A method for constructing a physiological signal and psychological reaction prediction model under solar radiation

Abstract

The application provides a solar radiation physiological signal and psychological reaction prediction model construction method, comprising the following steps: deploying a multi-channel acquisition system to obtain original multi-channel time series data; performing time series alignment, signal noise reduction and outlier correction to obtain multi-modal synchronous time series data; constructing a solar radiation human body projection model, calculating physiological statistical features and individual activity levels; using the physiological statistical features and a GA-LSTM model, calculating and outputting model labels to obtain model multi-task targets; constructing a prediction model architecture, performing joint training to obtain model training weights; and designing a time lag adjustment layer to perform physiological-mental time alignment to obtain a physiological-mental prediction model. The application can comprehensively improve the prediction accuracy, generalization ability and interpretability of the thermal physiological-mental coupling, and can also solve the time difference between thermal physiology and psychological reaction, and improve the dynamic response capability and applicability of the model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of thermal sensation prediction, in particular to a physiological signal and psychological reaction prediction model construction method under solar radiation. BACKGROUND

[0002] The prediction of skin temperature (physiological signal) and thermal sensation (psychological reaction) under solar radiation refers to the estimation of the temperature change of the skin surface and the psychological state of thermal perception caused thereby by establishing a mathematical model or using a specific technology in the dynamic change environment of solar radiation. It can reveal the interaction mechanism of human body and solar radiation from a scientific perspective and provide quantitative decision-making basis for thermal safety protection, environmental design, health management and the like from an application perspective, especially under the background of global climate warming and increasing demand for outdoor activities. This research has important value for improving human survival comfort and environmental adaptability.

[0003] The existing prediction methods mainly predict through thermal equilibrium models and thermal sensation models, but the existing methods have the following problems: 1) insufficient dynamic response: most classical models are designed based on quasi-steady indoor environments, and when facing rapid changes or intermittent irradiation of solar radiation, the rapid response process of skin temperature cannot be accurately described, and there may be a time difference between thermal physiology and psychological reaction and thermal environment change. 2) weak psychological feedback mechanism: traditional PMV / PPD models often ignore the influence of individual thermal adaptability, instantaneous psychological fluctuations and subjective control. 3) difficult to finely distinguish the thermal response of each part of the body: local thermal sensation and comfort are difficult to model, especially under solar irradiation, the surface temperature distribution of the human body is extremely uneven. SUMMARY

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide a physiological signal and psychological reaction prediction model construction method under solar radiation, which can comprehensively improve the prediction accuracy, generalization ability and interpretability of thermal physiology- psychology coupling, and also solve the time difference between thermal physiology and psychological reaction, and improve the dynamic response ability and applicability of the model.

[0005] To achieve the above purpose, the present application provides the following scheme: a physiological signal and psychological reaction prediction model construction method under solar radiation, comprising:

[0006] Deploy a multi-channel acquisition system for synchronously collecting physiological- psychological- environmental parameters, use the multi-channel acquisition system to collect objective data, combine the objective data with subjective data to obtain original multi-channel time series data;

[0007] Perform time series alignment, signal denoising and outlier correction on the original multi-channel time series data to obtain multi-modal synchronous time series data;

[0008] A solar radiation human body projection model is constructed using the multi-modal synchronous time series data and a human body kinematics model, physiological statistical features are calculated according to the solar radiation human body projection model, and an individual activity level is calculated in combination with environmental variables to eliminate the influence of environmental disturbances;

[0009] A model label is calculated and output using the physiological statistical features and a GA-LSTM model, and a model multi-task target is obtained;

[0010] The GA-LSTM model is optimized to obtain a prediction model architecture, the prediction model architecture, the physiological statistical features and the model multi-task target are jointly trained to obtain model training weights;

[0011] A time delay adjustment layer is designed at the end of the prediction model architecture to perform physiological-mental time alignment, and a physiological-mental prediction model is obtained in combination with the prediction model architecture, the time delay adjustment layer, the model training weights and the model multi-task target.

[0012] Optionally, the multi-channel acquisition system includes a 16-point wireless skin temperature patch, a physiological side sensor assembly, an environmental side sensor assembly, a dual-spectrum solar radiation meter, a VL53L5CX multi-zone ToF sensor and a communication module; the physiological side sensor assembly includes a three-axis inertial measurement unit, a photoplethysmography sensor and an electrocardiogram analog front-end chip, and the environmental side sensor assembly includes an infrared carbon dioxide sensor and a thermal air speed sensor.

