Chemical Park Illegal Behavior Intervention and Tracking System Based on Edge Computing

Abstract

The present invention provides a chemical park violation intervention and tracking and tracing system based on edge computing, aiming to solve the problems of single perception capability, insufficient intelligent analysis, limited edge collaboration, passive intervention and weak tracing capability of existing systems. The system adopts improved ResNet‑FPN and Transformer architecture to realize multimodal data fusion, and deeply combines chemical process knowledge with dual-stream intent understanding framework and graph neural network to analyze violations; adopts adaptive edge node networking and federated learning mechanism to realize dynamic scheduling of computing resources and knowledge migration; the intervention decision model based on risk entropy supports predictive intervention, and the multi-objective optimization algorithm ensures that the intervention strategy balances safety and efficiency; adopts multimodal trajectory reconstruction technology and distributed evidence storage network to build a complete chain of evidence to realize the credible tracing of violations. The present invention breaks through the technical bottlenecks of traditional systems in professional scene understanding and real-time processing, and significantly improves the intelligent level of safety management in chemical parks.

Description

Technical Field

[0001] The present invention relates to the field of industrial safety and intelligent control technology, and in particular to a chemical park illegal behavior intervention and tracking and tracing system based on edge computing. Background Art

[0002] Chemical parks are core places for the production, storage and use of hazardous chemicals. Safety management is crucial for accident prevention and emergency response. The existing violation monitoring system faces the following technical bottlenecks under complex working conditions:

[0003] A) Single perception capability: Traditional systems mainly rely on single video surveillance and lack the ability to understand the intentions of operators, making it difficult to predict potential violations.

[0004] B) Insufficient intelligent analysis: The existing system lacks a deep understanding of chemical process knowledge and cannot accurately identify illegal operations in professional scenarios, resulting in high false alarm rates and high missed alarm rates.

[0005] C) Limited edge collaboration: With a centralized architecture, computing resources cannot be dynamically adjusted according to the complexity of the scenario, affecting the real-time performance and scalability of the system.

[0006] D) Passive intervention methods: lack of predictive intervention capabilities, measures can only be taken after violations occur, and early warning and proactive prevention of risks cannot be achieved.

[0007] E) Weak traceability: Evidence collection is scattered and storage is fragmented, making it difficult to establish a complete chain of evidence for violations, affecting the effectiveness of accountability and improvement.

[0008] Therefore, there is an urgent need for an intelligent intervention system based on edge computing, combining multimodal perception, trusted evidence and dynamic intervention technology to improve the ability to quickly identify, accurately trace and intelligently handle violations. Summary of the invention

[0009] The present invention provides a chemical park violation intervention and tracking and tracing system based on edge computing, which aims to solve the technical problems of single perception, insufficient analysis, limited coordination, passive intervention, and weak tracing in the existing park security system. The system includes the following components:

[0010] 1. Multimodal perception unit

[0011] The multimodal perception unit is responsible for achieving all-round perception and preliminary analysis of violations. It integrates a multi-source heterogeneous perception network of high-definition video cameras, infrared thermal imagers, millimeter-wave radars and gas concentration sensors to collect multi-dimensional perception data. It uses an improved residual network-feature pyramid network (ResNet-FPN) to extract features, and improves the ability to capture key features in dynamic scenes through hierarchical feature enhancement and temporal correlation encoding. It is equipped with a cross-modal feature alignment network based on the Transformer architecture, and uses hierarchical attention mechanism and spatial position encoding to achieve alignment and fusion of multi-source data. The fused feature data is transmitted to the edge intelligent processing unit for real-time analysis, and the original data is distributed to the traceability and evidence storage unit for evidence storage and reference.

[0012] 2. Edge Intelligent Processing Unit

[0013] The edge intelligent processing unit is the core of the system, responsible for the intention understanding and intelligent analysis of illegal behaviors. This unit adopts a dual-stream intention understanding framework to analyze illegal intentions from two dimensions: behavior sequence and environmental context. Among them, the behavior sequence stream processing network uses a bidirectional GRU network to extract the temporal evolution characteristics of the behavior, and the environmental context stream processing network constructs the scene state vector, analyzes the correlation characteristics between environmental elements, and dynamically integrates the behavior sequence characteristics and environmental state characteristics through an adaptive fusion function; based on the complexity of the scene, an adaptive edge node networking method is constructed to form a distributed collaborative computing network to support dynamic scheduling and balanced allocation of computing resources; a federated learning framework with differential privacy protection is used to realize knowledge transfer and sharing between edge nodes, while ensuring data privacy while improving the system's model performance and overall analysis efficiency.

[0014] 3. Process intelligence analysis unit

[0015] The process intelligent analysis unit provides the system with chemical professional knowledge support and violation analysis capabilities, constructs a chemical process knowledge graph, and systematically describes the process flow, operating specifications, and safety constraints. It uses the reasoning mechanism of graph neural networks to deeply integrate real-time monitoring data with the knowledge graph to achieve accurate identification of illegal operations in complex scenarios. It constructs an intention-behavior probability graph model, combines it with a historical case library, and quantitatively evaluates the probability of violations, providing decision support for predictive intervention.

