A method for monitoring, simulating, dynamically predicting dust inhalation dose of dust-exposed workers and early warning of pneumoconiosis risk

Abstract

The application discloses a kind of monitoring analog synergistic driving dust worker dust inhalation dose dynamic prediction and pneumoconiosis risk early warning method, comprising the following steps: field monitoring is combined with analog breathing hemisphere area dust information, and microenvironment field of monitoring-simulation fusion is constructed;CFD-DPM numerical model and multi-source heterogeneous dust concentration field are constructed;LSTM-CNN-Attention prediction model based on DRL element controller optimization is constructed;inhalation dose is predicted by LSTM-CNN-Attention prediction model;component-particle size synergistic toxicity weighting and real-time prediction of morbidity probability;online closed-loop feedback and DRL model evolution.The application can improve the accuracy and real-time of dust exposure assessment of dust workers, provide technical support for reducing the risk of pneumoconiosis, protecting the health rights and interests of workers and improving the intelligent level of occupational health management.

Description

Technical Field

[0001] This invention relates to a method for dynamic prediction of dust inhalation dose and early warning of pneumoconiosis risk in dust-exposed workers through monitoring and simulation collaborative driving, belonging to the interdisciplinary field of occupational health monitoring and artificial intelligence. Background Technology

[0002] Currently, there are 2.772 million people employed in my country's coal mining and processing industry (according to the Fifth National Economic Census). The number of deaths from coal worker's pneumoconiosis far exceeds the number of deaths from coal mine production safety accidents, earning it the nickname "hidden mine disaster." The effectiveness of coal worker's pneumoconiosis prevention and control is directly related to the well-being of millions of miners' families and the sustainable development of the country's energy industry.

[0003] Existing technologies, such as studies assessing disease risk based on workers' basic information (e.g., length of service, age) and average dust concentration in the work environment, have the following limitations:

[0004] (1) Disconnect between monitoring and simulation: On-site monitoring is difficult, and only discrete dust concentrations can be obtained, which cannot reflect the concentration gradient in the respiratory hemisphere (0.5 m hemisphere space around the nose and mouth); although numerical simulation (CFD) can calculate continuous fields, it lacks real-time monitoring data for verification, and the simulation error needs to be reduced. Related monitoring and simulation have not been effectively integrated into research, resulting in low accuracy of individual exposure level prediction.

[0005] (2) Static models cannot adapt to dynamic environments: Traditional cumulative exposure dose models based on ICRP 66 or Logistic regression are offline static models that cannot adapt to dynamic changes such as ventilation adjustments in the work environment and process switching. They have large prediction errors when the environment changes abruptly.

[0006] (3) Insufficient individualization: Existing methods rely only on regional dust concentration monitoring data, without considering detailed factors such as dust particle size-component concentration gradient in the worker's breathing hemispherical microenvironment (a hemispherical space with a radius of 0.5 m around the nose and mouth), and without integrating spatiotemporal behavioral trajectories (pitch posture, breathing frequency, labor intensity), resulting in significant errors in individual dose estimation and limited universality.

[0007] (4) Lack of multi-component dust assessment mechanism: The toxicity of components such as coal dust, free SiO2, and silicates varies by 5 to 10 times. Existing models use uniform weights or static weights, which do not achieve coordinated dynamic weighting of component content fluctuations and particle size distribution. The risk is underestimated during high-risk periods (such as a sudden increase in SiO2 content after blasting).

[0008] (5) Low model adaptability: The hyperparameters of deep learning models such as LSTM-CNN-Attention rely on offline Grid Search tuning, which takes several hours to several days. After the environment changes, the parameters are frozen and cannot be updated online, the model performance degrades, and it cannot be corrected online in real time, resulting in model drift.

[0009] (6) Dose-risk mapping break: The current practice calculates "inhaled dose" and "risk of disease" separately: the former is first calculated by accumulating deterministic doses of dust exposure, and the latter is then calculated by static logistic mapping using a fixed dose-response coefficient. There is no biodynamic time lag or feedback study on individual susceptibility differences between the two, resulting in a logical break of "same dose - different outcomes" or "increased dose - no increase in risk", which makes it impossible to achieve early warning and timely intervention of risk. Summary of the Invention

[0010] This invention provides a method for dynamic prediction of dust inhalation dose and early warning of pneumoconiosis risk in dust-exposed workers through monitoring and simulation-driven collaborative methods. This method can improve the accuracy and real-time performance of dust exposure assessment for dust-exposed workers, and provide technical support for reducing the risk of pneumoconiosis, protecting workers' health rights, and improving the level of intelligent occupational health management.