[0013] Optionally, a multi-channel acquisition system for synchronously acquiring physiological-mental-environmental parameters is deployed, objective data is acquired using the multi-channel acquisition system, and objective data and subjective data are combined to obtain original multi-channel time series data, including:

[0014] Objective data is obtained by periodically and synchronously acquiring physiological-mental-environmental parameters using the multi-channel acquisition system; wherein the physiological-mental-environmental parameters include skin temperature data, heart rate data, electrocardiogram, IMU attitude data, environmental variables and solar radiation data;

[0015] A user terminal applet is designed to acquire user facial images, facial 52 feature points of the facial images are analyzed using a facial three-dimensional mesh reconstruction algorithm, semi-automatic mental state evaluation is performed, and subjective data is obtained;

[0016] The objective data and the subjective data are aggregated, uploaded and encrypted to obtain original multi-channel time series data.

[0017] Optionally, the original multi-channel time series data is time series aligned, signal denoised and outlier corrected to obtain multi-modal synchronous time series data, including:

[0018] Based on the original multi-channel time series data, a unified time axis is generated, and objective data missing processing rules and subjective data missing processing rules are set, channel interpolation and missing data processing are performed, and first time series data are obtained;

[0019] For objective data in the first time series data, classified signal denoising is performed according to different data types, and second time series data are obtained.

[0020] For objective data in the second time series data, different boundary conditions are set according to different data types to perform physiological-physical boundary correction, and through the way of calculating the median and the absolute median deviation, outlier correction is performed, and multi-modal synchronous time series data are obtained.

[0021] Optionally, for the objective data in the first time series data, classified signal denoising is performed according to different data types, including:

[0022] For skin temperature data, temporary linear filling, smoothing processing, high-frequency noise elimination and data missing recovery processing are performed;

[0023] For heart rate data, 2-order polynomial Savitzky-Golay filtering within a 5-second time window is performed and short-time fluctuation characteristics are retained;

[0024] For heart rate variability, 3-order polynomial Savitzky-Golay filtering based on RR interval within a 7-second time window is performed to highlight low-frequency-high-frequency change trend;

[0025] For electrocardiogram, cvxEDA algorithm is used to separate slow change and burst response, and low-pass filtering is performed on the slow change part for noise reduction;

[0026] For IMU attitude data, median filtering of a 3-second time window is used to perform motion perturbation and spike suppression;

[0027] For environmental variables, median filtering of a 5-second time window is used to eliminate occasional outliers;

[0028] For solar radiation data, direction weighted average is used to dynamically adjust the window.

[0029] Optionally, a solar radiation human body projection model is constructed using the multi-modal synchronous time series data and a human body kinematics model, physiological statistical characteristics are calculated according to the solar radiation human body projection model, and individual activity level is calculated in combination with environmental variables to eliminate environmental disturbance effects, including:

[0030] The IMU attitude data and human body kinematics model are used to model a skeleton chain to obtain an initial bone chain model, a standard human body mesh and 16 skin measurement point positions are defined in the initial bone chain model to calculate skin measurement point spatial coordinates and skin measurement point outer surface normal vectors;

[0031] According to the solar radiation data and GPS data, the solar elevation angle and azimuth angle are calculated to obtain a solar vector, and the solar vector of each skin measurement point is calculated to output an irradiance heat flux matrix to obtain a solar radiation human body projection model;

[0032] According to the skin temperature data, the temperature change rate of each skin measurement point in a time window is calculated to obtain skin temperature dynamic characteristics, according to the heart rate variability, the low-frequency band power and high-frequency band power in the time window are calculated to obtain heart rate frequency domain characteristics, and according to the electrocardiogram, the mean value of the tension component and the phase peak frequency in the time window are calculated to obtain electrocardiogram characteristics.

[0033] According to the environmental variables, the individual activity level is calculated by using the IMU attitude data to eliminate the influence of environmental disturbance; the calculation expression of the individual activity level is:

[0034]

[0035] wherein, is the individual activity level, is the IMU calculated chest angular velocity module value, is the attitude confusion entropy, is the sliding window mean value, is the calibration coefficient.

[0036] Optionally, the physiological statistical characteristics and the GA-LSTM model are used to calculate and output model labels to obtain model multi-task targets, including:

[0037] The GA-LSTM model is obtained by combining a genetic algorithm and a long short-term memory network, and dynamic hyperparameters are introduced into the GA-LSTM model; the dynamic hyperparameters include a physiological variable prediction lead time and a physiological-mental feeling response time difference;

[0038] According to the dynamic hyperparameters, the physiological statistical characteristics are used to calculate skin part temperature targets, skin uniform temperature targets, subjective cold and hot feeling targets and comfort time alignment targets of the GA-LSTM model to obtain model labels, and the model labels at each time are structured and output to obtain model multi-task targets.

