dry electrode sensors attached to the trainee's chest, collect skin conductance level signals through skin conductance sensors attached to the fingers, and collect respiratory rhythm signals through chest strap strain sensors. Combine ECG signals, skin conductance level signals, and respiratory rhythm signals to form a multimodal physiological signal sequence. Specifically, in embodiment S11, the trainee wore a physiological state monitoring vest and a smart ring; the Ag / AgCl dry electrode sensor embedded in the monitoring vest collected the trainee's electrocardiogram signal in real time at a sampling rate of 250 Hz; the microelectrode on the inside of the smart ring collected skin conductance level signals at a sampling rate of 50 Hz to reflect the minute resistance changes in sweat gland secretion caused by sympathetic nerve excitation; the chest band strain sensor captured the physical deformation caused by chest cavity fluctuations at a sampling rate of 20 Hz and converted it into a respiratory rhythm signal; the electrocardiogram signal, skin conductance level signal and respiratory rhythm signal combined to form a multimodal physiological signal sequence; S12. The ship's pitch acceleration sequence and roll angular velocity sequence are collected by a six-axis inertial measurement unit deployed in the communication compartment. The background noise sound pressure level sequence of the compartment is collected by a decibel meter. The ship's pitch acceleration sequence, roll angular velocity sequence and background noise sound pressure level sequence of the compartment are combined to form a ship physical environment parameter sequence. Specifically, in embodiment S12, an industrial-grade six-axis inertial measurement unit deployed below the communication control console outputs the ship's three-axis acceleration and three-axis angular velocity at a sampling rate of 100 Hz; the system extracts the ship's pitch acceleration sequence and ship's roll angular velocity sequence; simultaneously, a decibel meter fixed on the top of the cabin outputs the cabin background noise sound pressure level sequence at a frequency of 10 Hz; the ship's pitch acceleration sequence, ship's roll angular velocity sequence, and cabin background noise sound pressure level sequence quantify the physical causes that lead to vestibular dysfunction and auditory fatigue in trainees, forming a sequence of ship physical environment parameters; S13. Through the underlying driver interface of the communication terminal, count the number of effective keystrokes, message retransmission frequency, and error correction backspace times of the trainees within a unit of time, and combine the number of effective keystrokes, message retransmission frequency, and error correction backspace times into a sequence of job task load parameters. Specifically, in embodiment S13, the system background program is mounted on the underlying driver of the military communication keyboard; the system records the number of valid keystrokes, the number of backspace corrections due to input errors, and the frequency of message retransmissions due to checksum errors, with a statistical period of one minute; the number of valid keystrokes, the frequency of message retransmissions, and the number of backspace corrections constitute the job task load parameter sequence. S14. Using the absolute time of the network time protocol server as a reference, the multimodal physiological signal sequence, ship physical environment parameter sequence, and job task load parameter sequence with different sampling rates are resampled to a unified reference frequency using a cubic spline interpolation algorithm; the resampled multimodal physiological signal sequence, ship physical environment parameter sequence, and job task load parameter sequence are then spliced ​​together to form a multidimensional spatiotemporal synchronization state matrix. Specifically, in embodiment S14, since the ECG signal sampling rate is 250 Hz and the ship attitude signal sampling rate is 100 Hz, direct splicing would cause matrix dimension disorder. The system connects to the ship's local area network network time protocol server to add absolute timestamps to all sensor data packets. Subsequently, the system sets a unified reference frequency of 100 Hz, uses cubic spline interpolation algorithm to downsample and smooth the 250 Hz ECG signal, and upsamples and interpolates the 10 Hz noise signal. After processing, the system splices the physical quantities under the same absolute timestamp into a row vector with a dimension of 1×N, forming a multidimensional spatiotemporal synchronization state matrix over time, ensuring the uniformity of each physical quantity on the time axis during subsequent fusion calculations. The above embodiments synchronously collect data through multiple source sensors and the underlying interface, and use cubic spline interpolation and absolute timestamps to complete resampling and alignment, thereby constructing a multi-dimensional spatiotemporal synchronization state matrix, ensuring the consistency of multimodal data in the time dimension, and reducing the time deviation rate of subsequent feature fusion by more than 90% compared with unaligned data; S2. Extract the multimodal physiological signal sequence and the aligned ship physical environment parameter sequence from the multidimensional spatiotemporal synchronization state matrix; use the ship physical environment parameter sequence as the reference input signal of the adaptive filter, perform motion artifact removal and vestibular compensation heart rate subtraction calculation on the multimodal physiological signal sequence, and generate the purified physiological feature vector. The job task load parameter sequence and the purified physiological feature vector are input into the spatiotemporal attention state assessment model to output the real-time psychological load index and job stress state classification label of the trainee at the current time node. Please see Figure 2 The specific steps in S2 for generating the purified physiological feature vector include: S211. Extract the ship pitch acceleration sequence from the aligned ship physical environment parameter sequence as the reference input signal, and extract the electrocardiogram signal from the aligned multimodal physiological signal sequence as the main input signal. Specifically, in embodiment S211, the system extracts the original electrocardiogram signal from the multidimensional spatiotemporal synchronization state matrix as the main input signal and extracts the ship pitching acceleration sequence as the reference input signal. Since the violent pitching of the ship will cause the human vestibular nerve to compensate for the heart rate acceleration, this acceleration is not a purely psychological load caused by work pressure. If it is not removed, the evaluation model will misjudge the physiological reaction of seasickness as an extreme anxiety psychological reaction. S212. Construct a normalized minimum mean square adaptive filter and calculate the error signal between the main input signal and the reference input signal after processing by the filter weight vector. Specifically, in embodiment S212, the system constructs a normalized minimum mean square adaptive filter; the formula for calculating the error signal is as follows: ;in, e ( n ) indicates the current time step n The error signal is expressed in mV. d ( n ) indicates the current time step n The main input signal, i.e., the raw electrocardiogram signal, is measured in mV; w ( n () represents the current time step n The filter tap weight vector, where each weight coefficient is in mV·s. 2 / m, multiplied by the acceleration input, yields the disturbance estimate in mV; w T ( n ) is the transpose of the weight vector, used to perform a dot product operation with the reference input vector; x ( n ) indicates the current time step n The reference input signal, i.e., the ship's pitch acceleration sequence vector, is expressed in m / s². 2 Product term The physical meaning is the compensatory electrocardiogram fluctuation amplitude estimated by the system purely caused by the physical rolling of the ship; S213. Iteratively update the filter weight vector based on the error signal. When the filter weight vector converges, extract the continuous error signal as the pure electrocardiogram signal stripped of seasickness compensation response. Specifically, in embodiment S213, the iterative update formula for the filter weight vector is as follows: ;in, μ This represents the set step size factor; δ This represents the smallest positive constant that prevents the denominator from being zero, and its unit is m. 2 / s 4The value of the step size factor μ is determined in combination with the signal characteristics of the ship's pitch acceleration interference. In this embodiment, the value is 0.05, which is used to control the steady-state error while ensuring the filtering convergence speed, and is suitable for typical ocean scenarios with ship rolling frequencies of 0.1~0.5Hz. δ The value is much smaller than the minimum value of the input signal energy. In this embodiment, the