Method for collecting on-duty physiological indexes of nuclear power high-risk operating personnel

Abstract

The invention belongs to the technical field of power plant personnel health monitoring, and discloses a nuclear power high-risk operation personnel on-duty physiological index acquisition method, which comprises the following steps: synchronously processing heart impact and photoplethysmography signals, detecting heart beat characteristics, calculating heart rate, simultaneously performing quality control and abnormal heart rhythm screening, and calculating blood oxygen saturation. Measuring stability is ensured, non-invasive estimation of blood pressure is carried out, a result is output after calibration and smoothing, uncertainty is evaluated, multi-dimensional features are extracted, the blood glucose level is evaluated, risk grading is provided, respiratory modulation components are extracted from physiological signals, the respiratory rate is estimated, and a fatigue index is constructed in combination with heart rate variability; and fusing multiple physiological parameters to carry out standardization and abnormal scoring, implementing graded early warning, and completing data auditing and privacy protection. According to the application, synchronous, continuous and high-precision monitoring of key physiological indexes is realized, and dynamic assessment and real-time early warning of cognitive load, psychological pressure and operation risk of workers can be realized.

Description

Technical Field

[0001] This application belongs to the field of power plant personnel health monitoring technology, and in particular relates to a method for collecting physiological indicators of high-risk nuclear power workers during their work. Background Technology

[0002] In high-risk industrial sectors such as nuclear power plants, power line operations, and high-altitude work, the physiological state of workers directly impacts their cognitive performance, decision-making quality, and operational safety. Any operational error caused by fatigue, stress, or health abnormalities can lead to serious accidents. Therefore, effective monitoring of the physiological state of high-risk workers during their on-the-job activities is a crucial step in ensuring human resource reliability and production safety.

[0003] Currently, traditional physical examinations and pre-employment health assessments typically involve assessing workers' health status through manual consultations, paper questionnaires, or routine medical examinations (such as blood pressure monitors and blood glucose meters). However, these methods are mainly used for pre-employment testing and lack the ability to monitor workers' physiological state in real time during their work hours, failing to detect sudden fatigue or health risks that may occur during task execution.

[0004] Traditional wearable physiological monitoring devices utilize heart rate belts, finger-clip pulse oximeters, and wristbands to collect indicators such as heart rate and blood oxygen saturation. For example, Chinese patent CN202211438570.X discloses a wearable fatigue driving monitoring device and quantification method based on blood oxygen parameter classification. However, these solutions are mostly geared towards transportation, sports and fitness fields, lacking design and adaptation for the specific high-risk working group in nuclear power plants. Their shortcomings mainly include: limited monitoring indicators, failing to form a comprehensive physiological profile; data collection is mostly conducted in resting or exercise scenarios, lacking application verification combined with high-risk industrial operation scenarios; and existing systems are inadequate in data management, result feedback, and human-job matching evaluation, making it difficult to meet the high safety requirements of nuclear power plants.

[0005] Therefore, existing technologies have significant shortcomings in the on-the-job physiological monitoring of high-risk workers in nuclear power plants: on the one hand, they lack non-invasive, real-time multi-indicator data collection methods; on the other hand, they fail to integrate monitoring results with job suitability assessments. To address this issue, there is an urgent need to develop a method for collecting and monitoring physiological indicators for high-risk nuclear power plant workers, in order to improve the efficiency of physical and mental health assessments, reduce human error, and ensure nuclear power safety.

[0006] If high-risk workers pass the on-the-job physiological indicator monitoring, it indicates that their physical condition is good and they are able to perform their work tasks smoothly that day. However, if the on-the-job physiological indicator monitoring reveals that the physiological condition of high-risk workers deviates significantly from their baseline, it is recommended that relevant management personnel at the nuclear power plant determine whether they are suitable to continue working that day through observation and interviews, and make appropriate arrangements accordingly. Therefore, developing on-the-job physiological indicator monitoring tools for high-risk workers in nuclear power plants is of great significance for reducing human error and ensuring nuclear safety. Summary of the Invention

[0007] The main purpose of this application is to provide a method for collecting physiological indicators of high-risk nuclear power workers during their work, and to solve the problem of insufficient accuracy in the existing heart rate and respiratory cycle extraction.

[0008] Another objective of this application is to provide a method for collecting physiological indicators of personnel working in high-risk nuclear power plants, thereby solving the problem that single optical monitoring is susceptible to noise interference.