[0016] 4. Intervention control unit

[0017] The intervention control unit realizes intelligent intervention in illegal behaviors, adopts an intervention decision-making model based on risk entropy, combines the illegal intention identification results and process analysis output, and evaluates the optimal intervention time; generates intervention strategies through multi-objective optimization algorithms to minimize the impact on normal production while ensuring safety; establishes a distributed execution scheduling mechanism to coordinate the linkage response of various intervention equipment, and evaluates the intervention effect through real-time feedback to support dynamic optimization of strategies.

[0018] 5. Traceability evidence unit

[0019] The traceability and evidence reliability of the system are ensured by the traceability and evidence storage unit. Multimodal behavior trajectory reconstruction technology is used to associate discrete violation behavior fragments to form a complete behavior chain. The triggering conditions and evolution paths of violation behaviors are analyzed through the causal reasoning mechanism of the knowledge graph. A distributed evidence storage network is established to construct a multi-level evidence chain to achieve reliable data storage and efficient retrieval.

[0020] Beneficial effects of the present invention:

[0021] 1. Deepen perception and understanding: By combining multimodal perception and intent understanding framework, the system's ability to understand and predict violations can be improved.

[0022] 2. Enhanced intelligent analysis: Combine chemical process knowledge graph and graph neural network reasoning to achieve accurate identification of illegal operations in professional scenarios.

[0023] 3. Optimize edge collaboration: Improve the real-time performance and scalability of the system through adaptive edge computing architecture and federated learning framework.

[0024] 4. Improve intervention effectiveness: Adopt a predictive intervention decision-making model to achieve risk warning and proactive prevention, and reduce the probability of accidents.

[0025] 5. Enhanced traceability: Through the combination of multimodal trajectory reconstruction and distributed evidence storage network, it supports all-round traceability and responsibility determination of violations. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 The system architecture diagram of the present invention is shown;

[0027] Figure 2 The system working flow chart of the present invention is shown. DETAILED DESCRIPTION

[0028] Exemplary embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0029] Combination Figure 1 , the present invention provides a chemical park violation intervention and tracking and tracing system based on edge computing, which is composed of the following:

[0030] 1. Multimodal perception unit

[0031] The multimodal perception unit is used to achieve all-round perception and preliminary analysis of violations in chemical parks. The unit constructs a multi-source heterogeneous perception network, extracts multimodal features through an improved ResNet-FPN network, and uses a Transformer-based feature alignment network to complete multi-source feature fusion. The specific implementation is as follows:

[0032] A. Construction of multi-source heterogeneous perception network

[0033] The present invention constructs a multi-source heterogeneous sensing network in key areas of the chemical park to achieve collaborative collection of multi-dimensional data. , synchronous data acquisition is performed at each sampling time t∈T.

[0034] a) For visible light image data acquisition, the mathematical expression is:

[0035]

[0036] in, Represents the pixel value at the image coordinate (r, c) at time t; and Respectively represent the row and column coordinate sets of the image, , , H and W represent the image height and width respectively; ∈ means belonging to; represents a subset; N represents a natural number set. This data provides a visual feature representation of the scene.

[0037] b) For infrared thermal imaging data acquisition, the mathematical expression is:

[0038]

[0039] in, Represents the infrared radiation intensity value at the coordinate (r, c) at time t; and Respectively represent the row and column coordinate sets of the infrared image, , , and Respectively represent the height and width of the infrared image. This data provides the thermal radiation distribution characteristics of the target.

[0040] c) For millimeter wave radar data acquisition, the mathematical expression is:

[0041]

[0042] in, Represents the eigenvector of the i-th point; Represents the three-dimensional rectangular coordinate value of the i-th point in the point cloud; represents the velocity component of the point along the radial direction; Represents a point cloud index set, , Represents the total number of point clouds collected at time t. This data provides the spatial position and motion characteristics of the target.

[0043] d) For gas concentration data collection, the mathematical expression is:

[0044]

[0045] in, Indicates the concentration value of the kth gas at time t; k is the index number of the gas type; K is the total number of gas types monitored by the system. This data provides the gas concentration distribution characteristics in the area.

[0046] B. Multimodal feature extraction

[0047] The present invention adopts an improved ResNet-FPN network structure for feature extraction, introduces hierarchical feature enhancement and temporal correlation encoding mechanism, and specifically includes the following steps:

[0048] a) Improved residual network feature extraction

[0049] For input features , the residual mapping expression is:

[0050]

[0051] in, Represents the feature map output of the l-th layer network; represents the residual mapping function; Represents the set of learnable parameters of the l-th layer network; the specific form of the residual mapping function is:

[0052]

[0053] in, and Represent the weight parameters of the two convolutional layers respectively; BN represents the batch normalization operation, which is used to normalize the feature distribution; ReLU represents the linear rectification function, which is defined as ; Represents a convolution operation.