[0011] To achieve the above objectives, the present invention provides a method for dynamic prediction of dust inhalation dose and early warning of pneumoconiosis risk in dust-exposed workers based on monitoring and simulation-driven collaborative processes, comprising the following steps:

[0012] S1. By combining on-site monitoring with dust information from the simulated breathing hemisphere region, a monitoring-simulation fusion microenvironmental field is constructed;

[0013] S2. Construct a CFD-DPM numerical model and a multi-source heterogeneous dust concentration field;

[0014] S3. Construct an LSTM-CNN-Attention prediction model based on DRL meta-controller optimization;

[0015] S4. Predict the inhaled dose using the LSTM-CNN-Attention prediction model;

[0016] S5. Real-time prediction of component-particle size synergistic toxicity weighting and disease probability;

[0017] S6, Online Closed-Loop Feedback and DRL Model Evolution.

[0018] Furthermore, the specific process of S1 is as follows:

[0019] S1.1 On-site tracking of workers of different trades, while simultaneously deploying multiple miniature dust sensor arrays (PM2.5 and total dust, sampling frequency 1Hz) on the workers' shoulders, safety helmets, and chests to acquire concentration data of respirable dust and total dust at discrete monitoring points within the respiratory hemisphere. monitor (d, t), on-site monitoring of dust concentration for different work types;

[0020] S1.2 The microenvironmental field was calculated using Computational Fluid Dynamics-Discrete Phase Model (CFD-DPM) combined with User-Defined Typical Respiratory Function (UDF). UDF enabled accurate respiratory dynamics modeling that matched individual physiological characteristics. The CFD-DPM-UDF simulation method and its results were validated using field monitoring results. The three-dimensional concentration field C in the respiratory hemisphere of different occupations was studied. sim (d, x, y, z, t), particle size classification (PM2.5, PM10, total dust, etc.), different component particles (proportion of coal dust, silica dust and other dust), and partial respirable dust concentration data of individual respiratory hemisphere obtained by numerical simulation.

[0021] Furthermore, the specific process of S2 is as follows:

[0022] S2.1, Using monitoring data C monitor (d, t) Kalman filtering is used to correct the inlet boundary conditions and dust source term intensity of the simulated field, reducing the simulation error to an allowable value; simulation supplementary monitoring: in the blind zone of the monitoring point (e.g., 0.3m directly in front of the nose and mouth), the simulation results C are used. sim (d, x, y, z, t) fills in, generating a fusion concentration field:

[0023] C fuse (d, t) = w1′· C monitor (d, t) + w2′· C sim (d, x, y, z, t);

[0024] Where w1′ = (w1+w2) / w1; w2′ = (w1+w2) / w2; w1′ and w2′ are the normalized weights of the final weighted summation. and The original weights are represented by the monitoring data weight function and the simulated data weight function, respectively:

[0025] ; ;

[0026] Where α is the normalization coefficient, d is the distance to the nearest monitoring point (i.e., spatial attenuation), R(t) is the reliability index of the monitoring equipment at time t (i.e., time attenuation), β is the spatial attenuation coefficient, γ is the balance parameter, M(s, t) is the confidence score of the simulation model at location s and time t, and D(s, t) is the data sparsity index of location s near time t.

[0027] S2.2 Obtain worker coordinates (x, y, z) and attitude angle θ∈[0°,180°] through UWB positioning and inertial sensors;

[0028] From the fusion concentration field C fuse Concentration at the nasal and oral sites was extracted from (d, t) and combined with respiratory rate f breath (t), calculate individual exposure concentration:

[0029]

[0030] Among them, C i (d, t) represents the concentration of dust with particle size d at time t, P(d, θ) is the directional weighting function, θ is the angle relative to the worker, and f breath (t) represents respiratory rate data.