[0009] Another objective of this application is to provide a method for collecting physiological indicators of high-risk nuclear power workers during their work, solving the problem that traditional cuff-type blood pressure measurement cannot provide continuous monitoring.

[0010] Another objective of this application is to provide a method for collecting physiological indicators of high-risk nuclear power workers during their work, thereby addressing the problem that existing blood glucose testing methods rely on invasive techniques.

[0011] Another objective of this application is to provide a method for collecting physiological indicators of high-risk nuclear power workers during their work, thereby solving the problem that existing monitoring systems cannot perform comprehensive analysis of multidimensional physiological signals.

[0012] To achieve the above objectives, this application provides the following technical solution: A method for collecting on-the-job physiological indicators of high-risk nuclear power plant workers includes: Step 1: Simultaneously process cardiac impulse and photoplethysmography signals, detect cardiac characteristics and calculate heart rate, and perform quality control and abnormal heart rhythm screening at the same time; Step 2: Calculate blood oxygen saturation using dual-wavelength light absorption characteristics, and ensure measurement stability through signal separation, ratio method, and multi-pulse fusion; Step 3: Based on pulse conduction time (PTT), waveform characteristics, and individual parameters, a machine learning model is used to achieve non-invasive estimation of blood pressure. After calibration and smoothing, the results are output and the uncertainty is evaluated. Step 4: Extract multi-dimensional features using multi-wavelength optical signals, assess blood glucose levels using a deep learning model, and provide risk grading; Step 5: Extract respiratory modulation components from physiological signals, estimate respiratory rate, and construct a fatigue index by combining heart rate variability; Step 6: Integrate multiple physiological parameters for standardization and abnormality scoring, implement graded early warning, and complete data auditing and privacy protection.

[0013] As one feasible approach, step 1 includes: Step 1.1: Sampling PPG and BCG signals synchronously and adding timestamps; Step 1.2: Perform bandpass filtering on the signal to remove DC and trend signals, and then normalize the signal. Step 1.3: Use a triaxial accelerometer to identify motion artifacts and process them using appropriate strategies; Step 1.4: Locate the J wave of BCG and the pulse feature points of PPG using enhancement and peak detection methods; Step 1.5: Calculate the cross-correlation function of BCG and PPG for the short-time window, find the maximum correlation delay and use it to correct the event time of the two channels to obtain the cardiac timeline; Step 1.6: Estimate PSD using the Welch method for the same window, find the main heart rate peaks in the PSD and convert them to bpm, and verify the heart rate using the consistency of time-frequency domain results. Step 1.7: Use phase space nonlinear analysis to generate two-dimensional / three-dimensional Lorentz scatter plots and calculate the area, density, and asymmetry index of the scatter ellipse; Step 1.8: Based on the signal quality and multipath consistency test results, decide whether to output the final heart rate.

[0014] As an feasible approach, in step 1.6, if the difference between the two is ≤ ±5 bpm, output the heart rate; otherwise, mark it as low confidence and trigger resampling or quality detection.

[0015] As one feasible approach, step 2 includes: Step 2.1: Control the LED to emit red and infrared light alternately in a time-sharing manner to avoid signal crosstalk; Step 2.2: Separate the AC and DC components of each wavelength signal; Step 2.3: Calculate the AC / DC ratio of red light to infrared light, where AC represents the alternating current component and DC represents the direct current component. Obtain the dual-wavelength non-invasive blood oxygen saturation (SpO2) through the calibration curve. Step 2.4: Combine respiratory phase information to correct the pulse signal to reduce the impact of respiration; Step 2.5: Take the median of the measurement results of several consecutive qualified pulses as the final output; Step 2.6: Output a quality score, prompting for manual retesting if the confidence level is too low.

[0016] As an implementable approach, in step 2.2, DC is estimated for each wavelength signal using a sliding window and AC is obtained by removing DC; AC is calculated using bandpass filtering to preserve the pulse component.

[0017] As one feasible approach, step 3 includes: Step 3.1: Calculate the time difference between the BCG J wave and the PPG pulse foot as the pulse conduction time; Step 3.2: Extract various waveform morphological features from PPG and BCG signals; Step 3.3: Input PTT, waveform features, and individual parameters into the model; Step 3.4: Train a machine learning regression model using reference blood pressure as the label; Step 3.5: Perform short-time dynamic calibration, and use filtering methods to smooth the correction during operation; Step 3.6: Output the estimated value and confidence interval. Manual verification is recommended when the uncertainty is high.