[0054] b) Hierarchical feature enhancement

[0055] By constructing a feature pyramid network, adaptive fusion of multi-scale features is achieved:

[0056]

[0057] in, Represents the feature map of the l-th layer feature pyramid; represents an upsampling operation, which doubles the spatial size of the input feature map; represents the lateral connection mapping function, which is used to adjust the number of feature channels; Represents a 1×1 convolution operation, which is used for feature dimensionality reduction.

[0058] c) Temporal feature encoding

[0059] Introduce the temporal attention mechanism to achieve dynamic feature enhancement:

[0060]

[0061] in, Represents the temporal coding features at time t; Represents the attention weight of the feature at time t to the feature at time τ; represents the feature vector at time τ; T represents the length of the time window. The attention weight is calculated in the form of scaled dot product:

[0062]

[0063] in, represents the query vector at time t; represents the key vector at time τ; Represents feature dimension; represents the normalization function, which converts the input into a probability distribution; T represents the matrix transpose; Represents the square root operation.

[0064] C. Feature alignment and fusion

[0065] This paper realizes the alignment and fusion of multimodal features based on the improved Transformer architecture and innovatively introduces a hierarchical attention mechanism. The specific steps are as follows:

[0066] a) Feature Serialization

[0067] Map features of different modalities to a unified feature space:

[0068]

[0069] in, Represents a feature concatenation operation, which connects multiple feature tensors in the specified dimension; , , , Feature representations for visible light, infrared, point cloud, and gas data, respectively.

[0070] b) Position and modality encoding

[0071] Introduce position and modality information:

[0072]

[0073] in, Represents the encoded feature sequence; PE represents the position encoding matrix, which is used to characterize the position information of each element in the sequence; ME represents the modal encoding matrix, which is used to distinguish different data sources; + represents matrix addition operation.

[0074] c) Hierarchical Attention Fusion

[0075] Dynamic feature fusion is achieved through multi-head attention mechanism:

[0076]

[0077] in, represents the multi-head attention calculation function; Q, K, and V represent query, key, and value matrices respectively. The specific calculation process of multi-head attention is:

[0078]

[0079] in, represents the output of the i-th attention head; h represents the number of attention heads; represents the output mapping matrix; , , Respectively represent the parameter matrix of the i-th attention head; Represents the attention mechanism.

[0080] d) Feature fusion output

[0081] The final features are obtained through feedforward network and normalization:

[0082]

[0083] in, represents the final fusion feature; Representation layer normalization operation, used to stabilize feature distribution; represents a feedforward neural network, which is defined as:

[0084]

[0085] in, , represents the weight matrix; , represents the bias term; Represents the ReLU activation function.

[0086] D. Data distribution and storage

[0087] The fused features The original multimodal data is transmitted to the edge intelligent processing unit for analysis in real time. Sent to the traceability evidence unit for storage. 2. Edge intelligent processing unit The edge intelligent processing unit is the core processing module of the system, responsible for realizing the intention understanding and intelligent analysis of illegal behaviors. The unit adopts a dual-stream intention understanding framework to analyze the behavioral intentions of operators and the surrounding environment status, realizes distributed collaborative computing through adaptive edge node networking, and realizes knowledge sharing based on a federated learning framework. The specific implementation methods are as follows: A. Dual-stream intention understanding framework Construct a dual-stream intention understanding framework to analyze illegal intentions from two dimensions: the behavioral sequence of operators and the context of the working environment. Define discrete time series , perform intent analysis in each time window [tk, t], where k represents the length of the observation window, ranging from [30s, 60s], and the step length is 1s. a) Behavior sequence flow processing

[0088] A bidirectional GRU network is used to extract the temporal characteristics of the operator's behavior: in, represents the behavior representation vector at time t, d is the feature dimension, and its value is 256; represents the set of behavioral sequence features from tk to time t, where is the input feature vector at time τ, The input feature dimension is 128.

[0089] b) Environmental context flow processing

[0090] Construct the job scenario state vector:

[0091] in, represents the scene state vector at time t, The value is 256; Represents a set of environmental features, including:

[0092] Process parameters : is the reactor temperature, with a range of [0, 200]°C and an accuracy of 0.1°C; is the system pressure, with a value range of [0, 10] MPa and an accuracy of 0.01 MPa; is the feed flow rate, the value range is [0, 100] m³ / h, and the accuracy is 0.1 m³ / h;

[0093] Device Status : The feed pump status, with the value {0: stop, 1: run}; The state of the agitator, with the value {0: stop, 1: run}; is the cooling water valve opening, with a value range of [0, 100]% and an accuracy of 1%;

[0094] Spatial characteristics : The distance between the person and the equipment, with a value range of [0, 10] m and an accuracy of 0.1 m; is the distance between the personnel and the evacuation passage, with a value range of [0, 20] m and an accuracy of 0.1 m; It is the personnel density in the working area, with a value range of [0, 1] person / m² and an accuracy of 0.01 person / m².

