Safety management system and method for enterprise

Abstract

The invention discloses a safety management method for an enterprise, and belongs to the technical field of enterprise safety management. The method specifically comprises the following steps: S1, multi-source data acquisition: acquiring enterprise potential safety hazard full-life-cycle multi-source historical data; s2, standardized preprocessing: obtaining a structured and modelable standardized data set through data cleaning, heterogeneous data standardization and tagging; data is fully utilized and standardized: through multi-source data acquisition and standardized preprocessing, dispersed heterogeneous data is integrated, and the problems of low data quality and difficulty in centralized analysis in the prior art are solved; and the three-level vertical feature extraction system deeply mines basic information, association rules and time sequence trends in the data, so that maximum utilization of data values is realized. Risk prediction is accurate: based on a CNN-LSTM-Attention model of deep feature training, transformation from passive disposal to active prevention is realized, and the probability of occurrence of safety accidents is effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of enterprise safety management technology, and in particular to a safety management system and method for enterprises. Background Technology

[0002] Enterprise safety management is a core element in ensuring the continuous and stable operation of production and business activities. Its core objective is to identify and effectively address various safety hazards in a timely manner and prevent safety accidents. Currently, enterprise safety management generally faces the following prominent problems: First, hazard data is scattered across multiple systems (such as manual record sheets, equipment management systems, and Excel spreadsheets), with heterogeneous data formats and insufficient completeness, making centralized analysis difficult. Second, processing procedures rely on human experience, task allocation lacks scientific basis, and approval paths are fixed and rigid, resulting in long processing cycles and low efficiency. Third, the use of historical data is limited to statistical summarization, failing to explore deep correlations and time-series trends within the data. Risk prediction relies heavily on manual judgment, resulting in low accuracy and difficulty in achieving proactive prevention. Fourth, existing digital management systems mainly focus on the electronic presentation of processes, lacking deep integration with intelligent models, and cannot provide dynamic and accurate decision support for full-process management.

[0003] In existing technologies, some enterprises have attempted to introduce data analysis tools to optimize security management, but significant shortcomings remain: for example, feature extraction is often single-level, focusing only on basic attribute features and failing to capture the complex relationships and temporal evolution patterns between hazards and the environment, personnel, and processes, resulting in insufficient expressive power of model input features; models are often single-function (e.g., only predicting risk levels), lacking the ability to optimize processing flows and failing to form a closed-loop management system of "prediction-processing-feedback"; and the integration between models and management systems is low, with high data interaction latency, making it difficult to meet real-time management needs. Therefore, there is an urgent need for an enterprise security management method that can deeply integrate multi-source data, accurately extract hierarchical features, fuse multi-task models, and deeply collaborate with management systems. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a security management system and method for enterprises; it can solve the problems of scattered and heterogeneous data, lack of hierarchical feature extraction, low accuracy of risk prediction, rigid and inefficient processing procedures, and insufficient integration of models and systems in existing enterprise security management.

[0005] Technical Solution: To solve the above-mentioned technical problems, according to one aspect of the present invention, more specifically, a security management method for enterprises, specifically including the following steps:

[0006] S1. Multi-source data acquisition: Collect multi-source historical data on enterprise security risks throughout their entire lifecycle;

[0007] S2. Standardization Preprocessing: Through data cleaning, heterogeneous data standardization, and labeling, a structured and modelable standardized dataset is obtained;

[0008] S3. Deep Feature Extraction: Perform deep feature extraction based on standardized data to extract static and process basic features to form first-level features. Based on the first-level features, mine association rules and temporal evolution features to form second-level features. Based on the second-level features, extract deep semantic and long-term trend features to form third-level features.

[0009] S4. Hybrid Model Construction: Construct a hybrid safety management model that includes a risk level prediction sub-model and a process optimization sub-model.

[0010] S5. Phased Training: A phased training strategy of "pre-training-fine-tuning-verification" is adopted, combined with regularization and hyperparameter optimization, to obtain a security management model that meets the performance standards.

[0011] S6. Deep Integration of Model and System with Full-Process Application: The trained security management model is deeply integrated with the enterprise security management system based on .NET6, Vue 3, MySQL and Elsa workflow engine, covering the entire process of hazard entry, allocation, processing, acceptance and archiving. The model outputs risk assessment results, process optimization solutions and dynamic tracking suggestions in real time to achieve intelligent enterprise security management.

[0012] Furthermore, in step S1, the multi-source historical data includes: core business data, related supporting data, and environmental context data; core business data includes: hazard lifecycle data and hazard attribute data; related supporting data includes: organizational structure data, personnel role data, and process configuration data; environmental context data includes: equipment operation data, environmental monitoring data, and time-dimensional data; hazard lifecycle data includes: input information, allocation records, processing logs, acceptance results, and archived information; hazard attribute data includes: title, detailed description, occurrence area, designed equipment, and hazard type; organizational structure data includes: department level and job responsibilities; personnel role data includes: permissions and operation history of system administrators, approval managers, recorders, acceptance specialists, and processing personnel; process configuration data includes: approval rules, node settings, and timeout thresholds; equipment operation data includes: equipment model, runtime, and maintenance records; environmental monitoring data includes: temperature and humidity, dust concentration, and gas concentration; and time-dimensional data includes: seasons, holidays, and peak production period markers.