[0031] Furthermore, the DRL in S3 employs a near-end strategy to optimize the PPO algorithm, addressing the high variance problem in the continuous action space:

[0032] The state space, i.e., the real-time feature vector, is represented as:

[0033] ;

[0034] The action space, i.e., the discrete-continuous hybrid action, is represented as:

[0035] ;

[0036] , , ;

[0037] Among them, v air (t) represents the air velocity, Δkernel is the directly chosen size of the convolutional kernel, Δhidden is the dimension of the hidden layer, and Δheads is the number of attention heads. Adjustment of SiO2 dust feeding coefficient, ΔTF coal ∈[-0.07,0.07] represents the adjustment of the coal dust feeding coefficient, ΔTF other ∈[-0.01, 0.01] represents adjustments for other dust feeding coefficients;

[0038] Set the reward function R as follows: ;

[0039] Among them, D pred For the predicted pollutant concentration, D actual To measure the actual inhaled dose, |D pred -D actual |For prediction error, WHO limit = 0.05 mg / min, A t Let A be the action vector at time step t. t-1 Let be the action vector at time step t-1.

[0040] Furthermore, the specific process of S4 is as follows:

[0041] S4.1 Input the time series data X∈R^(T×F) into the LSTM-CNN-Attention prediction model, where T is the time step and F is the feature dimension;

[0042] S4.2 Dynamic Convolutional Feature Extraction Layer: The CNN layer adopts a configurable convolutional kernel group to extract multi-scale local exposure patterns in real time. The convolutional kernel configuration space is kernel∈{3, 5, 7}. The decision-making mechanism is dynamically issued by the deep reinforcement learning agent according to the complexity of the input features, and the output is a multi-scale representation of local temporal dependence.

[0043] S4.3. Based on the worker's length of service, memory capacity is dynamically allocated. The number of hidden units is hidden∈{32, 64, 96}. The memory depth is automatically matched by DRL based on the historical exposure duration. The adaptive balance of long short-term memory is achieved through LSTM layers.

[0044] S4.4 Adjustable multi-head attention mechanism: heads∈{2, 4, 6}, the attention layer automatically focuses on key time steps, weight α t = softmax(QK) T / √d k Dynamically focus on high-risk exposure time windows.

[0045] Furthermore, the specific process of S5 is as follows:

[0046] S5.1, The expression for the dynamic toxicity factor is constructed as follows: ;

[0047] in, Due to the toxicity of silica, TF coal (t) represents the combined toxicity of coal dust and its impurities, TF other (t) represents the mixed toxicity of non-silicon, non-coal dust, TF interaction(t) represents the synergistic additional toxicity among dust particles, and α1(t), α2(t), α3(t), and α4(t) represent the toxicity weights, respectively.

[0048] S5.2, The particle size toxicity weighted expression is constructed as follows: ;

[0049] in, Let be the median value (μm) of the i-th particle size interval. For different dust particle size toxicity weighting functions, The toxicity factors for different dust particles in this particle size range are α1, α2, α3, and α4, which are the toxicity weights, respectively.

[0050] S5.3, the real-time mapping expression for the 10-year disease probability is: ;

[0051] in, For the occupational time-varying disease probability, F J G is the dose-response function for occupational toxicity. J To account for the age effect correction of job experience, H J For the genetic-job interaction function, TF cumulative (t) represents the cumulative toxic exposure up to time t, where Genetics is the genetic susceptibility factor, and E J To reveal history and patterns, Age(t) represents the current age, and θ J S is the job parameter vector. J For the smoking function.

[0052] The 10-year disease probability real-time mapping model is a Logistic-dose coupling model. The parameters are validated based on big data of pneumoconiosis and individual physiological parameters such as workers' breathing rate and protective behavior are incorporated in real time to achieve minute-level dynamic mapping of the 10-year disease probability; the traditional static group assessment is transformed into a precise individual proactive early warning system.

[0053] Furthermore, the specific process of S6 is as follows: Deep Reinforcement Learning (DRL) relies on the continuous interaction between the policy network and the environment during training, and adjusts and optimizes the action policy based on the deviation between the predicted result and the actual target. This process adopts an iterative mechanism of "execution-collection-learning": after executing an action according to the current policy, the system records its state, action, reward obtained, and new state after transition, forming a complete interaction experience and storing it in the experience pool; during training, the system periodically extracts a set batch of historical interaction data from the experience pool, uses a temporal difference algorithm to estimate the value error, and then completes the parameter update of the policy network through gradient backpropagation. This training mechanism based on historical experience playback not only effectively improves data utilization efficiency, but also enhances learning stability by breaking the correlation between continuous samples, promoting the policy network to gradually achieve self-improvement and performance enhancement.