[0095] Represents a set of rule constraints, including:

[0096] Operating procedures constraints : is the operation step number, with a value range of [1, 20]; is the shortest execution time, in seconds; is a step parallel constraint, with values ​​{0: prohibited, 1: allowed}.

[0097] Security threshold constraints : is the upper temperature limit, 180°C; The upper limit of pressure is 8MPa : Warning area range, 2m.

[0098] Personnel qualification constraints : is the operator level, with values ​​of {1: primary, 2: intermediate, 3: senior}; It is the special operation requirement, and its value is {0: no requirement, 1: requirement}; It is the validity period of safety training, in days.

[0099] The calculation process of the state mapping function φ is: in, and Respectively represent the mapping matrices of environmental characteristics and rule constraints, and The average value is 128; and represents the weight matrix of the two-layer mapping network, The value is 128; is the corresponding bias term; ReLU represents the linear rectification function, defined as c) Feature Fusion

[0100] Achieve dynamic integration of behavioral intentions and environmental constraints: in, represents the fused feature representation, The value is 256; Represents the feature fusion function: in, represents the fusion mapping matrix; Denotes the bias term. B. Adaptive edge networking Based on the scene complexity, an adaptive edge node networking method is constructed to achieve dynamic scheduling of computing resources. a) Scene complexity evaluation: in, Indicates the scene complexity score at time t, with a value range of [0, 1]; represents the characteristic entropy, which is defined as: in, Represents the normalized probability value of the i-th dimension feature; Represents the node density: in, represents the number of active nodes at time t, Indicates the maximum node capacity of the system; Indicates the intensity of activity: Among them, A(τ) represents the activity at time τ, Indicates the maximum activity of the system, T indicates the length of the statistical window, which is 300s; α, β, γ are weight coefficients, and α+β+γ=1 is satisfied, which can be adjusted according to the actual application scenario. b) Resource Scheduling

[0101] Dynamically allocate computing resources based on scene complexity:

[0102] in, represents the amount of computing resources allocated to the i-th edge node; represents the scene complexity of the i-th node; M represents the total number of edge nodes; Represents the total system resources. The resource allocation constraints are: , in, Indicates the minimum resource guarantee for a single node to ensure basic computing capabilities. c) Load balancing process

[0103] Realize dynamic migration of computing tasks: in, Represents the load index of the i-th node, with a value range of [0, 1]; Indicates the actual resource usage at time τ; T indicates the statistical window length, which is 300s. When the node load exceeds the threshold, task migration is triggered: in, Indicates the load threshold, the value is 0.8; Represents the task migration decision from node i to node j at time t. C. Federated Learning Framework Federated learning is used to achieve knowledge migration and sharing between edge nodes, improving model performance while protecting data privacy. a) Local model update: in, represents the model parameters of the i-th node in the t-th round; η represents the learning rate, and an adaptive adjustment strategy is adopted: in, represents the initial learning rate, with a value of 0.001; and They represent the first-order and second-order momentum decay rates, taking values ​​of 0.9 and 0.999 respectively; Represents the local loss function: Among them, CE represents the cross entropy loss function; and They represent the true label and the predicted label respectively; λ represents the regularization coefficient, which is 0.0001; represents the L2 norm. b) Global model aggregation: in, represents the global model parameters of the t+1th round; represents the number of local samples of the i-th node; n represents the total number of samples of all nodes: c) Differential privacy protection: in, Represents the model parameters that add differential privacy protection; Indicates that the mean is 0 and the covariance matrix is Multidimensional Gaussian noise of, I is the unit matrix. Noise intensity Dynamically adjusted according to the privacy budget ε: Among them, ε represents the privacy budget, with a value range of [0.1, 10]; δ represents the failure probability, with a value of ; S represents parameter sensitivity: Where D and D' represent any two adjacent data sets, that is, data sets that differ by only one sample; Represents the model parameters trained on the dataset D. To ensure that the model performance does not degrade significantly due to noise injection, an adaptive clipping mechanism is introduced: Among them, B represents the clipping threshold, the initial value is 1, and it is dynamically adjusted during the training process: Where α represents the adjustment step size, which takes a value of 0.01; sign represents the sign function. D. Calculation result output The output of the edge intelligent processing unit includes the following: a) Intent understanding result: in: represents the behavior type label, K is the number of predefined behavior types; represents the violation probability score; Indicates the risk level, corresponding to no risk, low risk, medium risk, and high risk.

[0104] b) Scenario analysis results: in: represents the environment state vector; represents the set of activated constraints; Indicates the resource allocation scheme for each node.

[0105] c) Intermediate state set: in: The representation feature represents the sequence; Represents a snapshot of model parameters; Indicates calculation load statistics.

[0106] Intent understanding results And scenario analysis results The intermediate state will be sent to the process intelligence analysis unit for further processing. It will be sent to the traceability evidence unit for storage and auditing. At the same time, the system dynamically updates the resource allocation strategy and model parameters of the edge nodes based on the output results to ensure continuous optimization of system performance.