[0039] Optionally, the GA-LSTM model is optimized to obtain a prediction model architecture, and the prediction model architecture, the physiological statistical features, and the model multi-task target are jointly trained to obtain model training weights, including:

[0040] A genetic code genome of the GA-LSTM model is defined, and a long short-term memory network, a multi-attention mechanism, and a multi-output head composite neural network structure are fused according to the genetic code genome to generate an initial model architecture, and the initial model architecture is iteratively trained to obtain a prediction model architecture;

[0041] A multi-task joint loss function is defined, and a dynamic weight adjustment algorithm is used to automatically re-label the gradient norm of each task to complete the design of a multi-task joint loss dynamic weighting strategy; the calculation expression of the multi-task joint loss function is:

[0042]

[0043] wherein, is the mean square error of the skin 16-point temperature, is the skin average temperature, is the subjective cold and hot sensation ordered cross-entropy, is the subjective psychological comfort mean square error, all are dynamic weights;

[0044] AdamW is selected, and adaptive momentum and parameter decorrelation techniques are fused to complete the design of a regulation mechanism, and then the prediction model architecture, the physiological statistical features, and the model multi-task target are jointly trained based on the regulation mechanism and the multi-task joint loss dynamic weighting strategy to obtain an optimal epoch parameter.

[0045] Optionally, the genetic code genome includes a first genome for input window length, a second genome for defining LSTM layer number, a third genome for defining hidden size of each layer, a fourth genome for defining Attentionflag, a fifth genome for defining Feature subset mask, a sixth genome for synchronously delaying the dynamic hyperparameters, and a seventh genome composed of an optimizer, a learning rate, Dropout, and multi-task loss initial weight.

[0046] Optionally, a time delay adjustment layer is designed at the end of the prediction model architecture to perform physiological-mental time alignment, and then a physiological-mental prediction model is obtained in combination of the prediction model architecture, the time delay adjustment layer, the model training weights, and the model multi-task target, including:

[0047] Using the jointly trained prediction model architecture, stepwise sliding window inference is performed on the validation set to obtain the predicted subjective hot and cold and the actual subjective hot and cold corresponding to the predicted subjective hot and cold.

[0048] Using a sequence global alignment algorithm, dynamic time alignment is performed on the predicted subjective hot / cold temperatures and the actual subjective hot / cold temperatures, respectively, to calculate the optimal average alignment path and global offset, thus obtaining the optimal alignment time; the expression for calculating the optimal alignment time is:

[0049]

[0050] in, This is the optimal overall time delay constant compensation amount. In order to be in Search within the interval for the time delay constant compensation that minimizes the loss. For the model in The subjective outcome of the prediction at any given moment. Subjective label data that was actually collected;

[0051] By comparing the optimal alignment time with the physiological-psychological response time difference, a correction amount is obtained. Based on the correction amount, the predicted subjective cold and heat is shifted over time on a unified time axis to complete the design of the time lag adjustment layer.

[0052] By combining the prediction model architecture, the time delay adjustment layer, the model training weights, and the model multi-task objective, a physiological-psychological prediction model is obtained.

[0053] This invention discloses the following technical effects by providing a method for constructing a predictive model of physiological signals and psychological responses under solar radiation:

[0054] 1. Comprehensive and reliable data: It can simultaneously acquire 16 points of skin temperature, physiological signals, environmental signals, solar radiation, three-dimensional facial data and subjective data, realizing the simultaneous recording of multimodal data of objective data measurement and subjective data self-evaluation, providing an effective and reliable data foundation.

[0055] 2. High-dimensional feature extraction: 1) By constructing a solar radiation human body projection model, a 3D skeleton can be mapped to a dynamic solar vector in real time, and the actual photothermal exposure of skin points can be physically calculated, which is more accurate than simplified algorithms such as total radiation. 2) Through comprehensive extraction of physiological features, the model's generalization ability and prediction accuracy are enhanced. 3) By calculating individual activity levels and combining IMU angular velocity + attitude entropy, it can resist environmental disturbances and improve the correlation between physiological and psychological predictions.

[0056] 3. High model adaptability: 1) By setting hyperparameters, the physiological- psychological time difference and prediction step can be automatically searched and embedded as adjustable parameters in optimization, matching possible subjective lag characteristics. 2) By structuring a multi-task objective, a multi-task label is constructed, so that the model can not only finely predict the future temperature of each part of the skin, but also realize synchronous adaptive prediction of subjective cold-heat sensation-comfort, support multi-task collaborative prediction, and the model has stronger generalization ability.