[0013] Furthermore, in step S2, regarding missing data: missing value judgment rules are adopted, and key fields with a missing rate greater than or equal to the missing rate threshold are directly removed, while those with a missing rate less than the missing rate threshold are filled using the K-nearest neighbor algorithm. The time field format is unified by matching the time format regular expression; duplicate records are deleted based on the combination key of "hazard number + occurrence area + hazard type", and outliers in numerical fields such as processing time and impact range are identified and removed using the isolated forest algorithm.

[0014] Heterogeneous data standardization: Differentiated coding strategies are adopted for categorized data, including one-hot coding for unordered categorized data such as hazard type, department, and role; label coding is used for ordered categorized data such as regional level and risk level; time-based data is converted to Unix timestamp format, and derived features such as "whether it is a production period" and "number of days since the last occurrence of the same type of hazard" are generated; text-based data including hazard descriptions and handling notes are segmented, stop words are removed, and word vectors are generated using a BERT pre-trained model; numerical data including handling time and equipment running time are standardized using Z-Score to eliminate the influence of units.

[0015] Labeling and marking: Based on relevant regulations, a three-level risk labeling rule is established: High risk: May cause personal injury, major property damage, or production stoppage exceeding 24 hours; Medium risk: May cause minor property damage or production stoppage for 2-24 hours; Low risk: No property damage and no impact on production; Labeling is done independently by two or more safety management experts, and discrepancies are determined through collective review; At the same time, processing efficiency labels and recurrence labels are marked; Processing efficiency labels include: High efficiency, Average, Low efficiency; Recurrence labels include: Yes, No.

[0016] Furthermore, in step S3, the first-level feature extraction involves using a feature engineering toolkit to extract static basic features and process basic features based on a standardized dataset, forming a basic feature vector. Static basic features include: hazard type coding, region coding, department coding, statistical features of hazard description word vectors, equipment coding, and operating status features. Hazard type coding adapts to various hazard types; region coding supports three-level region division: workshop-production line-equipment; statistical features of hazard description word vectors include: word frequency, TF-IDF weight, and keyword frequency. Process basic features include: number of process nodes, statistical features of processing time at each node, role coding of recorder / processor / acceptance personnel, and operational proficiency features, approval pass rate, and first-time acceptance pass rate. Statistical features of processing time at each node include: minimum, maximum, median, and average; proficiency = number of times the same type of hazard was processed / total number of processing times.

[0017] Second-level feature extraction: Based on the first-level basic features, association rule mining and time series analysis algorithms are used to extract association rules and time series features. Association feature extraction uses the Apriori algorithm, including: the co-occurrence probability of different hazard types in the same area, the success rate of the same department handling hazards of different risk levels, the Pearson correlation coefficient between the average processing time of the same type of hazard handled by the same handler and the acceptance pass rate, the association strength between hazard type and involved equipment, and the suitability between department and handler. Time series feature extraction uses the sliding window method, including: the frequency of hazard occurrence in a certain area / department / equipment, the slope of the trend of processing time changes, the statistical characteristics of the time interval between repeated occurrences of the same type of hazard, and the distribution characteristics of peak periods of hazard occurrence. The statistical characteristics of the time interval between repeated occurrences of the same type of hazard include: average interval and interval variance.

[0018] Third-level feature extraction: Based on the second-level features, a feature extraction network is constructed using a deep learning framework to extract deep semantic and long-term trend features, forming the third-level features. Deep semantic features are extracted through a convolutional neural network (CNN), which contains three convolutional layers with 3×3 kernels (32, 64, 128 kernels respectively), two max-pooling layers with 2×2 kernels, and one Flatten layer. These layers perform convolution operations on the second-level associated features to mine the deep semantic associations between hazard type, region, department, and equipment, outputting convolutional features. Long-term trend features are extracted through a bidirectional long short-term memory (LSTM) network, which contains two hidden layers with 128 hidden units per layer and a dropout rate of 0.2. This network performs sequence modeling on the second-level temporal features to capture the long-term dependencies and temporal evolution trends of hazard occurrence and handling, outputting sequence features. The convolutional features and sequence features are concatenated dimensionally and normalized using a BatchNorm layer to obtain the third-level features.

[0019] Furthermore, in step S4, the safety management hybrid model includes a risk level prediction sub-model (CNN-LSTM-Attention hybrid network) and a processing flow optimization sub-model (XGBoost+LightGBM ensemble model). The risk level prediction sub-model takes the third-level features as input, strengthens the weights of key features through the Attention mechanism, and outputs the risk level and confidence level of the hidden danger. The processing flow optimization sub-model takes the third-level features and the risk level prediction results as input, and outputs the optimal handler code, the optimal number of processing nodes, the expected processing time of each node, the probability of processing timeout, and the optimal approval path.

[0020] Furthermore, in S5, the phased training specifically includes:

[0021] Dataset partitioning: The deep feature set was divided into training set, validation set, and test set in a 7:2:1 ratio using stratified sampling.

[0022] Model initialization: Initialize the CNN-LSTM-Attention network. Convolutional kernels are initialized using a He normal distribution, LSTM weights are initialized using orthogonal initialization, and the Attention layer uses a scaled dot product attention mechanism; initialize XGBoost parameters; initialize LightGBM parameters;

[0023] Pre-training phase: The training set is input into the model. The risk level prediction sub-model uses the cross-entropy loss function, and the processing flow optimization sub-model uses the mean squared error loss function to predict the processing time. The cross-entropy loss function is used to predict the person / path. The Adam optimizer is used for iterative training for 50 rounds, and a model snapshot is output every 10 rounds.

