Method and system for detecting the wearing of labor protection equipment by workers at oil and gas stations

Abstract

The present invention provides a method and system for detecting the wearing of labor protection equipment by workers at oil and gas stations, which relates to the field of object detection technology. The method includes real-time collection of worker image data through image acquisition equipment, calling a deep learning model to detect the wearing status of labor protection equipment, combining process parameters, equipment status and hazard source distribution data to assess the hazard level, generating a wearing standardization score and risk trend value, using a deep reinforcement learning algorithm to implement control measures, and continuously optimizing detection parameters and processing strategies to achieve intelligent and dynamic management of worker safety protection and improve the level of safe operation at stations.

Description

Technical Field

[0001] The present invention relates to article detection technology, and in particular to a method and system for detecting the wearing of labor protection articles by workers at oil and gas stations. Background Art

[0002] In the oil and gas industry, field operations present numerous hazards, including high temperatures, high pressures, flammable and explosive materials, and toxic and hazardous materials. The safety and protection of workers is paramount. Labor protection equipment (PPE) is a fundamental safeguard for workers, and its proper use is directly related to both life and production safety. Currently, oil and gas companies have generally established safety production management systems, requiring workers to wear PPE such as hard hats, protective clothing, protective gloves, protective glasses, and gas masks.

[0003] Traditionally, supervision of labor protection equipment worn by oil and gas station operators relies primarily on manual inspections, regular spot checks, and safety education. Safety managers conduct on-site inspections to verify workers' labor protection equipment, document any issues, and require rectification. Companies also regularly conduct safety education and training to strengthen workers' safety awareness and ensure they wear labor protection equipment correctly.

[0004] However, existing labor protection equipment wearing detection methods have obvious defects and shortcomings. First, the traditional manual inspection method has limited coverage and cannot achieve all-weather, all-round real-time monitoring, resulting in blind spots and time differences in supervision, making it difficult to promptly detect and correct irregular wearing behaviors. Secondly, the existing detection methods lack dynamic assessment of the actual hazard level of the working area, and are unable to adjust labor protection requirements according to actual conditions such as changes in process parameters, equipment status, and distribution of hazardous sources, resulting in a mismatch between protective measures and actual risks, posing safety hazards. Thirdly, the existing technology lacks an intelligent risk warning and management strategy generation mechanism, and cannot conduct forward-looking risk prevention and control based on the correlation between wearing behavior and environmental risks, making it difficult to achieve early identification and prevention of risks. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for detecting the wearing of labor protection equipment by workers at oil and gas stations, which can solve the problems in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a method for detecting whether workers at oil and gas stations are wearing labor protection products, comprising:

[0007] Image acquisition equipment deployed at oil and gas stations collects worker image data in real time, and uses a deep learning model to detect the wearing of labor protection equipment on the worker image data to obtain characteristic data on the worker's wearing status.

[0008] Based on the wearing status characteristic data, a correlation analysis is established in combination with the process parameters, equipment status and hazard source distribution data of the operating area to generate a real-time hazard level for the area where the operator is located; based on the real-time hazard level, an intelligent assessment is performed on the operator's current wearing status to generate a dynamic wearing compliance score, and a risk evolution trend value is calculated based on the dynamic wearing compliance score;

[0009] Based on the dynamic score of wearing norms and the risk evolution trend value, a deep reinforcement learning algorithm is used to generate an optimal control strategy and implement corresponding control measures;

[0010] The wearing status characteristic data, the real-time danger level, the dynamic wearing normative score and the control measures are stored in a database, an analysis report on the workers' wearing behavior of labor protection products is generated, and the wearing detection parameters and the graded processing strategy are continuously optimized based on the analysis report.

[0011] Based on the wear status characteristic data, combined with the process parameters, equipment status and hazard source distribution data of the work area, a correlation analysis is established to generate the real-time hazard level of the area where the operator is located, including:

[0012] Extracting wearing position features, wearing standardization features, and wearing integrity features based on the wearing state feature data to construct a wearing state feature vector; based on the wearing state feature vector, collecting process parameter data related to the position of the operator, and constructing the process parameter data into a process parameter spatiotemporal tensor;

[0013] According to the location of the operator, the operating status data of the relevant equipment in the operation area is obtained to construct an equipment status characteristic matrix; based on the location of the operator and the equipment status characteristic matrix, the distribution data of the corresponding hazard sources in the operation area is collected to establish a hazard source impact function;

[0014] Performing a tensor outer product operation on the wear state feature vector, the process parameter spatiotemporal tensor, the equipment state feature matrix, and the hazard source influence function to construct a spatiotemporal coupling tensor; inputting the spatiotemporal coupling tensor into a transformer architecture, and utilizing a self-attention mechanism to extract spatiotemporal correlation features between multi-source data to generate a feature correlation matrix;

[0015] Based on the spatiotemporal coupling tensor and the characteristic association matrix, a hazard level assessment value of the operating area is calculated, and a historical hazardous event case library is retrieved according to the hazard level assessment value, and the similarity with the historical cases is calculated to generate a decision basis for risk assessment; based on the hazard level assessment value and the decision basis, combined with the preset hazard level classification standard, a real-time hazard level of the area where the operating personnel is located is generated.

[0016] Based on the spatiotemporal coupling tensor and the characteristic association matrix, a hazard level assessment value of the operation area is calculated. According to the hazard level assessment value, a historical hazard event case library is retrieved, and similarity with historical cases is calculated to generate a decision basis for risk assessment, including:

[0017] The spatiotemporal coupling tensor and the feature association matrix are feature-concatenated to generate a fused feature vector; a hazard level assessment function is constructed based on the fused feature vector, the spatiotemporal coupling tensor and the feature association matrix are weighted by a weight matrix, and an initial hazard level assessment value is obtained by performing activation function processing;

[0018] Constructing a dynamic knowledge graph based on the initial risk level assessment value, inputting the initial risk level assessment value into the dynamic knowledge graph, extracting graph structure features through a graph convolutional network, and generating historical case retrieval features;

[0019] Calculating the cosine similarity between the initial risk level assessment value and the historical case retrieval feature to obtain feature-level similarity; optimizing the calculation result of the feature-level similarity using a contrastive learning method, and adjusting the contrast relationship between positive and negative samples using a temperature parameter to improve the accuracy of the similarity calculation;

[0020] Based on the feature-level similarity, similar historical cases are screened to construct a causal decision diagram; based on the causal decision diagram, a decision basis set is generated, wherein the decision basis set includes risk factors, impact relationships, and probability distribution information;

[0021] Calculate the statistical characteristics of the decision basis set and determine the decision reliability interval based on the confidence coefficient; optimize the decision basis set according to the decision reliability interval to generate the final risk assessment decision basis.

[0022] Based on the real-time hazard level, the operator's current wearing status is intelligently evaluated to generate a dynamic wearing standardization score, and a risk evolution trend value is calculated based on the dynamic wearing standardization score, including:

[0023] Determining an assessment weight coefficient for each labor protection product based on the real-time hazard level, collecting the real-time spatial coordinates of the labor protection product, performing a difference calculation between the real-time spatial coordinates and the preset standard wearing position coordinates to obtain a deviation distance, and performing a weighted sum operation on the assessment weight coefficient and the deviation distance to obtain a position deviation assessment value;

[0024] Collecting the wearing integrity information of the labor protection products, performing exponential decay processing on the position deviation evaluation value, and performing weighted combination with the wearing integrity information, and generating a wearing standardization dynamic score through a dynamic scoring function;

[0025] Establishing an association mapping function for the dynamic score of wearing normativeness and the real-time danger level, calculating an interaction coefficient between the two according to the association mapping function, and combining the interaction coefficient with the dynamic score of wearing normativeness to generate a comprehensive evaluation index;

[0026] Construct a fixed-length time window sequence, input the dynamic scoring data of wearing norms within a historical time period into the time window sequence in chronological order, and use an attention calculation method to extract the time series feature weights in the time window sequence;

[0027] The temporal feature weight is predicted based on the dynamic wearing normativeness score, and a risk evolution trend value is calculated according to the prediction result, where the risk evolution trend value represents a changing trend of the wearing state over time.

[0028] Establishing an association mapping function for the dynamic score of wearing norms and the real-time danger level, calculating the interaction coefficient between the two according to the association mapping function, and combining the interaction coefficient with the dynamic score of wearing norms to generate a comprehensive evaluation index includes:

[0029] Using a deep neural network to extract features from the operator's physiological characteristic data to obtain a fatigue feature vector, setting weight coefficients according to the importance of different features, and performing a weighted summation of the fatigue feature vector and the weight coefficients to obtain a fatigue assessment value;

[0030] Inputting the fatigue assessment value into the deep neural network to calculate the fatigue influence coefficient, and performing correction calculation on the dynamic wearing standardization score and the fatigue influence coefficient to obtain a corrected wearing standardization score;

[0031] collecting temperature data, humidity data, and illumination data of the working environment, determining the influence weight of the environmental factors based on the fatigue assessment value, and performing a weighted combination of the temperature data, the humidity data, and the illumination data to generate an environmental impact assessment value;

[0032] fusing and correcting the environmental impact assessment value with the real-time hazard level to obtain a corrected hazard level;

[0033] Establishing an initial mapping relationship based on the revised wearing compliance score and the revised hazard level, and modifying the initial mapping relationship using the operator's attention level data as an adjustment factor to generate a human factor mapping function;

[0034] The modified wearing standardization score is weightedly combined with the human factor mapping function, and the result of the weighted combination is adjusted to generate a comprehensive evaluation index.

