Industrial park safety production risk intelligent monitoring and early warning system

Abstract

This invention belongs to the field of safety production monitoring and early warning technology, and particularly relates to an intelligent monitoring and early warning system for safety production risks in industrial parks. It adopts a four-layer distributed architecture: perception layer, network layer, platform layer, and application layer. The perception layer collects safety production data from all elements of the park. The network layer constructs a high-speed, reliable communication network integrating multiple protocols. The platform layer first unifies the spatiotemporal reference of multi-source data through spatiotemporal grid coding, then uses a causal-driven cross-modal attention fusion mechanism to filter false associations, subsequently constructs a dynamic spatiotemporal causal graph to achieve spatiotemporal coupling modeling of risks, and finally completes risk propagation prediction and multi-dimensional root cause tracing based on a causal dynamic Bayesian network. The application layer relies on digital twins to realize risk visualization, intelligent emergency command, and automated report generation. This invention can solve the problem of poor risk early warning accuracy in existing monitoring and early warning methods for industrial parks.

Description

Technical Field

[0001] This invention belongs to the field of safety production monitoring and early warning technology, and in particular relates to an intelligent monitoring and early warning system for safety production risks in industrial parks. Background Technology

[0002] Industrial parks are characterized by a high concentration of enterprises, a concentration of hazardous sources, strong risk coupling, and significant chain reactions from accidents. Traditional safety monitoring mainly relies on single-sensor threshold alarms, manual inspections, and passive video monitoring, which suffers from problems such as data silos, weak fusion capabilities, insufficient identification of coupled risks, delayed early warnings, and poor interpretability.

[0003] In existing technologies, monitoring and early warning methods mostly employ single time series models (such as LSTM, GRU) or simple spatiotemporal models (such as ST). GCN and ASTGCN can only capture data correlations and cannot distinguish between causal relationships and spurious relationships, which easily leads to false alarms and missed alarms. In terms of the fusion of multi-source heterogeneous data (sensors, video, text, devices, meteorology), feature splicing or weighted fusion is mostly used, without establishing a unified spatiotemporal benchmark, resulting in low fusion accuracy. Risk propagation and tracing rely on human experience or static rules and cannot dynamically deduce the risk evolution path. The system architecture mostly adopts pure cloud centralized inference or simple edge computing, which makes it difficult to balance real-time performance and accuracy.

[0004] In summary, existing technologies cannot meet the needs of industrial parks for intelligent monitoring and early warning of safety production risks, which require high precision, early warning, low false alarms, interpretability, and strong real-time capabilities. There is an urgent need for a technology that can achieve unified spatiotemporal alignment, cross-modal causal fusion, spatiotemporally coupled risk modeling, dynamic source tracing, and edge detection. A cloud-based collaborative reasoning intelligent monitoring and early warning system. Summary of the Invention

[0005] The technical problem solved by this invention is to provide an intelligent monitoring and early warning system for safety production risks in industrial parks, so as to solve the problem of poor risk warning accuracy in existing monitoring and early warning methods for industrial parks.

[0006] The basic solution provided by this invention is an intelligent monitoring and early warning system for safety production risks in industrial parks, which includes a perception layer, a network layer, a platform layer, and an application layer. The perception layer deploys sensing devices at major hazard sources and key locations in the industrial park to collect various types of sensing data and generate multi-source datasets. The network layer constructs communication and transmission between the perception layer, platform layer, and application layer through a network architecture deployed in the industrial park. The platform layer receives multi-source datasets from the perception layer, performs preprocessing and data analysis to obtain risk warning information for the industrial park, and transmits it to the application layer. The application layer uses a digital twin of the industrial park to display the risk warning information, conduct emergency command and dispatch, and generate risk reports. Among these: The platform layer includes a data center and a computing platform. The data center receives multi-source datasets transmitted from the perception layer, stores and preprocesses them, and generates preprocessed data. The computing platform includes a spatiotemporal grid coding module, a cross-modal causal attention fusion module, a spatiotemporal coupled risk modeling module, and a dynamic risk propagation and tracing module. The spatiotemporal grid coding module divides the industrial park plane into spatiotemporal grid units and encodes them. The preprocessed data is then mapped to the spatiotemporal grid units. The cross-modal causal attention fusion module independently mines the intermodal direct causal relationships with time delay in each spatiotemporal grid cell through a preset cross-modal feature mining mechanism, calculates the average direct causal effect and constructs a spatiotemporal direct causal matrix that dynamically changes in spatial location and time, and fuses multimodal features through an end-to-end causal attention mechanism that embeds the spatiotemporal direct causal matrix to generate a cross-modal fusion feature vector. The spatiotemporal coupled risk modeling module constructs a real-time updated spatial risk map of the industrial park. It models the bidirectional causal interaction between spatial risk distribution and temporal evolution trend through a two-way spatiotemporal causal interaction layer, and introduces a risk accumulation dynamics model to calculate the final risk level value of each spatiotemporal grid unit. The dynamic risk propagation and tracing module constructs an adaptive causal dynamic Bayesian network, and calculates primary and secondary risks by combining a multi-dimensional root cause ranking algorithm based on posterior probability, causal betweenness centrality, and propagation time, generating a hierarchical risk propagation path diagram and risk warning information based on root cause tracing results.

[0007] Furthermore, the cross-modal causal attention fusion module includes a single-modal feature extraction unit, a spatiotemporal heterogeneous causal discovery unit, a direct causal effect calculation unit, and a causal attention fusion unit, wherein: The single-modal feature extraction unit is used to obtain grid-level feature vectors from each spatiotemporal grid cell. The feature vectors of m modalities are extracted and concatenated to obtain the initial feature vector of each spatiotemporal grid cell. ; The spatiotemporal heterogeneous causality discovery unit introduces a time delay parameter. The PC algorithm is used to mine causal pairs with time delay. ,in As a causal mode, The resulting mode; and a local causal graph for each spatiotemporal grid cell is constructed. Where V is the modality set, The set of causal edges is the union of all time delays, and each set of causal edges consists of causal edges. composition, The preset time delay amount; Indicates grid i, time t, and delay Below, mode k is paired with mode The existence of causal relationships; The direct causal effect calculation unit is used to calculate each causal edge. The backdoor adjustment formula is used to calculate mode k for modes. In the spacetime grid Placement, delay Average direct causal effect The expression is:

[0008] in, For causal budgeting, Adjust the variable set for the backdoor; For mathematical expectation; The direct causal effect calculation unit is also used to construct the spatiotemporal direct causality matrix and the time delay weight matrix. The expression is:

[0009] Time Delay Weight Matrix The expression is:

[0010] The time delay to generate the maximum ADE value; The causal attention fusion unit is used to fuse multimodal features through an end-to-end causal attention mechanism that embeds a spatiotemporal direct causal matrix to generate a cross-modal fused feature vector. Specifically: Initial eigenvectors for each spatiotemporal grid cell Perform a linear transformation to generate the query matrix. Key matrix Sum matrix , , , The weight matrix is ​​a learnable weight matrix; To calculate the time-delay-weighted causal attention score, the spatiotemporal direct causality matrix and the time-delay weight matrix are simultaneously embedded into the attention mechanism, expressed as:

[0011] in, The dimension of the key matrix, Hadamard product, causal matrix Filtering spurious associations, time delay weighting Strengthen the causal relationships that are time-sensitive; Calculate attention output:

[0012] Introducing a causal regularization term and performing end-to-end joint optimization, the total loss function is:

[0013]

[0014]

[0015] in, To predict cross-entropy loss for risk, To discover loss in causal discovery, A true causal matrix constructed for domain knowledge. For the sparseness loss of the causal matrix, , The regularization coefficient is used. Output cross-modal causal fusion feature vector:

[0016] in, This is a cross-modal causal fusion feature vector.

[0017] Furthermore, the spatiotemporal coupled risk modeling module includes a directed dynamic spatiotemporal causal graph construction unit, a bidirectional spatiotemporal causal interaction unit, and a risk calculation unit, wherein: The directed dynamic spatiotemporal causal graph construction unit is used to construct a directed dynamic spatiotemporal causal graph of an industrial park that is updated in real time, specifically as follows: Define the dynamic spatiotemporal causal graph of the industrial park as follows: , where the set of nodes For all spatiotemporal grid cells, the set of edges This represents the directed risk propagation relationship between spatiotemporal grid cells at time t; Calculate the directed edge weights from grid i to grid j at time t. It also considers geographical distance, real-time causal effects, wind direction influence, and process correlation, with the expression being:

[0018] in, Let be the Euclidean distance between grid i and grid j. Let i be the spatiotemporal direct causal matrix. and the direct causal matrix of spacetime j The strength of real-time causal association is obtained by calculating the difference between them; Wind direction is an influencing factor. For equipment and process correlation factors, , , , These are the initial learnable parameters; Construct the adjacency matrix at time t ,in If the preset constraint value is exceeded, then And recalculate the edge weights and adjacency matrix every minute to perform real-time dynamic updates of the graph structure; The bidirectional spatiotemporal causal interaction unit is used to model the bidirectional causal interaction between spatial risk distribution and temporal evolution trends, specifically: Cross-modal feature vector fusion Input a two-layer graph attention network (GAT), with each layer containing 8 attention heads, and output a 256-dimensional feature vector. ; The dilation rate of the temporal convolution is dynamically adjusted based on the spatial risk distribution at the current moment. , This is a rounding function; A three-layer multi-scale temporal convolutional network, MSTCN, is used as input, taking spatial feature sequences of 10 consecutive time windows as input, and outputting a 256-dimensional temporal feature vector. ; It is a spatial feature sequence of 10 consecutive time windows; By leveraging historical evolution trends, the weights of spatial graph attention can be dynamically adjusted. :

[0019] The original spatial attention coefficients, Let i be the rate of change of risk. A gating fusion mechanism is used to fuse spatial and temporal features:

[0020] For gating weights, and For the spatial and temporal characteristics after interaction, The spatiotemporal coupling characteristics after fusion; The risk calculation unit is used to calculate the final risk level value of each spatiotemporal grid cell through a risk accumulation dynamics model, specifically: The spatiotemporal coupling features are input into two fully connected layers, and the instantaneous risk value is output through the sigmoid activation function. The expression is:

[0021] in, , This is the weight matrix. , For bias terms, The symbol for the sigmoid activation function; By introducing a forgetting factor, the cumulative effect and decay characteristics of risk are modeled to obtain the cumulative risk amount. The expression is:

[0022] Forgetting factor, Let $t$ be the cumulative risk of grid $i$ at time $t-1$, with an initial cumulative risk of $0$. The risk outbreak threshold is dynamically determined based on the historical risk distribution of each grid, expressed as:

[0023] To adapt to the risk outbreak threshold, The historical risk average of grid i over the past 30 days. The standard deviation of grid i over the past 30 days. For safety factor; Output the final risk level value:

[0024] This is the final risk level value.

[0025] Furthermore, the dynamic risk propagation and tracing module includes an adaptive structural causal dynamic Bayesian network construction unit, a risk propagation path dynamic prediction unit, and a multi-dimensional root cause tracing and hierarchical propagation unit, wherein: The adaptive structure causal dynamic Bayesian network building unit constructs a causal dynamic Bayesian network based on knowledge in the field of chemical safety and historical accident data. The causal dynamic Bayesian network consists of a set of risk factor nodes and a set of risk event nodes. The expression of the causal dynamic Bayesian network is as follows:

[0026] This is the initial Bayesian network structure. This is the initial conditional probability table. This is an initial conditional probability table across time slices, learned from historical time-series data based on the introduced time slices. and Conduct online learning updates; The dynamic prediction unit for risk propagation paths is used to predict the future propagation path, scope of impact, and arrival time of risks in real time based on a causal dynamic Bayesian network when an anomaly in a risk factor is detected, generating a risk propagation path map, specifically: When the final risk level value of a risk factor node exceeds the preset first risk threshold, it is determined to be an abnormal node, and its observation status is input into the causal dynamic Bayesian network. A joint tree inference algorithm is used for forward filtering and prediction to calculate the future. The posterior probability of risk factor nodes in each time slice; Based on the average propagation time between risk factor nodes in historical accidents, the risk propagation speed and arrival time are calculated. For each predicted time slice, the risk impact range is defined as the set of spatiotemporal grid cells corresponding to risk factor nodes with a posterior probability greater than a preset probability threshold. The geographical area of ​​this set and the number of devices involved are calculated to generate a risk propagation path map in the form of a directed acyclic graph. The multi-dimensional root cause tracing and hierarchical propagation unit is used to locate the root cause of risk factor anomalies based on causal dynamic Bayesian networks when anomalies are detected, distinguish between primary and secondary risks, and generate a hierarchical propagation path diagram, specifically as follows: When the final risk level value of a spatiotemporal grid cell is greater than the second risk threshold, the corresponding risk event node is taken as the target node. The joint tree backward reasoning algorithm is used to calculate the posterior probability of all potential root cause nodes and to calculate the causal betweenness centrality of potential root cause nodes. Based on posterior probability, causal betweenness centrality, and propagation time, a comprehensive score for root cause nodes is calculated, and primary and secondary risk nodes are identified to generate a hierarchical propagation path map.

