A wearable device-based dynamic monitoring method for health risks of professional groups

Abstract

The application provides a wearable device-based professional group health risk dynamic monitoring method, which synchronously collects physiological signals such as heart rate variability, skin electric response and blood oxygen saturation, and environmental parameters such as temperature and humidity, noise and air pressure, performs signal filtering, normalization and environmental discretization coding, and dynamically allocates physiological feature channel weights based on the constructed context gated attention network; time series modeling is performed in combination with a bidirectional LSTM model, an individual fatigue probability is output, environmental factors and fatigue score contribution degrees are associated by using a class activation mapping, and a hierarchical early warning and semantic explanation report is adaptively triggered; a dynamic feedback optimization mechanism can continuously incrementally train the model according to user feedback, improve recognition accuracy and environmental adaptability, and the application realizes high robustness, interpretable fatigue state recognition and intelligent early warning under the driving of the environment.

Description

Technical Field

[0001] This invention relates to the field of wearable health monitoring and fatigue recognition technology, and in particular to a method for dynamic monitoring of health risks of occupational groups based on wearable devices. Background Technology

[0002] In the current field of occupational health monitoring and fatigue identification, wearable devices are widely used to continuously collect users' multimodal physiological signals, such as heart rate variability (HRV), electrical skin response (EDA), and blood oxygen saturation (SpO2), to assess individual health status and fatigue levels. Mainstream solutions primarily focus on static physiological signal analysis, emphasizing indicators such as heart rate, eye movement, electromyography, or body movement. This is achieved through threshold discrimination algorithms, machine learning-based feature classification, or trend inference using temporal neural networks, resulting in automatic identification and graded early warning of fatigue states. In recent years, some advanced technologies have attempted to introduce deep learning algorithms (such as LSTM and GRU) to improve temporal modeling capabilities and optimize the prediction performance of short-term and long-term fatigue fluctuations. Trend-wise, wearable device hardware is evolving towards multimodal sensing, high integration, and low power consumption, while health big data platforms support remote health management for individuals and groups. To adapt to complex work scenarios, some technologies have gradually incorporated environmental parameter monitoring modules, such as temperature, humidity, noise, and air pressure sensors, to record the basic state of the external environment. However, current mainstream dynamic health risk monitoring and fatigue assessment technologies generally only use environmental parameters as auxiliary or background information, failing to delve into their interaction and modulation relationship with physiological signals. Environmental and physiological multidimensional signals are often simply spliced ​​together as input, or fixed weights are used to influence the judgment process. Traditional Bayesian inference, multi-source evidence fusion, or fixed-weighted methods, while possessing a certain degree of generalization, have limited ability to model the fatigue impact mechanisms of complex environments and work dynamics, making it difficult to dynamically and accurately reflect the actual effect of environmental changes on fatigue assessment. Current representative technologies can be applied to high-risk occupational scenarios such as industrial inspection, coal mining, and high-altitude operations. By wearing health monitoring devices, basic physiological indicators and static environmental conditions of workers are periodically collected, thereby supporting on-the-job safety management and early warning. However, due to the lack of modeling of the complex relationship between environmental context—such as special working conditions like high temperature, strong noise / low air pressure, etc.—and users' physiological responses, the accuracy of fatigue recognition under high occupational stress, variable environment, or sudden working conditions is significantly reduced. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a method for dynamic monitoring of health risks of occupational groups based on wearable devices.

[0004] The technical solution of this invention is implemented as follows: A method for dynamic monitoring of health risks among occupational groups based on wearable devices, comprising: S1: The wearable device synchronously collects physiological signals such as heart rate variability (HRV), skin conductance (EDA), and blood oxygen saturation (SpO2) through its built-in multimodal sensors, and acquires environmental parameter data from temperature and humidity sensors, noise meters, and barometers in real time to form the original physiological signal sequence and environmental parameter sequence. S2: Perform sliding time window segmentation normalization on the physiological signal sequence, perform moving average trend smoothing and discretization encoding on the environmental parameter sequence, and generate a structured physiological feature matrix and an environmental context vector. S3: Construct a Context Gated Attention Network (CGAN), which consists of three layers: the input layer receives the environment context vector, the middle gating layer generates a channel weight matrix through the Sigmoid activation function, and the output layer generates dynamic attention weights based on Softmax normalization. The weight matrix of the gating layer is iteratively optimized through backpropagation based on the cross-entropy loss of the environment vector and the labeled fatigue state. S4: Input the physiological feature matrix and the environmental context vector into the trained CGAN to generate physiological feature channel attention weights that are dynamically associated with environmental parameters, and output the weighted dynamic physiological feature matrix. S5: Input the dynamic physiological feature matrix into a bidirectional LSTM network for time-series modeling, extract time-dependent features through gated recurrent units of forward and backward hidden layers, and output an individualized fatigue probability sequence. S6: Based on the fatigue probability sequence and the gradient backpropagation parameters of the LSTM network, calculate the class activation mapping weights of each physiological feature channel, generate a heat map of the correlation between environmental parameters and physiological indicators, and label the contribution distribution of environmental factors such as high temperature and high noise to the fatigue score. S7: Determine whether the fatigue probability sequence exceeds a preset threshold. If it does, trigger a graded warning signal and generate a semantic explanation report containing environmental context labels. The report contains a structured description of 'increased HRV fluctuations under high temperature conditions leading to an increase in fatigue scores'. S8: Construct a dynamic feedback optimization mechanism. Based on the environment-physiology correlation heatmap data of user interaction feedback, incrementally train the gating layer parameters of the CGAN network and update the attention weight generation model.

[0005] The present invention provides a method for dynamic monitoring of health risks among occupational groups based on wearable devices, which has the following beneficial effects: (1) This invention constructs a context-gated attention network (CGAN) to achieve dynamic collaborative modeling of physiological signals and external environmental context, which significantly improves the accuracy and robustness of fatigue state identification under complex working conditions. Traditional methods usually use fixed weights to fuse multi-source physiological data or rely on prior rules to remove environmental interference, which is difficult to adapt to dynamic and changing actual working scenarios. The CGAN mechanism proposed in this invention can automatically identify key environmental triggering conditions in a learnable way and generate a time-varying attention weight matrix accordingly, so as to achieve adaptive weighted regulation of different physiological feature channels. For example, in typical stress environments such as high temperature and high noise, the model can actively enhance the attention to sensitive indicators such as heart rate variability (HRV), skin conductance (EDA) and blink frequency, so as to more accurately capture early fatigue signs induced by environmental pressure. This mechanism effectively overcomes the evaluation lag and misjudgment problems caused by static modeling, and has stronger environmental adaptability and prediction sensitivity in high-risk occupational scenarios such as power inspection and mine exploration. (2) By introducing a technical architecture that combines bidirectional LSTM temporal modeling with gradient class activation mapping (Grad-CAM), this invention not only achieves continuous dynamic tracking of individualized fatigue trends, but also significantly enhances the interpretability and decision transparency of the model output, providing managers with intuitive and reliable risk assessment basis. This invention can capture the evolution path of fatigue state in the time dimension and generate semantic explanation reports in combination with environmental context, such as "heart rate fluctuations increase under high temperature environment, leading to an increase in fatigue score" and "continuous high-intensity noise causes autonomic nervous system disorder". Such visualization analysis results not only help on-site personnel to take timely intervention measures, but also provide the background monitoring platform with multi-dimensional display tools such as environmental-physiological interaction heatmaps, supporting cross-time period and cross-regional health risk comparison analysis. In addition, the entire processing flow adopts a lightweight neural network design, which can realize real-time inference on edge devices, avoiding the privacy leakage risk and communication delay problem caused by uploading raw physiological data to the cloud, and meeting the deployment requirements of high-security industrial scenarios. Attached Figure Description