[0035] Based on the dynamic score of wearing norms and the risk evolution trend value, a deep reinforcement learning algorithm is used to generate an optimal control strategy, and corresponding control measures are implemented, including:

[0036] Constructing the wearing normative dynamic score into an N-dimensional state score vector, constructing the risk evolution trend value into an M-dimensional trend feature vector, and combining the N-dimensional state score vector and the M-dimensional trend feature vector to form a state input;

[0037] A deep Q network is constructed using a deep reinforcement learning algorithm. The state input is used as the network input, and the control measures are used as the network output. The improvement of the dynamic score of the wearing standardization and the decrease of the risk evolution trend value are constructed as reward signals. The target Q value is calculated based on the reward signal.

[0038] Passing the state input into the deep Q-network to generate an action value estimate, calculating a loss function between the action value estimate and the target Q value using a temporal difference algorithm, and optimizing the parameters of the deep Q-network based on the loss function;

[0039] An ε-greedy strategy is used to select an optimal control strategy from the action value estimates output by the deep Q network, and control measures corresponding to the optimal control strategy are executed.

[0040] The state input is passed into the deep Q network to generate an action value estimate, a temporal difference algorithm is used to calculate a loss function between the action value estimate and the target Q value, and the parameters of the deep Q network are optimized based on the loss function, including:

[0041] The state input is passed into a deep Q network having a multi-layer neural network structure, and the value estimate of each control action in the current state is calculated based on the preset network parameters of the deep Q network, and the control action with the largest value estimate is selected as the execution action;

[0042] Execute the control action and collect an immediate reward value, generate a new state input vector based on the immediate reward value, pass the new state input vector into the deep Q network to calculate the maximum action value estimate at the next moment, add the immediate reward value to the discounted value of the maximum action value estimate to obtain a target Q value;

[0043] Calculate the mean square error between the maximum action value estimate and the target Q value using a temporal difference algorithm, construct the mean square error as a loss function, and calculate the gradient value of the loss function with respect to the preset network parameters;

[0044] Multiplying the gradient value by a preset learning rate to obtain a parameter adjustment amount, updating and optimizing the preset network parameters based on the parameter adjustment amount, generating optimized network parameters, and using the optimized network parameters to update the deep Q network.

[0045] A second aspect of an embodiment of the present invention provides a system for detecting the wearing of labor protection equipment by workers at oil and gas stations, comprising:

[0046] The first unit is used to collect real-time image data of workers through image acquisition equipment deployed at oil and gas stations, call a deep learning model to detect the wearing of labor protection products by the workers' image data, and obtain characteristic data of the workers' wearing status;

[0047] The second unit is configured to establish a correlation analysis based on the wearing status characteristic data, combined with the process parameters, equipment status and hazard source distribution data of the operating area, to generate a real-time hazard level for the area where the operator is located; perform an intelligent assessment of the operator's current wearing status based on the real-time hazard level, generate a dynamic wearing compliance score, and calculate a risk evolution trend value based on the dynamic wearing compliance score;

[0048] The third unit is configured to generate an optimal control strategy using a deep reinforcement learning algorithm based on the dynamic wearing standardization score and the risk evolution trend value, and execute corresponding control measures;

[0049] The fourth unit is used to store the wearing status characteristic data, the real-time danger level, the dynamic wearing standardization score and the control measures in a database, generate an analysis report on the workers' wearing behavior of labor protection products, and continuously optimize the wearing detection parameters and hierarchical processing strategies based on the analysis report.

[0050] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0051] processor;

[0052] a memory for storing processor-executable instructions;

[0053] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0054] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0055] The beneficial effects of this application are as follows:

[0056] The method for detecting the wearing of labor protection equipment by workers at oil and gas stations provided by the present invention can automatically identify the wearing status of workers' labor protection equipment through real-time image acquisition and deep learning model analysis, realizing the intelligent and automated wearing detection, greatly improving supervision efficiency, and reducing the workload of manual inspections.

[0057] The present invention correlates and analyzes the wearing status of workers with process parameters, equipment status and hazard source distribution data to generate real-time hazard levels and dynamic scores of wearing standards. It can formulate differentiated labor protection requirements for work areas with different risk levels, realize precise and dynamic safety management, and effectively improve the level of station safety management.

[0058] The present invention adopts deep reinforcement learning algorithm to generate optimal management and control strategies, and continuously optimizes detection parameters and processing strategies through data storage and analysis reports, building a complete closed-loop management mechanism. It can not only timely discover and correct irregular behaviors in wearing labor protection clothes, but also fundamentally improve the safety awareness and standardized operating habits of operators, significantly reducing the incidence of safety accidents in stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a flow chart of a method for detecting the wearing of labor protection products by workers at oil and gas stations according to an embodiment of the present invention;

[0060] Figure 2 A bar chart comparing the performance of the hazard level assessment method for multi-source data association analysis according to an embodiment of the present invention;

[0061] Figure 3 This is a bar chart comparing the performance of the dynamic evaluation method for wearing normativeness in different scenarios according to an embodiment of the present invention;

[0062] Figure 4 Generate a flow chart for the deep reinforcement learning management and control strategy of an embodiment of the present invention;

[0063] Figure 5 This is a bar chart comparing the performance of the Deep Q network in different test scenarios according to an embodiment of the present invention. DETAILED DESCRIPTION

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0065] The technical solution of the present invention is described in detail below with reference to specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0066] Figure 1 FIG. 1 is a flow chart of a method for detecting the wearing of labor protection products by workers at oil and gas stations according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0067] Image acquisition equipment deployed at oil and gas stations collects worker image data in real time, and uses a deep learning model to detect the wearing of labor protection equipment on the worker image data to obtain characteristic data on the worker's wearing status.

[0068] Based on the wearing status characteristic data, a correlation analysis is established in combination with the process parameters, equipment status and hazard source distribution data of the operating area to generate a real-time hazard level for the area where the operator is located; based on the real-time hazard level, an intelligent assessment is performed on the operator's current wearing status to generate a dynamic wearing compliance score, and a risk evolution trend value is calculated based on the dynamic wearing compliance score;

[0069] Based on the dynamic score of wearing norms and the risk evolution trend value, a deep reinforcement learning algorithm is used to generate an optimal control strategy and implement corresponding control measures;

[0070] The wearing status characteristic data, the real-time danger level, the dynamic wearing normative score and the control measures are stored in a database, an analysis report on the workers' wearing behavior of labor protection products is generated, and the wearing detection parameters and the graded processing strategy are continuously optimized based on the analysis report.

[0071] In an optional embodiment, based on the wear status characteristic data, combined with the process parameters, equipment status and hazard source distribution data of the work area, a correlation analysis is established to generate a real-time hazard level of the area where the operator is located, including:

[0072] Extracting wearing position features, wearing standardization features, and wearing integrity features based on the wearing state feature data to construct a wearing state feature vector; based on the wearing state feature vector, collecting process parameter data related to the position of the operator, and constructing the process parameter data into a process parameter spatiotemporal tensor;

[0073] According to the location of the operator, the operating status data of the relevant equipment in the operation area is obtained to construct an equipment status characteristic matrix; based on the location of the operator and the equipment status characteristic matrix, the distribution data of the corresponding hazard sources in the operation area is collected to establish a hazard source impact function;

[0074] Performing a tensor outer product operation on the wear state feature vector, the process parameter spatiotemporal tensor, the equipment state feature matrix, and the hazard source influence function to construct a spatiotemporal coupling tensor; inputting the spatiotemporal coupling tensor into a transformer architecture, and utilizing a self-attention mechanism to extract spatiotemporal correlation features between multi-source data to generate a feature correlation matrix;

[0075] Based on the spatiotemporal coupling tensor and the characteristic association matrix, a hazard level assessment value of the operating area is calculated, and a historical hazardous event case library is retrieved according to the hazard level assessment value, and the similarity with the historical cases is calculated to generate a decision basis for risk assessment; based on the hazard level assessment value and the decision basis, combined with the preset hazard level classification standard, a real-time hazard level of the area where the operating personnel is located is generated.

[0076] The system acquires wear status characteristic data from workers, which is collected in real time by sensors on wearable devices. The system extracts three types of features from this data: wearing position features, wearing compliance features, and wearing integrity features. Wearing position features include the actual wearing coordinates of safety equipment such as helmets, protective glasses, and protective gloves. Wearing compliance features include parameters such as helmet tightness, face mask angle, and gas mask airtightness. Wearing integrity features include indicators such as the proportion of required equipment worn and the coverage of key protective areas.