[0057] 4. Good prediction performance and high accuracy: Based on the GA-LSTM model structure, 1) By genetic coding, the network structure, feature subset, time window, synchronization parameter and multi-task loss weight can be globally described to automatically obtain the optimal task synchronization and structure under the current data distribution. 2) Through global optimization of structure-feature-time lag, it is no longer dependent on artificial experience, greatly reducing the burden of artificial parameter adjustment, and avoiding human error or bias. 3) Through adaptive weight optimization of multi-task loss, the multi-objective can be efficiently weighed, and the non-synchronous multi-branch output is more friendly, which improves the model robustness and portability under multi-group and strong scene disturbance.

[0058] 5. Physiological-psychological time alignment: By deploying the time lag adjustment layer, the best physiological-psychological overall time lag is automatically inferred, the model output is corrected backward / forward, the subjective output and individual experience are accurately synchronized, and the problem that the thermal physiology and psychological response may have a time difference with the thermal environment change is solved.

[0059] The technical solutions of the present application will be further described in detail below with the help of the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0061] Figure 1 The method flowchart provided for the embodiments of the present application is shown in the figure.

[0062] Figure 2 The architecture diagram of the multi-channel acquisition system provided for the embodiments of the present application is shown in the figure.

[0063] Figure 3 The model joint training flowchart provided for the embodiments of the present application is shown in the figure. DETAILED DESCRIPTION

[0064] With reference to the drawings and specific embodiments, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0065] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0066] As shown in the drawings, Figure 1 The present application provides a physiological signal and psychological reaction prediction model construction method under solar radiation, characterized by comprising the following steps:

[0067] Step 1, deploy a multi-channel acquisition system for synchronously acquiring physiological- psychological- environmental parameters, use the multi-channel acquisition system to acquire objective data, combine the objective data with subjective data to obtain original multi-channel time series data. Specifically, it includes:

[0068] As shown in the drawings, Figure 2 The multi-channel acquisition system comprises:

[0069] 1) 16-point wireless skin temperature patch: each measurement point is uniformly distributed on the forehead, chest, shoulder, arm, back of hand, leg and foot, etc.

[0070] 2) Physiological side sensor assembly:

[0071] Three-axis inertial measurement unit: three-axis IMU, high-stability Euler angle output, 6-axis sampling;

[0072] Photoelectric plethysmogram sensor: PPG, heart rate and HRV acquisition, light path structure prevents environmental interference;

[0073] Electrocardiogram analog front-end chip: EDA, 128Hz dynamic skin electric sensing;

[0074] 3) Environmental side sensor assembly:

[0075] Infrared carbon dioxide sensor (CO2+ temperature and humidity), thermal wind speed sensor:

[0076] 4) Dual-spectrum solar radiation meter: distinguish between global and direct two types of different bandwidth data;

[0077] 5) VL53L5CX multi-zone ToF sensor: arranged as a hemispherical shell, fixed on the head or shoulder with a flexible bracket, dynamically acquiring incident light intensity in each direction of the sky;

[0078] 6) Communication module: STM32H7 level controller is adopted, and each sensor data is collected at high speed through SPI and integrated, and is packaged and broadcast to the gateway through BLE.

[0079] 1.1 Periodically synchronously collecting physiological- psychological- environmental parameters by using the multi-channel acquisition system to obtain objective data; wherein the physiological- psychological- environmental parameters include skin temperature data, heart rate data, electrocardiogram, IMU attitude data, environmental variables and solar radiation data.

[0080] 1.2 By designing a user-side applet, integrating a front camera to collect user facial images, using a facial three-dimensional grid reconstruction algorithm to analyze facial 52 feature points of the facial images, performing semi-automatic psychological state evaluation to obtain subjective data.

[0081] 1.3 For the objective data and the subjective data, data aggregation, uploading and encryption are performed to obtain original multi-channel time series data.

[0082] Step 2, time series alignment, signal denoising and outlier correction are performed on the original multi-channel time series data to obtain multi-modal synchronous time series data. Specifically, it includes:

[0083] 2.1 Based on the original multi-channel time series data, a unified time axis is generated, and objective data missing processing rules and subjective data missing processing rules are set, channel interpolation and missing processing are performed to obtain first time series data.

[0084] Objective data missing processing rules: continuous missing ≤ 3 seconds, linear interpolation, continuous missing > 3 seconds, marked as NaN, and a mask (1 for missing, 0 for valid) is also generated;

[0085] Subjective data missing processing rules: only fill the original value at the collection point (every 2 minutes / each subjective feedback), and fill forward at the non-collection point. If the whole point subjective data is also missing / invalid, it is strictly marked as NaN without interpolation. If the feedback interval is far more than 2 minutes, all the intervals exceeding the interval are marked as NaN to avoid incorrect psychological mapping.