[0054] This invention significantly improves the accuracy and real-time performance of dust exposure assessment for workers exposed to dust through a collaborative mechanism of individual tracking monitoring, CFD, and DPM simulation. Tracking monitoring data provides dynamic computational conditions for numerical simulation, and simulation results inversely optimize monitoring deployment, forming a two-way collaborative system. This effectively solves the problems of low spatial resolution and strong time lag in dust monitoring in complex working environments using traditional methods. The dynamic prediction model tracks worker location, work intensity, and individual respiratory parameters in real time, enabling hourly cumulative calculation and prediction of inhaled doses and identification of high-risk work processes and individual exposure characteristics. The pneumoconiosis risk early warning system constructed in this invention correlates cumulative inhaled doses with characteristics such as differences in individuals, different scenarios, and dust particle sizes and compositions, achieving a leap from "environmental concentration monitoring" to "dynamic prediction of health effects." By identifying critically at-risk populations and positions, it supports enterprises in implementing tiered management, optimizing the allocation of protective resources, and guiding workers to adjust their work behaviors, promoting the transformation of occupational disease prevention from "post-event diagnosis" to "pre-event prevention." This invention provides key technical support for reducing the risk of pneumoconiosis, protecting workers' health rights, and improving the intelligent level of occupational health management. Attached Figure Description

[0055] Figure 1 This is an overall framework diagram of the present invention;

[0056] Figure 2 This is a flowchart of the prediction method of the present invention;

[0057] Figure 3 This is a dust concentration curve diagram of different types of work monitored on-site in an embodiment of the present invention;

[0058] Figure 4 This is a partial numerical simulation curve of respirable dust concentration obtained in the embodiments of the present invention;

[0059] Figure 5This is a comparison chart of simulated and partial monitoring data in an embodiment of the present invention;

[0060] Figure 6 This is a schematic diagram illustrating the fusion of numerical simulation and monitoring results in an embodiment of the present invention;

[0061] Figure 7 This is a schematic diagram of the DRL-optimized LSTM-CNN-Attention model in an embodiment of the present invention;

[0062] Figure 8 This is a graph showing the performance change trend of different models in random training rounds in embodiments of the present invention;

[0063] Figure 9 This is a comparison chart of risk predictions for continuous mining machine drivers using different models in an embodiment of the present invention;

[0064] Figure 10 This is a graph showing the trend of risk changes over time for different types of work in this embodiment of the invention. Detailed Implementation

[0065] The invention will now be further described with reference to the accompanying drawings.

[0066] like Figure 1 and Figure 2 As shown, a method for dynamic prediction of dust inhalation dose and early warning of pneumoconiosis risk in dust-exposed workers, driven by monitoring and simulation, includes the following steps:

[0067] S1. By combining on-site monitoring with dust information from the simulated breathing hemisphere region, a monitoring-simulation fusion microenvironmental field is constructed;

[0068] S2. Construct a CFD-DPM numerical model and a multi-source heterogeneous dust concentration field;

[0069] S3. Construct an LSTM-CNN-Attention prediction model based on DRL meta-controller optimization;

[0070] S4. Predict the inhaled dose using the LSTM-CNN-Attention prediction model;

[0071] S5. Real-time prediction of component-particle size synergistic toxicity weighting and disease probability;

[0072] S6, Online Closed-Loop Feedback and DRL Model Evolution.

[0073] Example: On-site monitoring of dust concentration for different work types, such as... Figure 3 As shown; partial dust concentration data of the individual respiratory hemisphere obtained by numerical simulation are as follows. Figure 4 As shown in (a), (b), and (c), the comparison charts between the simulation data and some monitoring data are as follows: Figure 5As shown; numerical simulation and field monitoring research are combined in a two-way collaborative manner, and Kalman filtering is used to introduce IoT data fusion methods to construct multi-source heterogeneous dust concentration fields, etc. Figure 6 As shown; DRL optimizes the LSTM-CNN-Attention model as follows Figure 7 As shown; the performance comparison curves with different models after random training epochs are as follows. Figure 8 As shown, with increasing random training epochs, the performance of the LSTM-CNN-Attention model, the Random Forest model, the Gradient Boosting Tree model, and the Logistic Regression model all gradually improves, which is consistent with the trend. In the early stages of training (10-20 epochs), the performance difference between the models is small, but after more than 25 epochs, the performance of the LSTM-CNN-Attention model continues to improve efficiently, while the performance improvement of the comparison model lags behind. By 50 epochs, the LSTM-CNN-Attention model has a 1% performance improvement over the best comparison model (Gradient Boosting Tree model), and the gap continues to widen. The LSTM-CNN-Attention model is more advantageous in reducing prediction errors.