[0018] As one feasible approach, step 4 includes: Step 4.1: Acquire optical signals using a PPG sensor with multiple wavelengths; Step 4.2: Extract time-domain features, frequency-domain features, fractal / complexity features, HRV index and respiratory coupling index for each channel to form a high-dimensional candidate feature set; Step 4.3: Select the features most sensitive to blood glucose from physiological signal features and individual parameters by feature selection (preserving highly correlated features) or dimensionality reduction methods (compressing redundant feature dimensions); Step 4.4: Employ a deep learning model and train it based on large-scale paired data; Step 4.5: Perform individualized fine-tuning and estimate the uncertainty of each prediction result; Step 4.6: Output blood glucose assessment values, risk classification, and contribution analysis of key influencing factors.

[0019] As one feasible approach, step 5 includes: Step 5.1: Extract the low-frequency envelope from the PPG or BCG signal to reflect respiratory modulation; Step 5.2: Perform spectral analysis on the respiratory envelope to identify the dominant frequency and calculate the respiratory rate; Step 5.3: Fuse the estimated respiratory rates from multiple channels and weight them according to signal quality. If the difference between two channels is large, mark them as low confidence. Step 5.4: Combine HRV, respiratory rate, and respiratory variability to output fatigue level using a linearly weighted or trained classifier.

[0020] As an implementable approach, step 6 includes: Step 6.1: All output parameters are aligned by timestamp and normalized based on the individual baseline; Step 6.2: Calculate the comprehensive multivariate anomaly score or Mahalanobis distance to quantify the overall deviation; Step 6.3: Set the grading alarm thresholds and trigger different levels of warnings according to the comprehensive score; Step 6.4: Attach confidence level, key contributing factors, and original signal segments to each alarm; Step 6.5: Record the feedback of the management personnel on the alarm to form a data closed-loop for model optimization; Step 6.6: Encrypt the data of the entire process and retain the event logs, timestamps, and responsibility records.

[0021] As an implementable approach, in step 6.3, set the thresholds T1 and T2; If S < T1, it is in a normal state; If T1 ≤ S < T2, it is in a caution state, and it is recommended to observe, rest, or repeat the measurement; If S ≥ T2 or a single item triggers the severe threshold, it is in an emergency state, immediately notify the management personnel and recommend suspending work.

[0022] Compared with the prior art, the method for collecting physiological indicators during on-duty of high-risk operation personnel in nuclear power provided by this application has the following beneficial effects: By adopting the intelligent bracelet collection technical means of multi-sensor fusion, this application realizes the synchronous, continuous, and high-precision monitoring of key physiological indicators such as heart rate, blood oxygen saturation, blood pressure, blood glucose, and body surface temperature, thus realizing the dynamic assessment and real-time warning of the cognitive load, psychological stress, and operation risks of the staff; at the same time, by using the technical means of photoplethysmography, spectral absorption characteristic detection, and intelligent algorithm processing, it realizes the accurate analysis and abnormal recognition of multi-dimensional physiological signals, effectively overcoming the problems of single monitoring index, discontinuous data, and lagging assessment in the prior art, and ensuring the real-time, accuracy, and reliability in the scenarios of cognitive evaluation, psychological measurement, and high-risk operation.

[0023] Aiming at the problem of insufficient extraction accuracy of existing heart rate and respiratory cycle, this application proposes a technical means of combining the time-domain, frequency-domain, and non-linear characteristics of the ballistocardiogram (BCG) for analysis. Extract the J-J interval and amplitude change characteristics in the time domain dimension, calculate the power spectral density in the frequency domain dimension, and construct two-dimensional and three-dimensional Lorenz scatter distribution models in the non-linear dimension. Through the above multi-feature fusion, it is possible to accurately extract the heart rate, cardiac cycle, and respiratory cycle, and improve the reliability of physiological parameter interpretation.

[0024] To address the issue of noise interference in single optical monitoring, this application employs photoplethysmography (PPG) to measure heart rate and blood oxygen saturation, and combines this with BCG signal characteristics for cross-validation. BCG signals are acquired simultaneously with PPG signals, and noise resistance and measurement accuracy are improved through dual-signal consistency analysis and anomaly removal mechanisms.

[0025] To address the issue of continuous monitoring in traditional cuff-based blood pressure measurement, this application introduces a fusion modeling method that integrates key features of the pulse waveform with BCG kinetic indicators in blood pressure estimation. By extracting characteristic parameters such as the rise edge, peak time, and waveform width of the pulse waveform and combining them with the kinetic characteristics in the BCG signal to establish a fusion model, cuffless continuous blood pressure measurement can be achieved, effectively improving comfort and real-time performance.