[0086] 2.2 For the objective data in the first time series data, classified signal denoising is performed according to different data types to obtain second time series data; specifically, it includes:

[0087] For skin temperature data T, temporary linear filling, smoothing processing, high-frequency noise elimination and data missing recovery processing are performed;

[0088] For heart rate data HR, 2-order polynomial Savitzky-Golay filtering within a 5-second time window is performed and short-time fluctuation characteristics are retained;

[0089] For heart rate variability (HRV), a 3rd order polynomial Savitzky-Golay filter is applied to the 7s window of RR intervals to highlight the low-high frequency trends.

[0090] For electrodermal activity (EDA), the cvxEDA algorithm is used to separate the slow and phasic responses, and a low-pass filter is applied to the slow component for noise reduction.

[0091] For IMU pose data, a 3s window median filter is applied to suppress small motion perturbations and spikes.

[0092] For environmental variables, a 5s window median filter is applied to remove occasional outliers.

[0093] For solar radiation data, a directionally weighted average is applied to dynamically adjust the window.

[0094] 2.3 For objective data in the second time series data, different boundary conditions are set according to different data types to perform physiological-physical boundary correction, and through the calculation of median and absolute median deviation, outlier correction is performed to obtain multi-modal synchronous time series data.

[0095] Boundary conditions, for example:

[0096] Skin temperature: allowed range 28-42℃, instantaneous change rate <0.5℃ / s.

[0097] HR: allowed 30-200bpm, instantaneous change <5bpm / s.

[0098] Ambient temperature: 0-50℃; total solar irradiance 0-1367W / m 2 .

[0099] As long as the boundary is triggered, it is recorded as an outlier.

[0100] Step 3, as shown in Figure 3 , a solar radiation human projection model is constructed using the multi-modal synchronous time series data and human kinematics model, and physiological statistical features are calculated according to the solar radiation human projection model, and individual activity level is calculated in combination with environmental variables to eliminate the influence of environmental disturbance. Specifically, it includes:

[0101] 3.1 A skeleton chain model is constructed using the IMU pose data and human kinematics model to obtain an initial bone chain model, and a standard human mesh and 16 skin measurement point positions are defined in the initial bone chain model to calculate skin measurement point spatial coordinates and skin measurement point outer surface normal vectors.

[0102] 3.2 Calculate the solar elevation and azimuth angle according to the solar radiation data and GPS data, get the solar vector, and calculate the solar vector of each skin measurement point to output the irradiance heat flux matrix, and get the solar radiation human body projection model.

[0103] 3.3 According to the skin temperature data (16 skin measurement points are extracted independently), calculate the temperature change rate of each skin measurement point in the time window to get the skin temperature dynamic characteristics, calculate the low frequency band power and high frequency band power in the time window according to the heart rate variability to get the heart rate frequency domain characteristics, and calculate the tension component mean and phase peak frequency in the time window according to the electrocardiogram to get the electrocardiogram characteristics.

[0104] 3.4 According to the environmental variables, calculate the individual activity level using the IMU attitude data to eliminate the influence of environmental disturbance; the calculation expression of the individual activity level is:

[0105]

[0106] wherein, is the individual activity level, is the IMU calculated chest angular velocity module value, reflecting the dynamic intensity, is the attitude confusion entropy (quantitatively describing activity diversity), is the sliding window mean (30 seconds), is the calibration coefficient.

[0107] Step 4, as shown in Figure 3 , using physiological statistical characteristics and GA-LSTM model, calculating and outputting model labels to get model multi-task target. Specifically including:

[0108] 4.1 Combine genetic algorithm and long short-term memory network to get GA-LSTM model, and introduce dynamic hyperparameters in the GA-LSTM model.

[0109] The dynamic hyperparameters include:

[0110] Physiological variable prediction lead time: that is, the model input sequence predicts the physiological state at t+Δ time;

[0111] Physiological-psychological feeling response time difference: modeling the natural time lag relationship between subjective cold / hot / comfort psychological response and physiological change.

[0112] GA-LSTM allows dynamic hyperparameters to be continuous search parameters, rather than traditional fixed windows, and the model adaptively captures the optimal information lag structure when evolving / training.

[0113] 4.2 According to the dynamic hyperparameters, the skin site temperature target, the skin uniform temperature target, the subjective cold and hot feeling target, and the comfort time alignment target of the GA-LSTM model are calculated using the physiological statistical features, the model label is obtained, and the model multi-task target is obtained by structuring the output of the model label at each time.