[0024] Fine-tuning phase: Based on the validation set, a grid search method and early stopping mechanism are used to prevent model overfitting;

[0025] Integration optimization: The output results of XGBoost and LightGBM of the processing flow optimization sub-model are integrated using a weighted voting method. The weights are dynamically allocated based on the accuracy of the validation set, and the higher the accuracy, the greater the weight.

[0026] Performance Validation: The model is tested using a test set. The performance of the risk prediction sub-model is evaluated using accuracy, precision, recall, and F1 score. The performance of the processing time prediction is evaluated using mean absolute error (MAE) and coefficient of determination (R²). The model is considered to have completed training when the risk prediction accuracy, F1 score, processing time prediction MAE, and processing person matching accuracy all reach the preset values.

[0027] Furthermore, in step S6, the deep integration of the model and system and its full-process application specifically include:

[0028] Interface adaptation: Develop a standardized RESTful API for the model and the enterprise security management system, supporting real-time interaction between model input / output data and the system's MySQL database;

[0029] Hazard entry stage: Users enter hazard information through the system. Hazard information includes: title, description, area, and equipment involved. The system automatically collects environmental context data, and after standardized preprocessing in step S1 and deep feature extraction in step S2, it is input into the safety management model in real time. The model outputs risk level and confidence level, automatically recommends the optimal responsible department, handler and priority, and matches preset processing templates based on hazard type.

[0030] In the hazard allocation phase: the system reads the optimal number of processing nodes and approval path output by the model through the Elsa workflow engine, automatically arranges personalized processing flows, supports custom addition and removal of nodes, dynamically allocates tasks to the processing personnel accounts, and sends notifications simultaneously through system pop-ups, emails, and SMS; based on the model's prediction of the estimated processing time for each node, it sets timeout warning thresholds and writes them into the system parameter maintenance module.

[0031] During the hazard tracking phase: The system collects processing flow data in real time, updates feature data and inputs it into the model every preset time interval, and dynamically predicts the probability of processing timeout. When the timeout probability is greater than or equal to the timeout probability threshold, the system automatically sends an early warning to the approval manager and recommends resource allocation solutions, including: adding processing personnel and adjusting processing priority. Users can view the process progress, model prediction results and processing suggestions through the "Safety Hazard Tracking" module, and can click "Agree" to advance the process or "Revert" to go back to the previous node.

[0032] Acceptance and Archiving Phase: Acceptance specialists submit acceptance results through the system. These results include on-site photos and test data. The model, combined with the third-level characteristics of similar historical hazards and acceptance criteria, verifies the reasonableness of the current acceptance results. After verification, the model outputs archiving suggestions, and the system automatically archives the hazard data to the MySQL database, supporting queries by process number, hazard type, and handling result. If a hazard is found to be non-hazardous, the user can click the "Close" button, and the system marks it as "Closed" and archives it. Simultaneously, this data is included in the model's negative sample library for subsequent iterations.

[0033] Report generation and decision support phase: Based on the third-level features and model output results, the system generates multi-dimensional visualization reports, including a heat map of hazard risk distribution, a trend chart of processing efficiency optimization, and a ranking table of high-risk hazards. Among them, the hazard risk distribution heat map supports drill-down analysis by region, department, and hazard type; the processing efficiency optimization trend chart compares the processing time and acceptance pass rate before and after model application; it supports users to select filtering conditions such as region, department, and time range, and export reports in PDF / Excel format; the report data is synchronized to the management decision dashboard to assist in the formulation of safety management strategies.

[0034] According to another aspect of the present invention, a security management system for an enterprise is provided. The system is used to implement the security management method for an enterprise described above, including: a multi-source data acquisition module, a data standardization preprocessing module, a deep feature extraction module, a hybrid model construction and training module, and a model integration and full-process application module.

[0035] Multi-source data acquisition module: used to collect historical and real-time data from multiple sources throughout the entire lifecycle of enterprise safety hazards;

[0036] Data standardization preprocessing module: used to obtain a structured, modelable standardized dataset through data cleaning, heterogeneous data standardization, and labeling.

[0037] The deep feature extraction module is used to perform deep feature extraction based on standardized data, extracting static and process basic features to form first-level features. Based on the first-level features, association rules and temporal evolution features are mined to form second-level features. Based on the second-level features, deep semantic and long-term trend features are extracted to form third-level features.

[0038] Hybrid Model Construction and Training Module: Used to construct a hybrid safety management model that includes a risk level prediction sub-model and a process optimization sub-model; adopts a phased training strategy of "pre-training-fine-tuning-validation", combined with regularization and hyperparameter optimization, to obtain a safety management model with satisfactory performance;

[0039] Model Integration and Full-Process Application Module: This module is used to deeply integrate the trained security management model with an enterprise security management system based on .NET 6, Vue3, MySQL, and the Elsa workflow engine. It covers the entire process of hazard entry, allocation, processing, acceptance, and archiving. Through the model, it outputs risk assessment results, process optimization solutions, and dynamic tracking suggestions in real time, thereby realizing intelligent enterprise security management.