[0107] 3. Process Intelligence Analysis Unit The process intelligence analysis unit is responsible for providing the system with chemical professional knowledge support and violation analysis capabilities. This unit constructs a chemical process knowledge graph, uses a graph neural network to implement process knowledge reasoning, and evaluates violation risks through an intention-behavior probability graph model. The specific implementation methods are as follows: A. Chemical process knowledge graph construction Construct a chemical process knowledge graph in the form of a triple G=(V, E, A), where V represents a node set, E represents an edge set, and A represents an attribute set. a) Node set definition: The node set consists of four subsets: Process parameter node collection: Contains the following parameter types: temperature nodes: reactor temperature [0-200℃], heat exchanger temperature [0-150℃]; pressure nodes: system pressure [0-1.0MPa], pipeline pressure [0-1.2MPa]; flow nodes: feed flow [0-10m³ / h], circulation flow [0-15m³ / h]. Equipment node collection: Includes the following equipment types: storage equipment nodes: reactors, storage tanks; mass transfer equipment nodes: pumps, compressors; heat exchange equipment nodes: heat exchangers, condensers.

[0108] Operation node collection: Contains the following operation types: process operation nodes: feeding, heating; equipment operation nodes: start-stop, switching; emergency operation nodes: emergency stop, pressure relief. Rule node set: Contains the following rule types: process procedure nodes: temperature control, pressure control; safety rule nodes: personnel requirements, equipment spacing; emergency plan nodes: fire disposal, leakage disposal. b) Edge set definition: in, represents a directed edge from node i to node j, Represents the type of edge, and T represents the set of edge types.

[0109] The edge type set T contains:

[0110] Parameter association edge :Describes the correlation between process parameters. When the correlation coefficient When establishing an association edge; device connection edge :Describe the material flow and energy transfer relationship between equipment; operation sequence edge : Describe the pre-order, post-order, and parallel relationships between operations; rule constraints : Describes the constraints of rules on parameters and operations. c) Attribute set definition: Attribute vector representation: Node attribute vector: ; Edge attribute vector: ;in: Represents a numerical attribute vector with a dimension of 10; Represents the category attribute vector, with a dimension of 20; Represents a text word embedding vector with a dimension of 50; Represents vector concatenation operation. B. Process knowledge reasoning uses graph neural network to realize deep reasoning of process knowledge, and combines attention mechanism to realize multi-level feature extraction. a) Node representation update: Basic update equation: in: Represents the hidden state vector of node v at layer t; Represents the set of neighbor nodes of node v; the GNN calculation process includes the following steps: message generation: ,in Generate a weight matrix for the message, is the bias vector; attention calculation: in is the attention weight matrix, LeakyReLU is the activation function; state update: in Update the weight matrix for the state, is the bias vector. b) Sub-image feature extraction: in: represents the subgraph feature vector; Represents a subgraph node set; Indicates the maximum number of layers ( ); represents the pooling function, which is defined as: ,in is the pooling weight matrix, is the pooling bias scalar. c) Knowledge reasoning output: in: Represents the probability distribution of K kinds of inference results; represents the context feature vector; represents the output weight matrix; Represents the output bias vector; C. Intent-Behavior Probability Graph Model The present invention constructs an intent-behavior probability graph model for assessing violation risks. a) Intent space definition: in: represents the kth intention type; Indicates the total number of intention types; including the following intention types: normal operation intention: process operation, equipment maintenance, emergency response; illegal operation intention: process parameter exceeding the limit, equipment misoperation, and procedure violation. b) Behavior space definition: in: represents the mth behavior type; Indicates the total number of behavior types; including the following behavior types: Operational behavior: valve adjustment, equipment start and stop, parameter setting; Mobile behavior: area entry, path selection, stay time; Interactive behavior: human-computer interaction, personnel collaboration, information transmission. c) Probability graph construction: Joint probability distribution: in: Indicates the length of the time series; represents the intention transfer probability, and the calculation formula is: in represents the historical feature vector at time t-1, Represents the context feature vector at time t represents the probability of behavior generation, and the calculation formula is: ,in represents the intention vector at time t, Represents the behavior vector at time t-1. D. Violation risk assessment Violation risk assessment is implemented based on the inference results and probability model. a) Risk score calculation: in: Represents the comprehensive risk score (including intention risk, behavior risk, and context risk), with a value range of [0, 1]; Represents the weight coefficient, satisfying: and ; Calculation of sub-item risks: Intentional risk: ,in Represents the probability of the kth intention; behavioral risk: ,in Indicates the most likely intention given The conditional probability of the mth behavior is, Represents the risk weight of the mth behavior; contextual risk: ,in Indicates the number of rule nodes, Represents the risk assessment function. b) Risk level classification: The risk threshold parameters are: Indicates a low risk threshold; indicates the medium risk threshold; Indicates high risk threshold. c) Risk warning generation: Contains: violation type, identifying the specific violation type; risk level, corresponding risk level classification results; probability score, probability assessment value of the violation; occurrence time, estimated time of the violation; location information, possible location of the violation. E. Calculation result output The risk assessment result is transmitted to the intervention control unit for disposal, and the reasoning process and assessment basis are sent to the traceability evidence unit for storage. The output content includes: risk assessment results: risk level and warning information; knowledge reasoning process: node status update process; probability model parameters: intention transfer probability and behavior generation probability.