value is 0.001. It only plays a denominator protection role when the reference input signal energy is close to zero, so as to avoid numerical calculation overflow and will not have a substantial impact on the filtering result. As time goes by n The advancement, when the filter weight vector w ( n When the rate of change between two consecutive iterations is less than one-thousandth, the filter weight vector is considered to have converged; at this point, the output continuous error signal... e ( n This refers to the pure electrocardiogram signal stripped of the physical interference of ship rolling. S214. Detect the peak value of the R wave in the pure electrocardiogram signal, calculate the time interval sequence of adjacent R waves, and use an autoregressive model to estimate the power spectral density of the time interval sequence; calculate the integrated power of the low-frequency band and the integrated power of the high-frequency band, and divide the integrated power of the low-frequency band by the integrated power of the high-frequency band to obtain the sympathetic vagus nerve balance index. Extract skin conductance level signals, calculate the phase arousal peak value of skin conductance level signals, and combine the sympathetic vagus nerve balance index with the phase arousal peak value to form a physiological feature vector after purification; Specifically, in embodiment S214, the system uses the Pan-Tompkins algorithm to detect the R-wave peak of the pure electrocardiogram signal and extract the time interval sequence; the system uses an autoregressive model of order 16 to estimate the power spectral density of the time interval sequence; the system calculates the integral power of the low-frequency band and the integral power of the high-frequency band. The sympathetic-vagal balance index is calculated by dividing the integrated power of the low-frequency band by the integrated power of the high-frequency band. The system extracts skin conductance level signals from the multimodal physiological signal sequence and calculates the number of sharp rise peaks in the skin conductance level signal within a five-minute sliding window, which is used as the phase arousal peak value. The sympathetic-vagal balance index and the phase arousal peak value are combined to form a high-fidelity post-purification physiological feature vector; for example, 0.04Hz ≤ f <0.15Hz, the high-frequency band range is 0.15Hz ≤ f ≤0.4Hz; Through the above embodiments, combined with Figure 2The comparison between the original ECG waveform containing vestibular interference and the filtered pure ECG waveform shows that the present invention effectively removes the vestibular compensatory physiological artifacts caused by ship swaying through adaptive filtering technology, eliminates the pollution of physiological signals by physical environmental noise, and extracts feature vectors that only reflect the true psychological stress state. This solves the core problem of existing technologies misjudging seasickness physiological reactions as psychological load, and greatly improves the objectivity and accuracy of psychological state assessment in complex ship environments. Furthermore, the specific steps in S2 for outputting the real-time psychological load index and occupational stress state classification labels of trainees include: S221. Normalize the purified physiological feature vector and the job task load parameter sequence, and concatenate them into a fused feature tensor. S222. Input the fused feature tensor into the pre-trained spatiotemporal attention state evaluation model. The one-dimensional convolutional layer of the spatiotemporal attention state evaluation model extracts spatial dimension features from the fused feature tensor to generate a local physiological response feature map. Specifically, in embodiment S222, the spatiotemporal attention state evaluation model adopts a deep network architecture; the input layer receives a fusion feature tensor with a dimension of 1×8; the one-dimensional convolutional layer contains 32 convolutional kernels with a size of 1×3, uses the ReLU activation function, slides on the fusion feature tensor to extract features, captures the local nonlinear mapping relationship between the abnormally high sympathetic vagus nerve balance index and the surge in the number of error correction backspaces, and outputs a local physiological response feature map with a dimension of 32×6; the model as a whole adopts a three-level convolution-recurrent-attention architecture, with two one-dimensional convolutional layers, with convolutional kernel sizes of 1×3 and 1×2 respectively, both with the same stride, and using the same padding to ensure that the feature length remains unchanged; the long short-term memory network layer has a bidirectional structure, with 64 hidden units each in the forward and backward directions; the self-attention module has 4 attention heads, and the multi-head outputs are spliced ​​and then processed by layer normalization to stabilize the feature distribution; Each layer is connected in a serial manner. The outputs of two one-dimensional convolutional layers are directly input into a bidirectional long short-term memory (LSTM) network layer. The hidden state sequence output from the LSM network layer is input into a multi-head self-attention layer. The output of the self-attention layer is normalized and then connected to a fully connected hidden layer containing 128 neurons, ultimately resulting in two output branches: regression and classification. Dropout layers are applied after both the convolutional layers and the fully connected hidden layers, with a dropout ratio of 0.2 to suppress model overfitting. The ReLU activation function is used for the convolutional layers, the sigmoid and tanh activation functions are used for the LSM gating units, the linear activation function is used for the regression branch, and the softmax activation function is used for the classification branch. In this embodiment, the training dataset of the spatiotemporal attention state assessment model consists of multimodal physiological and task load data of 120 communication personnel collected during actual ship navigation. The label data is jointly labeled by three level-two or above psychological counselors, combining concurrent behavioral observations with scores from the anxiety self-rating scale and fatigue scale. The psychological load index is labeled in the range of 0 to 1, and the negative emotion labels are divided into three categories: anxiety, fatigue, and stability. The dataset is divided in chronological order, with the first 70% of the samples serving as the training set, the middle 20% as the validation set, and the last 10% as the test set, to avoid data leakage caused by random partitioning. All data were collected under real-world ocean voyage scenarios, covering calm navigation, moderate sea states, severe sea states, and different mission intensities. The cumulative effective data duration for each subject was no less than 8 hours. The original sampling frequencies for various signals were 250Hz for ECG, 50Hz for skin conductance, 20Hz for respiration, and 100Hz for ship attitude. After resampling to a uniform 100Hz, the data was input into the model. The dataset was divided according to time sequence. The first 70% of each subject's data was assigned to the training set, the middle 20% to the validation set, and the last 10% to the test set to avoid time-series data leakage caused by random partitioning. The spatiotemporal attention state evaluation model employs a multi-task joint loss function. The total loss is a weighted sum of the mean squared error loss of the regression branch and the cross-entropy loss of the classification branch. The weights of the two types of losses are dynamically calculated from the initial error proportions of the validation set, with initial weights of 0.6 and 0.4, respectively. The formula for the loss function is as follows: ;in The mean squared error loss of the regression branch. y i This represents the true psychological load index. y 1 i These are the model's predicted values; M This represents the total number of training samples used in a single loss calculation, i.e., the number of samples in a single training batch. For the cross-entropy loss of the classification branch, K The total number of categories represents the task of classifying work-related stress states. In this embodiment, negative emotions are divided into three categories: stable, anxious, and fatigued, corresponding to K=3. The number of categories can be adjusted according to actual classification needs; k is the category index variable. zik For the category of genuine emotions, z 1 ik The class probabilities predicted by the model; τ 1. τ 2 represents the weighting coefficients for the two types of losses, which are dynamically calculated from the initial error proportion of the validation set, with initial weights of 0.6 and 0.4 respectively. Training was performed using the Adam optimizer with an initial learning rate of 0.001. The validation set was trained when the relative decrease in total loss over 10 consecutive iterations was less than 1 × 10⁻⁶.