[0077] These features are organized into a wearing status feature vector, such as V = [0.95, 0.88, 1.0, 0.92, 0.75], which represent the standardization of the helmet, the integrity of the protective clothing, the wearing condition of the gloves, the air tightness of the respiratory protective equipment, and the wearing status of the protective boots, respectively.

[0078] Based on the operator's real-time location information, the system collects process parameter data related to that location. Process parameters include environmental indicators such as temperature, pressure, hazardous gas concentration, and noise level, as well as production parameters such as production line speed and chemical reaction progress. The system organizes these parameters into a spatiotemporal tensor of process parameters based on their time series and spatial distribution. For example, in a chemical plant environment, the system collects the following parameters for area A at time t: temperature 85°C, pressure 4.2 MPa, ammonia concentration 15 ppm, noise 85 dB, and reactor speed 120 rpm. This data is stored in a tensor structure according to time and space coordinates.

[0079] Based on the operator's location, the operating status data of relevant equipment is obtained, including information such as equipment operating frequency, load factor, maintenance status, and fault history. This data is constructed into an equipment status feature matrix. For example, in a production workshop, the operating status matrix for pressure vessel P001 includes parameters such as operating time 8760 hours, current load factor 83%, last maintenance 45 days ago, historical fault frequency 0.02 times / month, and vibration amplitude 2.5mm.

[0080] Combining the operator's location and equipment status matrix, the system collects data on the distribution of hazardous sources within the work area, such as the storage locations of flammable and explosive materials, the distribution of high-voltage equipment, and toxic gas leakage risk points. The system then establishes a hazard impact function, which describes the degree of impact of a hazard source at different locations. For example, the hazard impact of a chemical storage tank decreases with distance: at a distance of 5 meters, the hazard level is 0.8, at 10 meters, it is 0.5, and at 20 meters, it is 0.2.

[0081] A tensor outer product operation is performed on the wearing state feature vector, the spatiotemporal tensor of process parameters, the equipment state feature matrix, and the hazard source impact function to construct a spatiotemporal coupling tensor. This operation fuses data from different dimensions into a unified high-dimensional representation, capturing the interactions between various elements. For example, if the worker's gas mask is not airtight (score 0.75) and the area is exposed to high toxic gas concentrations (15 ppm), the combined risk value will be significantly higher than if each factor is considered separately.

[0082] The constructed spatiotemporal coupling tensor is input into the Transformer architecture, which uses a self-attention mechanism to extract spatiotemporal correlation features between multi-source data. The Transformer model automatically learns the importance weights between different features and identifies potential risk patterns. After processing, it generates a feature correlation matrix that quantifies the interplay between different risk factors. For example, the system identifies a pattern where the combination of missing protective equipment and abnormalities in specific process parameters leads to a sharp increase in risk levels.

[0083] Based on the spatiotemporal coupling tensor and characteristic correlation matrix, the system calculates a hazard level assessment for the work area. This calculation takes into account the weights and interactions of various risk factors. The system compares this assessment with records in a historical hazardous incident case database, calculates similarities, and generates a basis for risk assessment decisions. For example, if the current situation has an 87% similarity to a poisoning incident in the case database caused by insufficient airtightness of a protective mask, the system will use this case as a reference for risk warnings.

[0084] Based on the hazard level assessment and decision-making basis, combined with the preset hazard level classification standards, the real-time hazard level of the area where the operator is located is generated. The hazard level can be divided into five levels: low risk (0-0.3), medium-low risk (0.3-0.5), medium risk (0.5-0.7), high risk (0.7-0.9), and extremely high risk (0.9-1.0).

[0085] The real-time hazard level is displayed on the operator's wearable device and simultaneously transmitted to the safety monitoring center. For example, if the calculated hazard level is 0.82, the system will determine the current high-risk status, prompting the operator to immediately check their protective equipment and consider evacuating the area, while also notifying the safety supervisor to come and handle the situation.

[0086] This method uses multi-source data fusion and deep learning technology to achieve accurate assessment of industrial operating environment risks, provide real-time safety protection for operators, and effectively prevent the occurrence of industrial safety accidents.

[0087] Figure 2 This is a bar chart comparing the performance of the hazard level assessment method for multi-source data association analysis according to an embodiment of the present invention:

[0088] This figure compares the performance of three different methods (traditional threshold judgment, single feature fusion, and tensor-coupled self-attention) across five evaluation metrics. In terms of hazard source identification accuracy, the three methods achieved 67.2%, 78.6%, and 92.3%, respectively. In terms of risk level prediction accuracy, the three methods achieved 72.5%, 81.4%, and 89.7%. In terms of early warning lead time, the three methods achieved 55.8%, 73.2%, and 87.5%, respectively. In terms of multi-source data association efficiency, the accuracy rates were 61.3%, 75.8%, and 90.2%, respectively. In terms of abnormal state handling capability, the three methods achieved 63.7%, 72.5%, and 88.6%, respectively. The data shows that the tensor-coupled self-attention method exhibits significant advantages across all evaluation metrics, particularly in hazard source identification accuracy, achieving the highest performance of 92.3%. The traditional threshold judgment method performed relatively poorly across all metrics, particularly in terms of early warning lead time, which was only 55.8%. Although the single feature fusion method has improved compared to traditional methods, it still lags behind the tensor-coupled self-attention method. Overall, the tensor-coupled self-attention method demonstrates stronger comprehensive analysis capabilities and prediction accuracy.

[0089] In an optional embodiment, based on the spatiotemporal coupling tensor and the characteristic association matrix, a hazard level assessment value of the operation area is calculated, and according to the hazard level assessment value, a historical hazard event case library is retrieved, and similarity with historical cases is calculated to generate a decision basis for risk assessment, including:

[0090] The spatiotemporal coupling tensor and the feature association matrix are feature-concatenated to generate a fused feature vector; a hazard level assessment function is constructed based on the fused feature vector, the spatiotemporal coupling tensor and the feature association matrix are weighted by a weight matrix, and an initial hazard level assessment value is obtained by performing activation function processing;

[0091] Constructing a dynamic knowledge graph based on the initial risk level assessment value, inputting the initial risk level assessment value into the dynamic knowledge graph, extracting graph structure features through a graph convolutional network, and generating historical case retrieval features;

[0092] Calculating the cosine similarity between the initial risk level assessment value and the historical case retrieval feature to obtain feature-level similarity; optimizing the calculation result of the feature-level similarity using a contrastive learning method, and adjusting the contrast relationship between positive and negative samples using a temperature parameter to improve the accuracy of the similarity calculation;

[0093] Based on the feature-level similarity, similar historical cases are screened to construct a causal decision diagram; based on the causal decision diagram, a decision basis set is generated, wherein the decision basis set includes risk factors, impact relationships, and probability distribution information;

[0094] Calculate the statistical characteristics of the decision basis set and determine the decision reliability interval based on the confidence coefficient; optimize the decision basis set according to the decision reliability interval to generate the final risk assessment decision basis.

[0095] The collected spatiotemporal coupling tensor is concatenated with the feature correlation matrix to generate a fused feature vector. The spatiotemporal coupling tensor has dimensions (T, N, D), where T represents the time step, N represents the number of spatial nodes, and D represents the feature dimension. The feature correlation matrix has dimensions (N, M), where M represents the number of correlated features.

[0096] During feature concatenation, the spatiotemporal coupling tensor is compressed in the time dimension to obtain an (N, D)-dimensional matrix. This is then concatenated with the feature correlation matrix in the feature dimension, ultimately resulting in a fused feature vector of (N, D + M) dimensions. For example, for an operation area containing 8 spatial nodes, the spatiotemporal feature dimension is 12, and the correlation feature dimension is 5. The resulting fused feature vector has a dimension of (8, 17).

[0097] A hazard level assessment function is constructed based on the fused feature vector, and the spatiotemporal coupling tensor and the feature correlation matrix are weighted by a weight matrix. The dimension of the weight matrix is ​​(D+M, K), where K represents the number of hazard level categories, which is set to 5 levels in this embodiment.

[0098] Perform a matrix multiplication operation, multiplying the fused feature vector by the weight matrix. The result is nonlinearly transformed using the ReLU activation function to obtain the initial danger level assessment value, with dimensions (N, K). In the specific implementation, the initial value of the weight matrix is ​​obtained from the pre-trained model. For example, for spatial node 3, its fused feature vector is [0.75, 0.82, 0.91, 0.56, 0.63, 0.78, 0.85, 0.72, 0.69, 0.81, 0.77, 0.83, 0.62, 0.59, 0.76, 0.88, 0.71]. After weighting by the weight matrix and processing with the activation function, the initial danger level assessment value is [0.12, 0.24, 0.53, 0.08, 0.03], indicating that the danger level of this spatial node is level 3 (corresponding to the maximum value of 0.53).