[0074] Taking continuous mining machine operators as an example, the prediction of pneumoconiosis risk using different models is quite different. Figure 9 As shown, the trends in pneumoconiosis risk over time for different occupations are as follows: Figure 10 As shown, in predicting the risk of pneumoconiosis for continuous mining machine drivers, the prediction trends of the Logistic model and the comparative models (Weibull model, exponential model, and piecewise model) are roughly the same, which is consistent with the trend. The risk prediction percentage of the Logistic model is consistently between the maximum and minimum risks in the comparative models, indicating stable prediction, reliable data, and a more accurate reflection of the actual risk of pneumoconiosis for continuous mining machine drivers.

Claims

1. A method for dynamic prediction of dust inhalation dose and early warning of pneumoconiosis risk in dust-exposed workers through a monitoring and simulation-driven collaborative approach, characterized in that... Includes the following steps: S1. By combining on-site monitoring with dust information from the simulated breathing hemisphere region, a monitoring-simulation fusion microenvironmental field is constructed; S2. Construct a CFD-DPM numerical model and a multi-source heterogeneous dust concentration field; S3. Construct an LSTM-CNN-Attention prediction model based on DRL meta-controller optimization; S4. Predict the inhaled dose using the LSTM-CNN-Attention prediction model; S5. Real-time prediction of component-particle size synergistic toxicity weighting and disease probability; S6, Online Closed-Loop Feedback and DRL Model Evolution.

2. The method for dynamic prediction of dust inhalation dose and early warning of pneumoconiosis risk in dust-exposed workers driven by monitoring and simulation according to claim 1, characterized in that, The specific process of S1 is as follows: S1.1 On-site tracking of workers of different trades, while simultaneously deploying multiple miniature dust sensor arrays on the workers' shoulders, safety helmets, and chests to acquire concentration data of respirable dust and total dust at discrete monitoring points within the respiratory hemisphere. monitor (d, t), on-site monitoring of dust concentration for different work types; S1.2 The microenvironmental field was calculated using Computational Fluid Dynamics-Discrete Phase Model (CFD-DPM) combined with User-Defined Typical Respiratory Function (UDF). UDF enabled accurate respiratory dynamics modeling that matched individual physiological characteristics. The CFD-DPM-UDF simulation method and its results were validated using field monitoring results. The three-dimensional concentration field C in the respiratory hemisphere of different occupations was studied. sim (d, x, y, z, t), particle size classification, different component particles, and partial dust concentration data of the individual respiratory hemisphere obtained by numerical simulation.

3. The method for dynamic prediction of dust inhalation dose and early warning of pneumoconiosis risk in dust-exposed workers driven by monitoring and simulation according to claim 2, characterized in that, The specific process of S2 is as follows: S2.1, Using monitoring data C monitor (d, t) The entrance boundary conditions and dust source term intensity of the simulated field are corrected by Kalman filtering to reduce the simulation error to an allowable value; simulation supplementary monitoring: in the blind zone of the monitoring point, the simulation results C are used. sim (d,x,y,z,t) fills in, generating a fusion concentration field: C fuse (d, t)= w1′· C monitor (d, t)+ w2′· C sim (d, x, y, z, t); Where w1′ = (w1+w2) / w1; w2′ = (w1+w2) / w2; w1′ and w2′ are the normalized weights of the final weighted summation. and The original weights are represented by the monitoring data weight function and the simulated data weight function, respectively: ； ； Where α is the normalization coefficient, d is the distance to the nearest monitoring point (i.e., spatial attenuation), R(t) is the reliability index of the monitoring equipment at time t (i.e., time attenuation), β is the spatial attenuation coefficient, γ is the balance parameter, M(s, t) is the confidence score of the simulation model at location s and time t, and D(s, t) is the data sparsity index of location s near time t. S2.2 Obtain worker coordinates (x, y, z) and attitude angle θ∈[0°, 180°] through UWB positioning and inertial sensors; From the fusion concentration field C fuse Concentration at the nasal and oral sites was extracted from (d, t), and combined with respiratory rate. breath (t), calculate individual exposure concentration: ； Among them, C i (d, t) represents the concentration of dust with particle size d at time t, P(d, θ) is the directional weighting function, θ is the angle relative to the worker, and f breath (t) represents respiratory rate data.