[0006] Figure 1 This is a flowchart of a method for dynamic monitoring of health risks of occupational groups based on wearable devices according to the present invention; Figure 2 This is a sub-flowchart of a method for dynamic monitoring of health risks of occupational groups based on wearable devices according to the present invention; Figure 3 This is another sub-flowchart of a method for dynamic monitoring of health risks of occupational groups based on wearable devices according to the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0009] like Figure 1 As shown, this invention provides a method for dynamic monitoring of health risks among occupational groups based on wearable devices, specifically including: S1: Synchronously collect the user's heart rate variability (HRV), skin conductance (EDA), and blood oxygen saturation (SpO2) physiological signals through the multimodal sensors built into the wearable device, and acquire environmental parameter data provided by temperature and humidity sensors, noise meters, and barometers in real time to form the original physiological signal sequence and environmental parameter sequence; S2: Perform sliding time window segmentation normalization on the physiological signal sequence, perform moving average trend smoothing and discretization encoding on the environmental parameter sequence, and generate a structured physiological feature matrix and an environmental context vector. S3: Construct a Context Gated Attention Network (CGAN), which consists of three layers: the input layer receives the environment context vector, the middle gating layer generates a channel weight matrix through the Sigmoid activation function, and the output layer generates dynamic attention weights based on Softmax normalization. The weight matrix of the gating layer is iteratively optimized through backpropagation based on the cross-entropy loss of the environment vector and the labeled fatigue state. S4: Input the physiological feature matrix and the environmental context vector into the trained CGAN to generate physiological feature channel attention weights that are dynamically associated with environmental parameters, and output the weighted dynamic physiological feature matrix. S5: Input the dynamic physiological feature matrix into a bidirectional LSTM network for time-series modeling, extract time-dependent features through gated recurrent units of forward and backward hidden layers, and output an individualized fatigue probability sequence. S6: Based on the fatigue probability sequence and the gradient backpropagation parameters of the LSTM network, calculate the class activation mapping weights of each physiological feature channel, generate a heat map of the correlation between environmental parameters and physiological indicators, and label the contribution distribution of environmental factors such as high temperature and high noise to the fatigue score. S7: Determine whether the fatigue probability sequence exceeds a preset threshold. If it does, trigger a graded warning signal and generate a semantic explanation report containing environmental context labels. The report contains a structured description of 'increased HRV fluctuations under high temperature conditions leading to an increase in fatigue scores'. S8: Construct a dynamic feedback optimization mechanism. Based on the environment-physiology correlation heatmap data of user interaction feedback, incrementally train the gating layer parameters of the CGAN network and update the attention weight generation model.

[0010] Step S1: The wearable device synchronously collects the user's heart rate variability (HRV), electrical skin response (EDA), and blood oxygen saturation (SpO2) physiological signals through its built-in multimodal sensors, and acquires environmental parameter data from temperature and humidity sensors, noise meters, and barometers in real time, forming raw physiological signal sequences and environmental parameter sequences. Specifically, this includes: S1.1: Based on the PPG and ECG sensors integrated in the wearable device, collect the user's heart rate variability (HRV) data to obtain time-domain and frequency-domain feature sequences, and output heart rate variability time-series features; Cardiovascular activity signals are acquired by integrating a photoplethysmography (PPG) sensor and an electrocardiogram (ECG) sensor into a wearable device. The sampling frequency of the two sensors is set in the range of 250Hz to 500Hz. The consistency of the sampling time reference is ensured by a synchronous triggering mechanism, and the full waveform acquisition of heart rate variability (HRV) data is realized. An analog front-end circuit with adaptive gain control (parameters: gain range 20dB to 60dB, input impedance greater than 10MΩ) is used to achieve amplitude modulation and noise suppression of cardiovascular pulse signals, and obtain amplitude-stable PPG and ECG raw voltage signal sequences. Furthermore, the analog signal is digitized by an analog-to-digital converter (ADC, resolution ≥12bit) to generate quantized PPG and ECG time-domain data, which serve as the input basis for subsequent time-domain and frequency-domain feature extraction. The Fast Fourier Transform (FFT) method (parameters: window length 1024, window function type Hanning window) is used to realize the frequency domain analysis function of HRV and obtain the power spectral density values ​​of low frequency (LF, 0.04–0.15Hz) and high frequency (HF, 0.15–0.4Hz). A sliding window statistical method (window length 5 minutes, step size 30 seconds) was used to perform time domain analysis on the digital signal to calculate indicators such as mean heart rate and standard deviation SDNN. An RR interval extraction algorithm based on peak detection and adaptive threshold adjustment (parameter: threshold adjustment step size 5ms) was used to analyze the ECG signal cycle by cycle to generate a continuous RR interval sequence, which was used as the original basis for HRV feature calculation. By combining time domain (such as SDNN, RMSSD) and frequency domain (such as LF / HF ratio) calculation methods, the results of the previous step are transformed into heart rate variability time series features, thereby achieving a quantitative characterization of the user's autonomic nervous system activity state. For example, in a power line inspection scenario, inspectors wear a wrist-worn wearable device with a PPG and ECG combined module. The PPG sampling frequency is set to 250Hz, and the ECG sampling frequency is set to 500Hz. The signal-to-noise ratio is improved by 40dB gain amplification through an analog front-end circuit in a high-noise environment. After ADC digitization, the HRV spectrum within a 5-minute window is calculated using Hanning window FFT, yielding an LF power of 860ms², an HF power of 420ms², and an LF / HF ratio of 2.05. Simultaneously, the time-domain SDNN index is calculated to be 55ms and the RMSSD index to be 42ms using a sliding window. Peak detection yields 300 consecutive RR intervals, with an average heart rate of 76 beats / minute. Related validation results demonstrate that under high-temperature and high-noise environmental conditions, this acquisition method can output a stable HRV time series, providing refined physiological data for subsequent contextual fusion fatigue assessment. S1.2: Perform bandpass filtering on the PPG signal in the heart rate variability time series features, and use a combination of Butterworth low-pass and high-pass filters to remove electromyographic interference and motion artifacts to obtain a clean PPG waveform. Output the filtered PPG signal sequence for subsequent RR interval extraction. S1.3: Based on the filtered PPG signal sequence, the peak detection algorithm is used to identify the peak point of each heartbeat cycle, calculate the time interval between adjacent peak points, i.e., the RR interval, to construct the original HRV time series, and output the RR interval sequence data for subsequent HRV feature extraction; S1.4: The user's skin conductance signal is acquired using the skin conductance response (EDA) module. The signal is processed by baseline drift correction and peak detection to extract the characteristics of skin conductance activity intensity and rate of change, and the original EDA signal sequence is output. The input conditions are: the user's skin conductance signal stream is acquired through the skin conductance response (EDA) module integrated into the wearable device. The signal includes changes in skin conductance generated by the activity of the autonomic nervous system and external environmental interference noise during the acquisition process; The constant voltage excitation method (parameters: constant bias voltage 0.5V, sampling frequency 8Hz) is used to continuously measure the conductivity of the skin surface at the sampling site. Furthermore, a baseline drift correction function is achieved through a digital high-pass filtering algorithm (parameter: cutoff frequency 0.05Hz), and a skin conduction signal sequence after low-frequency component suppression is obtained; Furthermore, a sliding window difference method (parameters: window length 2s, step size 0.5s) is used to achieve the local trend removal function and generate short-term fluctuation enhancement signals to provide high-resolution change information for subsequent event detection; Furthermore, a peak detection algorithm (parameter: threshold set to mean + 2 standard deviations) is applied to realize the transient event identification function of skin conductance response (SCR) and obtain SCR event timestamp and amplitude data; Furthermore, using the formula for calculating the rate of change:

[0011] in, This refers to the transient skin conductance amplitude. Inherent skin conductivity level, To measure the duration of an event, the rate of change in skin electrical activity intensity is quantitatively assessed, and a sequence of change rate characteristics is generated. Through the above chain algorithm processing, the constant voltage acquisition result of the previous step is transformed into the original EDA signal sequence, realizing a quantifiable characterization of the user's autonomic nervous activation state. For example, in a power line inspection scenario, workers wear skin conductance sensors with the sensor electrodes fixed to the palm of their non-dominant hand. The sampling frequency is set to 8Hz, and the constant bias voltage is 0.5V. The raw skin conductance values ​​collected are stable at 12μS in a quiet environment, but drift to 15μS after entering a high-temperature and high-humidity plant. After high-pass filtering, the low-frequency drift component is effectively suppressed, yielding a steady-state component of approximately 12.5μS. Sliding window differential processing results show five significant positive fluctuations within one minute, with amplitudes greater than 1.5μS. The peak detection algorithm, with a noise threshold set to the mean + 2 standard deviations, accurately identifies five SCR events, with a maximum amplitude of 2.3μS and an average event duration of 3.2s. Calculated using the rate of change formula, the maximum rate of change is 0.72μS / s, and the minimum rate of change is 0.31μS / s, showing a clear increasing trend in the high-temperature scenario. The original EDA signal sequence output, combined with amplitude and rate of change characteristics, can be used in the fatigue assessment model to significantly improve the sensitivity of fatigue identification in high-temperature environments. S1.5: Perform baseline drift correction processing on the original EDA signal sequence, use wavelet threshold denoising method to remove high-frequency noise interference in order to preserve the transient characteristics of SCR events, and output the corrected EDA signal sequence for subsequent EDA feature extraction; S1.6: The light absorption ratio of oxyhemoglobin to deoxyhemoglobin in the user's blood is collected based on the blood oxygen saturation (SpO2) sensor. The sensor adopts dual-wavelength photoelectric detection technology, calculates the SpO2 value by the absorption difference of red light and infrared light, and outputs the raw SpO2 time series data. S1.7: Perform sliding window mean filtering on the original SpO2 time series data to eliminate short-term fluctuations caused by exercise or respiration, and output a smoothed SpO2 signal sequence for subsequent SpO2 feature extraction; S1.8: Collect temperature and relative humidity data in the working environment through an embedded temperature and humidity sensor, perform temperature-humidity joint calibration and outlier removal processing to obtain the environmental temperature and humidity sequence, and output structured environmental parameter data; S1.9: Ambient temperature and relative humidity data are collected based on the temperature and humidity sensor in the wearable device. The sensor uses CMOS technology to integrate a temperature and humidity detection unit and outputs raw temperature and humidity numerical sequences to construct environmental context information. S1.10: Perform moving average filtering on the original temperature and humidity numerical sequence to eliminate measurement noise caused by short-term environmental fluctuations, and output a temperature and humidity sequence with smoothed trend for subsequent discretization encoding processing; S1.11: Ambient noise intensity data is collected based on a noise sensor in a wearable device. The sensor uses an electret condenser microphone to convert ambient sound pressure into a voltage signal and output the original noise intensity sequence for constructing environmental context features. S1.12: Perform A-weighted filtering and RMS calculation on the original noise intensity sequence to simulate the characteristics of human hearing perception and extract the effective sound pressure level, and output the corrected environmental noise intensity sequence for subsequent discretization coding processing; S1.13: Ambient air pressure data is collected based on the air pressure sensor in the wearable device. The sensor adopts a MEMS piezoresistive sensing unit and outputs a raw air pressure value sequence for altitude estimation and environmental context modeling. S1.14: Perform sliding window mid-range filtering on the original air pressure numerical sequence to suppress measurement errors caused by sudden air pressure fluctuations and output a stable air pressure sequence for subsequent altitude estimation and context modeling; S1.15: Based on the stable air pressure sequence, the corresponding altitude value is calculated using the international standard atmospheric model to convert the air pressure data into environmental altitude information and output the altitude sequence for subsequent environmental context fusion modeling. S1.16: Perform timestamp alignment and data synchronization processing on the acquired multimodal physiological signals and environmental parameter data to eliminate the differences in acquisition delay between sensors and generate the original physiological signal sequence and environmental parameter sequence under a unified time reference.

[0012] Step S2: Perform sliding time window segmented normalization on the physiological signal sequence, and perform moving average trend smoothing and discretization encoding on the environmental parameter sequence to generate a structured physiological feature matrix and an environmental context vector. Specifically, this includes: S2.1: The original physiological signal sequence is segmented based on a sliding time window. The continuous signal is divided into non-overlapping or partially overlapping segments with a fixed time window length to extract local dynamic features and obtain the segmented physiological signal segment sequence after time window slicing. Based on the raw physiological signal sequence acquired by multimodal sensors, a sliding time window segmentation method is used (parameter: window length). Step length This enables the slicing and decomposition of continuous time series in order to capture local dynamic features; Furthermore, by fixing the window length parameter The original signal is divided into several non-overlapping or partially overlapping signal segments on the time axis, and the original sampling rate of the time series is maintained in each segment to preserve the high-frequency transient and low-frequency trend characteristics. Furthermore, utilizing the window sliding step size The data is subjected to time-series iterative slicing, which results in overlapping portions between adjacent segments. This enhances the ability to capture abrupt fatigue characteristics and yields a set of physiological signal segments after time window slicing.

[0013] Furthermore, by adding a timestamp index to each segment, a precise correspondence between the segment sequence and the environmental parameter sequence is achieved, providing a synchronization benchmark for subsequent environmental context fusion; By using a sliding time window segmentation method, continuous raw physiological signals are transformed into a discrete set of fragmented signals, enabling explicit quantitative expression of local dynamic characteristics and providing structured input for subsequent normalization processing and feature matrix construction. For example, in a power line inspection scenario, the window length is used. =60 seconds, sliding step length The acquired HRV signal is sliced ​​using a 30-second sliding time window, resulting in each slice containing 720 sampling points (sampling rate 12Hz). Within each slice, the mean, standard deviation, and frequency domain power spectral density of the RR interval sequence are calculated to form a local slice feature vector. The slice segmentation process utilizes the formula...

[0014] The number of window slides is calculated, where n is the total sampling length, ensuring that the segment set completely covers the entire signal. Under high temperature (ambient temperature 38°C) and high humidity (relative humidity 85%) conditions, this segmentation method can capture transient features of increased heart rate fluctuations in multiple adjacent windows, providing high-quality local dynamic signal input for subsequent environment-driven fatigue identification, and significantly improving the sensitivity and reliability of fatigue assessment under complex working conditions; S2.2: Perform normalization on the segmented physiological signal fragment sequence, and use the Z-score normalization method to process each physiological signal fragment with zero mean and unit variance to eliminate data bias caused by differences in physiological indicators between individuals, and obtain a normalized physiological signal fragment matrix. S2.3: Perform moving average trend smoothing on the original environmental parameter sequence, filter the environmental data based on the arithmetic mean within the sliding window to eliminate the influence of short-term noise fluctuations, and obtain the trend smoothed environmental parameter sequence; S2.4: Perform discretization encoding on the trend-smoothed environmental parameter sequence, and map continuous values ​​to discrete category labels according to the preset environmental parameter interval threshold. For example, divide the temperature into three levels: low temperature, normal temperature, and high temperature, and generate a discrete state vector of environmental parameters. S2.5: Based on the normalized physiological signal fragment matrix and the discretized environmental parameter vector, construct a structured feature representation, and organize the physiological features and environmental context information into matrix and vector forms respectively, which serve as input data pairs for the subsequent context-gated attention network.

[0015] like Figure 2 As shown, step S3 involves constructing a Context-Gated Attention Network (CGAN), which comprises three layers: an input layer that receives the environment context vector, an intermediate gating layer that generates a channel weight matrix using a sigmoid activation function, and an output layer that generates dynamic attention weights based on softmax normalization. The weight matrix of the gating layer is iteratively optimized using a backpropagation algorithm based on the cross-entropy loss between the environment vector and the labeled fatigue state. Specifically, this includes: S3.1: Construct an input layer structure based on the environmental context vector. The input layer receives the discretized and encoded environmental context vector as the initial input of the CGAN network and outputs the environmental context embedding vector to support the subsequent gating layer's attention adjustment to the physiological feature channel. Based on the environment context vector after discretization and encoding, an embedding mapping method (parameters: vector dimension = number of environment parameter categories, embedding dimension = 64) is used to transform sparse discrete environment labels into dense low-dimensional representations, thereby improving the model's ability to continuously represent environmental features. Furthermore, by using a linear transformation method (parameters: weight matrix dimension = 64 × 128, bias vector dimension = 128), the embedded vector is mapped to a higher-dimensional feature space, and a mapped vector containing high-order interaction information of the environment is obtained. Furthermore, a batch normalization method is adopted (parameters: mean and variance are calculated within the batch, scaling factor γ and translation factor β are trainable) to standardize the numerical distribution of the high-dimensional mapping vector and generate an environment embedding vector with stable numerical distribution, thereby improving the convergence performance of subsequent gate layer calculations. Furthermore, by using a nonlinear activation method (parameter: activation function type = ReLU), nonlinear mapping is performed on the batch-normalized environment embedding vectors to enhance the feature representation ability of the model and generate environment context embedding vectors with sparsity and nonlinear discriminative characteristics. By using embedding mapping and nonlinear activation processing, the discretized environmental parameter vector from the previous step is transformed into an environmental context embedding vector that allows the gating layer to adaptively adjust the weights of physiological feature channels, thus realizing the basic input for environmental feature-driven attention weight generation. For example, in a power line inspection scenario, the input environmental context vector consists of four discrete labels: temperature level, humidity level, noise level, and altitude level. Each label is encoded using one-dimensional integers to form a vector of length 4 [2, 1, 3, 0], where temperature level "2" represents high temperature, humidity level "1" represents medium humidity, noise level "3" represents extremely high noise, and altitude level "0" represents low altitude. An embedding mapping method is used to map each label to a 64-dimensional dense vector; for example, temperature level "2" is mapped to... Humidity level "1" is mapped to Four embedding vectors are obtained sequentially; these vectors are then concatenated to form a combined vector of length 256. Using the linear transformation method, the calculation formula is as follows:

[0016] in, It is a 256×128 dimensional weight matrix. Given a 128-dimensional bias vector, we obtain a 128-dimensional high-order environment mapping vector V′. We then perform batch normalization on V′ and calculate the batch mean. With variance And scaling and translation, the normalization formula is as follows:

[0017] in, and To trainable scaling and translation parameters, the normalized high-order environment mapping vector is used. The environment context embedding vector is obtained by inputting the ReLU activation function. This vector is used to generate physiological feature channel weights in subsequent gating layer calculations. The model shows a significant improvement in HRV and blink frequency feature weights under high temperature and extremely high noise conditions, verifying the enhancement effect of embedding layer processing on environmental context sensitivity and model interpretability. S3.2: Construct an intermediate gating layer based on the environmental context embedding vector, and use a fully connected neural network and Sigmoid activation function to generate a channel weight matrix. The channel weight matrix is ​​used to characterize the importance distribution of each physiological feature channel under different environmental conditions, and outputs an environment-driven channel weight vector. The execution object of the intermediate gating layer constructed based on the environmental context embedding vector is the environmental context vector after discretization and encoding. This vector contains environmental state features such as temperature and humidity level, noise intensity category and altitude range, which are used as feature modulation control signals input to the fully connected neural network structure. A fully connected neural network mapping method (parameters: input dimension equals environment context vector dimension, number of hidden layer nodes set to 16~64 according to the complexity of environment features, weight initialization method is Xavier initialization) is adopted to realize the nonlinear mapping function from environment feature space to physiological feature channel weight space; Furthermore, by performing matrix multiplication between the weight matrix and the input environment vector, an original channel correlation score vector without activation function modulation is generated, and preliminary importance scores of each physiological feature channel under the current environmental conditions are obtained. Furthermore, the Sigmoid activation function is used (parameter: output compression range [0,1]) to achieve nonlinear compression modulation of the original channel correlation score and generate an environment-driven channel weight vector to ensure that the weight values ​​can be effectively processed by the subsequent Softmax normalization layer within a limited range. Furthermore, by backpropagating the gradient of the Sigmoid function, the differential signal between the environmental context and the physiological channel is fed back to the weight matrix of the fully connected network, thereby enabling the network to adaptively optimize the channel attention distribution under different environmental triggering conditions. By combining fully connected neural networks with Sigmoid activation, the environmental context embedding vector from the previous step is transformed into a weight vector that can directly represent the distribution of the importance of physiological feature channels under environmental conditions, thereby achieving the expected technical effect of context-driven channel attention modulation. For example, in a power line inspection scenario, the environmental context vector is set to a length of 3, corresponding to temperature levels (low temperature = 0, normal temperature = 1, high temperature = 2), noise levels (low noise = 0, medium noise = 1, high noise = 2), and altitude levels (low altitude = 0, medium altitude = 1, high altitude = 2). The input vector [2,2,0] represents high temperature, high noise, and low altitude conditions. The fully connected layer has an input dimension of 3, the hidden layer has 32 nodes, and the output dimension is equal to the number of physiological feature channels (5) (e.g., HRV, EDA, SpO2, blink frequency, head posture). Matrix multiplication is performed.

[0018] in, For the environment context vector, The weight matrix is ​​of size 5×3. This yields the original score vector. Then, Sigmoid modulation is performed, and the output vector values ​​are between 0 and 1, for example [0.92, 0.88, 0.34, 0.76, 0.65], corresponding to the HRV and EDA channels that should be given special attention under high temperature and high noise conditions. This channel weight vector is used for dynamic attention generation after subsequent Softmax normalization. The validation results show that the model's response strength to heart rate and skin conductance is significantly improved in hot and noisy environments, and the environmental adaptability of fatigue recognition is significantly enhanced. S3.3: Perform Softmax normalization based on the channel weight vector to generate dynamic attention weights, ensuring that each physiological feature channel has comparable attention intensity values ​​in different environmental contexts, and output a standardized attention weight vector; S3.4: Calculate the cross-entropy loss function based on the labeled fatigue state label and the current attention weight vector. The cross-entropy loss function is used to quantify the model prediction error, serve as the optimization objective of the backpropagation algorithm, and output the loss gradient signal. S3.5: Based on the loss gradient signal, perform backpropagation algorithm on the weight matrix of the gated layer, use stochastic gradient descent optimizer to iteratively update the weight parameters to minimize the model prediction error, and output the optimized channel weight generation model; The cross-entropy loss calculation result output by S3.4 is obtained based on the loss gradient signal and used as the driving data input for the iterative update of the weight matrix parameters; Backpropagation algorithm (parameter: learning rate) is used Loss gradient matrix This enables element-wise gradient propagation and backward error distribution on the fully connected weight matrix of the gated layer. Furthermore, gradient calculations are performed along the dimensions of each weight matrix using the chain rule, yielding the derivative with respect to the Sigmoid activation function. Perform element-wise multiplication to correct the amplitude of the gradient signal in the nonlinear mapping process; Furthermore, the weight matrix parameters are iteratively updated using a stochastic gradient descent (SGD) optimizer (parameters: batch size B, momentum coefficient μ), according to the formula...

[0019] in, Let be the parameters of the weight matrix after the t-th iteration optimization. For the weight parameter vector, For learning rate, This is the gradient of the loss function for the current batch of samples; Furthermore, through the momentum term The introduction of this method weights and accumulates historical gradient information using the formula.

[0020] To achieve smoothing of update direction and improve stability of training process, and to Add to the current weight parameters to obtain the optimized weight matrix; Through the above iterative training process, the loss gradient obtained by backpropagation is effectively transformed into the weight parameter update of the gating layer, realizing the adaptive optimization capability of the channel weight generation model under the changing environmental context. For example, in a high-temperature (ambient temperature 37°C) and high-noise (sound pressure level 85dB) environment, the input batch sample size B is 64, and the learning rate is... The initial weight matrix is ​​set to 0.001, the momentum coefficient μ is 0.9, and the initial weight matrix values ​​are randomly distributed within the interval [-0.05, 0.05]. This corresponds to the loss gradient matrix calculated from the batch sample cross-entropy loss. The dimension is [32×1]. During backpropagation, based on the Sigmoid derivative value... After scaling the gradient signal element by element, the SGD optimization formula is applied. Perform iterative updates and incorporate the momentum term formula. The update direction is stabilized, and the final output optimized weight matrix value range is concentrated in [-0.02, 0.03]. The corresponding environment-driven attention allocation weight increases on the HRV feature channel, while the value on the EDA feature channel tends to be stable, which significantly improves the accuracy and stability of fatigue state identification of the model under this complex environmental condition.

[0021] like Figure 3 As shown, step S4 involves inputting the physiological feature matrix and the environmental context vector into the trained CGAN to generate physiological feature channel attention weights dynamically associated with environmental parameters, and outputting a weighted dynamic physiological feature matrix. Specifically, this includes: S4.1: Based on the structured physiological feature matrix and the environmental context vector, an input feature tensor is constructed, wherein the physiological feature matrix is ​​used as the channel dimension input and the environmental context vector is used as the gating control signal input to the CGAN network to form an environment-driven feature weight modulation input. Based on the structured physiological feature matrix and the environmental context vector, a tensor construction method is used (parameters: the physiological feature matrix dimension is the number of channels × the number of time steps, and the environmental vector dimension is the number of environmental parameter categories) to formally organize the two types of data in a unified computation graph. Furthermore, through the feature mapping method (parameter setting: the physiological matrix is ​​linearly transformed and mapped to a dimension that matches the environment vector), the physiological feature channel data is converted into an embedding space representation that can be associated with the environmental context, and the physiological embedding matrix data is obtained. Furthermore, by using a gated signal injection method (parameter settings: environment vector is processed by embedding layer and normalization), the joint interface between the gated layer and the input layer of the context-gated attention network (CGAN) is realized, which inputs the environment context information as a control signal to the gated layer and the input layer, and generates an environment-driven control tensor. Furthermore, through tensor concatenation operations, the physiological embedding matrix and the environment-driven control tensor are connected in the feature dimension to construct an input feature tensor containing physiological-environment correlation as the computational input of CGAN, and generate a basic tensor structure that can be dynamically weighted. Through the tensor construction and signal injection processing described above, the physiological feature matrix and environmental context vector of the previous step are transformed into a high-dimensional input tensor with environmental-driven modulation function, thereby realizing the pre-data preparation for the calculation of the attention weight of the physiological feature channel by the environmental context. For example, in a power line inspection scenario, the physiological feature matrix includes three channels: HRV (20 time steps), EDA (20 time steps), and SpO2 (20 time steps), with a dimension of 3×20; the environmental context vector includes four parameter dimensions: temperature, altitude, humidity, and noise intensity. The physiological matrix is ​​linearly transformed and mapped to an 8-dimensional embedding space, using a weight matrix. Perform the transformation:

[0022] in This is the original physiological matrix. This is the embedded matrix. The environment vector is embedded through the matrix. Mapped to And perform Z-score normalization:

[0023] in It is the mean vector. This is the standard deviation vector. After normalization... As a gating control signal, and Generate by splicing along the feature dimensions The dimension is (8+8)×20. After this processing, the input tensor... CGAN enables dynamic modulation of channel weights under different environmental conditions, significantly improving fatigue recognition accuracy in harsh environments such as high temperature (>35°C) and high noise (>85dB); S4.2: Perform channel attention weight calculation on the input feature tensor, and use the Sigmoid activation function in the gated layer to perform nonlinear mapping on the environmental context vector to generate a preliminary channel attention weight vector, so as to achieve a preliminary importance score of each physiological feature channel under different environmental conditions; The input condition is an input feature tensor composed of a structured physiological feature matrix and an environmental context vector. The physiological feature matrix corresponds to the multimodal physiological parameters of each channel, and the environmental context vector corresponds to environmental factors such as temperature, humidity, noise intensity, and altitude after discretization and encoding. A matrix-vector parallel mapping method (parameters: input feature tensor, number of physiological channels m, environmental context dimension n) is adopted to input the environmental context vector into the fully connected network of the gated layer and establish a one-to-one mapping relationship with the physiological feature matrix channel dimension to realize the initialization of the nonlinear response of environmental factors to each channel; Furthermore, a mapping algorithm using the Sigmoid activation function (parameter: weight matrix) is employed. Bias vector A nonlinear transformation is performed on the linear combination result of the above parallel mapping to obtain a preliminary channel attention score. The score value is distributed between 0 and 1 to characterize the basic attention of each physiological feature channel under different environmental conditions. Furthermore, a batch normalization method (parameters: batch size B, normalization coefficients γ and β) is used to perform numerical stabilization processing on the initial attention weight vector output by Sigmoid, reducing the impact of input feature distribution differences on attention generation and ensuring that the output weights remain sensitive in the non-saturation range. Furthermore, the weighted sparsity constraint method (parameter: sparsity coefficient) is utilized. L1 regularization is introduced into the initial channel attention weight vector. By optimizing the objective function, the weights of non-critical channels are controlled to tend to zero, thereby strengthening the role of physiological channels that are significantly related to the environment in subsequent weighted calculations. By combining the above-mentioned Sigmoid nonlinear mapping with batch normalization and sparsity constraints, the input feature tensor of the previous step is transformed into a preliminary channel attention weight vector, realizing the output of the preliminary channel importance score index under environment-driven conditions, and providing a basic weight distribution for subsequent Softmax normalization allocation and channel weighted fusion. For example, considering the working environment of power inspection workers in summer with high temperatures and noise levels exceeding 85dB, the physiological feature matrix in the input feature tensor includes eight channels: HRV frequency domain index, EDA amplitude, SpO2 trend value, eye movement rate, blink frequency, skin temperature, respiratory rate, and body posture changes. The environmental context vector is discretized and encoded as follows: temperature category is high temperature (coded value 2), noise category is high noise (coded value 2), and altitude category is low altitude (coded value 0), forming an environment vector of length 3. When using a fully connected network for mapping, the weight matrix is ​​set as follows: Bias vector The mapped output is transformed into a preliminary attention vector with weights ranging from 0 to 1 using the Sigmoid function. After batch normalization, the attention values ​​for each channel are: HRV frequency domain index 0.83, EDA amplitude 0.80, SpO2 trend value 0.75, and eye movement rate 0.70, with the remaining channels below 0.4. By using L1 sparsity constraints to compress the channel weights below 0.4 to near zero, the output preliminary channel attention weight vector significantly improves the weight distribution of autonomic nervous system-related indicators (HRV and EDA) under high temperature and high noise conditions, achieving a priority scoring effect for fatigue sensitivity characteristics under environmental conditions. S4.3: Based on the preliminary channel attention weight vector, a standardized attention weight matrix is ​​generated through Softmax normalization to ensure that the sum of the weights of each channel is 1, thereby realizing the allocation of attention of the environmental context to the physiological feature channels; S4.4: Perform element-wise multiplication of the standardized attention weight matrix and the structured physiological feature matrix to perform weighted fusion of each physiological feature channel and output the weighted dynamic physiological feature matrix, which reflects the dynamic modulation effect of the environmental context on the individual's physiological state. S4.5: Perform feature channel normalization and residual connection operations on the weighted dynamic physiological feature matrix to enhance the nonlinear expressive power and training stability of the model, and output the optimized dynamic physiological feature matrix for subsequent time series modeling. The input is a weighted dynamic physiological feature matrix, which has been weighted and fused according to the output weights of the context-gated attention network in the channel dimension, and contains multi-channel physiological feature vectors modulated by environmental factors; A channel-by-channel normalization method (parameters: mean μ, variance σ) is used to unify the numerical range and eliminate the dimensions of each weighted feature channel. Specifically, this is achieved through the Z-score standardization calculation formula: ,in For eigenvalues, This represents the average value of that channel in the current batch. The normalized dynamic physiological feature matrix is ​​obtained by taking the standard deviation of the channel in the current batch. Furthermore, by employing the residual connection method (parameter: scaling factor α), the nonlinear expressive power of the features is enhanced while maintaining their original representational capabilities. Specifically, the normalized output is added element-wise to the unweighted dynamic physiological feature matrix, and a scaling factor can be optionally introduced. The adjustment is performed, and the calculation formula is as follows: ,in The result after normalization The input is the weighted feature matrix; Furthermore, through batch normalization operation (parameters: batch size b, momentum) To achieve stable feature distribution across batches, an exponentially weighted moving average is used to update the global mean and variance, thereby reducing gradient oscillations during training and improving convergence speed. Furthermore, the residual connection output is nonlinearly mapped by the activation function ReLU (parameter: lower threshold 0) to achieve feature sparsity and nonlinear separability enhancement, and negative features are set to zero to reduce noise interference. Through the above normalization and residual connection combination processing, the weighted dynamic physiological feature matrix is ​​transformed into a stable optimized dynamic physiological feature matrix with environmental context modulation information, thereby improving the numerical controllability and expressive power of subsequent bidirectional LSTM temporal modeling. For example, in a high-voltage power inspection scenario, the sampling batch size b is set to 64, and the normalization stage... Take the average value of each channel within this batch. Take the standard deviation of each channel within the batch and calculate the Z-score for each channel. For example, if the original value of a channel is 15.3, the mean is 10.2, and the standard deviation is 2.5, then the normalized value is... =2.04. Scaling factor for the residual connection stage. Setting the value to 0.5, the normalized output is added to the elements of the original weighted matrix and then multiplied by α to obtain an enhanced representation that retains some of the original features. Batch normalized momentum. Set to 0.9, the global mean and variance are updated using a sliding update to ensure consistency across batches. ReLU processing sets all negative features to zero; for example, a feature value of -0.8 becomes 0 after ReLU. In this scenario, the optimized dynamic physiological feature matrix, when processed by bidirectional LSTM, significantly improves the extraction efficiency of time-dependent features. The model can still stably output accurate fatigue probability sequences even in complex environments with both high temperature and high noise.