[0026] To address the issue of existing blood glucose testing methods relying on invasive techniques, this application proposes a blood glucose estimation method that combines multimodal optical signals with heart rate variability (HRV), blood oxygen saturation, and respiratory cycle characteristics. By acquiring multimodal optical signals through sensors and extracting parameters such as HRV, blood oxygen saturation, and respiratory cycle, a hyperglycemia risk assessment model is established based on machine learning algorithms, enabling non-invasive estimation and early warning of blood glucose levels.

[0027] To address the limitation of existing monitoring systems in performing comprehensive multidimensional physiological signal analysis, this application employs time-frequency analysis and nonlinear dynamic modeling to jointly interpret multi-source signals. The time-frequency and nonlinear characteristics of BCG, PPG, and respiratory signals are input into a unified analysis framework to establish a multidimensional physiological state assessment model. This technique not only provides dynamic monitoring of individual physiological states but can also be extended to high-risk scenarios such as psychological state assessment and industrial safety protection, further enhancing human factor reliability.

[0028] This method addresses the shortcomings in my country's monitoring of physiological indicators for high-risk workers in nuclear power plants. It provides a tool for monitoring these indicators, including heart rate, blood pressure, blood oxygen saturation, blood glucose, and body surface temperature. This method enables simplified monitoring of physiological indicators for high-risk workers in nuclear power plants. By collecting and monitoring these indicators, the immediate physiological state of employees can be determined, providing nuclear power plant management with recommendations on whether an employee is suitable to continue working. This reduces the risk of human error due to abnormal physiological indicators in high-risk workers, thus ensuring nuclear power safety. Attached Figure Description

[0029] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the technical description will be briefly introduced below.

[0030] Figure 1 A flowchart of the method for collecting on-the-job physiological indicators of high-risk nuclear power workers provided in this application; Figure 2 A schematic diagram illustrating the principle of blood oxygen saturation in existing technologies; Figure 3 A diagram of the BCG aortic arch provided for existing technology. Detailed Implementation

[0031] The following detailed description provides further details on specific implementation methods.

[0032] like Figures 1 to 3 As shown, this application provides a method for collecting on-the-job physiological indicators of high-risk nuclear power plant workers, including: Step 1: Extraction of cardiac information and calculation of heart rate.

[0033] Step 1.1: PPG and BCG are sampled synchronously, with a preferred sampling rate of 200~500Hz (250Hz in the example); the BCG channel uses a high-resolution ADC (18~24bit in the example) to ensure the detection of minute vibrations. Each sampling window carries a millisecond-level timestamp for accurate back-end alignment.

[0034] Step 1.2: Preprocessing. PPG is bandpass filtered at 0.5~8Hz (FIR or zero-phase IIR) to remove baseline drift and high-frequency noise; BCG is bandpass filtered at 0.5~20Hz, or extrapolated to 40Hz to preserve J-wave details; both signals are normalized after applying detrending and DC removal (high-pass with cut-off 0.3Hz).

[0035] Step 1.3: Motion Artifact Recognition. Instantaneous acceleration magnitudes are calculated using a triaxial accelerometer. When the mean or peak value within a short time window exceeds a threshold (example 0.15~0.25g), it is marked as a "high motion segment". The following strategy is applied to high motion segments: if the motion segment is short and has good signal-to-noise ratio, adaptive filtering is used; if the motion segment is long, it is marked as insufficient quality and the output is delayed.

[0036] Step 1.4: Heart rate peak / foot detection.

[0037] BCG: First, perform second-order difference or continuous wavelet transform (CWT) to enhance the J wave, and then use peak detection (the lowest peak interval corresponds to the upper limit of the maximum heart rate, such as 300ms) and prominence to determine the location of the J wave.

[0038] PPG: Detects the pulse foot and systolic peak using the first derivative and threshold method.

[0039] Step 1.5: Cross-correlation alignment. Calculate the cross-correlation function of BCG and PPG for a short time window (example 8s, 50% overlap), find the maximum correlation delay and use it to correct the event time of the two channels to obtain an accurate cardiac timeline.

[0040] Step 1.6: Frequency Domain Verification. Estimate the PSD (8s window length, 50% overlap, Hamming window) using the Welch method for the same window. Find the main heart rate peak in the PSD and convert it to bpm. Perform a consistency check with the time domain detection results. If the difference between the two is ≤ ±5 bpm, output the heart rate; otherwise, mark it as low confidence and trigger resampling or quality detection.