[0111] 4. Intervention control unit

[0112] The intervention control unit is responsible for implementing intelligent intervention on illegal behaviors. The unit adopts an intervention decision model based on risk entropy, generates intervention strategies through multi-objective optimization, and establishes a distributed execution scheduling mechanism. The specific implementation method is as follows:

[0113] A. Risk Entropy Intervention Decision Model A risk entropy model is constructed to assess the timing of intervention and form a hierarchical decision framework.

[0114] a) Risk entropy calculation:

[0115] in: represents the risk entropy value at time t; represents the probability distribution of the i-th risk state; represents the number of partitions of the risk state space;

[0116] Risk state transition probability:

[0117] in: Indicates that an intervention action is being performed From the state Transfer to state probability; Indicates the number of samples corresponding to state transitions in historical data; Indicates status Next action The total number of samples.

[0118] b) Assessment of timing for intervention:

[0119] in: represents the rate of change of risk entropy; represents the value-at-risk function; represents the weight coefficient of the i-th risk; Represents the assessment value of the i-th risk at time t.

[0120] Intervention trigger conditions:

[0121] in Indicates the intervention threshold of the corresponding dimension.

[0122] B. Multi-objective intervention strategy optimization A multi-objective optimization method is used to generate intervention strategies that balance safety and production efficiency.

[0123] a) Objective function construction:

[0124] in: Indicates security risk target: ; Express the production loss target: ; Denote the intervention cost target: . Constraints: ; ;

[0125] Where: represents the intervention strategy vector; Represents the weight coefficient of each target; represents the i-th type of security risk assessment function; represents the j-th type of production loss assessment function; represents the k-th type of intervention cost evaluation function; represents the constraint function; Represents the upper and lower bounds of the decision variables.

[0126] b) Pareto optimal solution search: The improved NSGA-III algorithm is used to solve multi-objective optimization problems:

[0127] Individual code: ,in represents the i-th intervention decision variable;

[0128] Fitness evaluation: ,in represents the non-dominated sorting level, represents the crowding distance;

[0129] Select an action:

[0130] in: Represents an individual Probability of being selected; Indicates the selection pressure parameter; Indicates the population size.

[0131] C. Distributed Execution Scheduling A distributed execution scheduling mechanism is established to coordinate the linkage response of multiple intervention devices.

[0132] a) Intervention task breakdown:

[0133] in: represents the set of intervention tasks; represents a single intervention task; Indicates the execution device; Indicates intervention action; Indicates execution time; represents the intervention parameter.

[0134] b) Scheduling priority calculation:

[0135] in: Representation Task Priority; Indicates the urgency of the task; Indicates execution efficiency; Indicates the impact of dependency; Represents the weight coefficient.

[0136] c) Real-time scheduling execution: Scheduling state vector: ,in Represents the execution status of device i at time t

[0137] Scheduling decision function: ,in Represents a matching pair of equipment and tasks; Indicates the number of tasks in the current scheduling cycle.

[0138] D. Intervention Effect Evaluation The present invention constructs a real-time feedback evaluation mechanism to support the dynamic optimization of intervention strategies.

[0139] a) Effect evaluation indicators:

[0140] in: Indicates the risk reduction effect, ; Indicates the degree of production impact, ; Indicates the resource consumption level, ;

[0141] Where: represents the risk value before and after the intervention; Indicates production indicators under normal and intervention conditions; represents the unit cost of the i-th type of resource; Indicates the usage of the i-th resource.

[0142] b) Strategy optimization feedback:

[0143] in: represents the strategy adjustment amount; represents the policy optimization function; Represents a set of optimization parameters.

[0144] E. Execution result output

[0145] The intervention execution status is transmitted to the traceability evidence unit, including: intervention decision-making process: risk entropy assessment results, strategy optimization process; execution scheduling information: task decomposition results, scheduling execution records; effect evaluation data: various evaluation indicators, optimization feedback results.

[0146] 5. Traceability evidence unit

[0147] The traceability evidence storage unit is responsible for ensuring the traceability of the system and the reliability of evidence. The unit uses multimodal behavior trajectory reconstruction technology, implements causal reasoning through knowledge graphs, and builds a distributed evidence storage network. The specific implementation is as follows:

[0148] A. Multimodal behavior trajectory reconstruction uses multi-source data fusion technology to reconstruct the complete violation behavior trajectory.

[0149] a) Space-time trajectory construction:

[0150] in: Represents trajectory event points; Indicates a timestamp; Represents the spatial position vector ; Represents event attribute vector; represents the total number of trajectory events;

[0151] Trajectory interpolation function:

[0152] in: represents the estimated position at time t; represents the time weight function, which is defined as: Represents the kernel function.

