LNG receiving station employee health risk prediction method and system based on deep learning

Abstract

The present invention relates to the technical field of health risk prediction, specifically to a method and system for predicting health risks for LNG receiving station employees based on deep learning. It includes: collecting multi-source heterogeneous data such as employees' physiological signs data, work task data, environmental parameter data, and individual health records; performing feature engineering on the collected data to extract effective features that are highly correlated with health risks, and constructing composite risk factors such as low-temperature frostbite risk index, fatigue accumulation index, chemical poisoning risk index, etc.; using features and risk factors to train and optimize deep learning models; based on model outputs, quantitatively assessing and making early and accurate predictions of specific health risks that may occur to employees in the future; and generating warnings and intervention recommendations when the predicted risks exceed the limit. The present invention can improve the intelligence level of occupational health monitoring at LNG receiving stations, realize individualized risk management, enhance inherent safety, and effectively protect employee health.

Description

Technical Field

[0001] The present invention relates to the technical field of health risk prediction, and specifically to a method and system for predicting health risks for LNG receiving station employees based on deep learning. Background Art

[0002] As a key link in the energy supply chain, liquefied natural gas (LNG) receiving stations operate in a unique and complex environment, posing multiple challenges to employee occupational health and safety. Current employee health risk management at LNG receiving stations faces the following challenges: First, there are multiple unique hazards within LNG receiving stations, including but not limited to: frostbite that can occur in low-temperature operating areas; the risk of leakage of toxic and hazardous gases such as hydrogen sulfide during the process; and some operations involving confined spaces, posing the risk of hypoxia and toxic gas accumulation. Furthermore, LNG receiving stations typically operate in a continuous mode, and shift work can easily disrupt employee circadian rhythms, leading to fatigue and health issues. The high-intensity physical labor and repetitive operations in some positions can lead to musculoskeletal injuries.

[0003] Secondly, existing health risk management methods have certain limitations. Currently, they mostly rely on traditional preventive measures such as regular occupational health examinations, training on safe operating procedures, wearing of personal protective equipment, and regular safety inspections. Although these measures are basic and necessary, they are insufficient in terms of real-time risk prediction, consideration of individual differences, and early identification of potential risks. Existing safety monitoring systems focus more on equipment status monitoring and over-limit alarms for environmental parameters, and are relatively weak in the ability to assess and predict dynamic health risks at the individual employee level. In addition, how to effectively integrate and analyze data from different sources and with different structures, and to extract deep patterns related to health risks, is a major challenge facing current technology. Summary of the Invention

[0004] To address the problems of existing LNG receiving station employee health risk management methods, such as weak predictability, insufficient consideration of individual differences, and difficulty in effectively integrating multi-source heterogeneous data for in-depth analysis and early warning, the present invention provides a deep learning-based LNG receiving station employee health risk prediction method and system. This method and system aims to achieve early, accurate prediction and timely warning of specific health risks that may occur to LNG receiving station employees in the future, thereby significantly improving the level of occupational health monitoring and inherent safety of LNG receiving stations, and protecting the life, health, and operational safety of employees.

[0005] The deep learning-based health risk prediction method for LNG receiving station employees includes:

[0006] S1. Collect multi-source heterogeneous data related to employee health risks from various sources within the LNG receiving terminal;

[0007] S2. Performing feature engineering on the collected multi-source heterogeneous data to extract effective features that are highly correlated with health risks, and constructing risk factors based on the effective features, wherein the risk factors include a low temperature frostbite risk index, a fatigue accumulation index, and a chemical poisoning risk index;

[0008] S3. Utilizing the effective features and the constructed risk factors, construct, train, and optimize a deep learning model for predicting health risks;

[0009] S4. Based on the input of the deep learning model, quantitatively assess the employee's health risks and output a prediction result;

[0010] S5. When the predicted health risk exceeds the preset threshold, generate risk warnings and intervention recommendations;

[0011] Preferably, the multi-source heterogeneous data related to employee health risks include physiological sign data from smart wearable devices worn by employees, work task data from the LNG receiving station operation management system, environmental parameter data from environmental monitoring systems or sensors, and individual health records and historical work injury data authorized by employees and after data desensitization; the physiological sign data include heart rate, heart rate variability, body surface temperature, blood oxygen saturation and sleep quality; the environmental parameter data include ambient temperature, humidity, air pressure, and harmful gas concentration.