-5 When the model converges, training is terminated and the optimal weight parameters are saved. During training, the total loss of the model is evaluated using the validation set after each iteration. An early stopping mechanism is adopted, and training is terminated if the loss fails to decrease for 10 consecutive iterations. The model weights with the lowest loss in the validation set are saved. The final performance metrics of the model on the test set are: mean absolute error ≤ 0.04 for the psychological load index regression task, accuracy ≥ 92% for the occupational stress state classification task, and F1 score ≥ 0.90. S223. The long short-term memory network layer of the spatiotemporal attention state assessment model receives local physiological response feature maps, calculates the hidden layer state sequence, and extracts the cumulative evolution features of psychological stress response in the time dimension. Specifically, in embodiment S223, a long short-term memory network layer containing 64 neurons processes local physiological response feature maps for 100 consecutive time steps; the forgetting gate and input gate mechanism inside the long short-term memory network layer extracts the evolutionary dynamic features of the psychological stress of trainees accumulating over time and outputs the hidden layer state sequence. S224. The self-attention mechanism layer of the spatiotemporal attention state assessment model assigns attention weights to the hidden layer state sequence to generate a weighted hidden layer state sequence; the weighted hidden layer state sequence is input into the fully connected layer; the regression branch of the fully connected layer outputs a continuous real-time psychological load index, and the classification branch of the fully connected layer outputs the task stress state classification label through the Softmax function; Specifically, in embodiment S224, the attention weight matrix is ​​calculated from the attention mechanism layer, and the hidden state that has recently deteriorated sharply is given a higher weight; the weighted hidden layer state sequence is input to a fully connected layer containing 128 neurons; the regression branch uses a linear activation function to output the trainee's current real-time psychological load index (range 0 to 1); the classification branch uses a Softmax activation function to output the most probable job stress state classification label, such as acute anxiety, severe fatigue, etc. The aforementioned spatiotemporal attention state assessment model architecture is improved for the specific scenario of psychological load assessment in a ship environment. Specifically, it uses one-dimensional convolution to extract the nonlinear correlation of local physiological-task features, capturing the instantaneous correspondence between skin arousal waves, sudden changes in heart rate variability, and surges in keystroke error rate, thus solving the problem of insufficient capture of local sudden stress features by a single LSTM architecture. It also incorporates a bidirectional long short-term memory network to extract cumulative evolutionary features in the time dimension, adapting to the physiological law of gradual accumulation of psychological load with sea conditions and task duration, and compensating for the weak ability of pure convolutional architecture to model temporal dependencies. Furthermore, it introduces a multi-head self-attention mechanism to assign higher weights to stress mutation nodes, focusing on state changes at key moments such as severe ship rolling and message peaks, while weakening invalid features during stable periods. Compared to existing single LSTM or single CNN psychological load assessment schemes, this architecture improves the state assessment accuracy by approximately 25% in complex ship interference scenarios and reduces the response delay to sudden high load states by 30%. The spatiotemporal attention state assessment model constructed in the above embodiments extracts spatiotemporal features through one-dimensional convolution and long short-term memory networks, and focuses on key stress nodes by combining self-attention mechanisms, outputting real-time psychological load index and emotion labels. Compared with the traditional single threshold judgment method, it improves the accuracy of state assessment in complex environments by about 25%. S3. Determine whether the real-time psychological load index is greater than the dynamic baseline warning threshold. If so, combine the real-time psychological load index with the occupational stress state classification label to form the current state variable. Input the current state variable into the reinforcement learning-based regulation strategy generation model. Based on the action value matrix in the regulation strategy generation model, match and output the initial adaptive training scheme with the largest expected reward value and the initial training guidance parameter set from the training action library. Otherwise, determine that the trainee's psychological state is normal, write the real-time psychological load index into the historical state database, and return to step S1. Please see Figure 3 The specific steps in S3 for determining whether the real-time psychological load index is greater than the dynamic baseline warning threshold include: S311. Retrieve the historical real-time psychological load index of the trainees during the past three consecutive normal flight days without communication tasks, and calculate the mean and standard deviation of the historical real-time psychological load index. Specifically, in embodiment S311, different trainees have different psychological stress tolerance thresholds; the system queries the historical status database to retrieve the trainee's historical real-time psychological load index during the past three days of no-task rest; for example, the mean of the historical real-time psychological load index is 0.25 and the standard deviation of the historical real-time psychological load index is 0.05. S312. The dynamic baseline warning threshold for trainees is calculated by adding a constant multiple of the standard deviation of the historical real-time psychological load index to the mean of the historical real-time psychological load index; the constant multiple is set to be between 1.5 and 2.0. Specifically, in embodiment S312, the constant multiple is calculated based on the historical physiological fluctuation variance of the trainees; specifically, in this embodiment, the system extracts the real-time psychological load index sequence of all normal rest periods of the trainees in the past 30 days, and calculates the skewness coefficient and kurtosis coefficient of the sequence; the absolute values ​​of the skewness coefficient and the absolute values ​​of the kurtosis coefficient are added together. If the sum is less than 1.0, it indicates that the historical data distribution is concentrated, and the constant multiple is set to a lower limit of 1.5; if the sum is greater than or equal to 1.0, it indicates that the historical data fluctuation is large, and the constant multiple is based on 1.5, plus the logarithmic transformation value of the sum of the skewness coefficient and the kurtosis coefficient, and the maximum limit is 2.0; If the system sets the constant multiple to 2.0, the calculated dynamic baseline warning threshold is 0.25 + 2.0 × 0.05 = 0.35. The dynamic baseline warning threshold represents the individual psychological tolerance threshold of the trainee. S313. Compare the real-time psychological load index at the current time point with the dynamic baseline warning threshold. When the real-time psychological load index is greater than the dynamic baseline warning threshold, the trainee is determined to be in a state of high psychological load. When the real-time psychological load index is not greater than the dynamic baseline warning threshold, the trainee is determined to be in a normal psychological state. Specifically, in embodiment S313, the dynamic baseline warning threshold is generated based on the individual's historical mean and standard deviation, combined with the aforementioned constant multiple; it reflects the maximum psychological load tolerance envelope of the trainee under the condition of not being stimulated by external task stressors after removing extreme outliers; when the real-time input value exceeds this envelope, it is determined from a mathematical and statistical perspective that it has deviated from the steady-state range of self-regulation. If the real-time psychological load index calculated in S224 is 0.82, which is greater than the calculated dynamic baseline warning threshold of 0.35, the system determines that the trainee is in a state of high psychological load and triggers the intelligent adjustment training process; if the real-time psychological load index is 0.30, which is not greater than 0.35, the system determines that the trainee is in a normal psychological state, the system does not execute training, only stores 0.30 in the historical state database, and returns to S1 to continue