Job safety early warning method and system based on big data driving

Abstract

This invention relates to the field of operational safety management technology, and discloses a big data-driven operational safety early warning method and system. This method employs field mapping fusion and multi-level annotation to process multi-source heterogeneous operational data. It extracts short, medium, and long-term time-series features through multi-scale time window sliding driven by the enterprise's operational cycle. A deep learning training sample set is input into a network architecture that integrates LSTM time-series coding, time decay enhancement, and multi-head attention. A risk prediction model is trained using a time decay enhancement focus loss function. A tiered early warning strategy is generated through attention weight tracing and log-odds transformation risk adjustment methods. The model is optimized using a multi-condition triggered incremental update mechanism. This invention can accurately capture the complex risk patterns of operational safety, providing strong protection for enterprise safe production.

Description

Technical Field

[0001] This invention relates to the field of work safety management technology, and more specifically, to a work safety early warning method and system based on big data. Background Technology

[0002] With the continuous expansion of industrial production scale, the safety management of high-risk operations such as hot work, high-altitude work, and confined space work has become a significant challenge for enterprises. Traditional work approval and safety management mainly rely on human experience and judgment. Approving personnel need to assess risks based on work application information and site conditions. This method suffers from high subjectivity, low efficiency, and difficulty in making full use of historical data.

[0003] Existing operational safety early warning methods typically employ a single, fixed time window for time-series analysis, failing to simultaneously capture short-term fluctuations, medium-term trends, and long-term patterns, resulting in incomplete risk feature extraction. Furthermore, existing methods have limited effectiveness in handling imbalanced sample classes, exhibiting insufficient ability to identify minority risk events and prone to missed detections. In addition, existing technologies lack modeling of the differences in the importance of time-series data, failing to highlight the guiding value of recent data for current predictions.

[0004] Therefore, there is a need for a work safety early warning method that can make full use of historical work data, accurately capture multi-scale time-series features, effectively handle sample imbalance problems, and provide interpretable early warning results, so as to improve the safety management level of enterprises and reduce the accident rate. Summary of the Invention

[0005] This invention provides a big data-driven method and system for early warning of operational safety, which solves the technical problems of incomplete extraction of time-series features, limited effect of sample imbalance processing, and insufficient modeling of the importance differences of time-series data in related technologies.

[0006] This invention provides a big data-driven operational safety early warning method, comprising the following steps: S1: Obtain multi-source heterogeneous job data, process it using field mapping fusion and multi-level annotation methods, and output a standardized job feature dataset; S2, based on a standardized operational feature dataset, extracts short, medium and long-term time series features through multi-scale time window sliding driven by the enterprise's operational cycle, and generates a deep learning training sample set; S3 inputs the deep learning training sample set into a network architecture that integrates LSTM temporal coding, temporal decay enhancement and multi-head attention, and trains the risk prediction model using the temporal decay enhancement focus loss function. S4. Use the trained risk prediction model to infer the characteristics of the work to be approved, and generate a graded early warning strategy through the risk adjustment method of attention weight tracing and log probability transformation. S5 provides visualized push notifications for tiered early warning strategies and collects operational feedback data. It optimizes the model using a multi-condition triggered incremental update mechanism and applies a hybrid approach of rule engine and transfer learning for cold start scenarios.

[0007] In a preferred embodiment, S1 includes: Mapping analysis is performed on fields from various data sources to identify fields with the same meaning but different names and unify them. A standard model for job data is established, and corresponding feature fields are extracted for different job types. Numerical features are normalized, categorical features are one-hot encoded, missing numerical features are filled with the historical average of the same job type, and missing categorical features are filled with the category with the highest frequency of occurrence. A multi-level labeling strategy is adopted to label the risk of the work results. Works that have accidents, anomalies, multiple rejections of approval, or serious violations found during on-site inspections are labeled as positive samples. For work that has been rejected multiple times but is completed safely after optimizing the work plan, the feature vector of the original application is extracted and labeled as a positive sample, and the feature vector of the optimized resubmission is extracted and labeled as a negative sample. Works that are completed normally without any anomalies are labeled as negative samples.

[0008] In a preferred embodiment, S2 includes: The lengths of short-term, medium-term, and long-term time windows are determined based on the company's actual operating cycle. The length of the short-term time window is determined based on the company's weekly work plan cycle, the length of the medium-term time window is determined based on the company's monthly maintenance cycle, and the length of the long-term time window is determined based on the company's quarterly production cycle. For the short-term window, the short-term feature vector is composed of the statistical operation frequency feature vector, short-term risk incidence rate, temperature distribution characteristics, and gas concentration characteristics. For the medium-term window, the medium-term feature vector is composed of the changing trend of risk incidence rate in each sub-interval, the fluctuation variance of risk factors, and the information entropy of operation type distribution. For the long-term window, the macro feature vector is composed of the seasonal feature coefficient, cumulative risk index, and production load change trend.

[0009] In a preferred embodiment, S2 further includes: For low-frequency job types, if the number of samples of a specific job type within a window is lower than the preset minimum sample threshold, a time-weighted interpolation method is used to find windows with sufficient data of the corresponding job type in the adjacent time periods before and after the corresponding window. The weight coefficients are calculated based on the time distance, and a weighted average is performed to obtain the interpolation features and construct a job type mask matrix. Short-term, medium-term, and macro-level feature vectors are concatenated along the feature dimension to form a multi-scale comprehensive feature vector. These vectors are then stacked in sliding order to form a time-series feature matrix. The jobs within the future prediction period corresponding to each time window are selected and labeled. The time-series feature matrix is ​​then uniformly processed into a fixed-length time series and divided into training, validation, and test sets according to time order.

[0010] In a preferred embodiment, S3 includes: In the feature embedding module, the input feature matrix is ​​mapped to the embedding space through a linear transformation to obtain the embedded feature tensor. The job type mask matrix is ​​then embedded and mapped. The feature embedding and type embedding are added element by element to obtain the initial representation tensor. In the temporal coding module, LSTM is used to temporally encode the initial representation tensor. The time series is processed step by step through the gating mechanism of forget gate, input gate, and output gate to obtain the temporally encoded hidden state sequence. In the time decay enhancement module, the time interval between each historical time step and the current prediction time is calculated, the decay coefficient is calculated using the exponential decay function, and the decay coefficient is multiplied element by element with the temporal encoded hidden state of the corresponding time step to obtain the enhanced hidden state sequence.

[0011] In a preferred embodiment, S3 further includes: The task feature vector at the target prediction time is extracted and mapped to the query vector Q. The enhanced hidden state sequence is mapped to the key vector sequence K and the value vector sequence V respectively. For each attention head, linear transformations are performed on Q, K and V respectively. The dot product of the query vector and the key vector at each time step is calculated to obtain the similarity score. The attention weight distribution is obtained by scaling and softmax normalization. The attention weights are weighted and summed with the corresponding value vectors to obtain the output vector of the corresponding attention head. After concatenating the output vectors of each attention head, feature fusion is performed through a two-layer fully connected network to obtain a comprehensive risk representation vector. The risk probability prediction value is then output through an output fully connected layer and a sigmoid activation function.

[0012] In a preferred embodiment, S3 further includes: The loss term consists of a basic cross-entropy term, a focus modulation term, and a time decay weight term. The focus modulation term dynamically adjusts the loss weight through a modulation factor. When the model has a high confidence level in predicting a sample, the loss weight of the corresponding sample is reduced. When the prediction confidence level is low, the loss weight of the corresponding sample is maintained. A balancing factor is introduced to apply different class weights to positive and negative samples. The time decay weight term is obtained by averaging the decay coefficients of all time steps within the corresponding time window of the sample to obtain the comprehensive time decay weight, which is introduced into the loss function as a sample-level weight coefficient. The loss of a single sample is the product of the time decay weight, the class weight, the focus modulation factor, and the basic cross-entropy loss. The total loss of the training batch is the mean of the losses of each sample.

[0013] In a preferred embodiment, S4 includes: The risk adjustment coefficient is determined based on the baseline risk rate of each type of operation and the environmental conditions of the current operation. The predicted risk probability value is converted into logarithmic probability form, and the risk adjustment coefficient is converted into logarithmic form. The two are added together and converted back into probability form through inverse transformation to obtain a comprehensive risk score. The risk level is determined based on the preset low risk threshold, medium risk threshold and high risk threshold. Analyze the attention weight distribution, select the historical time steps with the highest weight ranking, and calculate the similarity between the current operation and historical operations based on four dimensions: operation type matching degree, environmental parameter similarity, operation location relevance, and time and season similarity. Select the historical operations with the highest similarity ranking as similar historical cases, compare the feature distribution of the current operation with that of historical safe operations to identify abnormal feature items and generate a list of risk factors. Based on the risk factor list, retrieve matching safety measures from the mapping knowledge base to generate a list of safety measure suggestions, and generate a graded early warning strategy based on the risk level matching early warning strategy rules.

[0014] In a preferred embodiment, S5 includes: The system includes regular triggering conditions, emergency triggering conditions, performance triggering conditions, and data distribution triggering conditions. The regular triggering condition is when the amount of new data reaches a preset data volume threshold or when the time since the last training exceeds a preset time threshold. The emergency triggering condition is when a major security incident occurs. The performance triggering condition is when the prediction accuracy continuously declines beyond a preset performance decline threshold. The data distribution triggering condition is when the proportion of features with deviations exceeds a preset distribution deviation ratio threshold. When the amount of new data is less than the preset ratio threshold, incremental training is used; otherwise, full training is used. The rule engine and transfer learning hybrid method includes: when the number of historical samples is less than the preset cold start judgment threshold, the rule engine is activated to perform initial risk assessment, while loading the parameters of trained models of similar job types, adopting a hierarchical training strategy to freeze the bottom feature extraction layer and only fine-tune the top risk judgment layer, weighting and fusing the rule engine evaluation score and the model evaluation score, dynamically adjusting the weight ratio as data accumulates, and switching to model prediction when the amount of data reaches the preset model maturity threshold.