[0153] b) Behavior fragment association: fragment similarity calculation:

[0154] in: Indicates a behavior fragment; Indicates the similarity in time dimension; Represents spatial dimension similarity; Indicates attribute dimension similarity; Represents the weight coefficient.

[0155] Association graph construction:

[0156] in: Represents a set of behavior fragment nodes; Represents a similarity relationship edge set, satisfying: Represents the similarity threshold.

[0157] B. The causal reasoning mechanism realizes the causal chain reasoning of violations based on the knowledge graph.

[0158] a) Causal Modeling:

[0159] in: Represents a set of causal event nodes; represents the causal edge set; Represents the causal strength matrix; causal strength calculation: , where represents conditional probability; Represents mutual information: .

[0160] b) Causal chain reasoning: Reasoning goal:

[0161] in: represents the causal chain hypothesis; indicates observational evidence; represents the prior probability; represents the likelihood function.

[0162] Markov Monte Carlo sampling:

[0163] in: represents the transition probability kernel; represents the causal chain of the tth iteration.

[0164] C. Distributed Evidence Storage Network The present invention constructs a multi-level evidence chain to ensure reliable data storage.

[0165] a) Evidence data structure: Evidence = {evidence header, evidence content, digital signature}

[0166] Among them: the evidence header contains metadata: {evidence identification number, timestamp, type, source} the evidence content contains actual data: {original data, data hash value, parent evidence identification} the digital signature contains authentication information: {public key, signature value, certificate}

[0167] b) Evidence chain construction: The block structure is: {block number, timestamp, evidence set, forward block hash value, current block hash value}

[0168] Calculation of the current block hash value: in Represents a hash function

[0169] c) Consensus mechanism: Verification conditions:

[0170] Among them: integrity verification: ensure that the block data has not been tampered with; signature verification: ensure that the source of evidence is credible; timing verification: ensure that the block order is legal.

[0171] D. Retrieval service interface

[0172] Provides a multi-dimensional retrieval interface to support traceability queries.

[0173] Retrieval request format:

[0174] Search request = {search type, search conditions, time and space range, search options}

[0175] Among them: search types include: violation tracing, intervention record query, risk assessment history; search conditions include: keyword matching, threshold range, attribute constraints; time and space scope includes: time period setting, spatial area demarcation; search options include: sorting method, return quantity, and data desensitization.

[0176] Return result format:

[0177] Query result = {search status, result data, evidence storage certificate}

[0178] Among them: the search status reflects the query execution status and the permission verification results; the result data contains the core information that meets the search conditions; the evidence voucher provides proof of data credibility.

[0179] Combination Figure 2 As shown, the system workflow is as follows:

[0180] S1. Multi-source heterogeneous data acquisition: A multimodal perception unit is used to collect multi-dimensional perception data of the chemical park through high-definition video cameras, infrared thermal imagers, millimeter-wave radars and gas concentration sensors, including visible light images of the scene, thermal radiation distribution characteristics of the target, spatial position and motion characteristics, and gas concentration distribution data of the area.

[0181] S2. Multimodal feature fusion processing: First, the improved ResNet-FPN network is used to perform hierarchical feature extraction on the collected multimodal data. Then, the Transformer architecture with a hierarchical attention mechanism is introduced to achieve alignment and fusion of multi-source features, and finally a unified feature representation is generated to support subsequent intelligent analysis.

[0182] S3. Edge intelligent violation intention understanding: After receiving the fusion features, the edge intelligent processing unit analyzes the violation intention from the dimensions of behavior sequence and environmental context through a dual-stream intention understanding framework. At the same time, it dynamically schedules computing resources based on the complexity of the scenario and realizes knowledge transfer and sharing between edge nodes through federated learning.

[0183] S4. Process knowledge reasoning and risk assessment: The process intelligence analysis unit performs deep reasoning based on the pre-built chemical process knowledge graph, extracts multi-level features with the graph neural network, and quantitatively assesses the violation risk through the intention-behavior probability graph model, and outputs risk level and warning information.

[0184] S5. Intelligent intervention strategy execution: The intervention control unit first evaluates the optimal intervention time based on the risk entropy model, and then generates an intervention strategy that balances safety and production efficiency through a multi-objective optimization algorithm. It uses a distributed execution scheduling mechanism to coordinate the linkage response of various intervention devices, and continuously optimizes the intervention effect through real-time feedback.

[0185] S6. Tracing and Evidence Storage of Illegal Behaviors: The tracing and evidence storage unit adopts multimodal behavior trajectory reconstruction technology to associate discrete illegal behavior fragments to form a complete behavior chain, conducts causal reasoning analysis through knowledge graphs, builds a multi-level evidence chain based on a distributed evidence storage network, and adopts differential privacy technology to protect sensitive data.