monitoring; Combining the above embodiments Figure 3The changing trend of the mental stress index, the dynamic baseline threshold, and the high-load over-limit interval show that this invention abandons the limitations of a general fixed threshold and uses individual historical data to generate a personalized dynamic early warning baseline. It fully adapts to the differences in the psychological stress resistance baseline of different people, can accurately identify the state of psychological overload, effectively reduce false alarms and false negatives of abnormal states, and provide a reliable basis for subsequent intervention triggers. Furthermore, the specific steps in S3 for obtaining the initial adaptive training scheme and the initial training guidance parameter set include: S321. Discretize the real-time psychological load index into three discrete values: light load, moderate load, and heavy load; combine the discrete values ​​of the states with the work stress state classification labels to define the current state variable. Specifically, in embodiment S321, the boundary of the psychological load index interval is determined based on the output gradient distribution of the spatiotemporal attention state assessment model on the validation set; the system statistically analyzes the predicted psychological load index of all samples in the validation set and calculates the probability density function of the predicted psychological load index; the index value corresponding to the first extreme point where the first derivative of the probability density function turns from positive to negative is extracted as the boundary point between mild and moderate load; the index value corresponding to the inflection point where the second derivative reaches its maximum positive value is extracted as the boundary point between moderate and severe load, thereby achieving the objective delineation of the discretized interval; For example, the system defines the real-time psychological load index range [0.35, 0.6) as mild load, [0.6, 0.8) as moderate load, and [0.8, 1.0] as severe load; since the real-time psychological load index is 0.82, the state discrete value is severe load; the system combines severe load with the job stress state classification label (such as acute anxiety) to define the current state variable; S322. Read the action value matrix in the control strategy generation model. The row index of the action value matrix corresponds to the current state variable, and the column index of the action value matrix corresponds to all candidate adaptive training schemes in the training action library. Specifically, in embodiment S322, the system adopts a Q-Learning reinforcement learning architecture; reads the action value matrix stored in memory; the current row index of the action value matrix points to the current state variable, and the column index contains three candidate adaptive training schemes in the training action library: candidate adaptive training scheme one is simple breathing rhythm traction training, candidate adaptive training scheme two is breathing rhythm traction training combined with binaural beat auditory training, and candidate adaptive training scheme three is progressive muscle relaxation; Furthermore, the Q-Learning reinforcement learning architecture consists of a state space, an action space, and a reward mechanism. The state space contains a two-dimensional discrete variable set composed of a discrete psychological load index and negative emotion labels. The action space contains an orthogonal combination of breathing rhythm traction training, binaural beat auditory training, and progressive muscle relaxation. The dimension of the action value matrix is ​​the total number of states multiplied by the total number of actions, initialized as a zero matrix. The algorithm employs an offline strategy for temporal difference learning. After each training cycle, the Bellman equation is used to iteratively update the current state-action pair based on the immediate reward and the maximum expected action value of the next state. Q value; S323. Use an ε-greedy strategy to select actions in the action value matrix; generate a random number between zero and one, and determine whether the random number is less than the exploration probability threshold; if so, randomly select a candidate adaptive training scheme from the training action library as the initial adaptive training scheme; otherwise, find the column with the largest value in the row corresponding to the current state variable in the action value matrix, and use the candidate adaptive training scheme corresponding to the column with the largest value as the initial adaptive training scheme. Specifically, in embodiment S323, the exploration probability threshold is dynamically calculated based on the convergence degree of the action value matrix; the system statistically analyzes all data in the row corresponding to the current state variable in the action value matrix. Q The variance of the value is calculated, and the ratio of this variance to the historical maximum variance is taken. This ratio is then subtracted from 1 and multiplied by the base exploration coefficient of 0.2 to obtain the current exploration probability threshold. Q When the value difference is small, i.e., when convergence has not yet occurred, the exploration probability is high; when certain actions... Q When the value is significantly prominent, that is, when convergence has been achieved, the exploration probability automatically decays to near 0; If the system sets the exploration probability threshold to 0.1, the system generates a random number 0.85. Since 0.85 is not less than 0.1, the system executes the utilization strategy. It searches the row corresponding to the current state variable in the action value matrix and finds that the value of candidate adaptive training scheme two is the largest. The system uses candidate adaptive training scheme two as the initial adaptive training scheme. By discretizing continuous psychological load and combining it with negative emotion labels to form a two-dimensional state, this approach aligns with the multi-dimensional attributes of psychological states, providing a more accurate match for training needs compared to single-state descriptions. It utilizes a motion library to set up breathing traction, binaural rhythm, muscle relaxation, and their combinations, addressing training needs for different loads and emotion types, thus avoiding ineffective exploration of a generic motion space. A two-dimensional reward function combining task effectiveness and physiological improvement is designed, with weights dynamically adjusted based on task urgency, balancing personnel psychological state with job requirements. Exploration probability, learning rate, and discount factor are dynamically calculated based on scenario data, rather than fixed values, adapting to individual differences and state changes during long-haul missions. Compared to adaptive training schemes with fixed rules, this reinforcement learning architecture can continuously optimize strategies with training iterations, improving training effectiveness by over 30% for different individuals, with no significant attenuation of training effects in long-haul scenarios, demonstrating self-evolutionary capabilities. S324. Based on the initial adaptive training program, extract the corresponding initial training guidance parameter set from the parameter mapping table; when the initial adaptive training program includes respiratory rhythm traction training, the initial training guidance parameter set includes the initial scaling period of the visual respiratory halo and the initial beat frequency of the auditory guidance sound. Specifically, in embodiment S324, the system extracts the initial training guidance parameter set from the parameter mapping table; since the system detects that the trainee's current breathing rate is 22 times per minute, the system sets the initial scaling period of the visual breathing halo to 2.7 seconds, that is, 60 seconds divided by 22 times, and sets the initial beat frequency of the auditory guidance sound played by the bone conduction headphones to 22 BPM. The above embodiments generate dynamic baseline warning thresholds by calculating the variance of historical data, and when the threshold is exceeded, they use the action value matrix and dynamic exploration probability to match the initial adaptive training scheme, thereby achieving adaptive generation of training strategies and avoiding diminishing marginal utility caused by fixed training modes. S4. Execute the initial adaptive training program. During the execution of the initial adaptive training program, continuously calculate the real-time physiological feedback deviation rate of the trainees using a preset sliding time window; determine whether the real-time physiological feedback deviation rate is greater than the phase lock threshold. If so, based on the gradient direction of the real-time physiological feedback deviation rate, the variable step size optimization algorithm is used to update the guidance frequency parameter and stimulus