[0015] In a preferred embodiment, the big data-driven job safety early warning system is used to perform the steps in the above-described big data-driven job safety early warning method, including: The data fusion and annotation module is used to acquire multi-source heterogeneous operation data, and processes it using field mapping fusion and multi-level annotation methods to output a standardized operation feature dataset. The temporal feature construction module is used to extract short, medium and long-term temporal features based on a standardized operational feature dataset by sliding a multi-scale time window driven by the enterprise's operational cycle, and to generate a deep learning training sample set. The model training module is used to input the deep learning training sample set into a network architecture that integrates LSTM temporal coding, temporal decay enhancement and multi-head attention, and trains the risk prediction model using the temporal decay enhancement focus loss function. The risk reasoning and early warning module is used to reason about the characteristics of the work to be approved using the trained risk prediction model, and generates a graded early warning strategy through the risk adjustment method of attention weight tracing and log probability transformation. The feedback optimization module is used to visually push the tiered early warning strategy and collect operation feedback data. It adopts a multi-condition triggered incremental update mechanism to optimize the model and applies a hybrid method of rule engine and transfer learning for cold start scenarios.

[0016] The beneficial effects of this invention are as follows: The method employs a multi-scale time window sliding extraction driven by the enterprise's operational cycle to extract short, medium, and long-term time series features. The short-term window captures immediate risk fluctuations, the medium-term window captures cyclical risks of equipment maintenance, and the long-term window captures seasonal environmental impacts. The three methods work together to achieve comprehensive risk feature extraction, which can more accurately capture the complex risk patterns of operational safety compared to a single fixed time window method. The designed time decay enhanced focus loss function incorporates the time decay mechanism into the loss calculation process. It can not only solve the problem of imbalance between positive and negative samples, but also dynamically adjust the contribution of samples to model training according to the time distance of historical data. This achieves synergistic optimization of sample imbalance handling and time-series importance modeling, and improves the ability to identify minority risk events. Attached Figure Description

[0017] Figure 1 This is a flowchart of the big data-driven operational safety early warning method of the present invention; Figure 2 This is a flowchart of the big data-driven operation safety early warning method of the present invention. Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses a big data-driven operational safety early warning method, such as... Figures 1 to 2 As shown, it includes the following steps: S1: Obtain multi-source heterogeneous job data, process it using field mapping fusion and multi-level annotation methods, and output a standardized job feature dataset; The system retrieves all records from historical work application forms, approval process tables, and accident event tables from the enterprise production management system database to obtain basic application and process information for the work. It also extracts environmental monitoring records corresponding to the work time and location from the enterprise production management system, including temperature, pressure, and combustible gas detection data for the work area. The system supports access to an IoT monitoring platform to obtain real-time monitoring data to enhance data acquisition capabilities. Furthermore, it retrieves metadata information for safety measure photos and videos uploaded to the work site from the file storage system, including upload time, number of files, and file type. Finally, it performs mapping analysis on fields from each data source, identifying and unifying fields with the same meaning but different names.

[0020] Establish a standard model for operational data, defining a set of core fields including operational identifier, operational type, time, location, personnel, environment, measures, and results. For hot work data, extract information from the operational application information such as hot work level, ignition source type, operational area identifier, combustible material category, planned hot work duration, and number of supervisors. Extract information from the approval process information such as the number of approval steps, total approval time, and number of rejections. Extract information from environmental monitoring data such as average temperature, maximum temperature, and peak combustible gas concentration during the operational period. Query the accident event table to see if the operation is associated with any accident records and tag it accordingly. For high-altitude operations data, extract feature fields such as operational height, protective measure type code, number of personnel, safety belt usage indicator, wind force level, and rainfall indicator. For confined space operations data, extract feature fields such as space type code, oxygen concentration detection value, hydrogen sulfide concentration detection value, carbon monoxide concentration detection value, ventilation duration, and respirator equipment indicator.

[0021] The extracted numerical features are normalized, and the minimum and maximum values ​​of each feature in the entire dataset are calculated, linearly mapping the feature values ​​to the 0-1 interval. Categorical features are transformed using one-hot encoding: job type is converted to a four-dimensional binary vector, fire level to a three-dimensional binary vector, and location information to a region encoding vector. For missing feature fields, if the feature is numerical, the historical average of the same job type is used to fill the gaps; if the feature is categorical, the category with the highest frequency is used. Periodic features are extracted from the timestamp field: the hourly value is calculated and converted to a 24-dimensional periodic code; whether it is a workday is converted to a binary identifier; and the month information is extracted and converted to a 12-dimensional seasonal code.

[0022] Risk labels are applied to work results using a multi-level labeling strategy to accurately reflect the true risk status of the work. The incident table is queried to determine if a safety accident occurred, approval records are checked to determine if the work was repeatedly rejected or marked as abnormal, and on-site inspection records are checked to determine if serious violations were found. Works involving accidents, abnormalities, multiple rejections of approval, or serious violations found during on-site inspections are labeled as positive risk events and assigned a value of 1. For work that was repeatedly rejected in the approval record or marked as high-risk by approvers but was ultimately completed safely after supplementing safety measures and optimizing the work plan, the feature vector from the original application is extracted and labeled as a positive sample, while the feature vector from the optimized resubmission is extracted and labeled as a negative sample. This allows the subsequently trained model to learn the improvement effects of risk factors and the effectiveness of safety measures. Works that were successfully approved, executed according to regulations, and completed normally without abnormalities are labeled as negative samples and assigned a value of 0.

[0023] After statistical analysis of the labeled samples, it was found that the number of positive samples (i.e., risk events) was far less than that of negative samples, indicating a significant sample imbalance. To address the impact of sample imbalance on model training, the ratio of positive to negative samples was recorded. In subsequent model training, a focus loss function was adopted. This function dynamically adjusts the loss weights of different samples to achieve a sample weighting mechanism, improving the model's ability to identify minority risk events. The processed feature fields of each job type were concatenated according to the field order of the standard model. Fields not involved in a certain job type were filled with zero values ​​to form a fixed-dimensional job feature vector. All job records were converted into a set of feature vectors of a uniform dimension and stored along with timestamps and risk labels to construct a complete standardized job feature dataset. This standardized job feature dataset is organized in the form of a two-dimensional matrix, with each row corresponding to a job record, containing a fixed-dimensional feature vector, a job occurrence timestamp, and a risk label. The dataset is stored as a CSV file or a relational database table for easy reading and processing in subsequent steps.

[0024] S2, based on a standardized operational feature dataset, extracts short, medium and long-term time series features through multi-scale time window sliding driven by the enterprise's operational cycle, and generates a deep learning training sample set; The standardized operational feature dataset is sorted in ascending order according to the operation occurrence time to obtain a time-ordered operation sequence. Existing time series analysis methods typically use a single fixed time window, which cannot simultaneously capture short-term fluctuations, medium-term trends, and long-term patterns. This invention innovatively adopts a multi-scale time window aligned with the enterprise's operating cycle. The short-term window captures immediate risk fluctuations, the medium-term window captures cyclical risks related to equipment maintenance, and the long-term window captures seasonal environmental impacts. These three elements work together to achieve comprehensive risk feature extraction. The length of the multi-scale time window is determined based on the enterprise's actual operating cycle. The short-term time window length is determined according to the enterprise's weekly work plan cycle, with a default value of 7 days, which can be adjusted within the range of 3 to 14 days based on actual conditions. This is used to capture recent work intensity and immediate risk fluctuation characteristics. The medium-term time window length is determined according to the enterprise's monthly maintenance cycle, with a default value of 30 days, which can be adjusted within the range of 15 to 60 days based on actual conditions. This is used to capture cyclical risks related to equipment maintenance and trends in work load changes. The length of the long-term time window is determined based on the company's quarterly production cycle and seasonal variation patterns. The default value is set to 90 days, which can be adjusted within the range of 60 to 180 days according to actual conditions. This is used to capture the impact patterns of seasonal environmental factors such as high temperatures in summer and low temperatures in winter on operational safety. Select a time window sliding step of 1 day, and slide the window forward from the start time of the dataset.

[0025] For each window location, the immediate risk characteristics within the short-term window are statistically analyzed. The total number of operations occurring within the window is calculated, and the frequency of each type of operation is counted to obtain an operation frequency feature vector. The number of operations marked as risk events within the window is calculated, and divided by the total number of operations to obtain the short-term risk incidence rate. The ambient temperature feature values ​​of all operations within the window are extracted, and the average temperature, maximum temperature, and temperature standard deviation are calculated to obtain temperature distribution characteristics. The gas concentration feature values ​​of all operations within the window are extracted, and the average concentration and peak concentration are calculated. The longest consecutive period of high-risk operations within the window and the percentage of high-risk operations are statistically analyzed. These statistical characteristics are combined to form a short-term feature vector.

[0026] Then, the trend risk characteristics within the medium-term window are calculated. The medium-term window is divided into several sub-intervals, and the risk incidence rate of each sub-interval is calculated. The trend of risk incidence rate changes between adjacent sub-intervals is compared to determine whether the risk is on the rise. The variance of the volatility of various risk factors within the window is calculated; the larger the variance value, the higher the instability of the factor. The distribution of job types within the window is statistically analyzed, and the information entropy of the job type distribution is calculated. The entropy value reflects the degree of concentration or dispersion of job types. The proportion of jobs whose approvals were rejected within the window and the median approval time are extracted. These trend characteristics are combined to form a medium-term feature vector.