[0040] Beneficial Effects: Full and Standardized Data Utilization: By integrating scattered and heterogeneous data through multi-source data collection and standardized preprocessing, the problems of low data quality and difficulty in centralized analysis in existing technologies are solved. A three-level vertical feature extraction system deeply mines the basic information, correlation patterns, and temporal trends in the data, maximizing the utilization of data value. Accurate Risk Prediction: Based on a CNN-LSTM-Attention model trained with deep features, the system achieves a shift from "passive response" to "proactive prevention," effectively reducing the probability of security incidents. Intelligent and Efficient Processing: The processing workflow optimization sub-model outputs the optimal handler, processing node, and approval path. Combined with the Elsa workflow engine, it enables personalized orchestration and dynamic allocation of processes, improving security management efficiency.

[0041] High system integration: The model is deeply integrated with the enterprise safety management system, covering the entire lifecycle of hazards. It achieves closed-loop management of data collection, feature extraction, model prediction, process execution, and report generation, and is compatible with multi-role collaboration (system administrators, approval managers, processing personnel, etc.) and fine-grained access control requirements. Strong adaptability and scalability: Supports various hazard types and adapts to a three-level area division of workshop-production line-equipment, meeting the safety management needs of enterprises of different sizes and industries. The model iteration mechanism ensures that it can adapt to the dynamic changes in enterprise safety management needs and maintain its practical value in the long term. Strong decision support: Multi-dimensional visual reports and management decision dashboards provide data support for the formulation of safety management strategies, helping enterprises optimize resource allocation (such as increasing the frequency of inspections in high-risk areas and strengthening training on specific types of hazards) and improve the overall level of safety management. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the method. Detailed Implementation

[0043] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] Example:

[0045] S1: Multi-source data acquisition

[0046] The patented multi-source data acquisition module integrates scattered data within the enterprise, including: core business data, related supporting data, and environmental context data. Core business data includes: hazard lifecycle data and hazard attribute data. Related supporting data includes: organizational structure data, personnel role data, and process configuration data. Environmental context data includes: equipment operation data, environmental monitoring data, and time-dimensional data. Hazard lifecycle data includes: input information, allocation records, processing logs, acceptance results, and archived information. Hazard attribute data includes: title, detailed description, location, designed equipment, and hazard type. Organizational structure data includes: departmental levels and job responsibilities. Personnel role data includes: permissions and operation history of system administrators, approval managers, recorders, acceptance specialists, and processing personnel. Process configuration data includes: approval rules, node settings, and timeout thresholds. Equipment operation data includes: equipment model, runtime, and maintenance records. Environmental monitoring data includes: temperature and humidity, dust concentration, and gas concentration. Time-dimensional data includes: seasons, holidays, and peak production period markers.

[0047] Data collection method: Unstructured data is manually entered and supplemented through system interfaces connecting to the equipment management system and environmental monitoring sensors.

[0048] S2: Standardized Preprocessing

[0049] 2.1 Data cleaning

[0050] Missing value handling: Set the missing rate threshold of key fields (hidden danger type, occurrence area, involved equipment) to 20%. For a hidden danger with the "involved equipment" field missing and a missing rate of 12% < 20%, the K-nearest neighbor algorithm is used to fill it based on the common equipment of hidden dangers in the same area (Line 2 of the stamping workshop) and of the same type (equipment failure). For 1 hidden danger with both the "occurrence area" and "hidden danger type" missing (missing rate 100%), it is directly removed.

[0051] Duplicate record deletion: Based on the combination key of "hidden danger number + occurrence area + hidden danger type", 3 duplicate input records are identified and deleted.

[0052] Outlier handling: The Isolation Forest algorithm is used to identify outliers in the processing duration. A total of 17 abnormal data are removed, and finally 3,180 valid data are retained.

[0053] 2.2 Heterogeneous data standardization

[0054] Categorical data encoding: For unordered categorical data (hidden danger type: equipment failure / operation violation / environmental overstandard; department: equipment department / safety department / production department), one-hot encoding is used. For example, "equipment failure" is encoded as [1, 0, 0]; for ordered categorical data (area level: workshop / production line / equipment; risk level: low / medium / high), label encoding is used. For example, "equipment-level area" is encoded as 2, and "high risk" is encoded as 2.

[0055] Time data processing: Convert all time fields (entry time, start time of processing, etc.) to Unix timestamps, and derive features such as "whether it is the production period" (1 = production period, 0 = non-production period) and "number of days since the last occurrence of the same type of hidden danger" (for example, the interval between the current equipment failure and the last occurrence is 15 days).

[0056] Text data processing: For texts such as hidden danger descriptions and processing remarks, use the BERT pre-trained model for word segmentation (for example, "hydraulic system pressure insufficient" is segmented into "hydraulic system / pressure / insufficient") and stop word removal (delete words such as "de", "le", etc.).

[0057] Numerical data standardization: For numerical data such as processing duration, equipment operation duration, gas concentration, etc., use Z-Score standardization to eliminate the influence of dimensions.

[0058] 2.3 Labeling and annotation

[0059] Risk level labeling: Independently labeled by two safety management experts, with discrepancies reviewed collectively by the safety committee. The potential hazard of "leakage in the hydraulic system of the stamping machine" could cause the production line to shut down for 36 hours (more than 24 hours), and is labeled as "high risk"; the potential hazard of "operational violation of not wearing a safety helmet" has no property damage and does not affect production, and is labeled as "low risk".