[0041] Step 1.7: Nonlinear / Phase Space Analysis. Generate two-dimensional / three-dimensional Lorenz scatter plots using the delayed embedding method (delay τ is taken as the median RR / 2 or an empirical value of 0.25~0.5s), calculate the area, density, and asymmetry index of the scatter ellipse, and use them to identify arrhythmias (such as atrial fibrillation) as part of the decision rule.

[0042] Step 1.8: Output Rules. Output the final heart rate only if the short-time window signal quality index (SQI) is ≥ the threshold (Example 0.7) and the time domain / frequency domain / nonlinearity are consistent; otherwise, output "insufficient quality" or suggest remeasurement.

[0043] In the physiological signal processing and heart rate monitoring workflow, SQI stands for Signal Quality Index. It is a quantitative indicator that assesses the reliability of physiological signals (such as PPG and BCG signals in this workflow), typically ranging from 0 to 1. The closer the value is to 1, the less the signal is affected by noise, motion artifacts, electromagnetic interference, etc., and the higher the data quality. The closer the value is to 0, the higher the proportion of signal noise and the lower the data reliability.

[0044] In step 1.8, the window SQI is a quality index calculated for physiological signals within a short processing window, and its core function is: As the first screening threshold for data validity: only if SQI ≥ the set threshold (such as in Example 0.7) is the signal within the window considered usable for heart rate calculation; The system provides dual protection through multi-dimensional consistency verification: by combining the consistency of analysis results in the time domain, frequency domain, and nonlinear domain, it ultimately determines whether to output the heart rate, thus avoiding erroneous results caused by low-quality signals.

[0045] Step 2: Blood oxygen saturation (SpO2) measurement.

[0046] Step 2.1: Dual-wavelength alternating emission: The LED alternately emits 660nm (red light) and 940nm (infrared light) according to time slots, with a sampling rate of 100~250Hz for each wavelength; crosstalk is avoided by using synchronous detection or time multiplexing methods.

[0047] Step 2.2: AC / DC separation: Estimate the DC component (DC) for each wavelength signal using a sliding window (Example 1~2s) and obtain the AC component (AC) by removing the DC; the AC can be calculated using a bandpass filter (0.5~8Hz) to preserve the pulse component.

[0048] Step 2.3: Ratio method and calibration curve: Calculate the ratio R=(AC_red / DC_red) / (AC_ir / DC_ir), and obtain SpO2=AB·R (A and B are calibration coefficients) through the manufacturer / clinical calibration curve; the equipment can be calibrated once or periodically when the hardware is replaced.

[0049] Step 2.4: Respiratory / BCG phase correction: Identify the respiratory-affected segment using BCG or respiratory phase, and perform a respiratory phase-weighted average of the AC values ​​of SpO2 or remove affected pulses.

[0050] Step 2.5: Multi-pulse steady-state output: Before outputting SpO2, at least N consecutive qualified heartbeats are required (example N=3~5), and the median of these N window results is used as the final output. Hampel filtering is used to remove outliers.

[0051] Step 2.6: Quality Control and Confidence Level: Output includes SQI. If SQI < threshold (Example 0.6), mark it as "low confidence level" and prompt for manual retesting.

[0052] Step 3: Blood pressure estimation.

[0053] Step 3.1: PTT (Pulse Transmission Time) Measurement: Use the J wave time of BCG as the cardiac ejection reference time t_BCG; use the pulse foot time of PPG t_PPG_foot as the peripheral arrival time; calculate PTT = t_PPG_foot - t_BCG (note that cross-correlation has been used for alignment to ensure timing accuracy).

[0054] Step 3.2: Pulse waveform morphology features: Extract from PPG: upstroke slope, time to peak, half-peak width (FWHM), and augmentation index, etc.; extract cardiac dynamic features (short-time energy and Lorenz statistic) from BCG.

[0055] Step 3.3: Individualization and demographic features: Combine PTT, waveform features and individual parameters (age, sex, weight, height, vascular elasticity estimate or historical baseline value) as model inputs.

[0056] Step 3.4: Regression Model and Training: Machine learning regression models (examples: Ridge Regression, Random Forest, XGBoost, or Lightweight Neural Networks) are used to model systolic blood pressure (SBP) and diastolic blood pressure (DBP) separately. During training, cuff-based blood pressure readings or clinical reference measurements collected simultaneously are used as labels, and K-fold cross-validation and independent test sets are employed to evaluate performance.