[0027] Next, macroeconomic risk characteristics within the long-term window are extracted. The monthly distribution of risk events within the long-term window is analyzed to identify seasonal fluctuation patterns and extract seasonality coefficients. The cumulative total of high-risk operations and cumulative total of risk events within the long-term window are calculated to obtain the cumulative risk index. The changing trend of operation frequency within the long-term window is analyzed to determine the increase or decrease in production load. These long-term characteristics are then combined to form a macroeconomic feature vector.

[0028] This study analyzes the temporal distribution characteristics of various work types, statistically analyzing the frequency of hot work, high-altitude work, confined space work, and temporary power supply work over historical periods. It reveals an imbalance in the temporal distribution of different work types, with hot work occurring multiple times daily on average due to frequent equipment maintenance needs, while confined space work, due to stringent operating conditions, occurs only a few times per week on average. For low-frequency work types such as confined space work, the sample size within a time window is limited. If the sample size for a specific work type within a window falls below a set minimum sample threshold, a time-weighted interpolation method is used to supplement the features. Windows with sufficient data for that work type within adjacent time periods are identified, and weight coefficients are calculated based on time distance, with closer windows having higher weights. The interpolated features for the current window are obtained by weighted averaging of the feature values ​​from neighboring windows. Each type of work occurring within a window is identified, and a work type mask matrix is ​​constructed. The rows of the matrix correspond to the time window, the columns to the work type, and the matrix elements represent the occurrence identifiers of that type of work within that window.

[0029] Short-term, medium-term, and long-term feature vectors are concatenated along their feature dimensions to form a multi-scale comprehensive feature vector for a single time window. Following the sliding order of the time windows, the comprehensive feature vectors of each window are stacked into a time-series feature matrix, where the row dimension corresponds to the time step and the column dimension corresponds to the multi-scale features. To construct training samples, tasks occurring within the future prediction period corresponding to each time window are selected. The length of this prediction period is determined based on the average approval cycle of the enterprise, with a default value set to the range of 3 to 7 days; this embodiment uses 3 days. If a risk event occurs during the prediction period, the sample corresponding to that time window is marked as a positive sample; otherwise, it is marked as a negative sample. The time-series feature matrix is ​​used as the sample input, and the corresponding risk label is used as the sample output to construct a time-series sample dataset for model training. To ensure consistency of model input, the time-series feature matrix is ​​uniformly processed into a fixed-length time series, with the time step length set to be consistent with the long-term time window length (default value 90 steps). For sequences with insufficient historical data, zero values ​​are padded at the front; for sequences exceeding this length, the most recent time step is truncated as the model input. The dataset is divided into training, validation, and test sets in chronological order, ensuring that the training set is earlier than the validation and test sets to avoid data leakage.

[0030] S3 inputs the deep learning training sample set into a network architecture that integrates LSTM temporal coding, temporal decay enhancement and multi-head attention, and trains the risk prediction model using the temporal decay enhancement focus loss function. The overall architecture of the job risk prediction network is constructed, which includes a feature embedding module, a temporal coding module, a temporal decay enhancement module, a multi-head attention module, a feature fusion module, and a risk prediction output module.

[0031] In the feature embedding module, the embedding layer dimension is set, with a default value of 128 dimensions, which can be adjusted within the range of 64 to 256 dimensions depending on the feature complexity. The input feature matrix in the time-series sample dataset is mapped to the embedding space through a linear transformation to obtain the embedded feature tensor. The job type mask matrix is ​​also embedded and mapped to obtain the embedded representation of the job type. The feature embedding and type embedding are then added element-wise to obtain the initial representation tensor that incorporates the job type information.

[0032] In the temporal encoding module, a Long Short-Term Memory (LSTM) network is used to temporally encode the initial representation tensor. LSTM, through its memory units and gating mechanism, can effectively capture long-term dependencies in job safety data, making it particularly suitable for learning seasonal risk patterns and periodic job patterns. The hidden layer dimension of the LSTM is set, with a default value of 256 dimensions, which can be adjusted within the range of 128 to 512 dimensions depending on the temporal complexity. The LSTM processes features progressively from the start of the time series to the current prediction time. A forget gate determines which historical information to discard, an input gate determines which new information to update to the memory units, and an output gate determines which information to output at the current time step. At each time step, the LSTM takes the current feature vector and the hidden state from the previous time step as input, updates the state of the memory units, and outputs the current hidden state. After progressively processing the entire historical time series, a temporally encoded hidden state sequence is obtained, which contains temporal evolution information from history to the present and long-term risk patterns.

[0033] In the time decay enhancement module, a time decay mechanism is introduced to address the uneven temporal distribution of different job types. The time interval between each historical time step and the current prediction time is calculated in days, and a decay weight is calculated based on this time interval to highlight the importance of recent job data for current risk prediction. The decay half-life parameter is determined using two methods: the preferred method is to statistically analyze the company's historical maintenance records and calculate the average value of minor maintenance cycles as the initial reference value for the half-life, because the job risk status is relatively stable within a minor maintenance cycle, while equipment status and the working environment may change after the minor maintenance cycle. The general method, when the company lacks clear maintenance cycle records, is to test the model performance corresponding to different half-life values ​​on a validation set using cross-validation, selecting the half-life value that gives the highest F1 score on the validation set as the final parameter. The system can automatically select the optimal value based on the validation set performance without manual adjustment. For the company in this embodiment, the half-life parameter is determined based on the average value of minor maintenance cycles. The default value can be set to a value within the range of 10 to 20 days; this embodiment uses 14 days, which reflects the effective memory period of the job safety status.

[0034] The half-life parameter is industry-adaptable. For the petrochemical industry, the default value is typically set to 10-20 days; for the construction industry, it can be set to 5-10 days; and for the power industry, it can be set to 15-30 days. Different companies can adjust the half-life parameter according to their actual maintenance cycles or through cross-validation. Historical data with time intervals shorter than the half-life retains high weight, while the weight of historical data with time intervals longer than the half-life decays exponentially. An exponential decay function is used to calculate the decay coefficient for each time step. Specifically, the time interval is divided by the half-life to obtain the normalized time distance, and then the decay coefficient is calculated by raising the normalized time distance to a power of 0.5. The decay coefficient is then multiplied element-wise by the temporal encoded hidden state of that time step to obtain the enhanced hidden state sequence adjusted for time decay. This enhancement mechanism makes the model more focused on the risk patterns of recent operational data, especially for low-frequency operation types such as confined space operations. It can effectively utilize limited recent samples for risk assessment and reduce the interference of sparse data from the distant future.

[0035] In the multi-head attention module, the number of attention heads is set based on the number of key dimensions affecting job safety. The default value is four, which can be adjusted from two to eight depending on the actual application scenario. The default dimension for each attention head is 64. The reason for setting the number of attention heads to four is that the key dimensions affecting job safety include job type, environmental parameters, time period, and spatial location. Each attention head is responsible for learning the correlation pattern of one dimension. Different head configurations were tested during the model design phase. Too few heads cannot fully cover the features of multiple dimensions, while too many heads are redundant and increase computational overhead. The four-head configuration achieves the best balance between feature coverage and computational efficiency. The multi-head attention mechanism allows the model to learn the risk correlation patterns between historical jobs and the target prediction time from different dimensions. During training, no dimensional constraints are pre-imposed on the four attention heads; instead, they freely learn the feature patterns they focus on through a data-driven approach.

[0036] After training, by analyzing the weight distribution characteristics of each attention head, the actual attention dimensions learned by each head can be identified. The method for identifying the attention dimensions of each attention head is as follows: For each attention head, the attention weight distribution characteristics on all test samples are statistically analyzed. First, the average weight of the attention head for different types of historical tasks is calculated. If the average weight for the same task type is higher than that for different task types and the difference exceeds 0.3, the similarity of the task type is determined. Second, the correlation coefficient between the attention head's weight and the difference in environmental parameters is calculated. If the weight is negatively correlated with temperature differences and gas concentration differences and the correlation coefficient is less than -0.5, the similarity of the environmental parameters is determined. Third, the temporal distribution pattern of the attention head's weight is analyzed. If the weight for historical tasks in the same month or time period shows periodic fluctuations, the time periodicity of the attention is determined. Fourth, the relationship between the attention head's weight and spatial distance is calculated. If the weight for historical tasks in the same or adjacent areas is higher, the spatial location correlation is determined. This post-hoc interpretive analysis reveals that different attention heads do indeed automatically differentiate their attention to different aspects of operational safety.

[0037] During training, for each training sample's target prediction time step, the corresponding task feature vector is extracted and mapped to a query vector Q through a linear layer. The enhanced hidden state sequence is mapped to a key vector sequence K and a value vector sequence V through linear layers, respectively. For each attention head, the corresponding query transformation matrix, key transformation matrix, and value transformation matrix are used to perform linear transformations on Q, K, and V, respectively, to obtain the query vector, key vector sequence, and value vector sequence of that head. The dot product of the query vector and the key vector at each time step is calculated to obtain a similarity score sequence, which reflects the correlation between the target prediction time step and each historical time step in the attention dimension of that attention head. The similarity score is scaled by dividing by the square root of the key vector dimension to avoid excessively large values ​​that could lead to gradient vanishing. A softmax normalization function is applied to the scaled score sequence to obtain the attention weight distribution, where the sum of the weight values ​​is 1. A larger weight value indicates a higher reference value for the target prediction at the corresponding historical time step. The attention weights are weighted and summed with the value vectors at the corresponding time steps to obtain the output vector of that attention head. The above calculation process is repeated for the remaining three attention heads. Each attention head uses an independent transformation matrix to learn the risk correlation pattern between historical task data and the target prediction time from different dimensions. The output vectors of the four attention heads are concatenated to obtain a 256-dimensional multi-head attention output vector.