intensity parameter in the initial training guidance parameter set in real time, and generate a dynamic adjustment instruction sequence to be sent to the execution terminal; otherwise, the initial training guidance parameter set is kept unchanged, and the initial adaptive training scheme is continued to be executed until the set training time ends. Please see Figure 4 S4 includes the following steps: S411. When the initial adaptive training program includes respiratory rhythm traction training, set the sliding time window length to ten seconds and the step size to one second, and extract the real-time respiratory rhythm signal within the sliding time window from the multi-dimensional spatiotemporal synchronization state matrix. S412. Perform Hilbert transform on the real-time respiratory rhythm signal within the sliding time window, calculate the analytic signal of the real-time respiratory rhythm signal, and extract the imaginary part and the real part of the analytic signal. Specifically, in embodiment S412, the system not only controls the respiratory rate but also the phase of the respiratory rhythm; the system performs a Hilbert transform on the intercepted real-time respiratory rhythm signal; the calculation formula is as follows: ;in, H ( RSP ( t )) represents the output signal after the real-time respiratory rhythm signal has undergone Hilbert transform, i.e., the imaginary part of the analytic signal; RSP ( τ ) represents the integral variable τ The real-time respiratory rhythm signal at any given moment, i.e., the real part of the analytic signal; t Indicates the current time point; the system will RSP ( t As the real part of the analytic signal, H ( RSP ( t ()) is the imaginary part of the analytic signal; The integration interval of the continuous form Hilbert transform described above is (-∞ to +∞). For the discrete respiratory rhythm signals actually acquired, this embodiment uses Fast Fourier Transform (FFT) to implement the discrete Hilbert transform: first, an N-point FFT is performed on the discrete respiratory signal to obtain the spectrum, and then the positive frequency components are multiplied by... j negative frequency components multiplied by - j The DC component is set to zero, and then an inverse FFT is performed to obtain the analytic signal. The real part is the original breathing signal, and the imaginary part is the result of the Hilbert transform, which can efficiently calculate the instantaneous phase; furthermore, j The imaginary unit satisfies j 2 =-1; Multiply the positive frequency component by in the frequency domain j negative frequency components multiplied by - j Essentially, it applies a 90° phase shift to each frequency component of the signal, which is the frequency domain implementation principle of the Hilbert transform that shifts all frequency components of the original signal by 90°. S413. Calculate the ratio of the imaginary part of the analytic signal to the real part of the analytic signal using the four-quadrant arctangent function to obtain the actual instantaneous breathing phase of the trainee. Specifically, in embodiment S413, the formula for calculating the actual instantaneous respiratory phase is as follows: ; where Φactual ( t ) indicates that the trainees are in t The actual instantaneous phase of breathing at any given moment, measured in rad; for example, if the actual instantaneous phase of breathing is 0.8π radians, it means that the trainee is in the middle to late stage of inhalation. S414. Extract the instantaneous phase of the visual breathing halo sent to the execution terminal from the initial training guidance parameter set; Specifically, in embodiment S414, the system reads the code variables of the visual breathing halo animation currently playing on the display and extracts the guiding instantaneous phase Φ of the visual breathing halo. guide ( t If the instantaneous phase of the inhalation is 0.53 radians, it indicates that the inhalation and ascent phase is underway. S415. Calculate the absolute difference between the actual instantaneous breathing phase and the guided instantaneous phase, divide the absolute difference by twice pi, and calculate the real-time physiological feedback deviation rate. Specifically, in embodiment S415, the formula for calculating the real-time physiological feedback deviation rate is as follows: ;in, Error phase The real-time physiological feedback deviation rate is expressed as a percentage; for example, subtracting the 0.8π mentioned in S413 from the 0.3π mentioned in S414, and dividing by 2π, the real-time physiological feedback deviation rate is 25%. Furthermore, the complete cycle of the breathing phase corresponds to 2π radians. Dividing the absolute phase difference by 2π can normalize the deviation to the [0,1] interval. Its physical meaning is the proportion of the phase deviation to the complete breathing cycle. Using the normalized deviation rate instead of the absolute radian value can eliminate the influence of individual differences in breathing frequency, making the phase lock threshold universal across individuals and scenarios, and making it easier to uniformly determine the following state. The above embodiments utilize Hilbert transform to calculate the absolute difference between the actual instantaneous breathing phase and the guided instantaneous phase, quantifying the real-time physiological feedback deviation rate and providing a high-precision mathematical basis for determining the trainee's following status. Furthermore, the specific steps in S4 for generating the dynamic control instruction sequence and sending it to the execution terminal include: S421. Determine whether the real-time physiological feedback deviation rate is greater than the preset phase lock threshold. S422. When the real-time physiological feedback deviation rate is greater than the phase lock threshold, it is determined that the trainee has failed to follow the guided rhythm; calculate the derivative of the actual respiratory instantaneous phase with respect to time to obtain the trainee's actual respiratory rate. Specifically, in embodiment S422, the system presets a phase lock threshold of 15%; since the calculated real-time physiological feedback deviation rate of 25% is greater than 15%, the system determines that rhythm discontinuity has occurred; the system calculates the time derivative of the actual instantaneous respiratory phase, and finds that the trainee's current actual respiratory rate is 20 breaths per minute. S423. Forcefully modify the guidance frequency parameter in the initial training guidance parameter set to the actual breathing frequency, generate a synchronous follow command and send it to the execution terminal, so that the scaling period of the visual breathing halo completely matches the actual breathing frequency of the trainee and establishes a physiological resonance state. Specifically, in embodiment S423, the system generates a synchronous following instruction, forcibly modifying the guidance frequency parameter to match the cycle of 20 times per minute as in embodiment S422, and instantly and smoothly jumping the guidance instantaneous phase of the visual breathing halo to the actual breathing instantaneous phase; at this time, the halo fluctuations are perfectly in sync with the trainee's breathing, establishing a physiological resonance state; S424. After maintaining the physiological resonance state for three respiratory cycles, the guiding frequency parameter is gradually reduced by a variable step size optimization gradient of 0.5 breaths per minute, and a slow descent traction command is generated and sent to the execution terminal until the guiding frequency parameter drops to the set deep relaxation target frequency. When the real-time physiological feedback deviation rate is not greater than the preset phase lock threshold, no synchronous follow command or slow descent traction command is generated. Specifically, in embodiment S424, the phase-locking threshold is calculated based on the short-time coefficient of variation of the trainee's natural breathing frequency. Specifically, the system extracts the sequence of adjacent respiratory cycle time intervals of the trainee's breathing signal within 3 minutes before training, calculates the ratio of the standard deviation to the mean of the adjacent respiratory cycle time interval sequence, and obtains the short-time coefficient of variation. The short-time coefficient of variation is multiplied by a correction factor of 1.2 to obtain the phase-locking threshold. In this embodiment, the calculated phase-locking threshold is 15%, and the threshold defines the maximum mathematical deviation between the trainee's tolerable guided phase and the actual phase. The system generates a slow descent traction command, gradually reducing the guidance frequency parameter until it drops to the set deep relaxation target frequency, such as 6 times per minute. If the real-time physiological feedback deviation rate does not exceed the phase lock threshold of 15%, the system maintains the current parameters and continues to guide. Through the above embodiments, combined with Figure 4The synchronous convergence