[0060] Auxiliary labeling: processing efficiency label (based on the average processing time of similar hazards, this time 24 hours < average 30 hours, labeled "highly efficient"); recurrence label (this hazard is the first time it has occurred, labeled "no"). The final result is a standardized dataset containing three levels of risk labels, processing efficiency labels, and recurrence labels.

[0061] S3: Deep Feature Extraction

[0062] 3.1 First-level characteristics (static and process-based basic characteristics)

[0063] Feature extraction is performed using feature engineering toolkits (such as Scikit-learn), as shown in the following example:

[0064] Static basic features: Hazard type code [1,0,0], area code (stamping workshop=1, line 2=2, stamping machine 3=3), department code (equipment department=2), hazard description word vector TF-IDF weight, involved equipment code (stamping machine 3=13), equipment operating status (normal=1).

[0065] Basic process characteristics: 4 processing nodes (entry-assignment-processing-acceptance), processing time statistics for each node (entry 30 minutes, assignment 30 minutes, processing 24 hours, acceptance 60 minutes, minimum 30 minutes, maximum 24 hours, median 90 minutes, average 6.5 hours), operator proficiency, approval pass rate 100%, and acceptance first-time pass rate 100%.

[0066] 3.2 Second-level features (association rules and temporal evolution features)

[0067] Association features (Apriori algorithm): Co-occurrence probability of equipment failure and hydraulic system problems in the same area (Stamping Workshop Line 2) 0.35; Success rate of the Equipment Department in handling high-risk hazards 0.92; Pearson correlation coefficient between the average time Li Si takes to handle hydraulic system hazards and the acceptance pass rate 0.88; Association strength between hazard type and involved equipment 0.79.

[0068] Temporal characteristics (sliding window method, window size 3 months): The frequency of hidden dangers on the No. 2 line of the stamping workshop was 12 times, the slope of the change trend of the handling time was -0.2 (gradually shortening), the average interval of repeated occurrence of the same hidden danger was 45 days, and the peak time of hidden danger occurrence was 10:00-11:00 (accounting for 30%).

[0069] 3.3 Third-level features (deep semantics and long-term trend features)

[0070] Deep semantic features: The deep semantic association between "equipment failure - stamping workshop - hydraulic system - No. 3 stamping machine" is mined by extracting deep semantic features through CNN (3 convolutional layers, 3×3 convolutional kernels, number of layers 32 / 64 / 128; 2 max pooling layers, 2×2 pooling kernels; 1 Flatten layer) and outputting convolutional features.

[0071] Long-term trend characteristics: By extracting bidirectional LSTM, the temporal evolution trend of equipment failures in the stamping workshop over the past year is captured, and the sequence characteristics are output.

[0072] Feature fusion: Convolutional features and sequence features are concatenated according to their dimensions, and then normalized by the BatchNorm layer to obtain the third-level feature vector.

[0073] S4: Hybrid Model Construction

[0074] The twin model described in the patent is constructed with the following specific configuration:

[0075] Risk level prediction sub-model (CNN-LSTM-Attention): Input the third-level features, the Attention mechanism strengthens the weights of key features such as "downtime" and "equipment importance", and outputs the risk level (low / medium / high) and confidence level.

[0076] Process optimization sub-model (XGBoost+LightGBM integration): Input the third-level feature + risk level prediction result, output the optimal handler code, the optimal number of processing nodes (4), the estimated processing time for each node (20 minutes for data entry, 20 minutes for allocation, 20 hours for processing, and 40 minutes for acceptance), the processing timeout probability (3.2%), and the optimal approval path (Equipment Department Manager → Safety Director).

[0077] S5: Phased Training

[0078] 5.1 Dataset Partitioning

[0079] Stratified sampling was used to divide the feature data into a training set, a validation set of 636 records, and a test set in a ratio of 7:2:1 (ensuring that the sample proportions for each risk level were consistent).

[0080] 5.2 Model Initialization

[0081] CNN-LSTM-Attention: The convolutional kernels are initialized using a He normal distribution, the LSTM weights are initialized using orthogonal distribution, and the attention layer uses a scaled dot product attention mechanism.

[0082] XGBoost: max_depth=6, learning_rate=0.1, n_estimators=200, subsample=0.8.

[0083] LightGBM: num_leaves=31, learning_rate=0.08, n_estimators=200, feature_fraction=0.8.

[0084] 5.3 Phased Training

[0085] Pre-training phase: The training set is input into the model. The risk level prediction sub-model uses the cross-entropy loss function, the processing flow optimization sub-model uses the mean squared error loss function for processing time prediction and the cross-entropy loss function for processing person / path prediction. The Adam optimizer (learning rate 0.001) is trained iteratively for 50 rounds, and a model snapshot is output every 10 rounds. The performance is optimal in the 30th round snapshot.

[0086] Fine-tuning phase: Hyperparameters are optimized using a grid search method based on the validation set, combined with an early stopping mechanism (stopping if there is no performance improvement after 10 consecutive rounds) to prevent overfitting, and finally determine the optimal combination of hyperparameters.

[0087] Integration optimization: XGBoost validation set accuracy was 0.93, and LightGBM accuracy was 0.94. Weighted voting method was used for integration (weights were 0.49 and 0.51, respectively).

[0088] Performance verification: The test set evaluation results are as follows, all of which meet the preset thresholds (accuracy ≥ 0.9, F1 ≥ 0.85, MAE ≤ 2 hours, matching accuracy ≥ 0.9), and the model training is complete.