[0038] Attention weight analysis of the trained model revealed that the first attention head tends to assign higher weights to historical operations with similar job types, indicating that it has learned the risk correlations based on job type similarity. The second attention head tends to assign higher weights to historical operations with similar environmental parameters, such as temperature and gas concentration, indicating that it has learned the risk impact patterns based on environmental parameter similarity. The third attention head tends to assign higher weights to historical operations in the same season or at the same time of day, indicating that it has learned seasonal risks and diurnal differences based on time periodicity. The fourth attention head tends to assign higher weights to historical operations in the same or adjacent areas, indicating that it has learned regional risk transmission relationships based on spatial location correlation. This multi-dimensional risk correlation learning enables the model to comprehensively capture the complex influencing factors of operational safety.

[0039] A nonlinear transformation is applied to the multi-head attention output vector, and feature fusion is performed through a two-layer fully connected network. The first fully connected layer maps the multi-head attention output vector to an intermediate dimension, applies a ReLU activation function to introduce nonlinearity, and then uses a Dropout layer to randomly discard some neurons with a set probability to prevent overfitting; the default Dropout probability is set to 0.3. The second fully connected layer maps the intermediate dimension vector to the final representation dimension, obtaining a comprehensive risk representation vector. This vector integrates the temporal evolution information of historical task data and the feature information of the current task, and highlights the historical risk patterns most relevant to the current task in four dimensions—type, environment, time, and space—through the multi-head attention mechanism. The synergistic effect of the time decay mechanism and the multi-head attention mechanism is reflected in the following ways: the time decay mechanism assigns time-sensitivity weights to historical data at the feature level, giving higher numerical weights to the feature representations of recent data; the multi-head attention mechanism learns the correlation between historical data and the current prediction at the association level, giving higher attention weights to historical data with high correlation. The two factors work together to form a dual weighting mechanism: for recent and highly relevant historical data, both time decay and attention weights are applied, giving them a dominant influence on the prediction results; for distant or less relevant historical data, the weights are doubly suppressed to reduce their interference. This synergistic mechanism allows the model to both learn the evolution of risk using long-term historical data and highlight the guiding value of recent relevant data for current predictions.

[0040] In the risk prediction output module, the comprehensive risk representation vector is input to and output to a fully connected layer, which maps the vector to a single numerical value. A sigmoid activation function is applied to this value, compressing the output value to between 0 and 1, yielding the predicted probability of a risk event occurring during the operation. Simultaneously, the attention weight distribution calculated in the multi-head attention module is extracted, and the weights of multiple attention heads are averaged to obtain the comprehensive attention weight for each historical time step in relation to the current prediction. Analyzing the attention weight distribution identifies the time steps with the highest weight values, and the corresponding operational data for these time steps is extracted, providing a basis for subsequent risk factor tracing and measure generation.

[0041] The risk prediction network is trained using the training set data, employing a mini-batch stochastic gradient descent method. A certain number of samples are randomly selected from the training set each time to form a batch. The default batch size is 32 samples, which can be adjusted between 16 and 128 samples depending on computational resources. The temporal characteristics of the samples within each batch are input into the network for forward propagation calculations to obtain the predicted risk probability value for each sample.

[0042] To address two core challenges in operational safety early warning scenarios—severe sample class imbalance and differences in the importance of time-series data—a customized time-decay-enhanced focus loss function was designed. This loss function innovatively integrates a time decay mechanism into the loss calculation process, enabling it not only to resolve the imbalance between positive and negative samples but also to dynamically adjust the contribution of samples to model training based on the temporal proximity of historical data. This achieves synergistic optimization of sample imbalance handling and time-series importance modeling.

[0043] The design of the time-decay-enhanced focus loss function comprises three core components: the basic cross-entropy loss term, the focus modulation term, and the time-decay weight term. For each sample in the training batch, the basic binary cross-entropy loss for that sample is calculated. Let p be the probability that the model predicts a sample as a risk event, and y be the true label of the sample, where y = 1 represents a risk event and y = 0 represents normal operation. Then, the basic cross-entropy loss is calculated as the negative natural logarithm of p when y = 1, and the negative natural logarithm of 1 minus p when y = 0. The mathematical expression is: .

[0044] A focus modulation term is introduced on top of the basic cross-entropy loss to address the sample imbalance problem. The probability that the model predicts a sample belongs to its true class is defined as... When the sample is a positive sample, i.e., y equals 1 The value is equal to p when the sample is a negative sample, i.e., y equals 0. Equals 1 minus p. The focus modulation factor is designed as follows: ,in The modulation factor is used to control the modulation intensity by focusing on parameters. Its mechanism is that when the model predicts the samples correctly and with high confidence, The modulation factor is close to 1 Approaching 0 significantly reduces the loss weight of easily classified samples, which is beneficial when the model mispredicts a sample or has low confidence. Smaller modulation factor A larger weight is used to maintain the loss weight for the difficult-to-classify sample, making the model training process focus more on samples that are difficult to classify correctly, especially minority class risk event samples. A balance factor is introduced. Different class weights are applied to positive and negative samples. When a sample is positive, the class weight is... When the sample is a negative sample, the class weight is By setting A value greater than 0.5 can further increase the weight of positive samples, i.e., risky events, in the loss function. (Definition) The class weights corresponding to the samples, when y equals 1 equal When y equals 0 equal .

[0045] In one embodiment of the present invention, a time decay weight term is further introduced, incorporating the time decay coefficient calculated by the time decay enhancement module in S3 into the loss function design. For each historical time step in the training samples, a decay coefficient has been calculated based on the time interval between that time step and the target prediction time of the sample. This decay coefficient reflects the timeliness value of historical data. The overall time decay weight of the sample is obtained by averaging the decay coefficients of all time steps within the corresponding time window. The weight value ranges from 0 to 1, with samples from more recent times having a weight closer to 1 and samples from more distant times having a smaller weight. This time decay weight is then applied. Introducing the loss function as a weighting coefficient at the sample level makes the model training process pay more attention to the risk patterns of recent operational data, and reduces the interference of risk pattern drift caused by changes in the working environment, equipment status updates, and management system adjustments in long-term data on the current model training.

[0046] Combining the three components mentioned above, the complete expression for the time-decay enhanced focus loss function is: the loss of a single sample equals the negative time-decay weight multiplied by the class weight multiplied by the focus modulation factor multiplied by the base cross-entropy loss, mathematically expressed as: For a training batch containing N samples, the total loss is the sum of the losses of each sample divided by N and then normalized, expressed as: .

[0047] The optimal combination of hyperparameters in the loss function is determined through validation set experiments. (Focusing on parameters) Control the degree of attention given to difficult-to-classify samples and test on the validation set. Model performance for different values, select the one that gives the highest F1 score on the validation set. The value can be set to 2.0 by default. Too small a value... The value is insufficient in moderating sample imbalance; an excessively large value... An excessively high value can cause the model to overemphasize a very small number of difficult samples while ignoring the overall distribution pattern. Balance factor Control the class weight ratio of positive and negative samples. Considering the significant disparity in the ratio of positive to negative samples, test on the validation set. Model performance under different values, aiming to achieve the best balance between accuracy and recall. The default value can be set to 0.75. The half-life parameter of the time decay weight has been determined in the time decay enhancement module. This parameter is consistent with the time decay weight term in the loss function to ensure the coordination and consistency of the time decay mechanism during forward and backward propagation.

[0048] This customized loss function implements a three-tiered collaborative optimization mechanism. The first tier is a focus modulation mechanism to address the sample class imbalance problem. By dynamically adjusting the loss weights of easily classified and difficult-to-classify samples, the model can effectively learn the risk characteristics of a minority of positive samples even in a training environment with a large number of negative samples. The second tier is a time decay mechanism to address the difference in importance of time-series data. By assigning higher training weights to recent samples, the model pays more attention to the risk patterns in the current working environment, improving its adaptability to changes in operational safety status. The third tier is a synergistic effect with a multi-head attention mechanism. The loss function guides the model to learn to value recent data through time decay weights, while the multi-head attention mechanism learns the correlation patterns between historical data and current predictions through attention weight distribution. Together, they enable the model to capture both long-term risk evolution trends and highlight recent risk status characteristics. Models trained using the time decay enhanced focus loss function of this invention show further improved risk event recall, enhancing the ability to identify minority risk events while maintaining overall prediction accuracy.

[0049] Backpropagation is performed to calculate the gradient of the loss function with respect to the network parameters. The Adam optimization algorithm is then used to update the parameters of the embedding layer, LSTM, attention layer, and fully connected layer in the network. The learning rate is set to a default value of 0.001, the momentum parameter is set to a default value of 0.9, and the number of training epochs is determined based on the performance on the validation set, with a default value of 50 epochs. After each epoch, the model performance is evaluated using the validation set, and metrics such as prediction accuracy, recall, and F1 score are calculated. The network parameters corresponding to the epoch with the highest F1 score on the validation set are selected as the final model parameters and saved.