[0099] A dynamic knowledge graph is constructed based on the initial hazard level assessment value. In the knowledge graph, nodes represent spatial locations, edges represent spatial associations, and node features include the initial hazard level assessment value. The adjacency matrix A of the knowledge graph has dimensions (N, N) and represents the connection relationship between nodes. In this embodiment, the adjacency matrix is ​​calculated based on spatial distance. When the Euclidean distance between two nodes is less than a preset threshold (set to 50 meters), the corresponding adjacency matrix element is set to 1, otherwise it is set to 0.

[0100] The initial hazard level assessment is fed into a dynamic knowledge graph, and graph structural features are extracted using a two-layer graph convolutional network. The first layer of graph convolution outputs 64-dimensional features, while the second layer outputs 32-dimensional features, ultimately generating a historical case retrieval feature with a dimension of (N, 32).

[0101] Calculate the cosine similarity between the initial hazard level assessment value and the historical case retrieval features to obtain the feature-level similarity. The specific process is to represent the initial hazard level assessment value of the current scene as a vector P, and the cases in the historical case library as a vector set Q={Q1,Q2,...,Q m}, where m is the number of historical cases. For each historical case Q i , calculate its cosine similarity S with P i , S i The value range of is [-1, 1], and the larger the value, the higher the similarity. In this embodiment, the historical case library contains 500 cases. When the similarity between a case and the current scenario is calculated to be 0.87, it indicates that the two are highly similar.

[0102] A contrastive learning method is used to optimize feature-level similarity calculations. A temperature parameter, τ (set to 0.5), is introduced to adjust the contrast between positive and negative samples. Cases with similarity greater than a threshold (set to 0.75) are considered positive samples, while the rest are considered negative samples. The InfoNCE loss function is used to calculate contrastive loss, optimizing feature representation and improving the accuracy of similarity calculations. After optimization, the similarity of positive samples is further improved. For example, the original similarity of 0.87 is improved to 0.92 after optimization, while the similarity of negative samples decreases accordingly.

[0103] Based on feature-level similarity, similar historical cases were screened and the top 10 cases were selected to construct a causal decision graph. In the causal decision graph G=(V,E), the node set V represents risk factors, and the edge set E represents impact relationships. Each node has attributes including the risk factor name, risk level, and probability of occurrence. Edge attributes include impact strength and causal type. For example, the "Working at Heights" node has a risk level of 4 and a probability of 0.85. There is a causal relationship between this node and the "Inadequate Safety Protection Measures" node, with an impact strength of 0.79 and a causal type of "promotion."

[0104] Generate a decision basis set based on the causal decision diagram, including risk factors, impact relationships and probability distribution information. For each risk factor node v i , extract its risk level r i , probability of occurrence p i and the influence strength w of all edges connected to it ij The generated decision basis set D={d1,d2,...,d n}, where d i =(v i ,r i ,p i ,{(v j ,w ij )}), represents the risk level, occurrence probability and its impact relationship with other factors of risk factor vi.

[0105] Calculate the statistical characteristics of the decision-making basis set, including the average risk level, the highest risk level, the concentration of risk factors, and the degree of causal correlation. Determine the decision reliability interval based on the confidence coefficient α (set to 0.95). For the average risk level μ and standard deviation σ, the reliability interval is [μ-zα / 2·σ / ,μ+zα / 2·σ / ], where zα / 2 is the critical value of the standard normal distribution and n is the sample size. In this example, the calculated average risk level is 3.7, the standard deviation is 0.8, the sample size is 10, and the confidence level is 0.95, which corresponds to zα / 2 of approximately 1.96. Therefore, the decision reliability interval is [3.2, 4.2].

[0106] The decision basis set is optimized based on the decision reliability interval, outliers outside the interval are removed, and the weights of boundary values ​​are adjusted to generate the final risk assessment decision basis. For example, for a factor with a risk level of 2.0, since it is below the lower limit of the reliability interval of 3.2, its weight is reduced by 50%; for a factor with a risk level of 3.5, its original weight remains unchanged. The final risk assessment decision basis includes the optimized risk factor set, the impact relationship network, and corresponding risk control recommendations, providing security managers with a clear decision-making reference.

[0107] In an optional embodiment, intelligently evaluating the current wearing status of the operator based on the real-time hazard level, generating a dynamic wearing compliance score, and calculating a risk evolution trend value based on the dynamic wearing compliance score include:

[0108] Determining an assessment weight coefficient for each labor protection product based on the real-time hazard level, collecting the real-time spatial coordinates of the labor protection product, performing a difference calculation between the real-time spatial coordinates and the preset standard wearing position coordinates to obtain a deviation distance, and performing a weighted sum operation on the assessment weight coefficient and the deviation distance to obtain a position deviation assessment value;

[0109] Collecting the wearing integrity information of the labor protection products, performing exponential decay processing on the position deviation evaluation value, and performing weighted combination with the wearing integrity information, and generating a wearing standardization dynamic score through a dynamic scoring function;

[0110] Establishing an association mapping function for the dynamic score of wearing normativeness and the real-time danger level, calculating an interaction coefficient between the two according to the association mapping function, and combining the interaction coefficient with the dynamic score of wearing normativeness to generate a comprehensive evaluation index;

[0111] Construct a fixed-length time window sequence, input the dynamic scoring data of wearing norms within a historical time period into the time window sequence in chronological order, and use an attention calculation method to extract the time series feature weights in the time window sequence;

[0112] The temporal feature weight is predicted based on the dynamic wearing normativeness score, and a risk evolution trend value is calculated according to the prediction result, where the risk evolution trend value represents a changing trend of the wearing state over time.

[0113] The collected work scene data is used to determine the real-time hazard level, conduct an intelligent assessment of the current wearing status of the operator, generate a dynamic wearing compliance score, and calculate the risk evolution trend value.

[0114] The system intelligently assesses the operator's current wear status based on the real-time hazard level, using multi-sensor data fusion technology to collect working environment parameters, including temperature, humidity, hazardous gas concentration, noise level, and other environmental factors. For example, in a chemical plant environment, if the detected hazardous gas concentration reaches 25ppm (below the hazard threshold of 30ppm), the system will set the real-time hazard level to medium level 3 (assuming the hazard level range is 1-5).

[0115] Based on the determined real-time hazard level of 3, the system assigns different assessment weights to each labor protection item. For example, at this hazard level, the weight for gas masks is 0.45, the weight for hard hats is 0.25, the weight for protective gloves is 0.15, and the weight for safety shoes is 0.15. The system collects the real-time spatial coordinates of these labor protection items using positioning tags on wearable devices or camera-based visual recognition systems. The actual wearing position coordinates of the gas mask are (125, 180, 45), while the standard wearing position coordinates are (120, 175, 40), resulting in a calculated deviation distance of 8.7 units. Similarly, the deviation distances for other labor protection items are calculated: 5.2 units for hard hats, 12.3 units for protective gloves, and 3.8 units for safety shoes. These deviation distances are weighted and summed with the corresponding weight coefficients, resulting in an estimated position deviation value of 7.9 units.

[0116] The system collects wear integrity information for labor protection equipment and uses sensors and image recognition technology to determine the wear status of each item, such as whether the gas mask is properly sealed and whether the hard hat is securely fastened. In this example, the gas mask seal is detected to be 92%, the hard hat is secured at 98%, the protective gloves are worn completely at 85%, and the safety shoes are laced 100%. The system applies exponential decay to the position deviation assessment value, using a decay coefficient of 0.8, resulting in a position deviation impact value of 6.32 units after decay.

[0117] This value is weightedly combined with the wearing integrity information. Specifically, the integrity index of each labor protection product is multiplied by its weight and the sum is calculated. It is then comprehensively calculated with the attenuated position deviation impact value. The dynamic scoring function is used to generate a dynamic wearing standardization score of 85.4 points (out of 100 points).

[0118] A correlation mapping function was established between the dynamic wearing compliance score and the real-time hazard level. At hazard level 3, when the wearing compliance score was below 90, the interaction coefficient between the two was calculated to be 0.78. Combining this interaction coefficient with the dynamic wearing compliance score yielded a comprehensive evaluation index of 66.6, indicating that the worker's wearing compliance status presents a certain risk in the current hazardous environment.

[0119] To predict risk evolution trends, the system constructs a time window sequence of 10 lengths, storing dynamic dress compliance score data for the past 10 time points. Assume that the score data for the past 10 time points (each time point is 5 minutes apart) is: [87.2, 86.5, 85.8, 86.2, 85.7, 85.4, 84.9, 84.5, 85.0, 85.4]. The system uses an attention calculation method to extract the weights of temporal features in the time window sequence by calculating the importance weight of the data at each time point. For example, the weights for the most recent time points are: [0.07, 0.08, 0.09, 0.10, 0.11, 0.11, 0.12, 0.13, 0.14, 0.15], reflecting the greater contribution of recent data to the prediction.