[0012] Preferably, feature engineering is performed on the collected multi-source heterogeneous data to extract feature data, and risk factors are constructed based on the feature data, specifically including:

[0013] Extracted from physiological sign data: mean, standard deviation, and change trend of heart rate; time domain characteristics and frequency domain characteristics of heart rate variability; mean, minimum, maximum, and change rate of body surface temperature; mean and minimum values ​​of blood oxygen saturation; total sleep duration and proportion of each stage;

[0014] Extracted from work task data: continuous working hours, total working hours, number of night shifts, number of consecutive night shift days, shift regularity indicators;

[0015] Extracted from environmental parameter data: ambient temperature, humidity; wind chill index; heat stress index; time-weighted average concentration of specific gas concentration, compliance with short-term exposure limit, peak concentration and duration of limit violation; equivalent continuous A sound level of noise;

[0016] As well as composite risk factors constructed based on LNG process characteristics and occupational health knowledge base, including low-temperature frostbite risk index, fatigue accumulation index and chemical poisoning risk index.

[0017] Preferably, the method for constructing the low temperature frostbite risk index is:

[0018] Data collection: Collect ambient temperature data and record the real-time temperature values ​​of the environment to which employees are exposed And the temperature fluctuation range within a certain time interval ; Determine the continuous working time of employees in low temperature environment from work task data and work intensity level The work intensity level can be comprehensively assessed based on factors such as the amount of operational tasks per unit time and the degree of physical exertion; the employee's body surface temperature can be obtained from physiological sign data. and its rate of change , and blood oxygen saturation ;

[0019] Calculate the environmental impact factor: Based on the ambient temperature and fluctuation range, combined with the relevant data on the impact of low temperature environments on the human body in the occupational health knowledge base, calculate the environmental impact factor:

[0020] ;

[0021] in is the reference value at room temperature, and is the weight coefficient;

[0022] Calculate work impact factor: Based on working hours and work intensity, calculate work impact factor:

[0023] ;

[0024] in The maximum safe working time for employees in low temperature environment, and is the weight coefficient;

[0025] Calculate physiological impact factors: Calculate physiological impact factors based on body surface temperature, its rate of change and blood oxygen saturation:

[0026] ;

[0027] in is the reference value of normal body surface temperature, is the reference value of normal blood oxygen saturation, 、 and is the weight coefficient;

[0028] Constructing the low-temperature frostbite risk index: The above three influencing factors are weighted and combined to obtain the low-temperature frostbite risk index:

[0029] ;

[0030] in 、 and is the weight coefficient.

[0031] Preferably, the fatigue accumulation index is constructed by:

[0032] Data collection: Obtain employees’ continuous working hours from work task data Total working hours , number of night shifts , consecutive night shift days and shift regularity indicators ; Extract the mean heart rate from physiological sign data , standard deviation in the time domain characteristics of heart rate variability , total sleep duration and the proportion of each sleep stage ;

[0033] Calculate the fatigue factor related to working hours: Based on the number of consecutive working hours and the total working hours, calculate the fatigue factor related to working hours:

[0034] ;

[0035] in is the maximum continuous working time threshold set, is the maximum threshold of total working hours in a specific period, and is the weight coefficient;

[0036] Calculate the night shift-related fatigue factor: Calculate the night shift-related fatigue factor based on the number of night shifts and the number of consecutive night shift days:

[0037] ;

[0038] in The maximum number of consecutive night shift days that can be tolerated is: and is the weight coefficient;

[0039] Calculate the shift regularity fatigue factor: Calculate the shift regularity fatigue factor based on the shift regularity index:

[0040] ;

[0041] in is the weight coefficient;

[0042] Calculate physiological fatigue factor: Calculate physiological fatigue factor based on heart rate, heart rate variability and sleep status:

[0043] ;

[0044] in The reference value of normal heart rate. is the normal reference value of the standard deviation of heart rate variability, The reference value for normal sleep duration is: The sum of sleep stages that are not conducive to fatigue recovery;

[0045] Constructing the fatigue accumulation index: The above fatigue factors are weighted and integrated to construct the fatigue accumulation index:

[0046] ;

[0047] in 、 、 and is the weight coefficient.