[0050] S4. Use the trained risk prediction model to infer the characteristics of the work to be approved, and generate a graded early warning strategy through the risk adjustment method of attention weight tracing and log probability transformation. When an employee submits a new work request through the system, the system receives the various fields from the work request form, including the requested work type, planned work time, work location, work description, and applicant information. It queries the enterprise's production management system or IoT monitoring platform for the most recent environmental monitoring data corresponding to the work location, obtaining real-time environmental parameters such as current temperature, pressure, and gas concentration. It also queries the weather forecast system to obtain weather forecast data for the planned work period, including predicted temperature, wind speed, and rainfall probability. The system compares real-time and predicted environmental parameters to analyze trends in environmental conditions. It calculates the rate of temperature rise; if the predicted temperature for the work period is higher than the current temperature, it marks a risk of temperature rise. It calculates the historical fluctuation range of gas concentration to assess the potential range of gas concentration changes during the work period. Following the feature extraction rules defined in S1, it extracts the corresponding feature fields for the work type from the work request information. For hot work requests, it extracts information such as hot work level, ignition source type, work area, combustible material category, and planned duration, converting them into feature values. Real-time environmental monitoring parameters, predicted environmental parameters, and environmental change trend characteristics are all incorporated into the feature vector to form a comprehensive environmental feature that integrates the current state and future predictions. The extracted features undergo the same normalization and encoding processes as in S1 to ensure that the feature format is consistent with the training data.

[0051] The processed current job feature vector is input into the risk prediction network trained by S3. The network's layer parameters are loaded from the model storage file, including the embedding layer weight matrix, the gating weight matrix and memory unit weight matrix of the LSTM layer, the query key-value transformation matrix of the attention layer, and the weights and bias parameters of the fully connected layers. Historical job data for a certain period prior to the current application time is extracted from the standardized job feature dataset. The time period length is consistent with the long-term time window length. A historical time-series feature sequence is constructed according to the S2 method, containing time window features at three scales: short-term, medium-term, and long-term, forming a historical time-series feature matrix consistent with the model input requirements. The current job features and the historical time-series feature sequence are input into the network for forward inference. The network sequentially undergoes calculations in the modules of feature embedding, temporal encoding, time decay enhancement, multi-head attention, feature fusion, and output prediction, ultimately outputting the predicted risk probability value for the current job. Simultaneously, the attention weight distribution of the current job at each historical time step is extracted from the multi-head attention module. This weight distribution reflects the reference value of historical job data for current risk prediction.

[0052] The baseline risk rate for each job type was statistically analyzed from historical operational data. The proportion of historical risk events for hot work, confined space work, work at height, and temporary electrical work was calculated. Based on the obtained baseline risk rates and the risk level classification standards for different job types in industry safety regulations, a risk adjustment coefficient was determined for each job type. For hot work, due to the involvement of open flames and flammable materials, the historical risk rate is relatively high, and a risk adjustment coefficient of 1.2 was set. For confined space work, due to the enclosed space and complex environment, the risk adjustment coefficient was set at 1.15. For work at height, due to the risk of falls, the risk adjustment coefficient was set at 1.1. For temporary electrical work, the relative risk is relatively low, and a risk adjustment coefficient of 1.05 was set. Dynamic adjustments were made based on the current environmental conditions. If the current ambient temperature exceeds 30 degrees Celsius and the job type is hot work, the risk adjustment coefficient was increased from 1.2 to 1.5 to reflect the additional risks of hot work in high-temperature environments. If rainfall is predicted during the work period and the work type is high-altitude work or temporary power supply work, the risk adjustment coefficient will be increased accordingly.

[0053] A log-odds transformation-based method is employed to fuse the predicted risk probability with the risk adjustment coefficient, ensuring that the adjusted risk score remains within the valid range of 0 to 1. To avoid numerical calculation anomalies, boundary cases are handled: when the model's output risk probability p is less than 0.01, p is set to 0.01; when p is greater than 0.99, p is set to 0.99. This ensures that the division and logarithmic operations in the log-odds transformation are within the valid numerical range. The processed predicted risk probability p is converted to log-odds form, and the log-odds value is calculated as the natural logarithm of the quotient of p divided by 1 minus p. Then, the risk adjustment coefficient k is converted to logarithmic form, and the adjustment amount is calculated as the natural logarithm of k. The original log-odds value is added to the adjustment amount to obtain the adjusted log-odds value. Finally, an inverse transformation is used to convert the adjusted log-odds value back to probabilistic form, and the comprehensive risk score is calculated as an exponential function of 1 divided by 1 plus the negative of the adjusted log-odds value. The mathematical expression of this method is that the adjusted probability equals 1 divided by 1 plus the negative of e (original logarithmic probability plus logarithmic adjustment raised to the power of e), i.e. This logarithmic probability-based adjustment method ensures that regardless of the adjustment coefficient, the adjusted risk score remains between 0 and 1. Furthermore, the risk score increases when the adjustment coefficient is greater than 1 and decreases when the adjustment coefficient is less than 1, meeting the actual needs of risk adjustment. Risk level judgment thresholds are set based on the company's risk tolerance and historical risk distribution. In a preferred embodiment, the default value for the low-risk threshold is set to 0.3, the default value for the medium-risk threshold is set to 0.6, and the default value for the high-risk threshold is set to 0.8. A comprehensive risk score below the low-risk threshold is classified as low risk, between the low-risk and medium-risk thresholds as medium risk, between the medium-risk and high-risk thresholds as high risk, and above the high-risk threshold as extremely high risk. Companies can adjust these thresholds according to their actual safety management requirements.

[0054] The process of tracing risk factors and generating corrective measures transforms the abstract predictive results of the model into decision support information that is understandable and actionable for approvers. Attention weight distribution is analyzed, and historical time steps are sorted from highest to lowest weight. The top 10 historical time steps with the highest attention weights are selected, each corresponding to a sliding window position in S2 when constructing time-series features. All historical work records occurring within the time windows corresponding to these time steps are retrieved from the database, with a focus on work where risk events occurred within the window. Detailed information such as basic work information, approval process records, on-site execution status, accident records, corrective measures records, and lessons learned summaries are extracted.

[0055] The similarity between the current operation and each historical operation is calculated, taking into account four dimensions: operation type matching, environmental parameter proximity, operation location relevance, and time / seasonal similarity. For operation type matching, the matching score is 1 if the current and historical operations are exactly the same type; 0.5 if they are different types but belong to the same risk level (e.g., both are Level 1 operations); otherwise, the matching score is 0. For environmental parameter proximity, the differences in environmental parameters such as temperature, gas concentration, and pressure between the current and historical operations are calculated. The absolute value of each parameter difference is divided by the maximum possible range of that parameter for normalization, resulting in a normalized difference value. The average of all normalized difference values ​​is then calculated, and the environmental parameter proximity is subtracted from 1 to obtain the environmental parameter proximity score. The closer this score is to 1, the more similar the environmental parameters are.

[0056] For job location relevance, the relevance is 1 if the current job and historical jobs are located in the same plant area, 0.6 if they are in adjacent plant areas, and 0.3 if they are in different and non-adjacent plant areas. For time-seasonal similarity, the difference between the current job month and the historical job month is calculated. Considering the cyclical nature of months, the smaller of the difference value and 12 minus the difference value is taken as the actual month difference. The time-seasonal similarity is obtained by subtracting the actual month difference from 1 and dividing by 6. The closer this value is to 1, the more similar the seasons. The weighted sum of the similarities of the four dimensions is used to obtain the comprehensive similarity. The weight of each dimension is determined according to its importance to risk assessment. In a preferred embodiment, the default weight for job type matching is set to 0.4, the default weight for environmental parameter proximity is set to 0.3, the default weight for job location relevance is set to 0.2, and the default weight for time-seasonal similarity is set to 0.1. The job type has the highest weight because job type is the most important factor determining risk characteristics. Enterprises can adjust the weights of each dimension according to their actual situation. Historical tasks were sorted from highest to lowest based on overall similarity, and the top 5 most similar tasks were selected as similar historical cases to generate a list for approvers' reference. For historical cases in the list where risk events occurred, key information such as the cause of the accident, risk factors, emergency response process, and post-accident corrective measures recorded in the accident report were extracted and used as risk warning content for the current task.

[0057] By comparing the current operation characteristics with the distribution of characteristics from historical safe operations, identify features in the current operation that exceed the normal range. Calculate the mean and standard deviation of each feature from historical safe operation data, and compare the current operation's feature value with the mean to calculate the degree of deviation. For numerical features such as ambient temperature, gas concentration, and operation duration, if the current value exceeds the historical mean plus twice the standard deviation, it is marked as an abnormal feature. For categorical features such as hot work level and combustible material category, if the current category has a high frequency of occurrence in historical risk events, it is marked as a high-risk feature. Summarize all features marked as abnormal or high-risk to generate a risk factor list. For each risk factor in the list, label its specific value, degree of exceedance, and historical risk correlation, etc. For example, if the gas concentration in the current operation is 12 ppm, while the average gas concentration in historical safe operations is 5 ppm, then the risk factor list should be marked as "Gas concentration 12 ppm, exceeding the safe average of 5 ppm, deviation degree 140%, this concentration level has historically caused 3 risk events."