[0114] Step 5, as shown in Figure 3 The GA-LSTM model is optimized to obtain a prediction model architecture, and the prediction model architecture, the physiological statistical features, and the model multi-task target are jointly trained to obtain model training weights. Specifically, it includes:

[0115] 5.1 Define the genetic code genome of the GA-LSTM model, and according to the genetic code genome, fuse the long short-term memory network, multi-attention mechanism, and multi-output head composite neural network structure to generate an initial model architecture, and then iteratively train the initial model architecture to obtain a prediction model architecture.

[0116] The genetic code genome includes:

[0117] The first genome g1: the input window length, the longer the step, the more distant it captures;

[0118] The second genome g2: the number of LSTM layers, the number of layers: 1-4;

[0119] The third genome g3: the hidden size of each layer, 32-256, increasing by 32, different layers can be different;

[0120] The fourth genome g4: Attention flag, control whether to introduce a multi-head self-attention sublayer;

[0121] The fifth genome g5: Feature subset mask;

[0122] The sixth genome g6: synchronization delay the dynamic hyperparameters;

[0123] The seventh genome g7: optimizer, learning rate, Dropout, multi-task loss initial weight.

[0124] 5.2 Define a multi-task joint loss function, and use a dynamic weight adjustment algorithm to automatically re-label the gradient norm of each task to complete the design of the multi-task joint loss dynamic weighting strategy; the calculation expression of the multi-task joint loss function is:

[0125]

[0126] wherein, is the mean square error of the skin 16-point temperature, To equalize skin temperature, For the ordered cross-entropy of subjective hot and cold sensations, The mean square error represents the subjective psychological comfort level. All weights are dynamic.

[0127] 5.3 AdamW is selected and integrated with adaptive momentum and parameter decorrelation techniques to complete the design of the regulation mechanism. Then, based on the regulation mechanism and the dynamic weighting strategy of multi-task joint loss, the prediction model architecture, the physiological statistical features and the multi-task objective of the model are jointly trained to obtain the optimal epoch parameters, i.e., the model training weights.

[0128] Step 6: Design a time-lag adjustment layer at the end of the prediction model architecture to perform physiological-psychological time alignment. Then, combine the prediction model architecture, the time-lag adjustment layer, the model training weights, and the model's multi-task objective to obtain the physiological-psychological prediction model. Specifically, this includes:

[0129] 6.1 Using the jointly trained prediction model architecture, progressive sliding window inference is performed on the validation set to obtain the predicted subjective temperature and the corresponding actual subjective temperature. The predicted subjective temperature and the actual subjective temperature include the temperature observation sequence and the comfort sequence.

[0130] 6.2 Using a sequence global alignment algorithm, dynamic time alignment is performed on the predicted subjective hot / cold weather and the actual subjective hot / cold weather, respectively, to calculate the optimal average alignment path and global offset, and obtain the optimal alignment time; the expression for calculating the optimal alignment time is:

[0131]

[0132] in, This is the optimal overall time delay constant compensation amount. In order to be in Search within the interval for the time delay constant compensation that minimizes the loss. For the model in The subjective outcome of the prediction at any given moment. The data consists of real, collected subjective labels.

[0133] 6.3 By comparing the optimal alignment time with the physiological-psychological response time difference, a correction amount is obtained. Based on the correction amount, the predicted subjective cold and heat is time-shifted on a unified time axis to complete the design of the time lag adjustment layer.

[0134] 6.4 Combining the prediction model architecture, the time delay adjustment layer, the model training weights, and the model multi-task objective, a physiological signal-psychological response prediction model is obtained.

[0135] Therefore, the present application can comprehensively improve the prediction accuracy, generalization ability and interpretability of the thermal physiological-psychological coupling, and can also solve the time difference between thermal physiology and psychological response, and improve the dynamic response ability and applicability of the model.

[0136] The various embodiments are described in a progressive manner in the specification, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be mutually referred to.

[0137] The principles and implementation manners of the present application are described by applying specific examples in the specification, and the above description of the examples is only used to help understand the method of the present application and its core idea; meanwhile, for those skilled in the art, the specific implementation manners and application ranges can be changed according to the idea of the present application. In summary, the content of the specification should not be understood as a limitation on the present application.