[0048] Preferably, the chemical poisoning risk index is constructed by:

[0049] Data collection: Obtain real-time concentrations of specific toxic gases from environmental parameter data , and the time-weighted average concentration of the gas in a certain period of time , short-term exposure limit compliance, peak concentration and duration of overrun Determine the use of protective equipment by employees while working in areas with toxic gases from work task data ;

[0050] Calculate the risk factor for exposure to a single gas: Calculate the risk factor for exposure to a single gas based on the concentration, exposure time, and toxicity characteristics of the gas:

[0051] ;

[0052] in is the occupational exposure limit for the gas, is the time-weighted average permissible concentration limit, is the peak concentration limit, The duration of the entire monitoring period;

[0053] Calculation of comprehensive exposure risk factors for multiple gases: Based on the synergistic effect of multiple gases when they exist simultaneously, the comprehensive exposure risk factors for multiple gases are calculated:

[0054] ;

[0055] in is the number of toxic gas types monitored, For the The synergistic effect coefficient of the gases, For the Risk factors for exposure to the gases;

[0056] Calculate the impact factor of protective equipment: Calculate the impact factor of protective equipment based on the use of protective equipment:

[0057] ;

[0058] Constructing a chemical poisoning risk index: A chemical poisoning risk index is constructed by integrating multiple gas exposure risk factors and protective equipment influencing factors:

[0059] ;

[0060] in is the comprehensive gas exposure risk factor, It is the impact factor of protective equipment.

[0061] Preferably, the deep learning model is selected from at least one of a recurrent neural network, a long short-term memory network, a gated recurrent unit, a Transformer model, or a multimodal fusion network using an attention mechanism.

[0062] Preferably, the predicted health risk includes at least one of frostbite risk in low-temperature working areas, chemical poisoning risk, and excessive fatigue risk.

[0063] The deep learning-based LNG receiving station employee health risk prediction system is characterized by including a data acquisition module, a data processing and analysis module, and a risk warning and visualization display module;

[0064] The data acquisition module is used to collect multi-source heterogeneous data related to employee health risks;

[0065] The data processing and analysis module is used to perform feature engineering on the collected data, extract effective features that are highly correlated with health risks, construct risk factors based on the effective features, accommodate and execute the trained deep learning model to predict health risks based on the effective features and the constructed risk factors, and quantitatively evaluate the health risks and output prediction results;

[0066] The risk warning and visualization display module is used to generate risk warnings and provide intervention suggestions when the predicted health risk exceeds a preset threshold.

[0067] Preferably, a model management and update module is also included to manage the deep learning model deployed in the system throughout its life cycle, including model version control, continuous monitoring and evaluation of online / offline performance. When the model performance is detected to have dropped to a certain level or a sufficient amount of new data has been accumulated, the module supports retraining, tuning and iterative updating of the model to ensure the continued effectiveness and accuracy of the prediction model.

[0068] Compared with the prior art, the advantages of the present invention are:

[0069] Improve the accuracy and timeliness of health risk predictions: By using deep learning technology to conduct comprehensive analysis and pattern mining of data from multiple dimensions such as physiology, environment, and tasks, it is possible to identify and predict individual employees' potential health risks earlier and more accurately than traditional methods that rely on experience or lagging indicators.

[0070] Realize individualized health risk management: The present invention can fully consider the individual physiological status differences of different employees, the current working environment and the workload they undertake, and conduct differentiated and dynamic health risk assessment and early warning, thereby realizing truly individualized occupational health monitoring.

[0071] Providing data-driven intelligent decision support: The system generates quantitative health risk data, risk trend analysis reports, and risk factor contribution analysis, which can provide LNG terminal managers with more scientific and objective decision-making basis in formulating safety management strategies, optimizing resource allocation, and improving operating processes, thereby promoting the development of intelligent and refined safety management. DETAILED DESCRIPTION

[0072] Example 1: Prediction of frostbite risk for employees in low-temperature working areas of LNG receiving stations

[0073] At LNG receiving stations, some areas (such as the LNG unloading area and the area near the BOG compressor room in the tank area) experience extremely low temperatures, putting employees at risk of frostbite when working in these areas. This example aims to predict this risk.

[0074] Data collection:

[0075] The focus is on collecting data from sensors integrated into specific parts of smart gloves, smart sleeves or smart work clothes worn by employees (such as fingertips and wrists), and monitoring skin surface temperature and its rate of change, as well as blood oxygen saturation data in real time.