[0058] A knowledge base mapping risk factors to safety measures was established, built upon the company's historical rectification records and industry safety standards. The knowledge base stores effective safety measures corresponding to various risk factors, including technical, management, and emergency measures. Based on the identified risk factor list, matching safety measures were retrieved from the knowledge base. For risk factors involving excessive gas concentrations, retrieved measures included "ensuring adequate replacement ventilation to reduce gas concentrations to safe levels," "continuously monitoring gas concentrations during operations and equipping portable gas detectors," and "establishing a restricted area to prohibit unauthorized personnel from entering." For risk factors involving excessively long operation times, retrieved measures included "shortening the duration of single operations and switching to segmented operations," "increasing personnel rotation to avoid fatigue," and "extending rest intervals between operations." For risk factors involving hot work in high-temperature environments, retrieved measures included "adjusting operation times to lower temperatures in the early morning or evening," "increasing on-site cooling measures such as spraying and ventilation," and "equipping with heatstroke prevention supplies and emergency medications." All retrieved safety measures were prioritized based on factors such as effectiveness, implementation difficulty, and cost. Generate a structured list of security measure recommendations, with each measure in the list including a detailed description of the measure's content, key implementation points, and expected effects.

[0059] Based on the risk level determined by the comprehensive risk score, corresponding early warning strategy rules are matched and tiered early warning content is generated. For low-risk levels, a routine work safety reminder is generated, including standard safe operating procedures, basic protection requirements, and routine monitoring arrangements. This indicates to approvers that the risk of the operation is within a controllable range and approval can be obtained according to standard procedures. For medium-risk levels, a key monitoring early warning is generated, including a list of risk factors, summaries of similar historical cases, targeted monitoring suggestions, and a list of safety measure suggestions. This indicates to approvers that they need to pay close attention to the risk factors listed, confirm that the applicant has developed corresponding safety measures, and suggest clearly specifying monitoring requirements in the approval comments. For high-risk levels, a strict control early warning is generated, including a detailed quantitative analysis of risk factors, a complete record of similar historical cases, especially accident cases, a list of mandatory safety measure requirements, suggestions for on-site monitoring personnel configuration, emergency plan preparation requirements, and a continuous monitoring plan for the work process. This indicates to approvers that the operation has a high risk and requires a strict review of the completeness of the work plan and the adequacy of safety measures. If necessary, the applicant should be required to supplement safety measures or optimize the work plan before approval. If the risk level is extremely high, a recommendation to postpone approval is generated. This recommendation includes the specific items and degree of risk exceeding the limits, lessons learned from similar high-risk operations in the past, the main problems with the work plan, suggestions for optimizing the work plan, and requirements for redoing the risk assessment. The approver is notified that the operation is too risky and that approval should be postponed. The applicant is required to reassess the work plan, implement more effective risk control measures, or adjust the work time and methods to reduce risk before resubmitting for approval. The risk level identifier, predicted risk probability, comprehensive risk score, list of similar historical cases, list of risk factors, list of safety measure recommendations, and approval decision recommendations are organized into a structured early warning strategy to provide approvers with comprehensive decision support information.

[0060] S5 provides visualized push notifications for tiered early warning strategies and collects operational feedback data. It optimizes the model using a multi-condition triggered incremental update mechanism and applies a hybrid approach of rule engine and transfer learning for cold start scenarios. S5 includes processing flows for two scenarios: For regular scenarios with sufficient historical data, it executes processes such as visual early warning push, job feedback collection, model performance monitoring, and continuous updates; For cold start scenarios with insufficient historical data, such as new job types or newly built device areas, it adopts a hybrid early warning strategy that combines rule engines and transfer learning to ensure the continuity and reliability of the early warning function.

[0061] Based on the configuration information of the work approval process, query the approval node of the current work application and the corresponding approver's ID. Retrieve the department and superior management information of the work applicant from the organizational structure database. Query the corresponding on-site safety supervisor and local safety management personnel based on the work location. Compile the user ID, mobile device ID, contact information, and other information of the above personnel into a list of personnel receiving early warning information.

[0062] Design a visual display interface for early warning information, using red to indicate extremely high risk, orange to indicate high risk, yellow to indicate medium risk, and green to indicate low risk. Place risk level indicators prominently at the top of the interface, displaying both the predicted risk probability and the specific numerical values ​​of the comprehensive risk score. Present the risk factor list in a structured format, labeling each risk factor with its specific value, degree of exceedance, and historical risk correlation, and marking each risk factor with a warning icon. Display a list of similar historical cases, highlighting those with past accidents in red; clicking on a case allows viewing detailed accident reports and lessons learned. For operations involving multi-dimensional risk factors, create a radar chart to display the normalized value distribution of each dimension. Each axis of the radar chart corresponds to a different risk factor dimension, such as operation type risk, environmental parameter risk, personnel configuration risk, time factor risk, and spatial location risk. Points on the axes represent the normalized risk value of that factor, and the size of the radar chart visually reflects the overall risk level of the operation. The system queries historical databases to retrieve recent risk trend data for this job type, creating a time-series line chart to display the changing trend of risk scores. The current job's risk score is marked on the chart to help approvers determine the relative position of the current risk level within the historical context. A list of recommended safety measures is displayed in a tiered format, indicating the priority, implementation details, and expected effects of each measure. Approvers can directly copy these recommendations into their approval comments. Approval decision recommendations are prominently displayed at the bottom of the interface, with a bold red text indicating "Approval Recommended to Postpone" for extremely high-risk jobs.

[0063] Warning information is pushed to recipients' mobile clients via the mobile application's push notification interface. The push notification includes the risk level and a brief risk warning; clicking the notification leads to a detailed warning page. A warning notification window pops up on the approver's PC system interface via the web system's message center interface, displaying complete visual warning content. For high-risk and extremely high-risk operations, additional SMS alerts are sent to key approvers and safety managers via the SMS gateway interface to ensure timely delivery of warning information. A risk warning display area is embedded in the operation approval interface, allowing approvers to simultaneously view risk prediction information and suggested measures while processing the approval process, using the warning information as a supplementary reference for approval decisions.

[0064] The system automatically collects complete execution data of operations, extracts actual result labels, and marks any associated accident records or anomalies as risk events; otherwise, it marks them as normal completion. The predicted risk level is compared with the actual result. If a high-risk prediction is made and an event actually occurs, the warning is considered accurate; if a high-risk prediction is made but the operation is normal, it is considered a false alarm; if a low-risk prediction is made but an event actually occurs, it is considered a missed alarm. Evaluation metrics such as warning accuracy, false alarm rate, and missed alarm rate are statistically analyzed over a period of time. When the missed alarm rate exceeds a set safety threshold, the system automatically adjusts the risk level judgment threshold, lowering the threshold for high and extremely high risks, making the model more conservative in its warnings, preferring to over-report rather than under-report, ensuring the bottom line of safety management. Regular model performance monitoring reports are generated, including trends in warning accuracy over different time periods, comparisons of prediction effects for different operation types, false alarm and missed alarm case analyses, and model performance improvement suggestions, for review and decision-making reference by enterprise safety management personnel. Complete operation data containing actual result labels is converted into standardized feature vectors according to the S1 processing flow and added to the standardized operation feature dataset to provide training data for subsequent model updates.

[0065] A model update trigger strategy is set, employing a multi-condition trigger mechanism to adapt to update needs in different scenarios. The regular trigger conditions are: the accumulated new job data reaching a set data volume threshold or the time elapsed since the last model training exceeding a set time threshold. The default data volume threshold is 1000 records, and the default time threshold is 30 days. Meeting either condition initiates the model update evaluation process. The emergency trigger condition is a major safety incident occurring within the enterprise. Regardless of the amount of new data, an emergency model update is immediately triggered to incorporate the incident case into the model's learning to prevent similar incidents from recurring. The performance trigger condition is when the system detects a continuous decline in model prediction accuracy exceeding a set performance degradation threshold. The default performance degradation threshold is 5%, indicating model performance degradation and requiring model diagnosis and updates. The data distribution trigger condition is when a significant deviation in the feature distribution of newly added data is detected from the feature distribution of training data, determined using a feature statistics change monitoring method. The specific monitoring method is as follows: For each feature, calculate the mean and standard deviation of the newly added data, and compare them with the mean and standard deviation of that feature in the training data. If the absolute value of the difference between the mean of the newly added data and the mean of the training data divided by the standard deviation of the training data is greater than a set mean offset threshold, or if the absolute value of the difference between the standard deviation of the newly added data and the standard deviation of the training data divided by the standard deviation of the training data is greater than a set variance offset threshold, then the distribution of that feature is considered to have shifted significantly. The default value for the mean offset threshold is 2, and the default value for the variance offset threshold is 0.5. Calculate the proportion of features with significant shifts out of the total number of features. If this proportion exceeds a set distribution offset proportion threshold (default value is 30%), then the overall data distribution is determined to have shifted, triggering a model update to address changes in operational patterns or the emergence of new risk factors.

[0066] When a model update is triggered, the proportion of newly added data to the original training set is evaluated. If the amount of newly added data is less than the set proportion threshold of the training set and no major incident has occurred (default threshold is 10%), incremental training is used. The model parameters are fine-tuned using the new data, with a small learning rate set to avoid destroying learned knowledge and preventing catastrophic forgetting. The default fine-tuning learning rate is 0.0001. If the amount of newly added data is greater than or equal to the proportion threshold of the training set, or a major security incident has occurred, full training is used. The model parameters are reinitialized using the updated complete dataset, and full training is performed to fully learn new risk patterns and data distributions. The time-series feature construction process of S2 is re-executed using the updated complete dataset to generate a new training sample set. The risk prediction network of S3 is trained using the new training sample set. After training, the performance metrics of the new model are evaluated on an independent test set, calculating key indicators such as accuracy, recall, F1 score, false positive rate, and false negative rate.