[0027] Furthermore, the spatiotemporal grid coding module divides the industrial park's plane into spatiotemporal grid cells and encodes them. The preprocessed data is then mapped into these spatiotemporal grid cells as follows: The industrial park's layout is divided into square grid cells of equal size, and a unique spatial code is generated for each grid cell. Divide the continuous time axis into non-overlapping fixed-length time windows and generate a unique time code for each time window; By combining spatial and temporal codes, a globally unique spatiotemporal grid code is generated; The preprocessed multi-source heterogeneous data is mapped to corresponding spatiotemporal grid cells according to its geographical location and collection time, and feature aggregation is performed to generate standardized grid-level feature vectors.

[0028] Furthermore, the application layer includes a digital twin construction module, an early warning and display module, an emergency command and dispatch module, and a risk report generation module, wherein: The digital twin construction module constructs a three-dimensional geographic base for the industrial park based on high-precision terrain data and satellite imagery data, imports park drawings and BIM models, generates a digital model of the industrial park, and connects to real-time data of all elements to obtain a digital twin of the industrial park. The early warning and display module is used to receive risk warning information pushed by the platform layer and generate a visual warning effect in the digital twin of the industrial park; The emergency command and dispatch module calls the pre-built emergency plan library, matches it with the risk warning information, generates the optimal emergency plan, and assigns it to the corresponding operation and maintenance personnel; The risk report generation module is used to establish a standardized report template library. It automatically fills in the standardized report templates based on risk warning information to generate risk reports.

[0029] The technical principle of this invention lies in the following: The technical solution of this application adopts a four-layer distributed architecture: perception layer, network layer, platform layer, and application layer. The perception layer realizes the collection of all elements of data; the network layer constructs a high-speed and reliable communication network integrating multiple protocols; the platform layer, as the core computing unit, completes the entire process from data to risk warning through four core modules; the application layer realizes visualization, emergency command, and report generation based on digital twins. Specifically: 1. The spatiotemporal grid coding module addresses the issue of inconsistent temporal and spatial scales in multi-source heterogeneous data by employing a dual coding mechanism of "spatial grid partitioning + temporal window coding" to generate spatiotemporal grid cells. It then maps multi-source heterogeneous data into these spatiotemporal grid cells, generating standardized grid-level feature vectors. 2. The cross-modal causal attention fusion module addresses the problems of traditional multimodal fusion methods, which can only capture correlations and are prone to introducing spurious associations. It proposes a three-level fusion mechanism: "spatiotemporal heterogeneous causal discovery - direct causal effect calculation - causal attention fusion". Spatiotemporal heterogeneous causal discovery: Independently mine intermodal causal relationships with time delays for each spatiotemporal grid cell, use an improved PC algorithm to test causal independence under different time delays, and construct a local causal graph exclusive to each grid cell to solve the problem that global static causal models cannot adapt to the characteristics of different areas of the park. Direct causal effect calculation: The average direct causal effect (ADE) is calculated using the backdoor adjustment formula based on the Do operator, eliminating indirect causal interference, and constructing a spatiotemporal direct causal matrix and a time delay weight matrix that only contain direct causal relationships; Causal attention fusion: Simultaneously embedding the spatiotemporal direct causal matrix and the time delay weight matrix into the attention mechanism, so that the attention weight is proportional to the intensity of the causal effect, realizing causal-driven multimodal feature fusion and effectively filtering out false associations.

[0030] 3. The spatiotemporal coupled risk modeling module addresses the problem that traditional models cannot simultaneously model the spatial transmission and temporal evolution of risks and neglect the cumulative effect of risks. It proposes a coupled modeling method of "dynamic spatiotemporal causal graph - two-way spatiotemporal causal interaction - risk accumulation dynamics": Directed dynamic spatiotemporal causal graph construction: Taking into account geographical distance, real-time causal effects, wind direction influence and equipment process correlation, the directed edge weights between grids are calculated, and the graph structure is updated at fixed periods to realize dynamic modeling of risk propagation relationships; Two-way spatiotemporal causal interaction: The spatial risk distribution dynamically adjusts the dilation rate of the temporal convolution, and the temporal evolution trend dynamically adjusts the weight of the spatial graph attention, realizing two-way causal interaction between space and time, rather than simple feature splicing. Risk accumulation dynamics modeling: Introducing a forgetting factor to model the cumulative effect and decay characteristics of risk, and adaptively calculating the risk outbreak threshold based on the historical risk distribution of each grid, thus solving the problem that instantaneous risk calculation cannot predict delayed outbreak accidents.

[0031] 4. The Dynamic Risk Propagation and Source Tracing Module addresses the "black box" nature of traditional deep learning models and their inability to provide interpretable early warnings. It employs a Causal Dynamic Bayesian Network (CDBN) to achieve dynamic risk projection and root cause tracing. CDBN Construction: Based on knowledge in the field of chemical safety and historical accident data, a causal dynamic Bayesian network is constructed, and a time-slice mechanism is introduced to model the temporal evolution of risks. The network parameters are updated through online learning. Risk propagation path prediction: A joint tree inference algorithm is used for forward filtering and prediction to calculate the risk probability of all nodes in multiple future time slices, and to quantify the risk propagation speed, arrival time and impact range. Multi-dimensional root cause tracing: Bayesian backward reasoning is used to calculate the posterior probability of potential root causes. A three-dimensional root cause comprehensive scoring system is constructed by combining causal betweenness centrality and propagation time. This system automatically distinguishes between primary and secondary risks and generates a hierarchical propagation path map.

[0032] In the application layer, to address the issues of unintuitive display of traditional early warning information and low efficiency of emergency command, a virtual-real mapping system for the park is constructed based on 3D digital twin technology. Construct a high-precision 3D digital twin of the park to achieve real-time mapping of the physical entity status and visualization of risk distribution; Establish an intelligent matching mechanism between the emergency response plan database and risk warning information to achieve visualized scheduling of emergency resources and dynamic tracking of the response process; It automatically extracts system operation data and risk event information, generates standardized security analysis reports, and provides data support for management decisions.