Claims

1. A method for constructing a predictive model of physiological signals and psychological responses under solar radiation, characterized in that, include: Deploy a multi-channel acquisition system to synchronously collect physiological, psychological, and environmental parameters. Use the multi-channel acquisition system to collect objective data, and combine the objective data with subjective data to obtain raw multi-channel time-series data. The original multi-channel timing data is subjected to timing alignment, signal denoising and outlier correction to obtain multimodal synchronous timing data; A solar radiation human body projection model is constructed using the multimodal synchronous time-series data and human kinematics model. Based on the solar radiation human body projection model, physiological statistical characteristics are calculated, and then combined with environmental variables, individual activity levels are calculated to eliminate the influence of environmental disturbances. Using physiological statistical features and the GA-LSTM model, the model labels are calculated and output to obtain the model's multi-task objectives; The GA-LSTM model is optimized to obtain a prediction model architecture. The prediction model architecture, the physiological statistical features, and the model's multi-task objective are then used for joint training to obtain the model training weights. A time-delay adjustment layer is designed at the end of the prediction model architecture to perform physiological-psychological time alignment. Then, by combining the prediction model architecture, the time-delay adjustment layer, the model training weights, and the model multi-task objective, a physiological-psychological prediction model is obtained. Using physiological statistical features and the GA-LSTM model, model labels are calculated and output to obtain the model's multi-task objectives, including: By combining genetic algorithms and long short-term memory networks, a GA-LSTM model is obtained, and dynamic hyperparameters are introduced into the GA-LSTM model. The dynamic hyperparameters include the prediction lead of physiological variables and the physiological-psychological response time difference. Based on the dynamic hyperparameters and using the physiological statistical features, the skin temperature target, skin average temperature target, subjective hot and cold sensation target, and comfort time alignment target of the GA-LSTM model are calculated to obtain model labels. The model labels at each time step are then output in a structured manner to obtain the model multi-task objectives. A time-lag adjustment layer is designed at the end of the prediction model architecture to perform physiological-psychological time alignment. Then, by combining the prediction model architecture, the time-lag adjustment layer, the model training weights, and the model's multi-task objective, a physiological-psychological prediction model is obtained, including: Using the jointly trained prediction model architecture, stepwise sliding window inference is performed on the validation set to obtain the predicted subjective hot and cold and the actual subjective hot and cold corresponding to the predicted subjective hot and cold. Using a sequence global alignment algorithm, dynamic time alignment is performed on the predicted subjective hot / cold temperatures and the actual subjective hot / cold temperatures, respectively, to calculate the optimal average alignment path and global offset, thus obtaining the optimal alignment time; the expression for calculating the optimal alignment time is: in, This is the optimal overall time delay constant compensation amount. In order to be in Search within the interval for the time delay constant compensation that minimizes the loss. For the model in The subjective outcome of the prediction at any given moment. Subjective label data that was actually collected; By comparing the optimal alignment time with the physiological-psychological response time difference, a correction amount is obtained. Based on the correction amount, the predicted subjective cold and heat is shifted over time on a unified time axis to complete the design of the time lag adjustment layer. By combining the prediction model architecture, the time delay adjustment layer, the model training weights, and the model multi-task objective, a physiological signal-psychological response prediction model is obtained.

2. The method for constructing a predictive model of physiological signals and psychological responses under solar radiation according to claim 1, characterized in that, The multi-channel acquisition system includes a 16-point wireless skin temperature patch, a physiological side sensor assembly, an environmental side sensor assembly, a dual-spectrum solar radiometer, a VL53L5CX multi-zone ToF sensor, and a communication module. The physiological side sensor assembly includes a triaxial inertial measurement unit, a photoplethysmography (PPG) sensor, and an electrocardiogram (ECG) simulation front-end chip. The environmental side sensor assembly includes an infrared carbon dioxide sensor and a thermal anemometer.

3. The method for constructing a predictive model of physiological signals and psychological responses under solar radiation according to claim 2, characterized in that, A multi-channel data acquisition system is deployed to synchronously collect physiological, psychological, and environmental parameters. This system is used to collect objective data, and by combining objective and subjective data, raw multi-channel time-series data is obtained, including: The multi-channel acquisition system is used to periodically and synchronously acquire physiological, psychological, and environmental parameters to obtain objective data; wherein, the physiological, psychological, and environmental parameters include skin temperature data, heart rate data, electrocardiogram, IMU posture data, environmental variables, and solar radiation data; By designing a user-facing mini-program, facial images of users are collected. A facial 3D mesh reconstruction algorithm is used to analyze 52 facial feature points in the facial images to conduct a semi-automatic psychological state assessment and obtain subjective data. The objective data and the subjective data are aggregated, uploaded, and encrypted to obtain the original multi-channel time-series data.

4. The method for constructing a predictive model of physiological signals and psychological responses under solar radiation according to claim 3, characterized in that, The original multi-channel timing data is subjected to timing alignment, signal denoising, and outlier correction to obtain multimodal synchronization timing data, including: Based on the original multi-channel time series data, a unified time axis is generated, and objective data missing processing rules and subjective data missing processing rules are set. Channel interpolation and missing processing are performed to obtain the first time series data. For the objective data in the first time series data, classification signal denoising is performed according to different data types to obtain the second time series data; For the objective data in the second time series data, different boundary conditions are set according to different data types to perform physiological-physical boundary correction, and outlier correction is performed by calculating the median and absolute median deviation to obtain multimodal synchronous time series data.