[0076] Collect ambient temperature, wind speed, and humidity data in the work area. These data can come from fixed-mounted environmental monitoring sensors or integrated environmental sensors on employee-worn devices.

[0077] The exact start time and duration of employees' exposure in low-temperature areas can be obtained through their electronic work order system, positioning system or NFC clock-in records.

[0078] Feature engineering and risk factor construction:

[0079] From the continuous skin temperature data, calculate the skin temperature drop rate and the lowest skin temperature value.

[0080] The wind chill index is calculated based on ambient temperature and wind speed, which can more accurately reflect the human body's perception of cold.

[0081] A low-temperature frostbite risk index is constructed by combining ambient temperature data, the duration of employees' continuous work in low-temperature areas, and their physiological and physical data.

[0082] Individual differences are considered, such as employees’ basal metabolic rate and recent history of cold exposure as auxiliary features.

[0083] Deep Learning Models:

[0084] A long short-term memory (LSTM) model is used. Input features include time series of skin temperature, ambient temperature, wind speed, humidity, calculated wind chill index, and low temperature frostbite risk index.

[0085] The goal of the model is to predict the probability that the employee's skin temperature will fall below a preset safety threshold (e.g., within a certain period of time in the future (e.g., within the next 10 minutes or 30 minutes). , the specific threshold is set according to medical standards and industry specifications), or directly output a frostbite risk level (such as: no risk, low risk, moderate risk, high risk).

[0086] The training dataset for the model includes historical low-temperature operation data, which includes normal operation data and data when minor frostbite or near-frostbite events occurred.

[0087] Risk warning and intervention suggestions:

[0088] When the predicted frostbite risk exceeds a preset threshold (e.g., high risk level, or skin temperature is predicted to be below The probability is greater than 80%), and the system sends vibration or sound and light reminders through the smart work clothes or smart gloves worn by employees.

[0089] At the same time, early warning information is sent to team leaders or supervisors through mobile applications or management platforms, clearly pointing out the employees at risk, locations and risk levels, and providing intervention suggestions at the same time.

[0090] Example 2: Integrated application of comprehensive health risk prediction and privacy protection technology for LNG receiving station employees

[0091] This embodiment aims to integrate multiple data to predict the comprehensive health status of employees (such as overall fatigue, acute poisoning risk, etc.), and illustrate the integrated application of privacy protection technology in the system.

[0092] Data Fusion:

[0093] Integrate physiological data from wearable devices (such as heart rate, HRV, sleep quality assessment, activity level).

[0094] Integrate real-time environmental parameters from environmental monitoring systems (such as 、 、 concentration, ambient temperature and humidity, and noise level).

[0095] Integrate data from work scheduling and tasking systems (e.g., continuous work hours, number of night shifts, task urgency and complexity).

[0096] Integrate employee-authorized health record data (such as chronic disease history and recent abnormal physical examination indicators).

[0097] Deep Learning Models:

[0098] Build a multimodal deep learning model. This model consists of multiple parallel subnetworks, each responsible for processing data from one modality or category. For example, use an LSTM to process physiological time series data and a simple feedforward network to process static environmental or task data.

[0099] Each sub-network extracts high-level feature representations, which are then effectively integrated through a fusion layer (e.g., using an attention mechanism or gating mechanism). The attention mechanism can dynamically learn the contribution weights of different modalities or features to the overall health risk in a specific context.

[0100] The fused features are finally input into one or more fully connected layers, which output scores of comprehensive health risks of employees in the future (such as overall fatigue level) or the occurrence of specific acute risks (such as probability of poisoning).

[0101] Integrated application of privacy protection technology:

[0102] During the model training phase, if the health risk prediction model requires training using employee data from multiple LNG receiving terminals or different departments within the same terminal, a federated learning framework can be deployed and applied to avoid the centralization and cross-domain transmission of raw sensitive data. In this framework, each data holder uses their own data to train a local model or update the parameters of a global model, uploading only the encrypted model parameter updates to a central coordination server. The central server aggregates the parameter updates uploaded by each party, generating a new global model, which is then distributed to each party for the next round of local training. This allows the model to learn a wider range of data patterns without exposing the original data of each party, improving its generalization capabilities.