[0024] Step S5: The dynamic physiological feature matrix is ​​input into a bidirectional LSTM network for time-series modeling. Time-dependent features are extracted through gated recurrent units in the forward and backward hidden layers, and an individualized fatigue probability sequence is output. Specifically, this includes: S5.1: Based on the weighted dynamic physiological feature matrix, construct the input sequence of the bidirectional LSTM network, where the input at each time step is the physiological feature vector at the current time, so as to construct an input format suitable for time series modeling; S5.2: Perform forward and backward propagation gated recurrent unit (GRU) operations on the dynamic physiological feature sequence input to the bidirectional LSTM network to extract the forward and backward temporal dependent features to obtain a bidirectional hidden state representation containing contextual information. The weighted dynamic physiological feature matrix is ​​used as the temporal input of a bidirectional LSTM network. A gated recurrent unit (GRU) structure is used to perform forward propagation computation at each time step of the input sequence (parameter: number of hidden units). Input dimension Using tanh as the candidate state activation function, dependency modeling is implemented along the forward temporal direction to obtain the forward hidden state sequence. ; Furthermore, backpropagation computation is performed on the same input sequence in reverse time sequence through a gated recurrent unit (parameter: number of hidden units). Input dimension Using tanh as the candidate state activation function, we model the context relationship along the reverse time direction to obtain the reverse hidden state sequence. ; Furthermore, an update gate function is used. With reset gate function Gated feature fusion is performed on the forward and backward hidden state sequences, and the update gate calculation formula is as follows: ,in The input vector is the current time step. For the input weight matrix, Here is the hidden state weight matrix. It is the sigmoid function; Furthermore, through the candidate state calculation formula This enables the extraction of new information at the current time step, where This represents the candidate hidden state at the current time step t. To reset the gate output, and These are the input and hidden weight matrices, respectively. Furthermore, the forward and backward hidden states are combined into a bidirectional hidden state representation through feature concatenation. It outputs a synthetic feature vector containing bidirectional contextual information; By using a gated loop structure and a bidirectional propagation mechanism, the dynamic physiological feature matrix of the previous step is transformed into a comprehensive hidden state representation containing positive and negative time-dependent features, thereby realizing global temporal modeling of fatigue state under complex working conditions. For example, in a power line high-altitude operation scenario, the dynamic physiological feature matrix contains 32-dimensional normalized physiological features sampled every 500ms. The forward and backward GRU units of the bidirectional LSTM network are each configured with 64 hidden units and an input dimension of 32. The update and reset gate weight matrices are initialized using Xavier. During forward propagation, based on the formula... The update gate output at each time step is calculated to obtain the trend of increased weighting of heart rate variability features under high temperature and humidity conditions. During backpropagation, the same parameter structure is used to model the sequence in reverse order, extracting the inverse dependence features of skin conductance response and blood oxygen level in the minutes before the end of the task. The forward and backward hidden states are concatenated to form a 128-dimensional bidirectional hidden representation, which can accurately capture the bidirectional impact of changes in working conditions on physiological state in actual verification. The output hidden state significantly improves the stability of fatigue probability estimation in subsequent Softmax classification and maintains a low false alarm level under different load conditions. S5.3: Based on the hidden layer output of the bidirectional LSTM network, calculate the fatigue state prediction probability corresponding to each time step, and use the Softmax classifier to map the hidden state vector to the posterior probability distribution of the fatigue state, forming an individualized fatigue probability sequence. Based on the hidden layer output of a bidirectional LSTM network, the Softmax classification algorithm (parameter: number of classification categories = 2, corresponding to fatigue and non-fatigue states) is used to map the hidden state vector to a probability distribution. Furthermore, the hidden state vector at each time step is processed using the Softmax function. With weight matrix Perform a linear mapping and calculate the unnormalized scores for each category. And the following formula is obtained:

[0025] in For category The corresponding linear score, This is the hidden state vector at the current time step; Furthermore, the softmax normalization function is used to convert the scores of each category into probability values, and the posterior probability of each category is calculated using the following formula:

[0026] in For time step Fatigue state as a category The probability, For the index score of category j, The sum of the scores for all categories; Furthermore, by iterating step by step over the time series, sequence data including the fatigue state probability at each time step is generated, resulting in an individualized fatigue probability sequence; Furthermore, by smoothing the probability sequence (parameters: window length = 3, smoothing method = moving average), the impact of instantaneous abnormal fluctuations on the evaluation results is reduced, and a stable fatigue probability sequence is generated for subsequent interpretability calculations. By using Softmax normalization and smoothing, the non-normalized scores of the bidirectional hidden layer are transformed into fatigue probability sequences with time indices, enabling dynamic tracking and analysis of the posterior probability of individual fatigue states at different time points. For example, in a power line inspection scenario, the system receives a bidirectional LSTM hidden layer output vector with a dimension of 128, and the Softmax classifier outputs a class with a number of 2. The weight matrix is ​​then used to... Size is Performing a linear mapping yields the scores for the fatigue and non-fatigue categories, respectively. and According to Softmax calculation, when time step At that time, the probability of fatigue state The probability of non-fatigue state is approximately 0.36, and the probability of fatigue state is approximately 0.64. After smoothing with a moving average of window length 3, the probability of fatigue at this time step is adjusted to approximately 0.34, thus maintaining the stability of judgment under varying environmental noise conditions and greatly improving the reliability of fatigue identification under complex conditions. S5.4: The cross-entropy loss function is used to calculate the error between the fatigue state prediction result and the actual labeled data, and an error gradient signal is generated to guide the backpropagation optimization of the network parameters in the future. S5.5: By using the update gate and reset gate mechanism of the gated loop unit, selective memory and forgetting of long-term dependent information are performed to optimize the temporal modeling performance of fatigue state recognition and improve the robustness and generalization ability of the model in complex environments.

[0027] Step S6: Based on the fatigue probability sequence and the gradient backpropagation parameters of the LSTM network, calculate the class activation mapping weights of each physiological feature channel, generate a heatmap linking environmental parameters and physiological indicators, and label the contribution distribution of environmental factors such as high temperature and high noise to the fatigue score. Specifically, this includes: S6.1: Based on the fatigue probability sequence and the gradient backpropagation parameters of the LSTM network, perform gradient weighted class activation mapping (Grad-CAM) calculation on each physiological feature channel to obtain the class activation weight matrix of different physiological features at each time step, and output the original class activation weight map. Based on the fatigue probability sequence and the gradient backpropagation parameters of the bidirectional LSTM network, the gradient weighted class activation mapping (Grad-CAM) algorithm (parameters: network hidden layer output node index, corresponding fatigue category label) is used to calculate the prediction decision contribution of each physiological feature channel at different time steps. Furthermore, by calculating the gradient of the posterior probability of the fatigue category with respect to the features of the LSTM hidden layer, and using the chain rule combined with the gradient values ​​obtained from backpropagation, a feature sensitivity matrix corresponding to each time step is generated to characterize the dependence of the prediction result on the input features. Furthermore, the time dimension of the gradient sensitivity matrix is ​​pooled using the global average pooling method (parameter: number of time steps) to obtain the global gradient weight of each physiological feature channel, which is used to measure the overall importance of the channel in fatigue state prediction. Furthermore, the channel global gradient weights are multiplied element-wise by channel with the bidirectional LSTM hidden layer feature maps to generate a weighted feature response matrix, thereby modulating the activation intensity of the original features by the gradient weights. By stacking the weighted feature response matrix along the time step dimension, the original class activation weight map is obtained, enabling visualization of the contribution distribution of each physiological feature channel at different time steps. For example, in a field operation detection under high temperature (ambient temperature 38°C) and high noise (85dB) conditions, the input fatigue probability sequence is 120 in length, and the LSTM hidden layer contains 64 nodes. The hidden layer output is calculated for each target fatigue category. For input features The gradient is obtained by backpropagation at each time step to get the gradient matrix. Perform global average pooling, multiplying the weight vector by the hidden layer feature matrix channel by channel to generate a weighted response matrix. The original class activation weight map is directly stacked and output along the time step dimension. In this scenario, the HRV channel under high temperature conditions... The response intensity of the blink frequency channel is significantly higher than that of other channels under noisy conditions, and it increases dramatically in the time step range of 75 to 90, which verifies the model's feature attention capability in complex environments. S6.2: Perform max pooling operation on the channel dimension on the original class activation weight map to compress spatial redundant information and generate a channel-compressed class activation response vector as an indicator of the relative contribution of each physiological feature in fatigue recognition. S6.3: Based on the environmental parameter sequence and the channel-compressed class activation response vector, construct a multidimensional environment-physiology interaction matrix, where each row corresponds to an environmental parameter dimension, each column corresponds to a physiological feature channel, and the matrix elements are the normalized correlation coefficients under the corresponding combinations; Based on the environmental parameter sequence and the channel-compressed class activation response vector, the Pearson correlation coefficient calculation method (parameters: environmental parameter time series, physiological feature contribution vector) is used to realize the linear correlation measurement between each environmental parameter and each physiological feature channel in the time dimension. Furthermore, by using a normalization method (parameters: row mean and standard deviation of the correlation coefficient matrix), the comparability of correlation coefficients between environmental parameters of different dimensions is achieved, and normalized correlation coefficient matrix data is obtained. Furthermore, through a matrix mapping algorithm (parameters: environmental parameter dimension index, physiological feature channel index), the normalized correlation coefficients are arranged by row corresponding to the environmental parameter dimension and by column corresponding to the physiological feature channel, generating a draft of a multidimensional environment-physiology interaction matrix; Furthermore, by using a missing value imputation algorithm (parameter: position of empty elements in the matrix, imputation method is column mean substitution), the integrity of the interaction data between environmental parameters and physiological characteristics is achieved, forming a structurally complete interaction matrix; Furthermore, by using a numerical stabilization method (parameter: matrix element range constrained to [-1,1]), extreme correlation coefficients are suppressed, generating the final multidimensional environment-physiology interaction matrix; By constructing a multidimensional environment-physiology interaction matrix, the channel compression activation response vector and environmental parameter sequence from the previous step are transformed into correlation data of environmental parameters and physiological indicators, thus providing an accurate numerical basis for subsequent heatmap mapping and environmental factor contribution labeling.