[0120] Based on the extracted time-series feature weights and the current dynamic wear compliance score, the system predicts the score's evolutionary trend over the next 30 minutes. Using the time-series prediction algorithm, the system calculated a risk evolution trend value of -0.42, indicating an average decrease of 0.42 points in the wear compliance score every five minutes. This negative trend indicates a risk of continued deterioration in the worker's wear. The system will generate an early warning, advising the worker to adjust their PPE equipment, particularly their respirator and gloves.

[0121] The system dynamically adjusts assessment parameters based on different scenarios. For example, in high-temperature environments, the weight of heat-resistant clothing is automatically increased; in noisy environments, the weight of hearing protection is correspondingly increased. The system also considers individual differences among workers, such as the impact of body shape differences on the standard wearing position, and improves assessment accuracy through personalized parameter calibration. When the risk evolution trend value exceeds the preset threshold, the system triggers different levels of intervention measures, ranging from reminders and adjustments to recommended suspension of operations, to ensure worker safety.

[0122] Figure 3 The following is a bar chart comparing the performance of the dynamic evaluation method for wearing normativeness in different scenarios according to an embodiment of the present invention:

[0123] This figure compares the accuracy of three different assessment methods (traditional static scoring, temporal feature evaluation, and dynamic mapping fusion) across five different test scenarios. In high-risk work environments, the three methods achieved accuracy rates of 73.5%, 82.4%, and 91.8%, respectively. In complex dynamic scenarios, the accuracy rates reached 68.2%, 79.6%, and 88.5%. In long-term continuous work, the accuracy rates were 70.8%, 85.3%, and 93.2%. In multi-person collaborative scenarios, the accuracy rates were 65.3%, 76.8%, and 85.4%. In extreme weather conditions, the accuracy rates were 58.7%, 72.1%, and 81.9%, respectively. The data shows that the dynamic mapping fusion method achieved the best assessment results in all test scenarios, particularly in the long-term continuous work scenario, achieving the highest accuracy rate of 93.2%. In contrast, the traditional static scoring method performed the worst, achieving only 58.7% accuracy in extreme weather conditions. The performance of the temporal feature evaluation method lies somewhere in between the two. Although it is significantly better than the traditional method, it still fails to reach the level of the dynamic mapping fusion method. Overall, the dynamic mapping fusion method demonstrates stronger environmental adaptability and evaluation accuracy.

[0124] In an optional embodiment, a correlation mapping function is established between the dynamic score of wearing norms and the real-time danger level, an interaction coefficient between the two is calculated based on the correlation mapping function, and the interaction coefficient is combined with the dynamic score of wearing norms to generate a comprehensive evaluation index, including:

[0125] Using a deep neural network to extract features from the operator's physiological characteristic data to obtain a fatigue feature vector, setting weight coefficients according to the importance of different features, and performing a weighted summation of the fatigue feature vector and the weight coefficients to obtain a fatigue assessment value;

[0126] Inputting the fatigue assessment value into the deep neural network to calculate the fatigue influence coefficient, and performing correction calculation on the dynamic wearing standardization score and the fatigue influence coefficient to obtain a corrected wearing standardization score;

[0127] collecting temperature data, humidity data, and illumination data of the working environment, determining the influence weight of the environmental factors based on the fatigue assessment value, and performing a weighted combination of the temperature data, the humidity data, and the illumination data to generate an environmental impact assessment value;

[0128] fusing and correcting the environmental impact assessment value with the real-time hazard level to obtain a corrected hazard level;

[0129] Establishing an initial mapping relationship based on the revised wearing compliance score and the revised hazard level, and modifying the initial mapping relationship using the operator's attention level data as an adjustment factor to generate a human factor mapping function;

[0130] The modified wearing standardization score is weightedly combined with the human factor mapping function, and the result of the weighted combination is adjusted to generate a comprehensive evaluation index.

[0131] Real-time data from wearable devices is collected and an initial wear compliance score is calculated. This score is calculated based on three dimensions: position deviation, tightness, and wear integrity. Position deviation is determined by comparing the distance difference between the actual position of the wearable device and the standard position. For example, if the helmet is offset by more than 3 cm, the deviation score is 0.7. Tightness is calculated by measuring the pressure between the wearable device and the human body. For example, when the seat belt pressure value is within the range of 15N to 20N, the tightness score is 0.9. Wear integrity is evaluated based on the wearing status of the specified wearable devices. For example, if the helmet, goggles, seat belt, and other equipment are all correctly worn, the integrity score is 1.0. The system assigns weights of 0.3, 0.3, and 0.4 to these three dimensions, respectively, and calculates the initial wear compliance score through weighted summation.

[0132] Detects hazardous factors in the working environment, including the height of aerial work, the operating status of surrounding equipment, and the concentration of hazardous substances. For aerial work, the system sets a hazard level based on the working height, such as Level 1 for work below 10 meters, Level 2 for work between 10 and 20 meters, and Level 3 for work above 20 meters. For equipment operating status, the system monitors parameters such as vibration frequency and temperature. When the vibration frequency exceeds 40Hz or the temperature exceeds 80°C, the hazard level increases. For hazardous substance concentration, the system monitors toxic gas or dust concentrations in real time. For example, when the carbon monoxide concentration exceeds 50ppm, the hazard level increases. The system comprehensively considers these hazard factors and generates a real-time hazard level, which is divided into levels 1 to 5, with level 5 indicating the highest hazard.

[0133] The correlation mapping between the dynamic wear compliance score and the real-time danger level is achieved through a deep neural network. This network consists of an input layer, multiple hidden layers, and an output layer. The input layer receives the wear compliance score and real-time danger level data. The hidden layer contains 64 neurons and uses the ReLU activation function to handle feature extraction and nonlinear transformation. The output layer generates the interaction influence coefficient. For example, when the wear compliance score is 0.85 and the real-time danger level is level 3, the calculated interaction influence coefficient is 0.72.

[0134] To assess worker fatigue, the system collects physiological data such as heart rate, blink rate, and facial expressions. Heart rate data is collected through wearable devices. The normal heart rate range is 60-100 beats per minute; a rate outside this range indicates fatigue. Blink rate is detected using a facial recognition algorithm. The normal blink rate is 15-20 times per minute, and a decrease in this rate indicates increased fatigue. Facial expression analysis can identify signs of fatigue such as yawning and half-closed eyes.

[0135] A deep neural network was used to extract features from this data, generating a fatigue feature vector. Based on the importance of heart rate, blink rate, and facial expression, weights were set to 0.4, 0.3, and 0.3, respectively, and a weighted sum was taken to calculate the fatigue assessment value. For example, when the heart rate was 110 beats / minute, the blink rate was 10 times / minute, and the facial expression showed obvious fatigue, the calculated fatigue assessment value was 0.75.

[0136] The fatigue assessment value is input into the deep neural network to calculate the fatigue impact coefficient. The network uses a three-layer structure, including an input layer containing 32 neurons, a hidden layer containing 16 neurons, and an output layer. Through the forward propagation algorithm, when the fatigue assessment value is 0.75, the calculated fatigue impact coefficient is 0.68. The system corrects the dynamic wearing compliance score and the fatigue impact coefficient by multiplying the wearing compliance score by the fatigue impact coefficient. For example, when the initial wearing compliance score is 0.85 and the fatigue impact coefficient is 0.68, the corrected wearing compliance score is 0.578.

[0137] Environmental sensors collect data on the temperature, humidity, and illumination of the work environment. The temperature sensor measures ambient temperature, with an optimal operating temperature range of 18°C ​​to 28°C. The humidity sensor measures relative humidity, with an optimal operating humidity range of 40% to 60%. The illumination sensor measures ambient light intensity, with an optimal operating illumination range of 300 to 500 lux.

[0138] The impact weights of environmental factors are determined based on the fatigue assessment value. The higher the fatigue level, the greater the impact weights of environmental factors. When the fatigue assessment value is 0.75, the impact weights of temperature, humidity, and illumination are set to 0.5, 0.3, and 0.2, respectively. The system performs a weighted combination of temperature, humidity, and illumination data to generate an environmental impact assessment value. For example, when the temperature is 32°C, the humidity is 70%, and the illumination is 200 lux, the calculated environmental impact assessment value is 0.65.

[0139] The environmental impact assessment value is combined with the real-time hazard level to produce a revised hazard level. This fusion method multiplies the real-time hazard level by the weighting factor of the environmental impact assessment value. For a real-time hazard level of 3 and an environmental impact assessment value of 0.65, the revised hazard level is 3.65, rounded to 4.

[0140] An initial mapping relationship is established based on the revised conformity score and the revised hazard level. This mapping relationship is trained using a deep learning model, using conformity scores and hazard levels from a historical dataset as input and accident rates as output labels. Once trained, the model is able to predict potential risks based on the conformity scores and hazard levels input.

[0141] The human factor mapping function is a functional relationship formed by modifying the initial mapping relationship by incorporating human factors such as worker cognition, behavior, and emotion. It reflects the complex relationship between a worker's internal psychological state and external safety behavior, enabling more accurate prediction of changing safety risk trends under varying human factors. The human factor mapping function is implemented using a multi-layer perceptron architecture, consisting of three hidden layers, with 32, 16, and 8 neurons in each layer, respectively. A sigmoid activation function is used to capture the nonlinear relationship between human factors and safety risks.