[0186] S7. System collaborative data transfer: Establish an efficient data transfer mechanism between the various units of the system to achieve real-time transmission and sharing of perception data, analysis results, risk assessment, intervention instructions and traceability information. The multimodal perception unit distributes data to the edge processing and evidence storage unit, and the analysis results are transmitted between the processing units. Finally, all execution records flow into the traceability evidence storage unit for unified management.

[0187] In summary, the present invention provides a chemical park violation intervention and tracking and tracing system based on edge computing, which innovatively solves the deep-seated technical problems in the safety management of chemical parks. By integrating multimodal perception and process knowledge graphs on the edge side, the bottleneck of insufficient understanding of professional scenarios by traditional systems has been broken; adaptive edge networking and federated learning mechanisms have been adopted to achieve real-time computing and knowledge transfer under complex working conditions; based on risk entropy-driven intervention decisions and multi-state trajectory reconstruction technology, a complete technical chain from prevention, disposal to tracing has been constructed. The various functional units of the system work together to enhance the intelligent disposal capabilities of violations while ensuring the credibility and integrity of the chain of evidence, providing an innovative technical solution for the inherent safety management of chemical parks.

[0188] Although this document shows the specific application of the present invention through specific embodiments, these embodiments are only for illustrative purposes and are not intended to limit the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims, and any appropriate modification, equivalent substitution or improvement based on the principle of the present invention should be deemed to fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be given the broadest interpretation to cover all such modifications, equivalent structures and functions.

Claims

1. A chemical park illegal behavior intervention and tracking and tracing system based on edge computing, characterized in that: include: Multimodal sensing unit, used to collect multi-dimensional sensing data of the chemical park, including high-definition video, infrared thermal imaging, millimeter-wave radar and gas concentration data; The multimodal perception unit uses an improved residual network-feature pyramid network ResNet-FPN to extract features from multi-source heterogeneous perception data. The improved ResNet-FPN network improves the ability to capture key features in dynamic scenes through hierarchical feature enhancement and temporal correlation encoding, and uses a Transformer structure with a hierarchical attention mechanism to achieve cross-modal feature alignment; The edge intelligent processing unit is used to receive the feature data extracted by the multimodal perception unit to realize the intention understanding and intelligent analysis of the illegal behavior; the edge intelligent processing unit adopts a dual-stream structure to construct a bidirectional GRU sequence network and a scene state vector network respectively, integrates the behavior sequence features and the environment state features through an adaptive fusion function, and automatically schedules the computing resources of the edge node according to the complexity of the scene; the edge intelligent processing unit realizes distributed collaboration by adopting a federated learning framework with differential privacy protection, ensuring data privacy while improving model performance; A process intelligence analysis unit is used to receive the output data of the edge intelligent processing unit and provide chemical professional knowledge support and violation analysis capabilities; the process intelligence analysis unit constructs a chemical process knowledge graph and uses a graph neural network to perform deep reasoning on real-time data, quantitatively evaluates violation risks by establishing an intention-behavior probability graph model, and triggers intervention signals according to risk entropy thresholds; An intervention control unit is used to receive the risk assessment results of the process intelligent analysis unit and implement intelligent intervention on illegal behaviors; The intervention control unit evaluates the intervention timing according to the risk entropy calculation results, generates a multi-objective balanced intervention strategy using Pareto optimal search, and coordinates multiple intervention devices to execute a linkage response through a distributed scheduling mechanism; The traceability evidence storage unit is used to receive the execution records of the intervention control unit and the original data of the multimodal perception unit to ensure the traceability of the system and the reliability of evidence; the traceability evidence storage unit associates and reconstructs discrete violation behavior fragments through a spatiotemporal trajectory construction algorithm, and uses a distributed evidence storage network to build a multi-level evidence chain including evidence headers, evidence content and digital signatures.

2. The system according to claim 1, characterized in that The edge intelligent processing unit supports a dynamic scheduling mechanism of computing resources based on multi-task decomposition, including: Node load monitoring and allocation adjustment process, dynamically allocating computing resources according to the complexity of the scenario; In the load balancing process, when the node load exceeds the threshold, task migration is triggered to ensure efficient resource utilization.

3. The system according to claim 1, characterized in that The process intelligent analysis unit dynamically updates the edge attribute weights of the chemical process knowledge graph through cross-analysis of the historical case library and real-time data, thereby improving the accuracy of knowledge reasoning.

4. The system according to claim 1, characterized in that The intervention control unit establishes a real-time feedback evaluation mechanism to dynamically evaluate the intervention effect through three dimensions: risk reduction effect, production impact degree and resource consumption level, and feeds the evaluation results back to the strategy optimization module to support the continuous improvement of the intervention strategy.

5. The system according to claim 1, characterized in that The traceability evidence unit adopts a causal reasoning mechanism to assist in tracing the source of violations, and searches for the causal chain hypothesis with the maximum a posteriori probability based on the Markov Monte Carlo sampling algorithm. The causal chain hypothesis is composed of edges in the knowledge graph, and the weights of the edges are determined by conditional probability and mutual information.

Images