[0089] S6: Deep Integration of Models and Systems and Full-Process Application

[0090] The system is developed based on .NET 6 (backend), Vue 3 (frontend), MySQL (database), and the Elsa workflow engine. It enables real-time interaction between the model and the system through a RESTful API, covering the entire application process.

[0091] 6.1 Hazard Entry Stage

[0092] User operation: Safety specialist Zhang San enters hazard information (title, description, area, and equipment involved) through the system front end.

[0093] System processing: Automatically collects environmental context data (device runtime, temperature, humidity, etc.), and inputs it into the model in real time after preprocessing and feature extraction.

[0094] Model output: Risk level "High Risk" (confidence 0.96), recommended responsible department "Equipment Department", optimal handler "Li Si" (matching accuracy 0.95), priority "Level 1", matching the preset "Equipment Hydraulic System Fault Handling Template" (shutdown → pressure relief → maintenance → pressure test → resumption of production).

[0095] 6.2 Hazard Allocation Phase

[0096] Workflow orchestration: The Elsa workflow engine reads the model output and automatically orchestrates personalized workflows (entry → equipment department manager approval → Li Si processing → Sun Qi acceptance), supporting manual addition and removal of nodes (not adjusted in this instance).

[0097] Task allocation: The system dynamically allocates tasks to Li Si's account and simultaneously sends notifications via system pop-ups, corporate email, and SMS ("You have a high-risk hazard handling task: the hydraulic system of the No. 3 stamping machine on line 2 of the stamping workshop is leaking. Please complete the handling within 20 hours.").

[0098] Timeout setting: Based on the model's predicted total duration of 21.3 hours, a timeout warning threshold of 25 hours is set and written into the system parameter maintenance module.

[0099] 6.3 Hazard Tracking Phase

[0100] Dynamic monitoring: The system collects process data every 2 hours (e.g., Li Si has received the task and started the pressure relief operation), updates the feature data and inputs it into the model to dynamically predict the timeout probability (1.8% probability of timeout in the 10th hour and 2.5% probability of timeout in the 18th hour).

[0101] Warning and Allocation: The probability of timeout is always less than the threshold (25%), and no warning is triggered; users can check the progress through the "Safety Hazard Tracking" module (such as "In Progress: Hydraulic oil pipe replacement completed, awaiting pressure test"), and can click "Agree" to advance or "Revert" to adjust.

[0102] 6.4 Acceptance and Archiving Stage

[0103] Acceptance and verification: Acceptance specialist Sun Qi submitted the acceptance results (3 on-site photos, pressure test data of 10MPa (meets the standard)). The model was compared with the third-level characteristics of similar hidden dangers in history and the acceptance standards, and the verification was passed.

[0104] Archiving: The system automatically archives the hazard data (including the handling process and acceptance results) to the MySQL database, and supports querying by hazard number, hazard type, handling result and other dimensions; this is a valid hazard, and there is no "close" operation.

[0105] 6.5 Report Generation and Decision Support Stage

[0106] Visualized reports: The system generates multi-dimensional reports, including: heat map of hidden danger and risk distribution; trend chart of processing efficiency optimization; and ranking table of high-risk hidden dangers.

[0107] Decision support: Reports are exported to Excel format and synchronized to the management decision dashboard to assist in strategy formulation (such as increasing the frequency of hydraulic system inspections in the stamping workshop and conducting special training on equipment maintenance).

[0108] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A safety management method for enterprises, characterized in that, Specifically, the following steps are included: S1. Multi-source data acquisition: Collect multi-source historical data on enterprise security risks throughout their entire lifecycle; S2. Standardization Preprocessing: Through data cleaning, heterogeneous data standardization, and labeling, a structured and modelable standardized dataset is obtained; S3. Deep Feature Extraction: Perform deep feature extraction based on standardized data to extract static and process basic features to form first-level features. Based on the first-level features, mine association rules and temporal evolution features to form second-level features. Based on the second-level features, extract deep semantic and long-term trend features to form third-level features. S4. Hybrid Model Construction: Construct a hybrid safety management model that includes a risk level prediction sub-model and a process optimization sub-model. S5. Phased training: A phased training strategy of "pre-training-fine-tuning-verification" is adopted, combined with regularization and hyperparameter optimization, to obtain a security management model with satisfactory performance. S6. Deep Integration of Model and System and Full-Process Application: The trained security management model is deeply integrated with the enterprise security management system based on .NET 6, Vue 3, MySQL and Elsa workflow engine, covering the entire process of hazard entry, allocation, processing, acceptance and archiving. The model outputs risk assessment results, process optimization solutions and dynamic tracking suggestions in real time to realize intelligent enterprise security management.

2. The safety management method for enterprises according to claim 1, characterized in that: In step S1, the multi-source historical data includes: core business data, related supporting data, and environmental context data; core business data includes: hazard lifecycle data and hazard attribute data; related supporting data includes: organizational structure data, personnel role data, and process configuration data; environmental context data includes: equipment operation data, environmental monitoring data, and time-dimensional data; hazard lifecycle data includes: input information, allocation records, processing logs, acceptance results, and archived information; hazard attribute data includes: title, detailed description, occurrence area, designed equipment, and hazard type; organizational structure data includes: department level and job responsibilities; personnel role data includes: permissions and operation history of system administrators, approval managers, recorders, acceptance specialists, and processing personnel; process configuration data includes: approval rules, node settings, and timeout thresholds; equipment operation data includes: equipment model, runtime, and maintenance records; environmental monitoring data includes: temperature and humidity, dust concentration, and gas concentration; and time-dimensional data includes: seasons, holidays, and peak production period markers.