[0142] This function uses the worker's attention level data as a modulating factor to modify the initial mapping relationship. This attention level data is collected using eye tracking technology, which uses specialized cameras to capture 50 frames of eye movement data per second, recording the frequency and distribution of changes in the worker's gaze. The system analyzes this data to assess attentional focus. Continuous gaze on the same area for more than 1.5 seconds is considered focused; attention is considered distracted if the gaze jumps more than 10 times within 3 seconds.

[0143] The attention level is quantified as a value between 0 and 1, with 1 indicating complete concentration and 0 indicating complete distraction. For example, when the attention level is 0.6, the human factor mapping function adjusts the weight distribution in the initial mapping relationship to enhance the influence of attention factors on safety risk prediction, thereby generating a revised mapping that takes attention factors into account. Specifically, when the attention level is above 0.8, the human factor mapping function reduces the weight of the hazard level in the risk prediction; when the attention level is below 0.4, the hazard level weight is increased to reflect the additional risk caused by inattention.

[0144] The revised wear compliance score is weightedly combined with the human factors mapping function. This combination is performed by substituting the revised wear compliance score into the human factors mapping function to obtain an intermediate value. This intermediate value is then weighted and summed with the revised wear compliance score in a ratio of 7:3. The weighted combination result is adjusted by setting an adjustment coefficient based on the complexity of the work environment. In simple environments, the adjustment coefficient is 0.9; in complex environments, the adjustment coefficient is 1.1.

[0145] The weighted combination result is multiplied by the adjustment coefficient to generate the final comprehensive evaluation index. This index ranges from 0 to 100 and is divided into five levels: 0-20 is extremely high risk, 21-40 is high risk, 41-60 is medium risk, 61-80 is low risk, and 81-100 is very low risk. For example, when the corrected wearing compliance score is 0.578, the intermediate value obtained by substituting it into the human factors mapping function is 0.65. After weighting it at a ratio of 7:3, it is 0.6006. Multiplying it by the adjustment coefficient of 1.1 in a complex environment, the final comprehensive evaluation index is 66.07, which is a low risk level.

[0146] In an optional embodiment, based on the dynamic score of wearing norms and the risk evolution trend value, a deep reinforcement learning algorithm is used to generate an optimal control strategy, and the corresponding control measures are executed, including:

[0147] Constructing the wearing normative dynamic score into an N-dimensional state score vector, constructing the risk evolution trend value into an M-dimensional trend feature vector, and combining the N-dimensional state score vector and the M-dimensional trend feature vector to form a state input;

[0148] A deep Q network is constructed using a deep reinforcement learning algorithm. The state input is used as the network input, and the control measures are used as the network output. The improvement of the dynamic score of the wearing standardization and the decrease of the risk evolution trend value are constructed as reward signals. The target Q value is calculated based on the reward signal.

[0149] Passing the state input into the deep Q-network to generate an action value estimate, calculating a loss function between the action value estimate and the target Q value using a temporal difference algorithm, and optimizing the parameters of the deep Q-network based on the loss function;

[0150] An ε-greedy strategy is used to select an optimal control strategy from the action value estimates output by the deep Q network, and control measures corresponding to the optimal control strategy are executed.

[0151] like Figure 4 As shown, the method includes:

[0152] The dynamic scoring of wearing compliance is constructed as an N-dimensional state score vector. For example, for a safety helmet wearing scenario, the state score vector can be set to 5 dimensions, representing the scores of helmet positioning correctness, chin strap tightness, helmet integrity, helmet cleanliness, and wearing stability. The score range for each dimension is 0-100, where 0 indicates complete non-compliance and 100 indicates complete compliance. For example, if a worker's helmet positioning correctness score is 85, chin strap tightness score is 70, helmet integrity score is 90, helmet cleanliness score is 60, and wearing stability score is 75, then the worker's state score vector is [85, 70, 90, 60, 75].

[0153] The risk evolution trend value is constructed as an M-dimensional trend feature vector. In this embodiment, M can be set to 3, corresponding to the short-term risk trend, medium-term risk trend, and long-term risk trend respectively. The value range of each trend is -1 to 1, where -1 indicates a significant decrease in risk, 0 indicates stable risk, and 1 indicates a significant increase in risk. For example, if a worker's short-term risk trend is 0.3 (slight increase), the medium-term risk trend is 0.1 (basically stable), and the long-term risk trend is -0.2 (slight decrease), then the trend feature vector of the worker is [0.3, 0.1, -0.2].

[0154] The N-dimensional state score vector and the M-dimensional trend feature vector are combined to form the state input. In this embodiment, the state input is a vector of dimension (N + M), which is the direct concatenation of the two vectors. For the above example, the state input is [85, 70, 90, 60, 75, 0.3, 0.1, -0.2], with a dimension of 8.

[0155] A deep Q-network is constructed using a deep reinforcement learning algorithm. In this embodiment, the deep Q-network consists of three fully connected layers. The first layer has 64 neurons and uses the ReLU activation function; the second layer has 32 neurons, also using the ReLU activation function. The number of neurons in the output layer is equal to the number of selectable control measures, K. For example, K can be set to 4, corresponding to the four control measures: "verbal reminder," "send safety warning SMS," "arrange safety training," and "suspend work permissions."

[0156] The network takes the state input as input and the control measures as output. The output of the Deep Q Network is a K-dimensional vector, where each element represents the estimated value of taking the corresponding control measure under the current state. For example, for the state input [85, 70, 90, 60, 75, 0.3, 0.1, -0.2], the Deep Q Network outputs [5.2, 7.8, 3.1, 1.5], indicating that sending a safety warning text message (value 7.8) is the most valuable control measure under the current state.

[0157] The reward signal is constructed based on the increase in the dynamic dress compliance score and the decrease in the risk evolution trend value. Specifically, the reward signal is calculated as: Reward = Increase in Dress Compliance Score × 0.6 + Decrease in Risk Trend × 0.4. For example, if, after implementing a certain control measure, the dress compliance score increases from 76 to 82 (an increase of 6 points) and the risk trend value decreases from 0.3 to 0.1 (a decrease of 0.2), the reward signal = 6 × 0.6 + 0.2 × 0.4 = 3.6 + 0.08 = 3.68.

[0158] The target Q-value is calculated based on the reward signal. In this example, the target Q-value is calculated by adding the current reward to the maximum Q-value in the next state. The discount factor is set to 0.9. For example, if the current reward is 3.68 and the maximum Q-value in the next state is 8.5, then the target Q-value = 3.68 + 0.9 × 8.5 = 11.33.

[0159] The state input is passed to the deep Q-network to generate an action-value estimate. A temporal difference algorithm is used to calculate the loss function between the action-value estimate and the target Q-value. In this example, the loss function uses mean squared error. For example, if the action-value estimate for a state-action pair is 9.2 and the target Q-value is 11.33, then the loss value is (9.2 - 11.33)² = 4.55.

[0160] The parameters of the deep Q-network are optimized based on the loss function. In this example, the Adam optimizer is used, with a learning rate of 0.001 and a batch size of 32. During each training cycle, 32 samples are randomly sampled from the experience replay buffer for training. The experience replay buffer is set to 10,000 and stores past states, actions, rewards, and next states. For example, after 1,000 training iterations, the action-value estimate is closer to the target Q-value, and the loss value drops below 0.5.

[0161] An ε-greedy strategy is used to select the optimal control strategy from the action-value estimates output by the deep Q-network. In this example, ε is initially set to 1.0 and decays gradually to 0.01 as training progresses, with a decay rate of 0.995. In practice, for example, when ε is 0.05, there is a 95% probability of selecting the action with the highest Q-value, and a 5% probability of selecting a random action. In the above example, the deep Q-network outputs [5.2, 7.8, 3.1, 1.5]. The action with the highest Q-value corresponds to "Send a safety warning text message," so there is a 95% probability of selecting this control measure.

[0162] Execute the control measures corresponding to the optimal control strategy. The system automatically records the time, target, and type of control measures implemented to facilitate subsequent evaluation of their effectiveness. For example, the system may record, "At 10:30 AM on X / X / 2023, the 'Send Safety Warning SMS' control measure was executed on Worker A, reminding him to adjust the position of his helmet and the tightness of his chin strap." During the next evaluation cycle, the system will recalculate the worker's dynamic wear compliance score and risk evolution trend value to assess the effectiveness of the control measure and determine whether further control actions are necessary.