[0033] The beneficial effects of this application are: 1. It solves the problems of inconsistent spatiotemporal benchmarks, insufficient fusion depth, and numerous spurious correlations in multi-source heterogeneous data: Through a spatiotemporal grid coding mechanism, for the first time, accurate alignment of multi-source heterogeneous data such as sensor, video, text, meteorological, and equipment data under a unified spatiotemporal benchmark has been achieved, significantly improving the accuracy of data fusion. A causal-driven cross-modal attention fusion method is adopted to replace the traditional feature splicing or weighted fusion, which effectively eliminates the vast majority of false associations and significantly reduces the false alarm rate of the system. The spatiotemporal heterogeneous causal discovery mechanism can adapt to the modal causal characteristics of different regions, effectively enhance the key modal features of different functional regions, and significantly improve the representational ability of fused features.

[0034] 2. It solves the problems of being unable to perform spatiotemporal joint modeling of risks coupled with multiple factors, and the difficulty in identifying the spatial transmission and temporal evolution patterns of risks: Real-time updated directed dynamic spatiotemporal causal graphs can accurately capture the impact of dynamic factors such as environmental changes and process correlations on risk propagation, significantly improving the accuracy of risk propagation path prediction. The two-way spatiotemporal causal interaction mechanism realizes deep coupling between space and time, significantly enhances the ability to identify coupled risks caused by the superposition of multiple risk factors, and can effectively identify major accident hazards that traditional single indicator threshold alarms cannot detect. The risk accumulation dynamics model solves the early warning problem of delayed-outbreak accidents, significantly extends the average early warning time, and buys a valuable window of opportunity for emergency response.

[0035] 3. It solves the problems of the "black box" nature of deep learning early warning models, their inability to provide the root causes and propagation paths of risks, and their poor interpretability: The causal dynamic Bayesian network-based source tracing method can not only provide risk warning results, but also accurately locate the root cause of the risk, significantly improving the accuracy of root cause tracing. The three-dimensional root cause comprehensive scoring system solves the problem of "high probability, low contribution" root cause misjudgment caused by traditional systems that rely solely on posterior probability, and significantly reduces the misjudgment rate of secondary risks; The layered risk propagation path map can clearly show the complete propagation process from the primary risk to the warning event, directly guiding emergency response personnel to take measures at the source and effectively shortening the emergency response time.

[0036] 4. It solves the problems of traditional systems, such as unintuitive display of early warning information, low efficiency of emergency command and coordination, and lack of data support for management decisions: The 3D visualization display based on digital twins makes the security situation of the park clear at a glance, significantly improving the intuitiveness and efficiency of security management. The intelligent emergency command and dispatch mechanism enables automatic matching of emergency plans, one-click dispatch of emergency resources, and real-time tracking of the handling progress, which greatly improves the collaborative efficiency of emergency response. The automated risk report generation function can comprehensively summarize the safety operation status of the park, provide managers with scientific and accurate decision-making basis, and promote the transformation of park safety management from experience-driven to data-driven. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the system architecture according to an embodiment of the present invention. Detailed Implementation

[0038] The following detailed description illustrates the specific implementation method: The basic implementation examples are as follows: Figure 1 As shown: The intelligent monitoring and early warning system for safety production risks in industrial parks includes a perception layer, a network layer, a platform layer, and an application layer. The perception layer deploys sensing devices at major hazard sources and key locations in the industrial park to collect various types of sensing data and generate multi-source datasets. Specifically: First, based on the functional zoning and risk level of the industrial park, the park is divided into five areas: major hazard source area, production equipment area, warehousing and logistics area, public auxiliary area, and personnel activity area. Sensing devices are deployed in a targeted manner to achieve full coverage without blind spots. Major hazard source areas, including tank areas, reaction vessel areas, and liquefied gas storage areas, are the areas with the highest safety risks in the park. The following sensing devices are deployed there: 1. Gas monitoring sensors: Explosion-proof combustible gas sensors, toxic gas sensors and oxygen sensors are deployed at the flanges, valves, pipe interfaces and other easy leakage points of each storage tank and reactor. The sampling frequency is set to 1Hz and the detection accuracy is not less than ±2% FS. 2. Process parameter sensors: Deploy temperature sensors, pressure sensors, liquid level sensors, and flow sensors to monitor the internal process parameters of the storage tank and reactor in real time, with the sampling frequency set to 1Hz; 3. Vibration sensors: Deployed on rotating equipment such as reactors, pumps, and compressors to monitor the vibration of the equipment during operation, with a sampling frequency set to 10Hz; 4. High-definition explosion-proof camera: Deploy one explosion-proof network camera with a resolution of 2 megapixels or higher every 50 meters, with a frame rate of 25fps and a resolution of 1920×1080, supporting nighttime infrared imaging and intelligent analysis; 5. Fire sensors: Deploy smoke sensors, flame sensors, and temperature sensors to achieve early fire detection, with a sampling frequency of 1Hz.

[0039] The production area includes production workshops, production lines, power distribution rooms, etc., and the following sensing devices are deployed there: 1. Environmental monitoring sensors: Deploy temperature sensors, humidity sensors, and dust concentration sensors, with a sampling frequency of 1 time / minute; 2. Equipment Status Monitoring Unit: Connects to the PLC system of the production equipment via an industrial bus to collect parameters such as operating current, voltage, speed, and power of the equipment. The sampling frequency is set to 10Hz. 3. Video surveillance cameras: Deploy high-definition network cameras in key locations such as production lines, control panels, and power distribution rooms to achieve real-time monitoring of the production process; 4. Access control sensors: Deploy access control sensors in important locations such as power distribution rooms and control rooms to record personnel entry and exit.

[0040] The warehousing and logistics area includes raw material warehouses, finished product warehouses, loading and unloading areas, etc., and the following sensing devices are deployed: 1. Environmental monitoring sensors: Deploy temperature sensors, humidity sensors, and gas concentration sensors, with a sampling frequency of 1 time / minute; 2. Video surveillance cameras: Deploy high-definition network cameras at warehouse entrances and exits, aisle walkways, and loading and unloading areas to achieve real-time monitoring of goods and personnel; 3. Electronic fence: Deploy electronic fences around the warehouse to prevent unauthorized personnel from entering; 4. Vehicle positioning equipment: Equip all work vehicles in the park with GPS positioning terminals to track vehicle location and driving trajectory in real time.