[0103] Data Storage and Access Control: All employee health-related data stored in the system is encrypted at rest using a strong encryption algorithm (e.g., AES-256). The system implements a strict, role-based data access control policy to ensure that only authorized personnel (e.g., employees can view their own data, their direct supervisors can view risk alerts under specific conditions, and HSE administrators can view anonymized statistical reports) have access to the data within their authorized scope. All data access is recorded in detail in audit logs for easy verification.

[0104] Data publishing and sharing: Differential privacy can be applied when publishing statistical analysis reports on employee health risks or sharing data across departments for research. By adding mathematically calculated random noise to query results, the published statistics maintain macroscopic accuracy while preventing the accurate derivation of any individual employee's private information.

[0105] In the above embodiments, the choice of deep learning model (such as LSTM, multimodal fusion network) is based on the characteristics of the data to be processed (temporal nature, multi-source heterogeneity) and the prediction task to be solved (frostbite risk, comprehensive health risk).

[0106] Input and output dimensions: The model's input dimensions depend on the number and type of features selected. For example, for an LSTM model, the input might be a time series data frame containing multiple physiological or environmental features. The output dimensions are determined by the prediction objective, such as one dimension for predicting a single risk probability or the number of categories for predicting multi-category risk levels.

[0107] Activation function: In the hidden layers of a neural network, ReLU is often used as the activation function to introduce nonlinearity and accelerate training. In the output layer, the activation function is selected based on the task type, such as the Sigmoid function for binary classification probability output and the Softmax function for multi-classification probability output.

[0108] Loss function: In supervised learning, the choice of loss function depends on the prediction task. For example, for binary risk prediction (e.g., "risky" / "no risk"), binary cross-entropy loss can be used; for multi-class risk level prediction, categorical cross-entropy loss can be used; for predicting continuous risk values ​​(e.g., skin temperature), mean squared error loss or mean absolute error loss can be used.

[0109] Optimizer selection: Common optimizers include Adam and RMSprop. The Adam optimizer is widely used due to its adaptive learning rate adjustment and good overall performance.

[0110] Construction and annotation of training datasets: High-quality training datasets are the cornerstone of deep learning model performance. The dataset should contain a sufficient number and diversity of samples to cover the various working conditions and health states that LNG terminal employees may encounter. Data annotation is crucial for supervised learning tasks. For example, in frostbite risk prediction, historical meteorological data, employee exposure data, and corresponding frostbite events (or simulated critical frostbite states) are required as labels. For scenarios lacking clear labels, unsupervised or semi-supervised learning methods can be considered for anomaly detection or pattern discovery.

[0111] Model evaluation metrics: Model performance is measured using a range of evaluation metrics. For classification tasks, commonly used metrics include accuracy, precision, recall, F1 score, and area under the receiver operating characteristic (ROC) curve. For regression tasks, commonly used metrics include root mean square error (RMSE) and mean absolute percentage error (MAPE). During model development, the dataset is divided into training, validation, and test sets to objectively assess the model's generalization ability.

[0112] Example 3:

[0113] Based on the deep learning-based health risk prediction method and system for LNG receiving station employees provided in Example 1 and Example 2, the detailed functions and interactions of each module of the system are described in this embodiment:

[0114] The data acquisition module continuously obtains data from various sensors and external systems, performs preliminary formatting and timestamp marking, and then sends it to the data transmission and storage module through a secure data transmission protocol (such as MQTT, HTTPS).

[0115] The data processing and analysis module performs preprocessing, feature engineering, and risk factor construction on the received data according to preset scheduling strategies or real-time triggers, then calls the corresponding deep learning model to predict health risks and outputs the prediction results (such as risk level, probability, and contribution factor).

[0116] The prediction results are sent to the risk warning and visualization module. This module determines whether to trigger an alert based on the results and pre-set rules. It then displays risk information, trend charts, and alert notifications to authorized users via user interfaces (e.g., web dashboards and mobile apps). If a high-level alert is triggered, it may also initiate a linkage via API calls to other systems (e.g., emergency response systems and communication systems).

[0117] The Model Management and Update Module periodically or on demand obtains new training data from the Data Transmission and Storage Module to evaluate and retrain the models in the Data Processing and Analysis Module. The updated model versions are then redeployed to the Analysis Module. This module is also responsible for model monitoring and maintenance.