[0163] In an optional embodiment, the state input is passed into the deep Q network to generate an action value estimate, a temporal difference algorithm is used to calculate a loss function between the action value estimate and the target Q value, and the parameters of the deep Q network are optimized based on the loss function, including:

[0164] The state input is passed into a deep Q network having a multi-layer neural network structure, and the value estimate of each control action in the current state is calculated based on the preset network parameters of the deep Q network, and the control action with the largest value estimate is selected as the execution action;

[0165] Execute the control action and collect an immediate reward value, generate a new state input vector based on the immediate reward value, pass the new state input vector into the deep Q network to calculate the maximum action value estimate at the next moment, add the immediate reward value to the discounted value of the maximum action value estimate to obtain a target Q value;

[0166] Calculate the mean square error between the maximum action value estimate and the target Q value using a temporal difference algorithm, construct the mean square error as a loss function, and calculate the gradient value of the loss function with respect to the preset network parameters;

[0167] Multiplying the gradient value by a preset learning rate to obtain a parameter adjustment amount, updating and optimizing the preset network parameters based on the parameter adjustment amount, generating optimized network parameters, and using the optimized network parameters to update the deep Q network.

[0168] A deep reinforcement learning control system is used to optimize network management strategies. The system feeds state inputs into a deep Q-network, generates action-value estimates, and calculates a loss function using a temporal difference algorithm to optimize the parameters of the deep Q-network.

[0169] The Deep Q Network uses a multi-layer neural network structure, including an input layer, multiple hidden layers, and an output layer. The input layer receives a vector representing the current network state, which may include features such as network traffic, congestion, and link utilization. In this embodiment, the state vector dimension is 32 and includes 15 network performance indicators and 17 user experience indicators. The hidden layer consists of three fully connected layers, with 128, 256, and 128 neurons respectively, and the activation function uses the ReLU function. The number of neurons in the output layer is equal to the number of optional control actions, which is 8 in this example, corresponding to different traffic scheduling strategies.

[0170] After the system receives the state input vector, it propagates it forward through a deep Q-network. For example, in one iteration, the input state vector s contains metrics such as an average network latency of 55ms, a packet loss rate of 0.2%, and a link utilization of 65%. This state vector is calculated through each layer of the deep Q-network, and the final output layer generates estimated Q values ​​for eight different control actions: [2.3, 1.7, 3.5, 2.1, 2.8, 2.6, 2.0, 3.2]. The system selects the action with the highest estimated value, the third action with a Q value of 3.5, as action a. This action corresponds to the priority queue scheduling strategy, allocating more bandwidth to critical business traffic.

[0171] After executing the selected control action, the system observes environmental changes and collects an immediate reward value. The reward function is designed based on network performance metrics, incorporating factors such as latency reduction, throughput improvement, and user experience improvement. In this example, after executing the action, the average network latency decreased to 45ms, the packet loss rate dropped to 0.1%, and user satisfaction increased by 10%. The calculated immediate reward value, r, is 4.2.

[0172] The system generates a new state s' after executing the action, which contains the updated network metrics. This new state vector is fed back into the deep Q-network to calculate the value estimates for each action at the next moment. For example, the value estimates for the eight actions in the new state are [2.5, 2.0, 3.7, 2.3, 3.9, 2.8, 2.2, 3.0]. The system selects the action with the highest value estimate, 3.9, which is the fifth action.

[0173] Based on the principles of temporal difference learning, the system calculates a target Q-value. The target Q-value is the sum of the immediate reward and the discounted estimate of the maximum action value in the next state. The discount factor γ is set to 0.95 to indicate the emphasis on future rewards. For the above example, the target Q-value is calculated as: the immediate reward of 4.2 plus the product of the maximum Q-value of the next state of 3.9 and the discount factor of 0.95, resulting in a target Q-value of 7.905.

[0174] The loss function is calculated using the temporal difference algorithm, which is the mean squared error between the estimated value Q(s,a) of executing action a in the current state and the target Q value. In this example, the estimated value of executing action a (the third action) is 3.5, and the target Q value is 7.905. The mean squared error is calculated as (3.5-7.905). 2 is equal to 19.3506. This error constitutes the loss function of the deep Q network.

[0175] Based on the loss function, the system uses a backpropagation algorithm to calculate the gradient values ​​of the preset network parameters. The deep Q network contains the weight matrix W1 (32×128) from the input layer to the first hidden layer, the weight matrix W2 (128×256) from the first hidden layer to the second hidden layer, the weight matrix W3 (256×128) from the second hidden layer to the third hidden layer, and the weight matrix W4 (128×8) from the third hidden layer to the output layer, as well as the corresponding bias vectors b1, b2, b3, and b4. The system calculates the gradient of the loss function with respect to these parameters, obtaining the gradient matrices ∇W1, ∇W2, ∇W3, and ∇W4 and the gradient vectors ∇b1, ∇b2, ∇b3, and ∇b4.

[0176] Set the learning rate α to 0.001 and multiply the gradient by the learning rate to get the parameter adjustment. For example, the adjustment for weight matrix W1 is α·∇W1, and the other parameters are similar. The system updates the network parameters based on the adjustment, with W1 updated to W1-α·∇W1. The other parameters are also updated accordingly to generate the optimized network parameters.

[0177] To improve training stability, the system employs an experience replay mechanism. The transfer samples (s, a, r, s') generated by each interaction are stored in an experience replay pool with a capacity of 10,000. During training, the system randomly samples small batches of data of size 64 for parameter updates. Furthermore, the system updates the target network parameters every 100 iterations, reducing the correlation of Q-value estimates and improving training stability.

[0178] Through multiple iterations of training, the Deep Q Network gradually optimizes and learns the optimal control strategy for different network states. Experimental results show that compared to traditional methods, the network control system based on deep reinforcement learning reduces average latency by 23.5%, increases throughput by 18.7%, and improves user experience satisfaction by 32.1%, effectively improving network performance.

[0179] Figure 5 This is a bar chart comparing the performance of the Deep Q network in different test scenarios according to an embodiment of the present invention:

[0180] The figure shows a performance comparison of three different methods (standard deep Q-network, optimized temporal difference algorithm, and adaptive parameter adjustment) in four test scenarios. In complex dynamic environments, the three methods achieved efficiencies of 78.5%, 85.7%, and 92.3%, respectively; in multi-task collaboration, the performance reached 65.2%, 79.4%, and 86.8%; in high-noise environments, the performance reached 72.8%, 77.5%, and 83.1%; and in real-time response scenarios, the efficiency reached 81.3%, 89.2%, and 94.5%, respectively. The data shows that the adaptive parameter adjustment method achieved the best performance in all test scenarios, generally outperforming the other two methods. This method's advantage was particularly pronounced in complex dynamic environments and real-time response scenarios, with efficiencies reaching 92.3% and 94.5%, respectively. In comparison, the standard deep Q-network performed relatively poorly, achieving only 65.2% efficiency in the multi-task collaboration scenario. The performance of the optimized temporal difference algorithm is between the two, but it is still significantly improved compared to the standard method, achieving a high efficiency of 89.2% in real-time response scenarios.

[0181] A second aspect of an embodiment of the present invention provides a system for detecting the wearing of labor protection equipment by workers at oil and gas stations, comprising:

[0182] The first unit is used to collect real-time image data of workers through image acquisition equipment deployed at oil and gas stations, call a deep learning model to detect the wearing of labor protection products by the workers' image data, and obtain characteristic data of the workers' wearing status;

[0183] The second unit is configured to establish a correlation analysis based on the wearing status characteristic data, combined with the process parameters, equipment status and hazard source distribution data of the operating area, to generate a real-time hazard level for the area where the operator is located; perform an intelligent assessment of the operator's current wearing status based on the real-time hazard level, generate a dynamic wearing compliance score, and calculate a risk evolution trend value based on the dynamic wearing compliance score;

[0184] The third unit is configured to generate an optimal control strategy using a deep reinforcement learning algorithm based on the dynamic wearing standardization score and the risk evolution trend value, and execute corresponding control measures;

[0185] The fourth unit is used to store the wearing status characteristic data, the real-time danger level, the dynamic wearing standardization score and the control measures in a database, generate an analysis report on the workers' wearing behavior of labor protection products, and continuously optimize the wearing detection parameters and hierarchical processing strategies based on the analysis report.