[0041] The public auxiliary area includes roads, pipelines, sewage treatment plants, etc., and the following sensing devices are deployed there: 1. Pipeline monitoring sensors: Deploy pressure sensors, flow sensors, and leak detection sensors at key nodes of the underground pipeline network to monitor the network's operating status in real time; 2. Weather station: Deploy automatic weather stations in open areas of the park to collect meteorological data such as temperature, humidity, wind speed, wind direction, rainfall, and air pressure. The sampling frequency is set to once per hour. 3. Video surveillance cameras: Deploy high-definition network cameras on the main roads and intersections of the park to achieve real-time monitoring of traffic and public areas.

[0042] The personnel activity area includes the work area, office area, canteen, etc., and the following sensing devices are deployed there: 1. UWB positioning base station: Deploy UWB positioning base stations in the work area and equip all workers with positioning tags. The positioning accuracy is better than 30cm and the update frequency is 1Hz. 2. Smart Safety Helmet: Equip high-risk workers with smart safety helmets that integrate a camera, microphone, gas sensor and one-button alarm function; 3. Video surveillance cameras: Deploy high-definition network cameras in densely populated areas to achieve real-time monitoring of people's behavior.

[0043] For different types of sensing devices, corresponding communication protocols and acquisition methods are adopted to convert raw data into standardized data in a unified format, generating multi-source datasets. Specifically, for sensor data, analog and digital signals from various sensors are acquired through edge gateways; for video data, real-time video streams from cameras are obtained through the RTSP protocol; for equipment operation data, equipment operation parameters are read from industrial control systems such as PLCs and DCSs through the OPC UA protocol; for personnel positioning data, signals from positioning tags are received through positioning base stations to calculate the real-time location coordinates of personnel and generate personnel trajectory data; and for text data, inspection records and hazard reporting records are collected through mobile inspection terminals, and equipment operation logs and maintenance records are obtained through API interfaces.

[0044] The raw data collected from multiple sources were then standardized, including metadata addition, format unification, and data classification. Metadata addition included adding device ID, device type, geographical location (latitude and longitude), collection timestamp (accuracy 1 second), and spatiotemporal grid code to each data sample. Format unification involved converting all data to JSON format and defining a unified data structure. Data classification involved dividing the standardized data into six categories: sensor data, video data, text data, device data, personnel data, and meteorological data, which were stored in the local database of the edge gateway to obtain a multi-source dataset.

[0045] The network layer constructs communication transmission between the perception layer, platform layer, and application layer through a network architecture deployed in the industrial park. In this embodiment, the network layer serves as the communication hub of the system, adopting a basic network architecture of "edge access - park backbone - cloud interconnection". It integrates multiple communication technologies such as 5G, LoRa, and industrial Ethernet to build a high-speed and reliable communication network covering the entire industrial park, enabling the uplink transmission of multi-source heterogeneous data from the perception layer to the platform layer, as well as the downlink distribution of control commands from the platform layer and application layer to the perception layer devices.

[0046] The network layer has a built-in multi-protocol conversion engine that supports the parsing and conversion of mainstream industrial protocols such as Modbus, OPC UA, and MQTT. It also has the ability to resume data transmission after interruption and flow control to ensure the integrity and real-time performance of data transmission. Through basic network security protection measures, it realizes device access authentication and encrypted data transmission to ensure the security and reliability of the communication process.

[0047] The platform layer receives multi-source datasets from the perception layer, performs preprocessing and data analysis to obtain risk warning information for the industrial park, and then transmits it to the application layer. The platform layer includes a data center and a computing platform. The data center receives the multi-source datasets transmitted from the perception layer, stores and preprocesses them, and generates preprocessed data. In this embodiment, the data center receives multi-source datasets transmitted from the network layer, performs preprocessing operations such as data cleaning, standardization, and alignment, and adopts a hybrid storage structure to achieve efficient storage of different types of data, providing high-quality standardized input data for the subsequent computing platform. The hybrid storage architecture includes: Time-series data storage: InfluxDB time-series database is used to store time-series data such as sensor data and device operation data, supporting high-concurrency writing and fast querying; Unstructured data storage: Use MinIO object storage to store unstructured data such as video frames, images, and documents; Relational data storage: MySQL database is used to store structured data such as user information, equipment information, emergency plans, and hazard records; Caching storage: Use Redis database to cache real-time monitoring data and hot data to improve system response speed.

[0048] The computing platform includes a spatiotemporal grid coding module, a cross-modal causal attention fusion module, a spatiotemporally coupled risk modeling module, and a dynamic risk propagation and attribution module, among which: The spatiotemporal grid coding module divides the industrial park's plane into spatiotemporal grid cells and encodes them. Preprocessed data is then mapped into these spatiotemporal grid cells. Specifically: The industrial park's plan is divided into equal-sized square grid cells, and a unique spatial code is generated for each cell. Specifically, the grid side length is first adaptively determined based on the park's risk level; for example, a recommended value is 10m for a chemical industrial park. The park's plan is then divided into several equal-sized square grid cells. Next, the GeoHash algorithm is used to spatially encode each grid cell, with the code length set to 7 bits (corresponding to a geographic precision of approximately 15m), generating a unique spatial identifier. Finally, a grid-entity mapping table is established to record the entity information such as sensors, cameras, and production equipment contained in each grid.

[0049] The continuous timeline is divided into non-overlapping fixed-length time windows, and a unique time code is generated for each time window; specifically, the timeline is divided into continuous non-overlapping 1-minute time windows, and a unique time code is generated using a Unix timestamp. :

[0050] in, is the Unix timestamp of the start time of the t-th time window.

[0051] Spatial and temporal codes are combined to generate a globally unique spatiotemporal grid code; specifically, a splicing method is used to generate a globally unique spatiotemporal grid code.

[0052] The preprocessed multi-source heterogeneous data is mapped to corresponding spatiotemporal grid cells based on its geographical location and acquisition time, and feature aggregation is performed to generate standardized grid-level feature vectors. Specifically: based on the geographical location and acquisition time of the data, all preprocessed data is mapped to corresponding spatiotemporal grid cells, and feature aggregation is performed on the multi-source data within each grid cell to generate standardized grid-level feature vectors.

[0053] in, For sensor characteristics, For video features, For text features, As a meteorological feature, These are equipment characteristics.