[0057] Step 3.5: Individual Calibration and Online Correction: Short-term calibration (e.g., 2-4 cuff measurements) is required for first-time use. These points are used to fit the individual calibration coefficients. During operation, a sliding window or Kalman filter is used to smooth the output and correct system drift in real time.

[0058] Step 3.6: Confidence and Manual Hints: For each estimated output confidence interval (e.g., based on model prediction variance or ensemble model standard deviation), if the uncertainty is greater than the threshold, prompt that manual retesting is required or that the measurement should be rolled back to the cuff measurement.

[0059] Step 4: Blood glucose assessment.

[0060] Step 4.1: Multi-wavelength acquisition: The smart bracelet is equipped with multi-wavelength PPG (example 6~8 wavelengths, covering visible light to near infrared, such as 520 / 560 / 660 / 810 / 880 / 940 / 1050 nm), with a sampling rate of 100~250Hz per channel.

[0061] Step 4.2: Feature set construction: Extract time-domain features (peak value, rise time, pulse area), frequency-domain features (PSD energy distribution), fractal / complexity features (Higuchi fractal dimension, spectral entropy), HRV indicators (SDNN, RMSSD, LF / HF) and respiratory coupling indicators for each channel; and construct a high-dimensional candidate feature set.

[0062] Step 4.3: Feature selection and dimensionality reduction: Select the features that contribute the most using LASSO, mutual information, or importance ranking based on a tree model, or use principal component analysis (PCA) / autoencoder for dimensionality reduction.

[0063] Step 4.4: Model and Training: Employ a temporal deep learning model, such as CNN-LSTM or a Transformer-based regression network. The input is multi-channel temporal features (or the original multi-channel PPG window), and the output is a blood glucose estimate or a hyperglycemic risk probability. Training data must be large-scale paired data (optical signal + finger prick blood / CGM). Data augmentation (adding motion noise, baseline drift, etc.) is performed during training to improve generalization.

[0064] Step 4.5: Individual fine-tuning and uncertainty estimation: After model deployment, perform a small number of individual fine-tunings (3-5 marker points) and output the uncertainty using MC-Dropout or ensemble methods; when the uncertainty exceeds the threshold, prompt that finger-prick blood verification is required.

[0065] Step 4.6: Interpretability and Risk Warning: The output is not just a blood glucose value, but also provides influencing factors (such as the weight of absorption at a certain wavelength and HRV index), and gives a "low / medium / high" risk classification and suggested actions (retesting, medical treatment, etc.).

[0066] Step 5: Calculate the respiratory cycle.

[0067] Step 5.1: Low-frequency envelope extraction: Low-pass filter the PPG / BCG (example 0.05~0.7Hz) or use Empirical Mode Decomposition (EMD) to extract intrinsic mode functions (IMFs) consistent with respiration.

[0068] Step 5.2: Dominant frequency identification: Estimate the dominant respiratory frequency and calculate the respiratory interval (seconds / breath) using the autocorrelation function or PSD (window length 30~60s, 50% overlap) on the extracted low-frequency envelope.

[0069] Step 5.3: Combined channel solution: Extract respiratory rates from PPG and BCG simultaneously and average them according to SQI (the channel with higher SQI has a higher weight). If the difference between the two channels is large, mark them as low confidence.

[0070] Step 5.4: Fatigue Index Construction: Combining HRV (time domain: SDNN, RMSSD; frequency domain: LF, HF, LF / HF) with respiratory rate and respiratory variability, a linearly weighted or trained classifier (e.g., random forest) is used to output the fatigue level (normal / mild / severe). Supervised learning is performed during training using cognitive assessment scores or subjective fatigue questionnaires as labels.

[0071] Step 6: Data fusion and early warning.

[0072] Step 6.1: Time Alignment and Normalization: All output parameters are aligned by timestamp, and the z-score (z = (x - μ) / σ) is calculated according to the individual baseline (historical mean μ and standard deviation σ) to eliminate individual differences. The baseline can be established from the first 5 - 15 minutes of resting data collection and updated periodically.

[0073] Step 6.2: Multivariate Anomaly Scoring: Define the combined scoring function S = Σw i ·z i (where w i are trainable or preset weights, and z i are the z-scores of each indicator), or calculate the multivariate deviation using the Mahalanobis distance. On the training set, w i can be learned through logistic regression or tree models, or directly learn the risk classifier.