[0186] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0187] processor;

[0188] a memory for storing processor-executable instructions;

[0189] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0190] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0191] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0192] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting the wearing of labor protection equipment by workers at oil and gas stations, characterized in that: include: Image acquisition equipment deployed at oil and gas stations collects worker image data in real time, and uses a deep learning model to detect the wearing of labor protection equipment on the worker image data to obtain characteristic data on the worker's wearing status. Based on the wearing status characteristic data, combined with the process parameters, equipment status and hazard source distribution data of the operating area, a correlation analysis is established to generate a real-time hazard level for the area where the operator is located; based on the real-time hazard level, an intelligent assessment is performed on the operator's current wearing status to generate a dynamic wearing standardization score, and a risk evolution trend value is calculated based on the dynamic wearing standardization score, including: Determining an assessment weight coefficient for each labor protection product based on the real-time hazard level, collecting the real-time spatial coordinates of the labor protection product, performing a difference calculation between the real-time spatial coordinates and the preset standard wearing position coordinates to obtain a deviation distance, and performing a weighted sum operation on the assessment weight coefficient and the deviation distance to obtain a position deviation assessment value; Collecting the wearing integrity information of the labor protection products, performing exponential decay processing on the position deviation evaluation value, and performing weighted combination with the wearing integrity information, and generating a wearing standardization dynamic score through a dynamic scoring function; Establishing an association mapping function for the dynamic score of wearing normativeness and the real-time danger level, calculating an interaction coefficient between the two according to the association mapping function, and combining the interaction coefficient with the dynamic score of wearing normativeness to generate a comprehensive evaluation index; Construct a fixed-length time window sequence, input the dynamic scoring data of wearing norms within a historical time period into the time window sequence in chronological order, and use an attention calculation method to extract the time series feature weights in the time window sequence; Predicting the time series feature weight based on the dynamic score of wearing norms, and calculating a risk evolution trend value according to the prediction result, wherein the risk evolution trend value represents a trend of change of the wearing state over time; Based on the dynamic score of wearing norms and the risk evolution trend value, a deep reinforcement learning algorithm is used to generate an optimal control strategy and implement corresponding control measures; The wearing status characteristic data, the real-time danger level, the dynamic wearing normative score and the control measures are stored in a database, an analysis report on the workers' wearing behavior of labor protection products is generated, and the wearing detection parameters and the graded processing strategy are continuously optimized based on the analysis report.

2. The method according to claim 1, characterized in that Based on the wear status characteristic data, combined with the process parameters, equipment status and hazard source distribution data of the work area, a correlation analysis is established to generate the real-time hazard level of the area where the operator is located, including: Extracting wearing position features, wearing standardization features, and wearing integrity features based on the wearing state feature data to construct a wearing state feature vector; based on the wearing state feature vector, collecting process parameter data related to the position of the operator, and constructing the process parameter data into a process parameter spatiotemporal tensor; According to the location of the operator, the operating status data of the relevant equipment in the operation area is obtained to construct an equipment status characteristic matrix; based on the location of the operator and the equipment status characteristic matrix, the distribution data of the corresponding hazard sources in the operation area is collected to establish a hazard source impact function; Performing a tensor outer product operation on the wear state feature vector, the process parameter spatiotemporal tensor, the equipment state feature matrix, and the hazard source influence function to construct a spatiotemporal coupling tensor; inputting the spatiotemporal coupling tensor into a transformer architecture, and utilizing a self-attention mechanism to extract spatiotemporal correlation features between multi-source data to generate a feature correlation matrix; Based on the spatiotemporal coupling tensor and the characteristic association matrix, a hazard level assessment value of the operating area is calculated, and a historical hazardous event case library is retrieved according to the hazard level assessment value, and the similarity with the historical cases is calculated to generate a decision basis for risk assessment; based on the hazard level assessment value and the decision basis, combined with the preset hazard level classification standard, a real-time hazard level of the area where the operating personnel is located is generated.

3. The method according to claim 2, characterized in that Based on the spatiotemporal coupling tensor and the characteristic association matrix, a hazard level assessment value of the operation area is calculated. According to the hazard level assessment value, a historical hazard event case library is retrieved, and similarity with historical cases is calculated to generate a decision basis for risk assessment, including: The spatiotemporal coupling tensor and the feature association matrix are feature-concatenated to generate a fused feature vector; a hazard level assessment function is constructed based on the fused feature vector, the spatiotemporal coupling tensor and the feature association matrix are weighted by a weight matrix, and an initial hazard level assessment value is obtained by performing activation function processing; Constructing a dynamic knowledge graph based on the initial risk level assessment value, inputting the initial risk level assessment value into the dynamic knowledge graph, extracting graph structure features through a graph convolutional network, and generating historical case retrieval features; Calculating the cosine similarity between the initial risk level assessment value and the historical case retrieval feature to obtain feature-level similarity; optimizing the calculation result of the feature-level similarity using a contrastive learning method, and adjusting the contrast relationship between positive and negative samples using a temperature parameter to improve the accuracy of the similarity calculation; Based on the feature-level similarity, similar historical cases are screened to construct a causal decision diagram; based on the causal decision diagram, a decision basis set is generated, wherein the decision basis set includes risk factors, impact relationships, and probability distribution information; Calculate the statistical characteristics of the decision basis set and determine the decision reliability interval based on the confidence coefficient; optimize the decision basis set according to the decision reliability interval to generate the final risk assessment decision basis.

4. The method according to claim 1, wherein Establishing an association mapping function for the dynamic score of wearing norms and the real-time danger level, calculating the interaction coefficient between the two according to the association mapping function, and combining the interaction coefficient with the dynamic score of wearing norms to generate a comprehensive evaluation index includes: Using a deep neural network to extract features from the operator's physiological characteristic data to obtain a fatigue feature vector, setting weight coefficients according to the importance of different features, and performing a weighted summation of the fatigue feature vector and the weight coefficients to obtain a fatigue assessment value; Inputting the fatigue assessment value into the deep neural network to calculate the fatigue influence coefficient, and performing correction calculation on the dynamic wearing standardization score and the fatigue influence coefficient to obtain a corrected wearing standardization score; collecting temperature data, humidity data, and illumination data of the working environment, determining the influence weight of the environmental factors based on the fatigue assessment value, and performing a weighted combination of the temperature data, the humidity data, and the illumination data to generate an environmental impact assessment value; fusing and correcting the environmental impact assessment value with the real-time hazard level to obtain a corrected hazard level; Establishing an initial mapping relationship based on the revised wearing compliance score and the revised hazard level, and modifying the initial mapping relationship using the operator's attention level data as an adjustment factor to generate a human factor mapping function; The modified wearing standardization score is weightedly combined with the human factor mapping function, and the result of the weighted combination is adjusted to generate a comprehensive evaluation index.

5. The method according to claim 1, wherein Based on the dynamic score of wearing norms and the risk evolution trend value, a deep reinforcement learning algorithm is used to generate an optimal control strategy, and corresponding control measures are implemented, including: Constructing the wearing normative dynamic score into an N-dimensional state score vector, constructing the risk evolution trend value into an M-dimensional trend feature vector, and combining the N-dimensional state score vector and the M-dimensional trend feature vector to form a state input; A deep Q network is constructed using a deep reinforcement learning algorithm. The state input is used as the network input, and the control measures are used as the network output. The improvement of the dynamic score of the wearing standardization and the decrease of the risk evolution trend value are constructed as reward signals. The target Q value is calculated based on the reward signal. Passing the state input into the deep Q-network to generate an action value estimate, calculating a loss function between the action value estimate and the target Q value using a temporal difference algorithm, and optimizing the parameters of the deep Q-network based on the loss function; An ε-greedy strategy is used to select an optimal control strategy from the action value estimates output by the deep Q network, and control measures corresponding to the optimal control strategy are executed.

6. The method according to claim 5, characterized in that The state input is passed into the deep Q network to generate an action value estimate, a temporal difference algorithm is used to calculate a loss function between the action value estimate and the target Q value, and the parameters of the deep Q network are optimized based on the loss function, including: The state input is passed into a deep Q network having a multi-layer neural network structure, and the value estimate of each control action in the current state is calculated based on the preset network parameters of the deep Q network, and the control action with the largest value estimate is selected as the execution action; Execute the control action and collect an immediate reward value, generate a new state input vector based on the immediate reward value, pass the new state input vector into the deep Q network to calculate the maximum action value estimate at the next moment, add the immediate reward value to the discounted value of the maximum action value estimate to obtain a target Q value; Calculate the mean square error between the maximum action value estimate and the target Q value using a temporal difference algorithm, construct the mean square error as a loss function, and calculate the gradient value of the loss function with respect to the preset network parameters; Multiplying the gradient value by a preset learning rate to obtain a parameter adjustment amount, updating and optimizing the preset network parameters based on the parameter adjustment amount, generating optimized network parameters, and using the optimized network parameters to update the deep Q network.

7. A labor protection equipment wearing detection system for oil and gas station workers, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to collect real-time image data of workers through image acquisition equipment deployed at oil and gas stations, call a deep learning model to detect the wearing of labor protection products by the workers' image data, and obtain characteristic data of the workers' wearing status; The second unit is configured to establish a correlation analysis based on the wearing status characteristic data, combined with the process parameters, equipment status and hazard source distribution data of the operating area, to generate a real-time hazard level for the area where the operator is located; perform an intelligent assessment of the operator's current wearing status based on the real-time hazard level, generate a dynamic wearing compliance score, and calculate a risk evolution trend value based on the dynamic wearing compliance score; The third unit is configured to generate an optimal control strategy using a deep reinforcement learning algorithm based on the dynamic wearing standardization score and the risk evolution trend value, and execute corresponding control measures; The fourth unit is used to store the wearing status characteristic data, the real-time danger level, the dynamic wearing standardization score and the control measures in a database, generate an analysis report on the workers' wearing behavior of labor protection products, and continuously optimize the wearing detection parameters and hierarchical processing strategies based on the analysis report.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images