[0054] The cross-modal causal attention fusion module independently mines intermodal direct causal relationships with time delays in each spatiotemporal grid cell through a pre-defined cross-modal feature mining mechanism. It calculates the average direct causal effect and constructs a spatiotemporal direct causal matrix that dynamically changes between spatial location and time. Finally, it fuses multimodal features through an end-to-end causal attention mechanism embedding the spatiotemporal direct causal matrix to generate a cross-modal fused feature vector. The cross-modal causal attention fusion module includes a single-modal feature extraction unit, a spatiotemporal heterogeneous causal discovery unit, a direct causal effect calculation unit, and a causal attention fusion unit, wherein: The single-modal feature extraction unit is used to obtain grid-level feature vectors from each spatiotemporal grid cell. The feature vectors of m modalities are extracted and concatenated to obtain the initial feature vector of each spatiotemporal grid cell. In this embodiment, the feature vector extraction method for different modalities is as follows: Sensor temporal features: Input the sensor temporal data of each grid into a bidirectional gated recurrent unit (BiGRU), set the hidden layer dimension to 128, and output a 128-dimensional temporal feature vector; Video visual features: Input video frames into the YOLOv8 model and extract 256-dimensional visual feature vectors containing human behavior, equipment status, and environmental anomalies; Text semantic features: Input the text token sequence into the BERT-base model, extract 768-dimensional semantic features, and then reduce the dimensionality to 128-dimensional through a fully connected layer; Meteorological and equipment characteristics: A two-layer multilayer perceptron (MLP) is used to map meteorological data and equipment operation data to a 128-dimensional feature space; The feature extraction methods for the different modalities described above are all existing mature technologies, and will not be elaborated upon in this application. Finally, the feature vectors of the above five modalities are concatenated to obtain the initial feature vector of each spatiotemporal grid cell. .

[0055] The spatiotemporal heterogeneous causality discovery unit introduces a time delay parameter. (Unit: minutes) indicates the delay in the effect of the causal mode on the outcome mode; The PC algorithm is used to independently mine causal pairs with time delays in each spatiotemporal grid cell. ,in As a causal mode, The outcome modality was determined; causal independence was assessed using the Fisher Z-test, with a significance level set at p < 0.05. Constructing a local causal graph for each spatiotemporal grid cell Where V is the modality set, These correspond to five modalities: sensor, video, text, meteorology, and equipment. The set of causal edges is the union of all time delays, and each set of causal edges consists of causal edges. composition, The preset time delay number, n=1,2,3,4,5; Indicates grid i, time t, and delay Below, mode k is paired with mode The existence of causal relationships; The direct causal effect calculation unit is used to calculate each causal edge. The backdoor adjustment formula is used to calculate mode k for modes. In the spacetime grid Placement, delay Average direct causal effect The expression is:

[0056] in, For causal budgeting, Adjust the variable set for the backdoor; For mathematical expectation; The direct causal effect calculation unit is also used to construct the spatiotemporal direct causality matrix and the time delay weight matrix. The expression is:

[0057] Time Delay Weight Matrix The expression is:

[0058] The time delay to generate the maximum ADE value; The causal attention fusion unit is used to fuse multimodal features through an end-to-end causal attention mechanism that embeds a spatiotemporal direct causal matrix to generate a cross-modal fused feature vector. Specifically: Initial eigenvectors for each spatiotemporal grid cell Perform a linear transformation to generate the query matrix. Key matrix Sum matrix , , , The weight matrix is ​​a learnable weight matrix; To calculate the time-delay-weighted causal attention score, the spatiotemporal direct causality matrix and the time-delay weight matrix are simultaneously embedded into the attention mechanism, expressed as:

[0059] in, The dimension of the key matrix, Hadamard product, causal matrix Filtering spurious associations, time delay weighting Strengthen the causal relationships that are time-sensitive; Calculate attention output:

[0060] Introducing a causal regularization term and performing end-to-end joint optimization, the total loss function is:

[0061]

[0062]

[0063] in, To predict cross-entropy loss for risk, To discover loss in causal discovery, A true causal matrix constructed for domain knowledge. For the sparseness loss of the causal matrix, , The regularization coefficient is used. Output cross-modal causal fusion feature vector:

[0064] in, This is a cross-modal causal fusion feature vector.

[0065] The spatiotemporal coupled risk modeling module constructs a real-time updated spatial risk map of the industrial park. It models the bidirectional causal interaction between spatial risk distribution and temporal evolution trends through a two-way spatiotemporal causal interaction layer, and introduces a risk accumulation dynamics model to calculate the final risk level value of each spatiotemporal grid unit. The spatiotemporal coupled risk modeling module includes a directed dynamic spatiotemporal causal graph construction unit, a bidirectional spatiotemporal causal interaction unit, and a risk calculation unit, wherein: The directed dynamic spatiotemporal causal graph construction unit is used to construct a directed dynamic spatiotemporal causal graph of an industrial park that is updated in real time, specifically as follows: Define the dynamic spatiotemporal causal graph of the industrial park as follows: , where the set of nodes For all spatiotemporal grid cells, the set of edges This represents the directed risk propagation relationship between spatiotemporal grid cells at time t; Calculate the directed edge weights from grid i to grid j at time t. It also considers geographical distance, real-time causal effects, wind direction influence, and process correlation, with the expression being:

[0066] in, Let be the Euclidean distance between grid i and grid j. Let i be the spatiotemporal direct causal matrix. and the direct causal matrix of spacetime j The strength of real-time causal association is obtained by calculating the difference between them; The wind direction influence factor is 1.5 if j is downwind of i, otherwise it is 0.5. This is the equipment process correlation factor. If the two grids belong to the same production process unit, the value is 2.0; otherwise, it is 1.0. , , , These are the initial learnable parameters; Construct the adjacency matrix at time t ,in If it exceeds the preset constraint value, for example if ,but And recalculate the edge weights and adjacency matrix every minute to perform real-time dynamic updates of the graph structure; The bidirectional spatiotemporal causal interaction unit is used to model the bidirectional causal interaction between spatial risk distribution and temporal evolution trends, specifically: Spatial feature extraction: fusing feature vectors across modalities Input a two-layer graph attention network (GAT), with each layer containing 8 attention heads, and output a 256-dimensional feature vector. ; Spatial-temporal interaction: Dynamically adjust the dilation rate of the temporal convolution based on the spatial risk distribution at the current moment. , This is a rounding function; Temporal feature extraction: A three-layer multi-scale temporal convolutional network (MSTCN) is used. The input is a spatial feature sequence of 10 consecutive time windows, and the output is a 256-dimensional temporal feature vector. The convolution kernel size is 3, and the dilation rates are 1, 2, and 3 respectively. It is a spatial feature sequence of 10 consecutive time windows.