3. The safety management method for enterprises according to claim 1, characterized in that: In step S2, regarding missing data: missing value judgment rules are adopted, and key fields with a missing rate greater than or equal to the missing rate threshold are directly removed. Key fields with a missing rate less than the missing rate threshold are filled using the K-nearest neighbor algorithm. The time field format is unified by matching the time format regular expression. Duplicate records are deleted based on the combination key "hazard number + occurrence area + hazard type". The isolated forest algorithm is used to identify and remove outliers in numerical fields such as processing time and impact range. Heterogeneous data standardization: Differentiated coding strategies are adopted for categorized data, including one-hot coding for unordered categorized data such as hazard type, department, and role; Ordered categorized data, including regional levels and risk levels, are tagged and encoded; time-based data is converted to Unix timestamp format, and derived features such as "whether it is a production period" and "number of days since the last occurrence of the same type of hazard" are generated; text-based data, including hazard descriptions and handling notes, are segmented, stop words are removed, and word vectors are generated using a BERT pre-trained model; numerical data, including handling time and equipment runtime, are standardized using Z-Score to eliminate the influence of units. Labeling and marking: Based on relevant regulations, a three-level risk labeling rule is formulated: High risk: May cause personal injury, major property damage or production stoppage for more than 24 hours; Medium risk: May cause minor property damage or production stoppage for 2-24 hours; Low risk: No property damage and no impact on production; Labeling is done independently by two or more safety management experts, and discrepancies are determined through collective review. Simultaneously label the processing efficiency and recurrence labels; The processing efficiency labels include: high efficiency, average efficiency, and low efficiency; the recurrence labels include: yes and no.

4. A safety management method for enterprises according to claim 1, characterized in that: In step S3, the first-level feature extraction involves using a feature engineering toolkit to extract static basic features and process basic features based on a standardized dataset, forming a basic feature vector. Static basic features include: hazard type coding, region coding, department coding, statistical features of hazard description word vectors, equipment coding, and operating status features. Hazard type coding adapts to various hazard types; region coding supports three-level region division: workshop-production line-equipment; statistical features of hazard description word vectors include: word frequency, TF-IDF weight, and keyword frequency. Process basic features include: number of process nodes, statistical features of processing time at each node, role coding of recorder / processor / acceptance personnel, and operational proficiency features, approval pass rate, and first-time acceptance pass rate. Statistical features of processing time at each node include: minimum, maximum, median, and average; proficiency = number of times the same type of hazard was processed / total number of processing times. Second-level feature extraction: Based on the first-level basic features, association rule mining and time series analysis algorithms are used to extract association rules and time series features. Association feature extraction uses the Apriori algorithm, including: the co-occurrence probability of different hazard types in the same area, the success rate of the same department handling hazards of different risk levels, the Pearson correlation coefficient between the average processing time of the same type of hazard handled by the same handler and the acceptance pass rate, the association strength between hazard type and involved equipment, and the suitability between department and handler. Time series feature extraction uses the sliding window method, including: the frequency of hazard occurrence in a certain area / department / equipment, the slope of the trend of processing time changes, the statistical characteristics of the time interval between repeated occurrences of the same type of hazard, and the distribution characteristics of peak periods of hazard occurrence. The statistical characteristics of the time interval between repeated occurrences of the same type of hazard include: average interval and interval variance. Third-level feature extraction: Based on the second-level features, a feature extraction network is constructed using a deep learning framework to extract deep semantic and long-term trend features, forming the third-level features. Deep semantic features are extracted through a convolutional neural network (CNN), which contains three convolutional layers with 3×3 kernels (32, 64, 128 kernels respectively), two max-pooling layers with 2×2 kernels, and one Flatten layer. These layers perform convolution operations on the second-level associated features to mine the deep semantic associations between hazard type, region, department, and equipment, outputting convolutional features. Long-term trend features are extracted through a bidirectional long short-term memory (LSTM) network, which contains two hidden layers with 128 hidden units per layer and a dropout rate of 0.

2. This network performs sequence modeling on the second-level temporal features to capture the long-term dependencies and temporal evolution trends of hazard occurrence and handling, outputting sequence features. The convolutional features and sequence features are concatenated dimensionally and normalized using a BatchNorm layer to obtain the third-level features.

5. A safety management method for enterprises according to claim 1, characterized in that: In step S4, the safety management hybrid model includes a risk level prediction sub-model (CNN-LSTM-Attention hybrid network) and a processing flow optimization sub-model (XGBoost+LightGBM ensemble model). The risk level prediction sub-model takes the third-level features as input, strengthens the weights of key features through the Attention mechanism, and outputs the risk level and confidence level of the hidden danger. The processing flow optimization sub-model takes the third-level features and the risk level prediction results as input, and outputs the optimal handler code, the optimal number of processing nodes, the expected processing time of each node, the probability of processing timeout, and the optimal approval path.