[0074] Step 6.3: Graded Alarm Strategy: Set thresholds T1 and T2: S < T1: Normal (green); T1 ≤ S < T2: Attention (yellow), it is recommended to observe / rest / repeat measurement; S ≥ T2 or a single item triggers a severe threshold (such as SpO2 < 90%, SBP outside the severe range): Emergency (red), immediately notify the management and recommend suspending work.

[0075] Step 6.4: Confidence and Explanability: Each alarm outputs the confidence (based on model probability or prediction uncertainty) and the key contributing factors (which indicator causes the alarm) at the same time, and attaches a snippet of the original signal for manual review.

[0076] Step 6.5: Human Interaction and Data Feedback: After receiving the alarm, the management can record the observation or interview results in the system interface, and these labeled data are used for subsequent model retraining and performance improvement.

[0077] Step 6.6: Audit and Privacy: The system retains event logs, timestamps, and responsibility records. The data is encrypted (such as AES - 128 / TLS) during transmission and storage, and provides permission and audit functions.

[0078] In summary, in high-risk operation scenarios, this method uses wearable devices to collect physiological parameters such as heart rate, blood oxygen saturation, blood pressure, blood glucose, and body surface temperature in real time, and combines time-domain, frequency-domain, and non-linear feature analysis of ballistocardiogram (BCG), photoplethysmogram (PPG), and respiratory signals to achieve quantitative assessment and early warning of the multi-dimensional physiological state of operators, thus effectively improving the comprehensiveness, real-time performance, and reliability of the health status monitoring of high-risk operators.

[0079] This method uses wearable smart bracelets to collect real-time physiological signals from multiple dimensions, such as heart rate, blood oxygen saturation, blood pressure, blood glucose, and body surface temperature, during cognitive assessments, psychological measurements, and high-risk operations. Combined with core detection technologies such as photoplethysmography, spectral absorption characteristics, and multi-sensor fusion, it achieves accurate monitoring and dynamic assessment of individual physiological states, thereby providing technical support for intelligent analysis and early warning of cognitive load, psychological stress, and operational risks.

[0080] The above description is only a specific embodiment of this application, but the protection scope of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application.

Claims

1. A method for collecting physiological indicators of high-risk nuclear power plant workers during their work, characterized in that, include: Step 1: Simultaneously process cardiac impulse and photoplethysmography signals, detect cardiac characteristics and calculate heart rate, and perform quality control and abnormal heart rhythm screening at the same time; Step 2: Calculate blood oxygen saturation using dual-wavelength light absorption characteristics, and ensure measurement stability through signal separation, ratio method, and multi-pulse fusion; Step 3: Based on pulse conduction time, waveform characteristics, and individual parameters, a machine learning model is used to achieve non-invasive estimation of blood pressure. After calibration and smoothing, the results are output and the uncertainty is evaluated. Step 4: Extract multi-dimensional features using multi-wavelength optical signals, assess blood glucose levels using a deep learning model, and provide risk grading; Step 5: Extract respiratory modulation components from physiological signals, estimate respiratory rate, and construct a fatigue index by combining heart rate variability; Step 6: Integrate multiple physiological parameters for standardization and abnormality scoring, implement graded early warning, and complete data auditing and privacy protection.

2. The method for collecting on-the-job physiological indicators of high-risk nuclear power plant workers according to claim 1, characterized in that, Step 1 includes: Step 1.1: Sampling PPG and BCG signals synchronously and adding timestamps; Step 1.2: Perform bandpass filtering on the signal to remove DC and trend signals, and then normalize the signal. Step 1.3: Use a triaxial accelerometer to identify motion artifacts and process them using appropriate strategies; Step 1.4: Locate the J wave of BCG and the pulse characteristic points of PPG using enhancement and peak detection methods; Step 1.5: Calculate the cross-correlation function of BCG and PPG for the short-time window, find the maximum correlation delay and use it to correct the event time of the two channels to obtain the cardiac timeline; Step 1.6: Estimate PSD using the Welch method for the same window, find the main heart rate peaks in the PSD and convert them to bpm, and verify the heart rate using the consistency of time-frequency domain results. Step 1.7: Use phase space nonlinear analysis to generate two-dimensional / three-dimensional Lorentz scatter plots and calculate the area, density, and asymmetry index of the scatter ellipse; Step 1.8: Based on the signal quality and multipath consistency test results, decide whether to output the final heart rate.

3. The method for collecting on-the-job physiological indicators of high-risk nuclear power plant workers according to claim 2, characterized in that, In step 1.6, if the difference between the two is ≤ ±5 bpm, output the heart rate; otherwise, mark it as low confidence and trigger resampling or quality detection.