[0067] Time-space interaction: Utilizing historical temporal evolution trends to dynamically adjust the weight of attention in the spatial graph. :

[0068] The original spatial attention coefficients, Let i be the rate of change of risk. A gating fusion mechanism is used to fuse spatial and temporal features:

[0069] For gating weights, and For the spatial and temporal characteristics after interaction, The spatiotemporal coupling characteristics after fusion; The risk calculation unit is used to calculate the final risk level value of each spatiotemporal grid cell through a risk accumulation dynamics model, specifically: The spatiotemporal coupling features are input into two fully connected layers, and the instantaneous risk value is output through the sigmoid activation function. The expression is:

[0070] in, , This is the weight matrix. , For bias terms, The symbol for the sigmoid activation function; By introducing a forgetting factor, the cumulative effect and decay characteristics of risk are modeled to obtain the cumulative risk amount. The expression is:

[0071] Forgetting factor, Let $t$ be the cumulative risk of grid $i$ at time $t-1$, with an initial cumulative risk of $0$. The risk outbreak threshold is dynamically determined based on the historical risk distribution of each grid, expressed as:

[0072] To adapt to the risk outbreak threshold, The historical risk average of grid i over the past 30 days. The standard deviation of grid i over the past 30 days. For safety factor; Output the final risk level value:

[0073] This is the final risk level value.

[0074] The dynamic risk propagation and tracing module constructs an adaptive causal dynamic Bayesian network and uses a multi-dimensional root cause ranking algorithm, incorporating posterior probability, causal betweenness centrality, and propagation time, to calculate primary and secondary risks, generating a hierarchical risk propagation path diagram and risk warning information based on root cause tracing results. The dynamic risk propagation and tracing module includes an adaptive causal dynamic Bayesian network construction unit, a dynamic risk propagation path prediction unit, and a multi-dimensional root cause tracing and hierarchical propagation unit, wherein: The adaptive structural causal dynamic Bayesian network building unit constructs a causal dynamic Bayesian network based on knowledge in the field of chemical safety and historical accident data. The causal dynamic Bayesian network consists of a set of risk factor nodes and a set of risk event nodes. Specifically, 25 risk factor nodes are first defined. (e.g., excessive combustible gas concentration, abnormal temperature, equipment failure, etc.) and 4 risk event nodes. (Leakage, fire, explosion, poisoning), all risk factor nodes and risk event nodes are represented by binary discrete states: , For nodes, 0 represents normal and 1 represents abnormal; Then, based on knowledge in the field of chemical safety and historical accident data, the PC algorithm is used to learn the initial Bayesian network structure. ; Next, the maximum likelihood estimation method is used to learn the initial conditional probability table. ; Finally, we introduce the 1-minute time slice and learn the cross-time slice transition probability table. Constructing a causal dynamic Bayesian network (CDBN): The expression is:

[0075] This is the initial Bayesian network structure. This is the initial conditional probability table. This is an initial conditional probability table across time slices, learned from historical time-series data based on the introduced time slices; it is updated using a Bayesian online learning method for every 100 newly added labeled samples. and The forgetting factor was set to 0.95; the above-mentioned causal dynamic Bayesian network construction process is a mature existing technology, and this application will not elaborate further.

[0076] The dynamic prediction unit for risk propagation paths is used to predict the future propagation path, scope of impact, and arrival time of risks in real time based on a causal dynamic Bayesian network when an anomaly in a risk factor is detected, generating a risk propagation path map, specifically: At the risk factor node When the final risk level value exceeds the preset first risk threshold, for example... When an abnormal node is identified, its observation status is changed. Input Causal Dynamic Bayesian Network (CDBN); A joint tree inference algorithm is used for forward filtering and prediction to calculate the future. The posterior probability of risk factor nodes in each time slice; wherein, the filtering step is to calculate the posterior probability of all nodes at time t in the observation The posterior probability under the given condition is expressed as:

[0077] in, Let be the filtered posterior probability, representing the probability of obtaining the outlier node at time t. Observations Under the condition of any risk factor node The probability of being in a certain state at time t; To observe the likelihood probability, we represent the probability when the risk factor node... When the state is at time t, an abnormal node is observed. Observations The probability is determined by the observation probability table of CDBN; For one-step prediction of prior probability, it means based solely on the sequence of all observations up to time t-1. Predicting risk factor nodes The probability of being in a certain state at time t. For all anomalous nodes from time slice 1 to time slice t-1 The sequence of observations; Here, is a normalization constant, representing the prediction of the observation at time t based on all observations up to time t-1. The probability of occurrence is used to ensure that the sum of the probabilities of all states is 1; The prediction steps are as follows: Based on the posterior probability at time t, recursively calculate the future probability. The node probability of each time slice is expressed as:

[0078] in, For predicting time slices; for The step-forward posterior probability represents the observations up to the current time t. Predict any risk factor node In the future A time slice (i.e.) The probability of being in a certain state at a given time. This represents the status of the risk factor node in the previous moment. Let be the state transition probability, representing the risk factor node. from State transition at time The probability of a state at a given time step is given by the cross-time-slice transition probability table of CDBN. Sure, for The step-forward posterior probability represents the observations up to the current time t. Predicting risk factor nodes exist The probability of being in a certain state at any given time is obtained through the previous round of recursion calculation; Based on the average propagation time between risk factor nodes in historical accidents, the risk propagation speed and arrival time are calculated; specifically: First, define the nodes. To the node Average propagation time Risks from historical accidents spread to If the median time is given, then the speed of risk transmission is:

[0079] Risk originates from abnormal nodes propagation to nodes The estimated arrival time is:

[0080] in, From arrive The shortest causal path; For each prediction time slice, the risk impact range is defined as the set of spatiotemporal grid cells corresponding to risk factor nodes with a posterior probability greater than a preset probability threshold, i.e. The set of spatiotemporal grid cells corresponding to the nodes is used to calculate the geographical area of ​​the set and the number of devices involved, generating a risk propagation path map in the form of a directed acyclic graph, where: The node colors gradually change from blue (0≤P<0.2), yellow (0.2≤P<0.5), orange (0.5≤P<0.8) to red (0.8≤P≤1) according to the risk probability; the edge thickness gradually changes from thin to thick according to the propagation speed; each node is labeled with the predicted risk probability and the estimated arrival time.