[0028] in, Let i be the environmental parameter value at the i-th time step. The activation response value for the channel compression class corresponding to the i-th time step. Let be the mean of their respective sequences. The calculated Pearson correlation coefficient; For example, in a power line inspection scenario, the environmental parameter sequence includes a temperature sequence (unit: °C, range: 28 to 40), a noise intensity sequence (unit: dB, range: 65 to 90), and an altitude sequence (unit: m, range: 100 to 420). The channel compression activation response vector has a length of 5, corresponding to five physiological characteristics: HRV, EDA, SpO2, blink frequency, and head posture. In this scenario, for the temperature and HRV channels, the Pearson correlation coefficient calculated using the above formula is 0.82, which is normalized to 0.79; the correlation coefficient between noise intensity and blink frequency is 0.76, which is normalized to 0.74; and the correlation coefficient between altitude and SpO2 is 0.65, which is normalized to 0.63. The above data are arranged in a multidimensional environment-physiology interaction matrix to form a 3×5 structured matrix. After normalization, the matrix elements are all constrained to the range of [-1,1]. The output matrix can significantly highlight the physiological indicators that are strongly correlated under high temperature and high noise conditions during subsequent heat map mapping, effectively supporting the interpretability analysis of fatigue state. S6.4: Perform heatmap mapping transformation on the multidimensional environment-physiology interaction matrix, use color gradient to visualize and encode the matrix elements, and generate a two-dimensional heatmap of the correlation between environmental parameters and physiological indicators, wherein the high contribution areas corresponding to environmental factors such as high temperature and high noise are highlighted with warm colors. S6.5: Based on the aforementioned heatmap and the preset semantic mapping rule base, semantic annotation is performed on key environment-physiological interaction patterns to generate a structured explanation report. The report includes a semantic description of 'increased HRV fluctuations under high temperature conditions lead to an increase in fatigue scores' to improve the understandability of the model output. Based on the associated heatmap and the preset semantic mapping rule base, a pattern matching algorithm (parameters: associated heatmap matrix element values, environmental parameter-physiological indicator combination patterns in the rule base) is used to realize pattern recognition and label extraction of significantly high contribution areas in the heatmap; Furthermore, through a rule matching engine (parameters: environmental parameter category labels, physiological indicator category labels, and contribution weight values), the semantic combination of the environment-physiological interaction pattern is generated, and a preliminary semantic description sequence containing environmental condition triggering descriptions and physiological response mechanisms is obtained. Furthermore, a semantic template filling algorithm (parameters: predefined sentence structure in the rule base, matched environmental parameter labels, matched physiological indicator labels, and contribution weight values) is used to achieve structured output of semantic descriptions and generate semantic report fragments for explanation. Furthermore, a causal relationship inference model (parameters: environmental parameter change trend sequence, physiological indicator dynamic response curve) is adopted to realize causal verification of the detected environment-physiological interaction pattern and generate highly reliable semantic interpretation entries after causal filtering. The report generator module transforms the semantic explanation entries from the previous step into a structured explanation report, enabling the output of highly readable and traceable explanation information, such as "the increased HRV fluctuations under high temperature conditions lead to an increase in fatigue scores." For example, in a high-temperature working environment scenario, the correlation heatmap shows that the normalized contribution of the temperature parameter to the HRV feature channel is 0.82, and the normalized contribution of the noise intensity parameter to the eye movement frequency feature channel is 0.76. The rule base contains relationship templates between the environmental parameter "high temperature" and the physiological indicator "HRV fluctuation", and between the environmental parameter "high noise" and the physiological indicator "increased eye movement frequency". The pattern matching algorithm detects that both of the above combinations exceed the contribution threshold of 0.8 (high temperature-HRV is slightly below the threshold but is confirmed to be significantly correlated through the causal verification model). Therefore, the semantic template filling algorithm is called to fill the environmental label with "high temperature environment" and the physiological label with "increased HRV fluctuation". The highly reliable relationship conclusions output by the causal inference model are included in the explanation report, generating results such as: the environmental factor "high temperature" leads to a decrease in the autonomic nervous system's regulatory ability, causing significant fluctuations in heart rate variability; the environmental factor "high noise" leads to increased visual fatigue, increased eye movement frequency, and further improved fatigue scores. Through this processing flow, the report is ultimately generated into structured entries, each of which includes a description of environmental conditions, physiological response mechanisms, contribution values, and causal verification status, thus significantly improving interpretability and credibility. Step S7: Determine whether the fatigue probability sequence exceeds a preset threshold. If it does, trigger a graded early warning signal and generate a semantic interpretation report containing environmental context labels. Specifically, this includes: S7.1: Perform dynamic threshold comparison processing on the fatigue probability sequence, and adaptively adjust the fatigue state trigger threshold based on the individual's historical fatigue baseline and the current environmental context vector to improve the personalization and environmental adaptability of the early warning. S7.2: Based on the comparison results, if the current fatigue probability value exceeds the adaptively adjusted threshold, a graded early warning signal is triggered, and an early warning event record containing fatigue level identifiers is generated to achieve differentiated responses to fatigue states of different severity. S7.3: Based on the environmental context vector and class activation mapping weights, calculate the contribution of each environmental factor to the current fatigue score, and generate an environmental-physiological interaction contribution vector to quantify the influence of different environmental parameters on the fatigue state. S7.4: Based on the environmental-physiological interaction contribution vector and fatigue probability sequence, a structured semantic explanation report is generated using a predefined semantic template library. The report contains an environmental-physiological causal description of 'increased HRV fluctuations under high temperature conditions leading to an increase in fatigue scores', in order to improve the interpretability of the model output and user trust. Based on the environment-physiology interaction contribution vector and fatigue probability sequence, a semantic template matching method (parameters: environmental label set, feature contribution value set) is used to extract the causal relationship between environmental conditions and key physiological indicators. Furthermore, by using a feature contribution threshold screening method (parameter: preset contribution threshold), the selection of high-impact environmental factors and their corresponding physiological characteristics is achieved, and a set of high-impact causal pairs is obtained. Furthermore, a rule-driven syntax generation algorithm (parameter: predefined semantic template library) is used to map causal sets to structured text descriptions and generate preliminary semantic explanation paragraphs. Furthermore, by using the fatigue probability sequence time window aggregation method (parameters: sliding window length, probability smoothing factor), the trend of fatigue state is summarized, and time series background labels corresponding to the interpretation report are generated; Furthermore, a jointly optimized report construction algorithm is used to fuse and combine causal relationship description paragraphs with temporal background labels to generate a complete structured semantic explanation report containing environmental context labels; By using semantic template matching and rule-driven generation, the relationship between high-impact environmental factors and physiological characteristics in the previous step is transformed into an environmental-physiological causal description, thereby improving the readability and professionalism of fatigue state explanations. For example, in a high-altitude work scenario, the barometric pressure sensor outputs an altitude of 1800m, the temperature and humidity sensor outputs a temperature of 37.5°C and a relative humidity of 45%, and the noise meter outputs an ambient noise level of 85dB. In the environment-physiology interaction contribution vector, temperature contributes 0.82 to HRV, and noise contributes 0.74 to EDA. In the fatigue probability sequence, the average value of the sliding window over the last 5 minutes is 0.87. Using a semantic template matching algorithm, the template pattern matching the high-temperature condition is found in the template library as: "Intensified fluctuations in <physiological indicators> under high-temperature conditions lead to an increase in fatigue scores," where physiological indicators are automatically replaced with HRV. By filtering with a contribution threshold of 0.7, low-contribution factors are eliminated, and the temperature-HRV causal pair is retained. Applying a syntactic generation algorithm, combined with the temporal background tag "fatigue scores continue to rise," an explanatory report paragraph is generated: "Intensified fluctuations in HRV under high-temperature conditions lead to an increase in fatigue scores, and the recent fatigue trend shows a continuous upward trend." After the report building algorithm integrates time series labels, it finally outputs a complete structured explanation report, which includes specific environmental parameters, descriptions of key physiological indicators and state trend information, significantly improving user understanding and the interpretability of model output; S7.5: The graded early warning signals and semantic interpretation reports are simultaneously pushed to the user's wearable device and the back-end management platform to support real-time intervention and remote supervision, and the early warning events are recorded in the database for subsequent analysis and model feedback optimization.