[0118] The privacy protection control module runs throughout the operation of all other modules. For example, it ensures data encryption during storage, performs permission verification during data access, and coordinates privacy protection operations across nodes during model training (such as federated learning). It interacts with each module to ensure that the entire system meets privacy protection requirements.

[0119] The interface protocols between modules can adopt standardized RESTful APIs, gRPC, or message queues to ensure decoupling and interoperability between modules. Data security and integrity must be ensured during data transfer, for example, through TLS / SSL encrypted transmission channels and end-to-end encryption of sensitive data.

[0120] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0121] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the content of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A deep learning-based health risk prediction method for LNG receiving station employees, characterized by: include: S1. Collect multi-source heterogeneous data related to employee health risks from various sources within the LNG receiving terminal; S2. Performing feature engineering on the collected multi-source heterogeneous data to extract effective features that are highly correlated with health risks, and constructing a composite risk factor based on the effective features, wherein the composite risk factor includes a low temperature frostbite risk index, a fatigue accumulation index, and a chemical poisoning risk index; S3. Utilizing the effective features and the constructed composite risk factors, construct, train, and optimize a deep learning model for predicting health risks; S4. Based on the input of the deep learning model, quantitatively assess the employee's health risks and output a prediction result; S5. When the predicted health risk exceeds the preset threshold, generate risk warnings and intervention recommendations; The multi-source heterogeneous data related to employee health risks includes physiological sign data from smart wearable devices worn by employees, work task data from the LNG terminal operation management system, environmental parameter data from environmental monitoring systems or sensors, and individual health records and historical work-related injury data authorized by employees and after data desensitization; The physiological sign data include heart rate, heart rate variability, body surface temperature, blood oxygen saturation and sleep quality; The environmental parameter data include ambient temperature, humidity, air pressure, and harmful gas concentration; The low temperature frostbite risk index data collection includes: collecting ambient temperature data and recording the real-time temperature value of the environment in which employees are exposed And the temperature fluctuation range within a certain time interval ; Determine the continuous working time of employees in low temperature environment from work task data and work intensity level ; Obtain employee's body surface temperature from physiological sign data and its rate of change , and blood oxygen saturation ; The low-temperature frostbite risk index is constructed by combining ambient temperature data, the duration of employees' continuous work in low-temperature areas, and their physiological signs data; The fatigue accumulation index data collection includes: obtaining the employee's continuous working hours from the work task data Total working hours , number of night shifts , consecutive night shift days and shift regularity indicators ; Extract the mean heart rate from physiological sign data , standard deviation in the time domain characteristics of heart rate variability , total sleep duration and the proportion of each sleep stage ; Combining work task data, heart rate and sleep data to build a fatigue accumulation index; The chemical poisoning risk index data collection includes: obtaining the real-time concentration of specific toxic gases from environmental parameter data , and the time-weighted average concentration of the gas in a certain period of time , short-term exposure limit compliance, peak concentration and duration of the overrun Determine the use of protective equipment by employees while working in areas with toxic gases from work task data ; A chemical poisoning risk index is constructed by combining the toxic gas concentration, duration of exceeding the limit, and the use of protective equipment in the environmental parameter data.

2. The method for predicting health risks of LNG receiving station employees based on deep learning according to claim 1 is characterized in that: The feature engineering process of the collected multi-source heterogeneous data to extract effective features highly correlated with health risks, and constructing a composite risk factor based on the effective features, specifically includes: Extracted from physiological sign data: mean, standard deviation, and change trend of heart rate; time domain characteristics and frequency domain characteristics of heart rate variability; mean, minimum, maximum, and change rate of body surface temperature; mean and minimum values ​​of blood oxygen saturation; total sleep duration and proportion of each stage; Extracted from work task data: continuous working hours, total working hours, number of night shifts, number of consecutive night shift days, shift regularity indicators; Extracted from environmental parameter data: ambient temperature, humidity; wind chill index; heat stress index; time-weighted average concentration of specific gas concentration, compliance with short-term exposure limit, peak concentration and duration of limit violation; equivalent continuous A sound level of noise; A composite risk factor is constructed based on LNG process characteristics and occupational health knowledge base.