4. The method for dynamic prediction of dust inhalation dose and early warning of pneumoconiosis risk in dust-exposed workers driven by monitoring and simulation according to claim 3, characterized in that, The DRL in S3 uses a near-end strategy to optimize the PPO algorithm to solve the high variance problem in the continuous action space. The state space, i.e., the real-time feature vector, is represented as: ； The action space, i.e., the discrete-continuous hybrid action, is represented as: ； , ， ； Among them, v air (t) represents the air velocity, Δkernel is the directly chosen size of the convolutional kernel, Δhidden is the dimension of the hidden layer, and Δheads is the number of attention heads. Adjustment of SiO2 dust feeding coefficient, ΔTF coal ∈[-0.07,0.07] represents the adjustment of the coal dust feeding coefficient, ΔTF other ∈[-0.01, 0.01] represents adjustments for other dust feeding coefficients; Set the reward function R as follows: ； Among them, D pred For the predicted pollutant concentration, D actual To measure the actual inhaled dose, |D pred -D actual |For prediction error, WHO limit = 0.05 mg / min, A t Let A be the action vector at time step t. t-1 Let be the action vector at time step t-1.

5. The method for dynamic prediction of dust inhalation dose and early warning of pneumoconiosis risk in dust-exposed workers driven by monitoring and simulation according to claim 4, characterized in that, The specific process of S4 is as follows: S4.1 Input the time series data X∈R^(T×F) into the LSTM-CNN-Attention prediction model, where T is the time step and F is the feature dimension; S4.2 Dynamic Convolutional Feature Extraction Layer: The CNN layer adopts a configurable convolutional kernel group to extract multi-scale local exposure patterns in real time. The convolutional kernel configuration space is kernel∈{3, 5, 7}. The decision-making mechanism is dynamically issued by the deep reinforcement learning agent according to the complexity of the input features, and the output is a multi-scale representation of local temporal dependence. S4.

3. Based on the worker's length of service, memory capacity is dynamically allocated. The number of hidden units is hidden∈{32, 64, 96}. The memory depth is automatically matched by DRL based on the historical exposure duration. The adaptive balance of long short-term memory is achieved through LSTM layers. S4.4 Adjustable multi-head attention mechanism: heads∈{2, 4, 6}, the attention layer automatically focuses on key time steps, weight α t = softmax(QK) T / √d k Dynamically focus on high-risk exposure time windows.

6. The method for dynamic prediction of dust inhalation dose and early warning of pneumoconiosis risk in dust-exposed workers driven by monitoring and simulation according to claim 5, characterized in that, The specific process of S5 is as follows: S5.1, The expression for the dynamic toxicity factor is constructed as follows: ； in, Due to the toxicity of silica, TF coal (t) represents the combined toxicity of coal dust and its impurities, TF other (t) represents the mixed toxicity of non-silicon, non-coal dust, TF interaction (t) represents the synergistic additional toxicity among dust particles, and α1(t), α2(t), α3(t), and α4(t) represent the toxicity weights, respectively. S5.2, The particle size toxicity weighted expression is constructed as follows: ； in, The median value (μm) of the i-th particle size interval. For different dust particle size toxicity weighting functions, The toxicity factors for different dust particles in this particle size range are α1, α2, α3, and α4, which are the toxicity weights, respectively. S5.3, the real-time mapping expression for the 10-year disease probability is: ； in, For the occupational time-varying disease probability, F J G is the dose-response function for occupational toxicity. J To account for the age effect correction of job experience, H J For the genetic-job interaction function, TF cumulative (t) represents the cumulative toxic exposure up to time t, where Genetics is the genetic susceptibility factor, and E J To reveal history and patterns, Age(t) represents the current age, and θ J S is the job parameter vector. J For the smoking function.

7. The method for dynamic prediction of dust inhalation dose and early warning of pneumoconiosis risk in dust-exposed workers driven by monitoring and simulation according to claim 6, characterized in that, The specific process of S6 is as follows: Deep reinforcement learning (DRL) relies on the continuous interaction between the policy network and the environment during the training process, and adjusts and optimizes the action policy based on the deviation between the prediction result and the actual target. It adopts an iterative mechanism of "execution-collection-learning": after executing an action according to the current policy, the system records its state, action, reward obtained and new state after transition, forming a complete interaction experience and storing it in the experience pool. During training, the system periodically extracts a set batch of historical interaction data from the experience pool, uses the temporal difference algorithm to estimate the value error, and then completes the parameter update of the policy network through gradient backpropagation.

Images