6. A safety management method for enterprises according to claim 5, characterized in that: In S5, the phased training specifically includes: Dataset partitioning: The deep feature set was divided into training set, validation set, and test set in a 7:2:1 ratio using stratified sampling. Model initialization: Initialize the CNN-LSTM-Attention network. Convolutional kernels are initialized using a He normal distribution, LSTM weights are initialized using orthogonal initialization, and the Attention layer uses a scaled dot product attention mechanism; initialize XGBoost parameters; initialize LightGBM parameters; Pre-training phase: The training set is input into the model. The risk level prediction sub-model uses the cross-entropy loss function, and the processing flow optimization sub-model uses the mean squared error loss function to predict the processing time. The cross-entropy loss function is used to predict the person / path. The Adam optimizer is used for iterative training for 50 rounds, and a model snapshot is output every 10 rounds. Fine-tuning phase: Based on the validation set, a grid search method and early stopping mechanism are used to prevent model overfitting; Integration optimization: The output results of XGBoost and LightGBM of the processing flow optimization sub-model are integrated using a weighted voting method. The weights are dynamically allocated based on the accuracy of the validation set, and the higher the accuracy, the greater the weight. Performance Validation: The model is tested using a test set. The performance of the risk prediction sub-model is evaluated using accuracy, precision, recall, and F1 score. The performance of the processing time prediction is evaluated using mean absolute error (MAE) and coefficient of determination (R²). The model is considered to have completed training when the risk prediction accuracy, F1 score, processing time prediction MAE, and processing person matching accuracy all reach the preset values.

7. A safety management method for enterprises according to claim 1, characterized in that: In step S6, the deep integration of the model and system and its full-process application specifically include: Interface adaptation: Develop a standardized RESTful API for the model and the enterprise security management system, supporting real-time interaction between model input / output data and the system's MySQL database; Hazard entry stage: Users enter hazard information through the system. Hazard information includes: title, description, area, and equipment involved. The system automatically collects environmental context data, and after standardized preprocessing in step S1 and deep feature extraction in step S2, it is input into the safety management model in real time. The model outputs risk level and confidence level, automatically recommends the optimal responsible department, handler and priority, and matches preset processing templates based on hazard type. In the hazard allocation phase: the system reads the optimal number of processing nodes and approval path output by the model through the Elsa workflow engine, automatically arranges personalized processing flows, supports custom addition and removal of nodes, dynamically allocates tasks to the processing personnel accounts, and sends notifications simultaneously through system pop-ups, emails, and SMS; based on the model's prediction of the estimated processing time for each node, it sets timeout warning thresholds and writes them into the system parameter maintenance module. During the hazard tracking phase: The system collects processing flow data in real time, updates feature data and inputs it into the model every preset time interval, and dynamically predicts the probability of processing timeout. When the timeout probability is greater than or equal to the timeout probability threshold, the system automatically sends an early warning to the approval manager and recommends resource allocation solutions, including: adding processing personnel and adjusting processing priorities. Users can view the process progress, model prediction results and processing suggestions through the "Safety Hazard Tracking" module, and can click "Agree" to advance the process or "Revert" to go back to the previous node. Acceptance and Archiving Phase: Acceptance specialists submit acceptance results through the system. These results include on-site photos and test data. The model, combined with the third-level characteristics of similar historical hazards and acceptance criteria, verifies the reasonableness of the current acceptance results. After verification, the model outputs archiving suggestions, and the system automatically archives the hazard data to the MySQL database, supporting queries by process number, hazard type, and handling result. If a hazard is found to be non-hazardous, the user can click the "Close" button, and the system marks it as "Closed" and archives it. Simultaneously, this data is added to the model's negative sample library for subsequent iterations. Report generation and decision support phase: Based on the third-level features and model output results, the system generates multi-dimensional visualization reports, including a heat map of hazard risk distribution, a trend chart of processing efficiency optimization, and a ranking table of high-risk hazards. Among them, the hazard risk distribution heat map supports drill-down analysis by region, department, and hazard type; the processing efficiency optimization trend chart compares the processing time and acceptance pass rate before and after model application; it supports users to select filtering conditions such as region, department, and time range, and export reports in PDF / Excel format; the report data is synchronized to the management decision dashboard to assist in the formulation of safety management strategies.

8. A security management system for enterprises, characterized in that: This system is used to implement a security management method for enterprises as described in any one of claims 1-7, comprising: a multi-source data acquisition module, a data standardization preprocessing module, a deep feature extraction module, a hybrid model construction and training module, and a model integration and full-process application module; Multi-source data acquisition module: used to collect historical and real-time data from multiple sources throughout the entire lifecycle of enterprise safety hazards; Data standardization preprocessing module: used to obtain a structured, modelable standardized dataset through data cleaning, heterogeneous data standardization, and labeling. The deep feature extraction module is used to perform deep feature extraction based on standardized data, extracting static and process basic features to form first-level features. Based on the first-level features, association rules and temporal evolution features are mined to form second-level features. Based on the second-level features, deep semantic and long-term trend features are extracted to form third-level features. Hybrid Model Construction and Training Module: Used to construct a hybrid safety management model that includes a risk level prediction sub-model and a process optimization sub-model; adopts a phased training strategy of "pre-training-fine-tuning-verification", combined with regularization and hyperparameter optimization, to obtain a safety management model with satisfactory performance; Model Integration and Full-Process Application Module: This module is used to deeply integrate the trained security management model with an enterprise security management system based on .NET 6, Vue 3, MySQL, and the Elsa workflow engine. It covers the entire process of hazard entry, allocation, processing, acceptance, and archiving. The model outputs risk assessment results, process optimization solutions, and dynamic tracking suggestions in real time, enabling intelligent enterprise security management.

Images