process of the guided frequency and the actual breathing frequency, and the change curve of the physiological feedback deviation rate from exceeding the limit to stabilizing below the threshold, show that the present invention breaks through the traditional one-way open-loop intervention mode. By actively adapting to the human rhythm to establish physiological resonance through instantaneous phase tracking, it then smoothly guides the target frequency for deep relaxation, avoiding the rhythm disjointness and psychological resistance caused by fixed rhythm reverse guidance. It achieves dynamic adaptive adjustment in accordance with the individual's immediate physiological state, significantly improving the compliance and intervention efficiency of respiratory regulation. Furthermore, in S4, when the initial adaptive training program includes progressive muscle relaxation, the calculation rules for the real-time physiological feedback deviation rate and the instruction generation rules are as follows: S431. Collect electromyographic signals by using surface electromyography sensors attached to the target muscle group of the trainee, and calculate the current root mean square amplitude of the electromyographic signal within the sliding time window. S432. Extract the initial root mean square amplitude of electromyography (EMG) of the trainee before training, subtract the current root mean square amplitude from the initial EMG amplitude to obtain the amplitude difference; divide the amplitude difference by the initial EMG amplitude to calculate the muscle relaxation rate; subtract the muscle relaxation rate from 100% to define the real-time physiological feedback deviation rate. S433. When the real-time physiological feedback deviation rate is greater than the set threshold, it is determined that the target muscle group has not reached the relaxation standard; a pause command and a reinforcement stimulation command are generated and sent to the execution terminal; the execution terminal stops playing the relaxation guidance for the next muscle group according to the pause command, and repeats the guidance for the current muscle group; the execution terminal outputs a binaural beat audio with a frequency difference of six Hz between the left and right ears to the bone conduction headphones according to the reinforcement stimulation command. The above embodiments calculate the muscle relaxation rate by measuring the difference in root mean square amplitude of electromyography. When the target is not met, repetitive guidance and binaural beat stimulation are automatically triggered, thus constructing a closed-loop correction mechanism for progressive muscle relaxation and ensuring the effectiveness of training movements. S5. After the training period ends, extract the job task load parameter sequence of the trainees after the training ends, and calculate the communication operation efficiency recovery rate; substitute the communication operation efficiency recovery rate and the decline of physiological indicators during the training period into the reward function formula to calculate the comprehensive reward value, and use the comprehensive reward value to update the action value matrix in the control strategy generation model. S5 includes the following steps: S51. After the set training duration of 5 minutes ends, the system stops issuing control commands; the system extracts the sequence of job task load parameters within 10 consecutive minutes after the training ends. S52. Calculate the communication operation efficiency recovery rate, using the following formula: ;in, P recovery Indicates the recovery rate of communication operation efficiency; K validThis indicates the effective number of keystrokes after training, measured in keystrokes per minute. R rts The message retransmission rate is represented as a decimal after training, i.e., the number of retransmitted messages / the total number of messages sent, with a value ranging from 0 to 1. P bl This represents the trainee's baseline operational efficiency under normal historical conditions. Specifically, it involves retrieving operational data from the trainee's past seven consecutive days during periods of calm sea conditions and without high-priority tasks. The average number of effective keystrokes per minute and the average message retransmission rate during these periods are calculated and then substituted into the same formula used for communication operational efficiency recovery rate. The resulting value is the trainee's individual baseline operational efficiency. P bl For example, the communication operation efficiency recovery rate reaches 1.15. S53. Extract the real-time psychological load index of trainees during the training execution period, extract the real-time psychological load index at the start of training and the real-time psychological load index at the end of training; subtract the real-time psychological load index at the end of training from the real-time psychological load index at the start of training to calculate the decrease in physiological indicators. S54. Substitute the communication operation efficiency recovery rate and the decline in physiological indicators into the reward function formula to calculate the comprehensive reward value; the reward function formula is: Among them, Re ward Indicates the overall reward value; α represents the performance weighting parameter; β Represents the physiological weighting parameter; Δ PLI This indicates the degree of decline in physiological indicators; for example, a comprehensive reward value of 0.894. Furthermore, the performance weight parameter α and the physiological weight parameter β are dynamically calculated based on the urgency index of the current job task. Specifically, the system reads the task queue of the communication terminal, counts the proportion of high-priority messages to the total number of messages, uses this proportion as the performance weight parameter α, and limits its minimum value to 0.5; calculates the value of 1 minus α as the physiological weight parameter β; in this embodiment, the proportion of high-priority messages is 60%, so the objective calculation yields α of 0.6 and β of 0.4. S55. The action value matrix in the model is generated by updating the control strategy using the comprehensive reward value; the system uses the Bellman equation to update the values ​​of state-action pairs. ;in, Q ( S t , A t ) represents the current state variable in the action value matrix. S t With the selected initial adaptive training program A t The corresponding value; ηγ represents the learning rate, which is 0.05 in this embodiment; γ represents the discount factor, which is 0.9 in this embodiment; max Q ( S t+1 , a ) represents the maximum expected action value in the next state; the training convergence termination condition of the action value matrix is ​​that within 20 consecutive training cycles, the proportion of the optimal action selection in the same state is ≥90%, and the maximum single-cycle update amount of all elements in the action value matrix is ​​less than 0.001. At this time, the strategy is determined to be converged, and the matrix update can be stopped; or, according to the system operation requirements, the online continuous update mode can be maintained to achieve the lifelong evolution of the strategy. Furthermore, the learning rate η is calculated based on the inverse decay function of the total number of training iterations; specifically, the initial learning rate is set to 0.1, and the current learning rate = initial learning rate / (1 + decay coefficient × cumulative training iterations); in this embodiment, the decay coefficient is set to 0.01; if the current cumulative training iterations are 100, the current learning rate is calculated to be 0.05, ensuring that the model converges quickly in the early stage of training and remains stable in the later stage; The discount factor γ is calculated based on the average duration of the transition in mental load state. Specifically, the system statistically analyzes the average number of time steps required for trainees to recover from severe load to normal state in the historical database. Discount factor = 1 - (1 / average time steps). If the average recovery time steps are 10 training cycles, the discount factor is calculated to be 0.9. The above setting balances the weight of immediate reward and long-term expected reward. The system substitutes the calculated comprehensive reward value into the Bellman equation and updates the action value matrix; through this step, the system achieves feedback-based closed-loop evolution, ensuring the efficient enhancement of people's psychological adaptability in complex environments; The above embodiments calculate the performance recovery rate by extracting effective keystrokes and retransmission frequency, calculate the comprehensive reward value by combining the decline in physiological indicators, and update the action value matrix using the Bellman equation, thereby realizing the closed-loop evolution of the control strategy and ensuring the continuous improvement of system training performance under long-haul conditions.