[0067] A comprehensive comparison of the new model's performance metrics with those of the currently running model is conducted. The new model must achieve accuracy, recall, and F1 score no lower than the current model, and a false negative rate no higher, to ensure the model update does not diminish the early warning effectiveness. If all performance metrics of the new model meet the requirements, it will proceed to the canary release phase for practical effect verification. During the canary release phase, newly submitted job applications are allocated using a random sampling method. A set proportion of job applications will use the new model for risk prediction; this proportion is set to a default of 20% to ensure sufficient sample size for evaluating the new model's performance while controlling the potential impact of risks. The remaining job applications will continue to use the current model for risk prediction. The prediction results of both models and the actual job results are recorded simultaneously in the system backend. A continuously set monitoring period is used to monitor the performance of the new model in practical applications; the default monitoring period is 7 days. The actual early warning accuracy, false positive rate, and false negative rate of the new model are statistically analyzed and compared with the current model. If the new model performs stably during the canary release and its performance metrics are better than or no worse than the current model, a full switch will be performed. The parameters of the new model will be exported and deployed to the online prediction service, replacing the original model. Simultaneously, the parameters of the original model will be backed up and stored to support rapid rollback. If the new model experiences anomalies during the canary release, such as an increase in the false negative rate, the canary release will be terminated, the current model will continue to be used, and training logs and anomalies will be recorded for subsequent analysis and improvement. If the performance metrics of the new model do not meet the standards, the current model will continue to run. The reasons for the insufficient performance of the new model will be analyzed, which may require adjusting the network structure, optimizing hyperparameters, or adding data preprocessing steps before retraining.

[0068] Through the aforementioned closed-loop feedback mechanism, the risk prediction model can continuously learn new operational data and risk patterns, the early warning strategy can be dynamically adjusted according to actual results, and the system's early warning accuracy continues to improve with data accumulation and model iteration.

[0069] To address the cold start issue caused by insufficient historical data for newly added work types or newly constructed equipment areas, the system employs a hybrid early warning strategy to ensure the continuity and reliability of the early warning function. When the number of samples for the work type or work area involved in the work application is less than the cold start judgment threshold in historical data (the default value for the cold start judgment threshold is 50), the system is judged as a cold start scenario and the hybrid early warning strategy is activated. When the number of samples is between the cold start judgment threshold and the model maturity threshold (the default value for the model maturity threshold is 500), the system is in a transition phase and continues to use the hybrid strategy, but dynamically adjusts the weight ratio of the rule engine and the model. When the number of samples reaches or exceeds the model maturity threshold, the system fully switches to model prediction, and the rule engine is only used for auxiliary verification. In cold start scenarios, the system first uses a rule engine based on safety specifications for initial risk assessment. The rule engine stores various work risk judgment rules stipulated in national safety production standards, industry safety technical specifications, and enterprise safety management systems, such as the safety conditions that must be met for hot work, the protective equipment that must be equipped for confined space work, and the fall prevention measures that must be taken for high-altitude work. The rule engine checks each item in the job application information to see if it meets the safety requirements. Items that do not meet the requirements are marked as risk points, and the rule evaluation score is calculated based on the number and severity of the risk points.

[0070] Simultaneously, the system searches for existing data similar to the current job type or area and uses transfer learning techniques for risk prediction. For newly added job types, it searches for existing job types with the most similar job characteristics; for example, the newly added "hot work on high-temperature equipment" can refer to the existing "conventional hot work" model. For newly built plant areas, it searches for existing plant areas with the most similar process flow, equipment configuration, and hazard source types. The system loads the parameters of trained models corresponding to similar job types or areas as the initial model and fine-tunes the model using the small amount of newly accumulated data. During fine-tuning, a layered training strategy is adopted, dividing the network into a bottom-level feature extraction layer and a top-level risk judgment layer. The bottom-level feature extraction layer includes a feature embedding layer, an LSTM temporal coding layer, and a time decay enhancement module. These layers are responsible for extracting general temporal evolution features and risk patterns from the original features, and their learned feature representation capabilities are transferable across different job scenarios. The top-level risk judgment layer includes a multi-head attention layer, a fully connected feature fusion layer, and an output prediction layer. These layers are responsible for learning the correlation between the current job and historical jobs, as well as scenario-specific risk judgment logic. During fine-tuning, all parameters of the bottom feature extraction layer are frozen and kept unchanged. Only the parameters of the top risk assessment layer are trained and updated. This allows for rapid adaptation to new scenarios with a small number of samples, while retaining the learned general feature extraction capabilities and avoiding overfitting in small sample situations. After fine-tuning, the model is used to predict the risk of the current task, and the model evaluation score is obtained.

[0071] The evaluation scores of the rule engine and the transfer learning model are weighted and fused to obtain a comprehensive risk score for the cold start scenario. In the early stages of data accumulation, the rule engine weight is set to a default value of 0.7, and the model weight to a default value of 0.3, relying more on explicit safety specifications for judgment. As data for this job type or region gradually accumulates, the weight ratio is dynamically adjusted. When the data volume reaches the first transition threshold, the weight is adjusted to 0.5 for both rules and 0.5 for the model (default value: 100 records). When the data volume reaches the second transition threshold, the weight is adjusted to 0.3 for rules and 0.7 for the model (default value: 200 records). When the data volume reaches the model maturity threshold, the system switches entirely to model prediction, with the rule engine serving only as an auxiliary verification tool. Through this hybrid strategy, the system can provide basic early warning functions during the cold start phase, avoiding situations where early warnings are impossible due to insufficient data. Simultaneously, as data accumulates, it gradually transitions to fully data-driven intelligent early warning, ensuring the continuity and reliability of the early warning system.

[0072] Application Examples This invention has been deployed and verified in a practical application at an energy and chemical enterprise, specifically for a hot work safety early warning scenario. The enterprise processed approximately 300 hot work applications through the system over a certain period.

[0073] The system collects historical hot work data from the company as training data. Examples of some historical work data are shown in Table 1: Table 1, Example of Historical Task Data

[0074] The system standardizes historical data according to method S1, encoding the hot work level, fire source type, and combustible material category, normalizing numerical features, extracting time-period features, and labeling risk events. Following method S2, it constructs multi-scale temporal features, setting short-term, medium-term, and long-term windows to extract temporal features such as operation frequency, risk occurrence rate, and environmental parameters. Using method S3, it trains the risk prediction network, constructing a network architecture that includes feature embedding, LSTM temporal encoding, time decay enhancement, multi-head attention, feature fusion, and risk prediction output. The network is trained using a time decay enhancement focus loss function to obtain the trained model parameters.

[0075] An applicant submitted a Level 1 hot work application to weld and repair a flammable liquid storage tank in Unit A, with a planned duration of 5 hours. Real-time monitoring data showed a current temperature of 26℃ and a gas concentration of 12ppm, while the weather forecast indicated the temperature would rise to 30℃ during the work period. The system loaded a pre-trained model using the S4 method for inference, outputting a predicted risk probability of 0.78. Attention weight distribution was extracted, revealing multiple attention heads focusing on similarity in work type, similarity in environmental parameters, temporal periodicity, and spatial location correlation. The highest weight was given to historical data on Level 1 hot work accidents in Unit A. The system determined a risk adjustment coefficient of 1.2 based on the baseline risk rate for hot work operations, dynamically increasing it to 1.5 due to the predicted temperature reaching 30℃. A logarithmic probability transformation method was used to fuse the risk probability of 0.78 and the adjustment coefficient of 1.5, resulting in a comprehensive risk score of 0.843, exceeding the high-risk threshold and classifying it as extremely high risk.

[0076] The system performs risk factor tracing and identifies abnormal characteristics of the current operation: gas concentration of 12 ppm exceeds the safe average by 5 ppm (deviation of 140%), operation duration of 5 hours exceeds the recommended 3 hours, and predicted temperature of 30℃ is close to the high-temperature standard. Safety measures are retrieved from the knowledge base: sufficient replacement ventilation to ensure gas concentration drops below 5 ppm, continuous monitoring of gas concentration, shortening operation time to 2 hours per operation, completing the operation over two days, adjusting operation time to 6-8 AM, increasing monitoring personnel to at least 3 people, equipping with fire extinguishers, setting up a warning area, and developing an emergency plan. An extremely high-risk warning is generated according to the S5 methodology and pushed to approvers, applicants, and safety management departments via mobile applications and web systems. The warning interface displays an extremely high-risk indicator in red, shows a risk score of 0.843, similar historical cases, a list of risk factors, and safety measure recommendations, with a "Recommendation to postpone approval" decision suggestion displayed at the bottom.

[0077] After reviewing the early warning, the approvers carefully analyzed the list of risk factors and historical accident cases, requiring the work unit to supplement safety measures. The work unit revised the plan: shortening the operation time from 5 hours to 2 hours per session, completing it over two days; adjusting the operation time to the early morning when temperatures are low; promising sufficient ventilation before operation to ensure gas concentrations drop below safe levels; and increasing monitoring personnel and emergency supplies. The revised plan was resubmitted, and the system again predicted a risk probability of 0.42 and a comprehensive risk score of 0.50, classifying it as medium risk. After approval, the operation was successfully completed over two days without incident. The system recorded complete data, labeling the original application features as positive samples and the optimized features as negative samples, adding them to the training dataset to allow the model to learn the risk improvement effects and the effectiveness of the measures.

[0078] Through the implementation of this invention, the company was able to issue early warnings for approximately 300 hot work operations within a certain period, successfully identifying several high-risk operations. Most of these operations were completed safely after optimization based on the early warning suggestions, a few were temporarily postponed due to excessive risk, and a very small number experienced minor anomalies due to unforeseen on-site factors despite the implementation of preventative measures. Statistical analysis shows that the early warning accuracy rate exceeded 85%, the recall rate exceeded 90%, and the F1 score exceeded 85%, with overall performance meeting industrial application requirements. The system accurately captures the complex risk patterns of operational safety through technological innovations such as multi-scale time windows, time decay mechanisms, multi-head attention mechanisms, time decay-enhanced focus loss functions, hybrid early warning strategies, and intelligent model updates, providing strong support for the company's safe production.