[0029] Step S8: Construct a dynamic feedback optimization mechanism. Based on the environment-physiological correlation heatmap data from user interaction feedback, incrementally train the gating layer parameters of the CGAN network to update the attention weight generation model. Specifically, this includes: S8.1: Obtain environment-physiology correlation heatmap data from user interaction feedback. The heatmap data includes labeled feedback information on the contribution of users to physiological indicators under specific environmental conditions such as high temperature and high noise, which is used as incremental training sample input for model optimization. S8.2: Based on the aforementioned environment-physiology correlation heatmap data, extract the environment context vector and original physiological feature matrix corresponding to the user feedback tags, and use them as structured input sample pairs for incremental training of the CGAN network to enhance the model's adaptability to user subjective perception. S8.3: Perform incremental learning operations on the gate layer parameters of the CGAN network, use the backpropagation algorithm to calculate the gradient change based on the cross-entropy loss function of the new samples, and update the weight matrix parameters of the Sigmoid activation function in the gate layer to optimize the generation mechanism of channel attention weights. S8.4: Based on the updated gate layer parameters, calculate the difference between the dynamic attention weights output by the CGAN network after incremental training and the original model output, and generate a model performance offset evaluation report to determine whether to trigger a model version update. S8.5: If the model performance deviation exceeds the preset threshold, the version update operation of the CGAN network attention weight generation model will be performed, and the incrementally trained parameters will be deployed to the health analysis module on the wearable device to improve the personalized adaptability and environmental generalization performance of the fatigue state recognition model.

[0030] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0031] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamic monitoring of health risks among occupational groups based on wearable devices, characterized in that, Includes the following steps: S1: The wearable device synchronously collects the user's physiological signals through its built-in multimodal sensors and acquires environmental parameter data in real time, forming the original physiological signal sequence and environmental parameter sequence. S2: Perform sliding time window segmentation normalization on the physiological signal sequence, perform moving average trend smoothing and discretization encoding on the environmental parameter sequence, and generate a structured physiological feature matrix and an environmental context vector. S3: Construct a context-gated attention network and train the context-gated attention network; S4: Input the physiological feature matrix and the environmental context vector into the trained context-gated attention network to generate physiological feature channel attention weights that are dynamically associated with environmental parameters, and output the weighted dynamic physiological feature matrix. S5: Input the dynamic physiological feature matrix into a bidirectional LSTM network for time-series modeling, extract time-dependent features through gated recurrent units of forward and backward hidden layers, and output an individualized fatigue probability sequence. S6: Based on the fatigue probability sequence and the gradient backpropagation parameters of the bidirectional LSTM network, calculate the class activation mapping weights of each physiological feature channel, generate a heatmap of the correlation between environmental parameters and physiological indicators, and label the contribution distribution of environmental factors to fatigue scores. S7: Determine whether the fatigue probability sequence exceeds a preset threshold. If it does, trigger a graded early warning signal and generate a semantic interpretation report containing environmental context labels.

2. The method for dynamic monitoring of occupational health risks based on wearable devices according to claim 1, characterized in that, Following step S7, the following is also included: S8: Construct a dynamic feedback optimization mechanism, based on the environment-physiological correlation heatmap data of user interaction feedback, to incrementally train the gating layer parameters of the context-gated attention network and update the attention weight generation model.

3. The method for dynamic monitoring of occupational health risks based on wearable devices according to claim 1, characterized in that, The physiological signals include heart rate variability, skin conductance, and blood oxygen saturation.

4. The method for dynamic monitoring of occupational health risks based on wearable devices according to claim 1, characterized in that, Step S2 specifically includes: The original physiological signal sequence is segmented and sliced ​​based on a sliding time window. The continuous signal is divided into non-overlapping or partially overlapping segments with a fixed time window length. Local dynamic features are extracted to obtain the segmented physiological signal segment sequence after time window slicing. Normalization is performed on the segmented physiological signal fragment sequence. The Z-score normalization method is used to process each physiological signal fragment with zero mean and unit variance to obtain a normalized physiological signal fragment matrix. The original environmental parameter sequence is processed by moving average trend smoothing. The environmental data is filtered based on the arithmetic mean within the sliding window to obtain the trend smoothed environmental parameter sequence. The trend-smoothed environmental parameter sequence is subjected to a discretization encoding operation. Based on a preset environmental parameter interval threshold, continuous values ​​are mapped to discrete category labels to generate a discretized environmental parameter vector. Based on the normalized physiological signal fragment matrix and the discretized environmental parameter vector, a structured feature representation is constructed, in which physiological features and environmental context information are organized into matrix and vector forms, respectively, as input data pairs for the context-gated attention network.

5. The method for dynamic monitoring of occupational health risks based on wearable devices according to claim 1, characterized in that, The context-gated attention network comprises an input layer, an intermediate gating layer, and an output layer.

6. The method for dynamic monitoring of occupational health risks based on wearable devices according to claim 5, characterized in that, The input layer receives the environmental context vector, the intermediate gating layer generates a channel weight matrix through the Sigmoid activation function, and the output layer generates dynamic attention weights based on Softmax normalization.

7. The method for dynamic monitoring of occupational health risks based on wearable devices according to claim 1, characterized in that, Step S4 specifically includes: Based on the structured physiological feature matrix and the environmental context vector, an input feature tensor is constructed, in which the physiological feature matrix is ​​used as the channel dimension input and the environmental context vector is used as the gating control signal input to the context-gated attention network to form the environment-driven feature weight modulation input. The input feature tensor is subjected to channel attention weight calculation. The environment context vector is nonlinearly mapped using the Sigmoid activation function in the gated layer to generate a preliminary channel attention weight vector. Based on the initial channel attention weight vector, a standardized attention weight matrix is ​​generated through Softmax normalization. The standardized attention weight matrix is ​​multiplied element-wise with the structured physiological feature matrix to perform weighted fusion of each physiological feature channel, and the weighted dynamic physiological feature matrix is ​​output. The weighted dynamic physiological feature matrix is ​​subjected to feature channel normalization and residual connection operations to output an optimized dynamic physiological feature matrix.

8. The method for dynamic monitoring of occupational health risks based on wearable devices according to claim 1, characterized in that, Step S5 specifically includes: Based on the weighted dynamic physiological feature matrix, the input sequence of the bidirectional LSTM network is constructed, where the input at each time step is the physiological feature vector at the current time, thus constructing an input format suitable for time series modeling. The forward and backward propagation gated recurrent unit operations are performed on the dynamic physiological feature sequence input to the bidirectional LSTM network to extract the forward and backward time-dependent features and obtain the bidirectional hidden state representation containing context information. Based on the hidden layer output of the bidirectional LSTM network, the fatigue state prediction probability corresponding to each time step is calculated. The hidden state vector is mapped to the posterior probability distribution of the fatigue state using a Softmax classifier, forming an individualized fatigue probability sequence.

9. The method for dynamic monitoring of occupational health risks based on wearable devices according to claim 1, characterized in that, Step S6 specifically includes: Based on the fatigue probability sequence and the gradient backpropagation parameters of the LSTM network, gradient weighted class activation mapping calculation is performed on each physiological feature channel to obtain the class activation weight matrix of different physiological features at each time step, and the original class activation weight map is output. Perform a maximum pooling operation on the channel dimension on the original class activation weight map to generate a channel-compressed class activation response vector; Based on the environmental parameter sequence and the class activation response vector, a multidimensional environment-physiological interaction matrix is ​​constructed; A heatmap mapping transformation is performed on the multidimensional environment-physiology interaction matrix, and the matrix elements are visualized and encoded using color gradients to generate a two-dimensional heatmap of the correlation between environmental parameters and physiological indicators. Based on the associated heatmap and the preset semantic mapping rule base, semantic annotation is performed on key environment-physiological interaction patterns to generate a structured explanation report.

10. A method for dynamic monitoring of occupational health risks based on wearable devices according to claim 9, characterized in that, In the multidimensional environment-physiology interaction matrix, each row corresponds to an environmental parameter dimension, each column corresponds to a physiological feature channel, and the matrix elements are the normalized correlation coefficients for the corresponding combinations.

Images