[0025] Furthermore, a psychological adaptability intelligent regulation system based on physiological feedback is provided to implement the aforementioned psychological adaptability intelligent regulation method based on physiological feedback. The system includes: The multidimensional spatiotemporal synchronization matrix construction module is used to acquire multimodal physiological signal sequences, ship physical environment parameter sequences, and job task load parameter sequences, and perform timestamp alignment to construct a multidimensional spatiotemporal synchronization state matrix. The anti-interference status assessment module is used to perform adaptive filtering calculations to generate a purified physiological feature vector by using the ship's physical environment parameter sequence as a reference input signal, and to output a real-time psychological load index using a spatiotemporal attention status assessment model. The reinforcement learning strategy matching module is used to output an initial adaptive training scheme and an initial training guidance parameter set by matching the action value matrix when the real-time psychological load index is greater than the dynamic baseline warning threshold. The physiological closed-loop training guidance module is used to calculate the actual instantaneous phase of respiration using Hilbert transform, calculate the real-time physiological feedback deviation rate, and generate a dynamic control command sequence using a variable step size optimization algorithm. The performance recovery reward update module is used to calculate the communication operation performance recovery rate and comprehensive reward value, and update the action value matrix; This invention uses the complex physical environment factors of a ship as deterministic noise input, and decouples physical seasickness reactions from real psychological stress through adaptive filtering to establish an anti-interference psychological state assessment benchmark. Based on this, it breaks through the traditional one-way open-loop training mode by introducing instantaneous phase tracking technology based on Hilbert transform to construct a microsecond-level closed-loop feedback between the training-guided rhythm and the actual physiological rhythm of the human body. When rhythmic discontinuity occurs, the system actively retreats to establish physiological resonance and then executes variable-step-size smoothing traction. Combined with a reinforcement learning mechanism, the iterative update of the training strategy is driven by the quantitative results of actual job performance recovery. This invention achieves precise assessment of psychological load under special conditions, solves the rhythm conflict problem caused by passive training, and constructs a closed-loop control system with data-driven and self-evolving capabilities, ensuring the operational efficiency of communication operators.

Claims

1. A method for intelligent regulation of psychological adaptability based on physiological feedback, characterized in that, Includes the following steps: S1. Obtain the multimodal physiological signal sequence, job task load parameter sequence, and ship physical environment parameter sequence of the trainees; S2. Based on the ship's physical environment parameter sequence, perform motion artifact removal and vestibular compensation heart rate subtraction calculation on the multimodal physiological signal sequence to generate purified physiological feature vectors. The purified physiological feature vector and the job task load parameter sequence are input into the spatiotemporal attention state assessment model, and the real-time psychological load index and job stress state classification label are output. S3. Determine whether the real-time psychological load index is greater than the dynamic baseline warning threshold. If so, input the real-time psychological load index and the occupational stress state classification label into the reinforcement learning-based regulation strategy generation model to obtain the initial adaptive training scheme and the initial training guidance parameter set; Otherwise, if the trainee's mental state is deemed normal, proceed to step S1. S4. Execute the initial adaptive training program and continuously determine whether the real-time physiological feedback deviation rate of the trainees is greater than the phase lock threshold. If so, based on the gradient direction of the real-time physiological feedback deviation rate, the variable step size optimization algorithm is used to update the guidance frequency parameter and stimulus intensity parameter in the initial training guidance parameter set in real time, and generate a dynamic control instruction sequence to be sent to the execution terminal; otherwise, the initial training guidance parameter set is kept unchanged, and execution continues until the training time ends. S5. Extract the sequence of job task load parameters of trainees after the training period ends, and calculate the communication operation efficiency recovery rate; based on the communication operation efficiency recovery rate and the decline of physiological indicators during the training period, update the action value matrix of the control strategy generation model through the reward function.

2. The intelligent regulation method for psychological adaptability based on physiological feedback according to claim 1, characterized in that, The specific steps for constructing the multidimensional spatiotemporal synchronization state matrix in S1 include: S11. Collect electrocardiogram signals, skin conductance level signals, and respiratory rhythm signals from trainees to obtain multimodal physiological signal sequences; S12. Collect the ship's pitch acceleration sequence, ship's roll angular velocity sequence, and cabin background noise sound pressure level sequence to obtain the ship's physical environment parameter sequence. S13. Statistically analyze the number of effective keystrokes, message retransmission frequency, and error correction backspace times of trainees within a unit of time to obtain the job task load parameter sequence. S14. Using the absolute time of the network time protocol server as a reference, the cubic spline interpolation algorithm is used to resample the multimodal physiological signal sequences, ship physical environment parameter sequences and job task load parameter sequences with different sampling rates to a unified reference frequency, thereby obtaining a multidimensional spatiotemporal synchronization state matrix.

3. The intelligent regulation method for psychological adaptability based on physiological feedback according to claim 1, characterized in that, The specific steps for generating the purified physiological feature vector in S2 include: S211. Extract the ship pitch acceleration sequence from the aligned ship physical environment parameter sequence as the reference input signal, and extract the electrocardiogram signal from the aligned multimodal physiological signal sequence as the main input signal. S212. Construct a normalized minimum mean square adaptive filter and calculate the error signal between the main input signal and the reference input signal after processing by the filter weight vector. S213. Iteratively update the filter weight vector based on the error signal. When the filter weight vector converges, extract the continuous error signal as the pure electrocardiogram signal stripped of seasickness compensation response. S214. Detect the peak value of the R wave in the pure electrocardiogram signal, calculate the time interval sequence of adjacent R waves, and use an autoregressive model to estimate the power spectral density of the time interval sequence to obtain the sympathetic vagus nerve balance index. The phase awakening peak value of the skin conductance level signal in the multimodal physiological signal sequence was calculated, and the sympathetic vagus nerve balance index and the phase awakening peak value were combined to form the purified physiological feature vector.

4. The intelligent regulation method for psychological adaptability based on physiological feedback according to claim 1, characterized in that, The specific steps in S2 for outputting the real-time psychological load index and occupational stress status classification labels of trainees include: S221. Normalize the purified physiological feature vector and the job task load parameter sequence, and concatenate them into a fused feature tensor. S222. Input the fused feature tensor into the spatiotemporal attention state evaluation model, and extract spatial dimension features from the fused feature tensor through a one-dimensional convolutional layer to generate a local physiological response feature map. S223. The long short-term memory network layer of the spatiotemporal attention state assessment model receives local physiological response feature maps, extracts the cumulative evolution features of psychological stress response in the time dimension, and calculates the hidden layer state sequence. S224. The self-attention mechanism layer of the spatiotemporal attention state evaluation model assigns attention weights to the hidden layer state sequence to generate a weighted hidden layer state sequence. The weighted hidden layer state sequence is input into the fully connected layer; the regression branch of the fully connected layer outputs a continuous real-time psychological load index, and the classification branch of the fully connected layer outputs the job stress state classification label through the Softmax function.