4. The method for collecting on-the-job physiological indicators of high-risk nuclear power plant workers according to claim 1, characterized in that, Step 2 includes: Step 2.1: Control the LED to emit red and infrared light alternately in a time-sharing manner to avoid signal crosstalk; Step 2.2: Separate the AC and DC components of each wavelength signal; Step 2.3: Calculate the AC / DC ratio of red light to infrared light, where AC represents the alternating current component and DC represents the direct current component. Obtain the dual-wavelength non-invasive blood oxygen saturation through the calibration curve. Step 2.4: Combine respiratory phase information to correct the pulse signal to reduce the impact of respiration; Step 2.5: Take the median of the measurement results of several consecutive qualified pulses as the final output; Step 2.6: Output a quality score, prompting for manual retesting if the confidence level is too low.

5. The method for collecting on-the-job physiological indicators of high-risk nuclear power plant workers according to claim 3, characterized in that, In step 2.2, DC is estimated for each wavelength signal using a sliding window and AC is obtained by removing DC; AC is calculated using bandpass filtering to preserve the pulse component.

6. The method for collecting on-the-job physiological indicators of high-risk nuclear power plant workers according to claim 1, characterized in that, Step 3 includes: Step 3.1: Calculate the time difference between the BCG J wave and the PPG pulse foot as the pulse conduction time; Step 3.2: Extract various waveform morphological features from PPG and BCG signals; Step 3.3: Take PTT, waveform features, and individual parameters as model inputs; Step 3.4: Use a machine learning regression model and train it with the reference blood pressure as the label; Step 3.5: Perform short-term dynamic calibration and use a filtering method for smoothing correction during operation; Step 3.6: Output the estimated value and confidence interval. It is recommended to perform manual review when the uncertainty is high.

7. The method for collecting on-the-job physiological indicators of high-risk nuclear power plant workers according to claim 1, characterized in that, Step 4 includes: Step 4.1: Use a PPG sensor with multiple wavelengths to collect optical signals; Step 4.2: Extract time-domain features, frequency-domain features, fractal / complexity features, HRV metrics, and respiratory coupling metrics for each channel to form a high-dimensional candidate feature set; Step 4.3: Through feature selection or dimensionality reduction methods, screen out the features that are most sensitive to blood glucose from physiological signal features and individual parameters; Step 4.4: Adopt a deep learning model and train it based on a large amount of paired data; Step 4.5: Perform individualized fine-tuning and estimate the uncertainty of each prediction result; Step 4.6: Output the blood glucose assessment value, risk classification, and contribution analysis of key influencing factors.

8. The method for collecting on-the-job physiological indicators of high-risk nuclear power plant workers according to claim 1, characterized in that, Step 5 includes: Step 5.1: Extract the low-frequency envelope from the PPG or BCG signal to reflect respiratory modulation; Step 5.2: Perform spectral analysis on the respiratory envelope, identify the main frequency, and calculate the respiratory rate; Step 5.3: Fusion the respiratory frequencies estimated from multiple channels and weight them according to the signal quality. If the difference between two channels is large, mark it with low confidence; Step 5.4: Combine HRV, respiratory frequency, and respiratory variability, and use linear weighting or a trained classifier to output the fatigue level.

9. The method for collecting on-the-job physiological indicators of high-risk nuclear power plant workers according to claim 1, characterized in that, Step 6 includes: Step 6.1: Align all output parameters according to the timestamp and standardize them based on the individual baseline; Step 6.2: Calculate the comprehensive multivariate anomaly score or Mahalanobis distance to quantify the overall deviation; Step 6.3: Set the grading alarm threshold and trigger different-level warnings according to the comprehensive score; Step 6.4: Attach the confidence level, key contribution factors, and original signal segments to each alarm; Step 6.5: Record the feedback of the management personnel on the alarm to form a data closed-loop for model optimization; Step 6.6: Encrypt the data of the entire process and retain the event log, timestamp, and responsibility record.

10. The method for collecting on-the-job physiological indicators of high-risk nuclear power plant workers according to claim 9, characterized in that, In Step 6.3, set the thresholds T1 and T2; If S < T1, it is in the normal state; If T1 ≤ S < T2, it is in the attention state. It is recommended to observe, rest, or repeat the measurement; If S ≥ T2 or a single item triggers the severe threshold, it is in the emergency state. Immediately notify the management personnel and recommend suspending work.

Images