5. The method for constructing a predictive model of physiological signals and psychological responses under solar radiation according to claim 4, characterized in that, For the objective data in the first time series data, classification signal denoising is performed according to different data types, including: For skin temperature data, temporary linear filling, smoothing, high-frequency noise elimination, and data loss recovery are performed. For heart rate data, a second-order polynomial Savitzky-Golay filter is applied within a 5-second time window to preserve short-term fluctuation characteristics; To address heart rate variability, a third-order polynomial Savitzky-Golay filter was applied within a 7-second time window based on the RR interval to highlight the low-frequency to high-frequency variation trend. For electrocardiograms, the cvxEDA algorithm is used to separate slow-varying and burst responses, and low-pass filtering is applied to the slow-varying part for noise reduction. For IMU attitude data, median filtering with a 3-second time window is used to suppress small motion perturbations and spikes. For environmental variables, median filtering with a 5-second time window is used to remove occasional outliers; For solar radiation data, a directional weighted average is used to dynamically adjust the window.

6. The method for constructing a predictive model of physiological signals and psychological responses under solar radiation according to claim 5, characterized in that, A solar radiation human body projection model is constructed using the aforementioned multimodal synchronous time-series data and human kinematic model. Based on this model, physiological statistical characteristics are calculated, and then, combined with environmental variables, individual activity levels are calculated to eliminate the influence of environmental disturbances, including: Using the IMU attitude data and human kinematics model, skeletal chain modeling is performed to obtain an initial skeletal chain model. In the initial skeletal chain model, a standard human body mesh and 16 skin measurement point positions are defined to calculate the spatial coordinates of the skin measurement points and the normal vector of the outer surface of the skin measurement points. The solar altitude angle and azimuth angle are calculated based on the solar radiation data and GPS data to obtain the solar vector. The solar vector of each skin measuring point is also calculated to output the irradiance heat flux matrix and obtain the solar radiation human body projection model. Based on the skin temperature data, the rate of temperature change at each skin measuring point within the time window is calculated to obtain the dynamic characteristics of skin temperature. Based on the heart rate variability, the low-frequency band power and high-frequency band power within the time window are calculated to obtain the heart rate frequency domain characteristics. Then, based on the electrocardiogram, the mean value of the tension component and the peak frequency of the phase within the time window are calculated to obtain the electrocardiogram characteristics. Based on the environmental variables, the individual activity level is calculated using the IMU attitude data to eliminate the influence of environmental disturbances; the expression for calculating the individual activity level is: in, For individual activity level, The magnitude of the chest angular velocity calculated by the IMU. For attitude disorder entropy, The mean of the sliding window. These are calibration coefficients.

7. The method for constructing a predictive model of physiological signals and psychological responses under solar radiation according to claim 1, characterized in that, The GA-LSTM model is optimized to obtain a prediction model architecture. Joint training is then performed using this prediction model architecture, the physiological statistical features, and the model's multi-task objective to obtain model training weights, including: Define the genetic coding genome of the GA-LSTM model, and based on the genetic coding genome, fuse the long short-term memory network, the multi-attention mechanism, and the multi-output head composite neural network structure to generate an initial model architecture. Then, iteratively train the initial model architecture to obtain the prediction model architecture. A multi-task joint loss function is defined, and a dynamic weight adjustment algorithm is used to automatically rescale the gradient norm of each task, thus completing the design of a dynamic weighting strategy for the multi-task joint loss. The calculation expression of the multi-task joint loss function is as follows: in, The mean square error of skin temperature at 16 points. To even out skin temperature, For the ordered cross-entropy of subjective hot and cold sensations, The mean square error represents the subjective psychological comfort level. All are dynamic weights; AdamW was selected, and adaptive momentum and parameter decorrelation techniques were integrated to complete the design of the regulation mechanism. Then, based on the regulation mechanism and the multi-task joint loss dynamic weighting strategy, the prediction model architecture, the physiological statistical features, and the model multi-task objectives were jointly trained to obtain the optimal epoch parameters.

8. The method for constructing a predictive model of physiological signals and psychological responses under solar radiation according to claim 7, characterized in that, The genetic coding genome includes a first genome for the input window length, a second genome for defining the number of LSTM layers, a third genome for defining the hidden size of each layer, a fourth genome for defining the attention flag, a fifth genome for defining the feature subset mask, a sixth genome for synchronizing and delaying the dynamic hyperparameters, and a seventh genome consisting of the optimizer, learning rate, Dropout, and initial weights for the multi-task loss.

Images