5. The intelligent regulation method for psychological adaptability based on physiological feedback according to claim 1, characterized in that, The specific steps in S3 to determine whether the real-time psychological load index is greater than the dynamic baseline warning threshold include: S311. Retrieve the historical real-time psychological load index of trainees during periods of normal navigation and without communication tasks, and calculate the mean and standard deviation of the historical real-time psychological load index. S312. Add a constant multiple of the standard deviation of the historical real-time psychological load index to the mean of the historical real-time psychological load index to obtain the dynamic baseline warning threshold for trainees. S313. Determine whether the real-time psychological load index at the current time point is greater than the dynamic baseline warning threshold; if so, determine that the trainee is in a state of high psychological load; otherwise, determine that the trainee is in a normal psychological state.

6. The intelligent regulation method for psychological adaptability based on physiological feedback according to claim 1, characterized in that, The specific steps in S3 to obtain the initial adaptive training scheme and the initial training guidance parameter set include: S321. Discretize the real-time psychological load index into discrete values ​​for light load, moderate load, and heavy load states; combine the discrete values ​​of states with the work stress state classification labels to define the current state variable; S322. Read the action value matrix in the control strategy generation model. The row index of the action value matrix corresponds to the current state variable, and the column index of the action value matrix corresponds to all candidate adaptive training schemes in the training action library. S323. Use an ε-greedy strategy to select actions from the action value matrix; generate a random number between zero and one. Determine whether the random number is less than the exploration probability threshold; if so, randomly select a candidate adaptive training program from the training action library as the initial adaptive training program; the initial adaptive training program includes at least one of breathing rhythm traction training, binaural beat auditory training, and progressive muscle relaxation. Otherwise, find the column with the largest value in the row corresponding to the current state variable in the action value matrix, and use the candidate adaptive training scheme corresponding to the column with the largest value as the initial adaptive training scheme. S324. Based on the initial adaptive training program, extract the corresponding initial training guidance parameter set from the parameter mapping table; when the initial adaptive training program includes respiratory rhythm traction training, the initial training guidance parameter set includes the initial scaling period of the visual respiratory halo and the initial beat frequency of the auditory guidance sound.

7. The intelligent regulation method for psychological adaptability based on physiological feedback according to claim 1, characterized in that, The specific steps for calculating the real-time physiological feedback deviation rate of trainees in S4 include: S411. Set a sliding time window; when the initial adaptive training program includes respiratory rhythm traction training, capture the real-time respiratory rhythm signal within the sliding time window. S412. Perform Hilbert transform on the real-time respiratory rhythm signal within the sliding time window, calculate the analytic signal of the real-time respiratory rhythm signal, and extract the imaginary part and the real part of the analytic signal. S413. Calculate the ratio of the imaginary part of the analytic signal to the real part of the analytic signal using the arctangent function to obtain the actual instantaneous breathing phase of the trainee; S414. Extract the instantaneous phase of the visual breathing halo sent to the execution terminal from the initial training guidance parameter set; S415. Calculate the absolute difference between the actual instantaneous breathing phase and the guided instantaneous phase, and divide the absolute difference by twice pi to obtain the real-time physiological feedback deviation rate.

8. The intelligent regulation method for psychological adaptability based on physiological feedback according to claim 7, characterized in that, The specific steps in S4 for generating a dynamic control instruction sequence and sending it to the execution terminal include: S421. Determine whether the real-time physiological feedback deviation rate is greater than the preset phase lock threshold. S422. When the real-time physiological feedback deviation rate is greater than the phase lock threshold, it is determined that the trainee has failed to follow the guided rhythm; calculate the derivative of the actual respiratory instantaneous phase with respect to time to obtain the trainee's actual respiratory rate. S423. Forcefully modify the guidance frequency parameter in the initial training guidance parameter set to the actual breathing frequency, generate a synchronous follow command and send it to the execution terminal, so that the scaling period of the visual breathing halo completely matches the actual breathing frequency of the trainee and establishes a physiological resonance state. S424. After maintaining the preset breathing cycle in the physiological resonance state, the guiding frequency parameter is gradually reduced by optimizing the gradient with variable step size, generating a slow descent traction command and sending it to the execution terminal until the guiding frequency parameter drops to the set deep relaxation target frequency. When the real-time physiological feedback deviation rate is not greater than the preset phase lock threshold, no synchronous following command or descent traction command is generated.

9. The intelligent regulation method for psychological adaptability based on physiological feedback according to claim 1, characterized in that, In S4, when the initial adaptive training program includes progressive muscle relaxation, the rules for calculating the real-time physiological feedback deviation rate and generating instructions are as follows: S431. Acquire electromyographic signals and calculate the current root mean square amplitude of the electromyographic signals within the sliding time window; S432. Extract the initial root mean square amplitude of electromyography (EMG) of the trainee before training, and subtract the current root mean square amplitude from the initial EMG amplitude to obtain the amplitude difference. The muscle relaxation rate is calculated by dividing the amplitude difference by the initial EMG root mean square amplitude; the real-time physiological feedback deviation rate is defined as 100% minus the muscle relaxation rate. S433. When the real-time physiological feedback deviation rate is greater than the set threshold, it is determined that the target muscle group has not reached the relaxation standard; a pause command and a strengthening stimulation command are generated and sent to the execution terminal. The execution terminal stops playing the relaxation instructions for the next muscle group according to the pause command, and repeats the instructions for the current muscle group. It also outputs binaural beat audio according to the reinforcement stimulation command.

10. A psychological adaptability intelligent regulation system based on physiological feedback, used to implement the psychological adaptability intelligent regulation method based on physiological feedback as described in any one of claims 1 to 9; the system comprises: The multidimensional spatiotemporal synchronization matrix construction module is used to acquire multimodal physiological signal sequences, ship physical environment parameter sequences, and job task load parameter sequences, and perform timestamp alignment to construct a multidimensional spatiotemporal synchronization state matrix. The anti-interference status assessment module is used to perform adaptive filtering calculations to generate a purified physiological feature vector by using the ship's physical environment parameter sequence as a reference input signal, and to output a real-time psychological load index using a spatiotemporal attention status assessment model. The reinforcement learning strategy matching module is used to output an initial adaptive training scheme and an initial training guidance parameter set by matching the action value matrix when the real-time psychological load index is greater than the dynamic baseline warning threshold. The physiological closed-loop training guidance module is used to calculate the actual instantaneous phase of respiration using Hilbert transform, calculate the real-time physiological feedback deviation rate, and generate a dynamic control command sequence using a variable step size optimization algorithm. The performance recovery reward update module is used to calculate the communication operation performance recovery rate and comprehensive reward value, and update the action value matrix.

Images