3. The method for predicting health risks of LNG receiving station employees based on deep learning according to claim 2 is characterized in that: The method for constructing the low temperature frostbite risk index is: Calculate the environmental impact factor: Based on the ambient temperature and fluctuation range, combined with the relevant data on the impact of low temperature environments on the human body in the occupational health knowledge base, calculate the environmental impact factor: ； in is the reference value at room temperature, and is the weight coefficient; Calculate work impact factor: Based on working hours and work intensity, calculate work impact factor: ； in The maximum safe working time for employees in low temperature environment, and is the weight coefficient; Calculate physiological impact factors: Calculate physiological impact factors based on body surface temperature, its rate of change and blood oxygen saturation: ； in is the reference value of normal body surface temperature, is the reference value of normal blood oxygen saturation, 、 and is the weight coefficient; Constructing the low-temperature frostbite risk index: The above three influencing factors are weighted and combined to obtain the low-temperature frostbite risk index: ； in 、 and is the weight coefficient.

4. The method for predicting health risks of LNG receiving station employees based on deep learning according to claim 2 is characterized in that: The fatigue accumulation index is constructed as follows: Calculate the fatigue factor related to working hours: Based on the number of consecutive working hours and the total working hours, calculate the fatigue factor related to working hours: ； in is the maximum continuous working time threshold set, is the maximum threshold of total working hours in a specific period, and is the weight coefficient; Calculate the night shift-related fatigue factor: Calculate the night shift-related fatigue factor based on the number of night shifts and the number of consecutive night shift days: ； in The maximum number of consecutive night shift days that can be tolerated is: and is the weight coefficient; Calculate the shift regularity fatigue factor: Calculate the shift regularity fatigue factor based on the shift regularity index: ； in is the weight coefficient; Calculate physiological fatigue factor: Calculate physiological fatigue factor based on heart rate, heart rate variability and sleep status: ； in The reference value of normal heart rate. is the normal reference value of the standard deviation of heart rate variability, The reference value for normal sleep duration is: The sum of sleep stages that are not conducive to fatigue recovery; Constructing the fatigue accumulation index: The above fatigue factors are weighted and integrated to construct the fatigue accumulation index: ； in 、 、 and is the weight coefficient.

5. The method for predicting health risks of LNG receiving station employees based on deep learning according to claim 2 is characterized in that: The chemical poisoning risk index is constructed as follows: Calculate the risk factor for exposure to a single gas: Calculate the risk factor for exposure to a single gas based on the concentration, exposure time, and toxicity characteristics of the gas: ； in is the occupational exposure limit for the gas, is the time-weighted average permissible concentration limit, is the peak concentration limit, The duration of the entire monitoring period; Calculation of comprehensive exposure risk factors for multiple gases: Based on the synergistic effect of multiple gases when they exist simultaneously, the comprehensive exposure risk factors for multiple gases are calculated: ； in is the number of toxic gas types monitored, For the The synergistic effect coefficient of the gases, For the Risk factors for exposure to the gases; Calculate the impact factor of protective equipment: Calculate the impact factor of protective equipment based on the use of protective equipment: ； Constructing a chemical poisoning risk index: A chemical poisoning risk index is constructed by integrating multiple gas exposure risk factors and protective equipment influencing factors: ； in is the comprehensive gas exposure risk factor, It is the impact factor of protective equipment.

6. The method for predicting health risks of LNG receiving station employees based on deep learning according to claim 1 is characterized in that: The deep learning model is selected from at least one of a recurrent neural network, a long short-term memory network, a gated recurrent unit, a Transformer model, or a multimodal fusion network using an attention mechanism.

7. The method for predicting health risks of LNG receiving station employees based on deep learning according to claim 1 is characterized in that: The predicted health risks include at least one of frostbite risk in low-temperature working areas, chemical poisoning risk, and excessive fatigue risk.

8. A deep learning-based LNG receiving station employee health risk prediction system, used to implement the method according to claims 1-7, characterized in that: It includes data collection module, data processing and analysis module and risk warning and visualization display module; The data acquisition module is used to collect multi-source heterogeneous data related to employee health risks; The data processing and analysis module is used to perform feature engineering on the collected data, extract effective features that are highly correlated with health risks, construct risk factors based on the effective features, accommodate and execute the trained deep learning model to predict health risks based on the effective features and the constructed risk factors, and quantitatively evaluate the health risks and output prediction results; The risk warning and visualization display module is used to generate risk warnings and provide intervention suggestions when the predicted health risk exceeds a preset threshold.

Images