[0079] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A job safety early warning method based on big data driving, characterized in that, Includes the following steps: S1: Obtain multi-source heterogeneous job data, process it using field mapping fusion and multi-level annotation methods, and output a standardized job feature dataset; S2, based on a standardized operational feature dataset, extracts short, medium and long-term time series features through multi-scale time window sliding driven by the enterprise's operational cycle, and generates a deep learning training sample set; S3 inputs the deep learning training sample set into a network architecture that integrates LSTM temporal coding, temporal decay enhancement and multi-head attention, and trains the risk prediction model using the temporal decay enhancement focus loss function. S4. Use the trained risk prediction model to infer the characteristics of the work to be approved, and generate a graded early warning strategy through the risk adjustment method of attention weight tracing and log probability transformation. S5 provides visualized push notifications for tiered early warning strategies and collects operational feedback data. It optimizes the model using a multi-condition triggered incremental update mechanism and applies a hybrid approach of rule engine and transfer learning for cold start scenarios.

2. The big data-driven operational safety early warning method according to claim 1, characterized in that, S1 includes: Mapping analysis is performed on fields from various data sources to identify fields with the same meaning but different names and unify them. A standard model for job data is established, and corresponding feature fields are extracted for different job types. Numerical features are normalized, categorical features are one-hot encoded, missing numerical features are filled with the historical average of the same job type, and missing categorical features are filled with the category with the highest frequency of occurrence. A multi-level labeling strategy is adopted to label the risk of the work results. Works that have accidents, anomalies, multiple rejections of approval, or serious violations found during on-site inspections are labeled as positive samples. For work that has been rejected multiple times but is completed safely after optimizing the work plan, the feature vector of the original application is extracted and labeled as a positive sample, and the feature vector of the optimized resubmission is extracted and labeled as a negative sample. Works that are completed normally without any anomalies are labeled as negative samples.

3. The big data-driven operational safety early warning method according to claim 1, characterized in that, S2 includes: The lengths of short-term, medium-term, and long-term time windows are determined based on the company's actual operating cycle. The length of the short-term time window is determined based on the company's weekly work plan cycle, the length of the medium-term time window is determined based on the company's monthly maintenance cycle, and the length of the long-term time window is determined based on the company's quarterly production cycle. For the short-term window, the short-term feature vector is composed of the statistical operation frequency feature vector, short-term risk incidence rate, temperature distribution characteristics, and gas concentration characteristics. For the medium-term window, the medium-term feature vector is composed of the changing trend of risk incidence rate in each sub-interval, the fluctuation variance of risk factors, and the information entropy of operation type distribution. For the long-term window, the macro feature vector is composed of the seasonal feature coefficient, cumulative risk index, and production load change trend.

4. The big data-driven operational safety early warning method according to claim 3, characterized in that, S2 further includes: For low-frequency job types, if the number of samples of a specific job type within a window is lower than the preset minimum sample threshold, a time-weighted interpolation method is used to find windows with sufficient data of the corresponding job type in the adjacent time periods before and after the corresponding window. The weight coefficients are calculated based on the time distance, and a weighted average is performed to obtain the interpolation features and construct a job type mask matrix. Short-term, medium-term, and macro-level feature vectors are concatenated along the feature dimension to form a multi-scale comprehensive feature vector. These vectors are then stacked in sliding order to form a time-series feature matrix. The jobs within the future prediction period corresponding to each time window are selected and labeled. The time-series feature matrix is ​​then uniformly processed into a fixed-length time series and divided into training, validation, and test sets according to time order.

5. The big data-driven operational safety early warning method according to claim 1, characterized in that, S3 includes: In the feature embedding module, the input feature matrix is ​​mapped to the embedding space through a linear transformation to obtain the embedded feature tensor. The job type mask matrix is ​​then embedded and mapped. The feature embedding and type embedding are added element by element to obtain the initial representation tensor. In the temporal coding module, LSTM is used to temporally encode the initial representation tensor. The time series is processed step by step through the gating mechanism of forget gate, input gate, and output gate to obtain the temporally encoded hidden state sequence. In the time decay enhancement module, the time interval between each historical time step and the current prediction time is calculated, the decay coefficient is calculated using the exponential decay function, and the decay coefficient is multiplied element by element with the temporal encoded hidden state of the corresponding time step to obtain the enhanced hidden state sequence.

6. The big data-driven operational safety early warning method according to claim 1, characterized in that, S3 further includes: The task feature vector at the target prediction time is extracted and mapped to the query vector Q. The enhanced hidden state sequence is mapped to the key vector sequence K and the value vector sequence V respectively. For each attention head, linear transformations are performed on Q, K and V respectively. The dot product of the query vector and the key vector at each time step is calculated to obtain the similarity score. The attention weight distribution is obtained by scaling and softmax normalization. The attention weights are weighted and summed with the corresponding value vectors to obtain the output vector of the corresponding attention head. After concatenating the output vectors of each attention head, feature fusion is performed through a two-layer fully connected network to obtain a comprehensive risk representation vector. The risk probability prediction value is then output through an output fully connected layer and a sigmoid activation function.

7. The big data-driven operational safety early warning method according to claim 1, characterized in that, S3 further includes: The loss term consists of a basic cross-entropy term, a focus modulation term, and a time decay weight term. The focus modulation term dynamically adjusts the loss weight through a modulation factor. When the model has a high confidence level in predicting a sample, the loss weight of the corresponding sample is reduced. When the prediction confidence level is low, the loss weight of the corresponding sample is maintained. A balancing factor is introduced to apply different class weights to positive and negative samples. The time decay weight term is obtained by averaging the decay coefficients of all time steps within the corresponding time window of the sample to obtain the comprehensive time decay weight, which is introduced into the loss function as a sample-level weight coefficient. The loss of a single sample is the product of the time decay weight, the class weight, the focus modulation factor, and the basic cross-entropy loss. The total loss of the training batch is the mean of the losses of each sample.

8. The big data-driven operational safety early warning method according to claim 1, characterized in that, S4 includes: The risk adjustment coefficient is determined based on the baseline risk rate of each type of operation and the environmental conditions of the current operation. The predicted risk probability value is converted into logarithmic probability form, and the risk adjustment coefficient is converted into logarithmic form. The two are added together and converted back into probability form through inverse transformation to obtain a comprehensive risk score. The risk level is determined based on the preset low risk threshold, medium risk threshold and high risk threshold. Analyze the attention weight distribution, select the historical time steps with the highest weight ranking, and calculate the similarity between the current operation and historical operations based on four dimensions: operation type matching degree, environmental parameter similarity, operation location relevance, and time and season similarity. Select the historical operations with the highest similarity ranking as similar historical cases, compare the feature distribution of the current operation with that of historical safe operations to identify abnormal feature items and generate a list of risk factors. Based on the risk factor list, retrieve matching safety measures from the mapping knowledge base to generate a list of safety measure suggestions, and generate a graded early warning strategy based on the risk level matching early warning strategy rules.

9. The big data-driven operational safety early warning method according to claim 1, characterized in that, S5 includes: The system includes regular triggering conditions, emergency triggering conditions, performance triggering conditions, and data distribution triggering conditions. The regular triggering condition is when the amount of new data reaches a preset data volume threshold or when the time since the last training exceeds a preset time threshold. The emergency triggering condition is when a major security incident occurs. The performance triggering condition is when the prediction accuracy continuously declines beyond a preset performance decline threshold. The data distribution triggering condition is when the proportion of features with deviations exceeds a preset distribution deviation ratio threshold. When the amount of new data is less than the preset ratio threshold, incremental training is used; otherwise, full training is used. The rule engine and transfer learning hybrid method includes: when the number of historical samples is less than the preset cold start judgment threshold, the rule engine is activated to perform initial risk assessment, while loading the parameters of trained models of similar job types, adopting a hierarchical training strategy to freeze the bottom feature extraction layer and only fine-tune the top risk judgment layer, weighting and fusing the rule engine evaluation score and the model evaluation score, dynamically adjusting the weight ratio as data accumulates, and switching to model prediction when the amount of data reaches the preset model maturity threshold.

10. A big data-driven operational safety early warning system, used to execute the steps of the big data-driven operational safety early warning method as described in any one of claims 1-9, characterized in that, include: The data fusion and annotation module is used to acquire multi-source heterogeneous operation data, and processes it using field mapping fusion and multi-level annotation methods to output a standardized operation feature dataset. The temporal feature construction module is used to extract short, medium and long-term temporal features based on a standardized operational feature dataset by sliding a multi-scale time window driven by the enterprise's operational cycle, and to generate a deep learning training sample set. The model training module is used to input the deep learning training sample set into a network architecture that integrates LSTM temporal coding, temporal decay enhancement and multi-head attention, and to train the risk prediction model using the temporal decay enhancement focus loss function. The risk reasoning and early warning module is used to reason about the characteristics of the work to be approved using the trained risk prediction model, and generates a graded early warning strategy through the risk adjustment method of attention weight tracing and log probability transformation. The feedback optimization module is used to visually push the graded early warning strategy and collect operation feedback data. It adopts an incremental update mechanism triggered by multiple conditions to optimize the model and applies a hybrid method of rule engine and transfer learning for cold start scenarios.

Images