[0081] The multi-dimensional root cause tracing and hierarchical propagation unit is used to locate the root cause of risk factor anomalies based on causal dynamic Bayesian networks when anomalies are detected, distinguish between primary and secondary risks, and generate a hierarchical propagation path diagram, specifically as follows: When the final risk level value of the spatiotemporal grid cell is greater than the second risk threshold, that is... The corresponding risk event nodes As the target node, its observation state is set to 1; The joint tree backward reasoning algorithm is used to calculate the posterior probability of all potential root cause nodes:

[0082] in, Let be the posterior probability of the root cause, representing the probability under known warning events. Given that the event has already occurred, the potential root cause node r is the probability that it is the root cause of the warning event. The larger the value, the higher the probability that r is the root cause of the accident. As a potential root cause node, Let be the likelihood probability, representing the probability that a warning event will occur when the potential root cause node r is in an abnormal state. The conditional probability of occurrence. This value comes from the conditional probability table (CPT) of the causal dynamic Bayesian network (CDBN), learned through historical incident data; The prior probability of the root cause node represents the probability that the potential root cause node r will occur abnormally in the absence of any observational information. This value is obtained from historical safety data of the park and reflects the inherent frequency of this risk factor within the park. The normalization constant represents the warning event. The marginal probability of occurrence is used to ensure that the sum of the posterior probabilities of all potential root cause nodes is 1; And calculate the causal betweenness centrality of the potential root cause nodes:

[0083] in, Let be the number of all causal paths from risk factor node s to risk event node t. denoted as the number of causal paths passing through node r; causal betweenness centrality measures the criticality of a node in a risk propagation network, with a larger value indicating that the node is more likely to become a hub for risk propagation.

[0084] Based on posterior probability, causal betweenness centrality, and propagation time, the comprehensive score of the root cause node is calculated:

[0085] in, The propagation time from the root cause r to the warning event. For the maximum prediction time slice, , , These are the weighting coefficients; Primary risk nodes are identified, and nodes that meet the following three conditions are considered primary risk nodes: 1. This node has no parent node in CDBN, or the risk value of its parent node is less than 0.2; 2. The anomaly of this node occurred earlier than the warning event and earlier than all other anomaly nodes; 3. This node's overall score ranks among the top 3 of all potential root cause nodes.

[0086] Based on causal dependence, the risk transmission path is divided into three layers: First layer: Primary risk nodes; The second layer consists of nodes directly affected by the primary risk node (i.e., the direct child nodes of the primary risk node). Layer 3 and above: Secondary risk nodes (i.e., nodes affected by nodes in Layer 2 and above).

[0087] Finally, a visualized hierarchical propagation path diagram is generated.

[0088] The application layer uses a digital twin of the industrial park to display risk warning information and conduct emergency command and dispatch, and generates risk reports. The application layer includes a digital twin construction module, an early warning and display module, an emergency command and dispatch module, and a risk report generation module, among which: The digital twin construction module builds a 3D geographic base for the industrial park based on high-precision terrain data and satellite imagery data. It then imports park drawings and BIM models to generate a digital model of the industrial park and integrates real-time data from all elements to obtain a digital twin of the industrial park. Specifically: First, a 3D geographic base is constructed. A digital elevation model (DEM) with a resolution of 0.5m and satellite imagery data with a resolution of 0.2m are collected for the park. A 3D geographic base is built based on CesiumJS to achieve a realistic restoration of the terrain. Then, the park's CAD master plan is imported, and coordinate registration and vectorization are performed to generate 3D outline models of the park's infrastructure, such as roads, buildings, pipelines, and walls. After completion, texture mapping and lighting rendering are performed on the geographic base to support real-time simulation of weather effects such as rain, snow, fog, and day and night.

[0089] Next, the physical models are imported and mapped: BIM models of production equipment such as storage tanks, reactors, pumps, and compressors in the park are imported, with a model accuracy better than 0.1m, retaining the key structural and attribute information of the equipment; a unique identifier code is assigned to each physical entity, and a one-to-one mapping relationship table of "physical entity - digital model" is established to record the entity's name, type, location, and process unit to which it belongs; a three-dimensional model of the park's underground pipeline network is constructed, and parameters such as the material, diameter, pressure, and flow direction of the pipeline network are labeled to achieve a transparent display of the underground pipeline network.

[0090] Then, real-time data access for all elements is implemented: a WebSocket real-time data channel is established to receive real-time data pushed by the platform layer, including sensor monitoring data, equipment operating status data, personnel location data, meteorological data, risk level data, etc.; the real-time data is mapped onto the corresponding digital model, and the physical status is displayed intuitively through color, size, animation, etc.: green is displayed when the equipment is running normally, yellow is displayed when it is abnormal, and red is displayed when it is faulty; personnel location is updated in real time through a blue person icon, and personnel trajectory playback is supported; gas concentration is displayed through a gradient color cloud map, with the color becoming darker as the concentration increases.

[0091] Finally, the interactive functions developed include enabling basic interactive operations such as free roaming, scaling, rotation, and sectioning of the digital twin, supporting switching between first-person and third-person perspectives; providing layer management functions, which can individually show / hide different layers such as equipment layer, pipeline layer, personnel layer, risk layer, and video layer; and supporting clicking on any digital model to view detailed information about the entity, including equipment parameters, historical data, maintenance records, and associated risks.

[0092] The early warning and display module receives risk warning information pushed by the platform layer and generates visualized warning effects in the digital twin of the industrial park. Specifically, it first receives risk warning information pushed by the platform layer in real time through the message queue Kafka, including fields such as warning ID, warning level, warning type, occurrence time, geographical location, scope of impact, risk description, and suggested handling measures. Then, it divides the warnings into four levels according to the risk level value, including: Blue alert (general risk): 0.2 ≤ <0.4; Yellow alert (significant risk): 0.4 ≤ <0.6; Orange alert (major risk): 0.6 ≤ <0.8; Red Alert (Extremely Serious Risk): 0.8 ≤ <1.0.

[0093] Next, the risk area is highlighted in the digital twin, and a flashing warning icon of the corresponding color is displayed according to the warning level. A red warning is accompanied by a full-screen flashing prompt. A warning details window pops up, displaying the basic information of the warning, the root cause analysis results, the risk propagation path, and the suggested response measures. A risk distribution heat map and a propagation path map are overlaid, with arrows of different colors indicating the direction of risk propagation and the thickness of the arrows indicating the propagation intensity. Real-time video from all surveillance cameras in and around the risk area is automatically retrieved, and a multi-window video monitoring interface pops up in the digital twin, supporting zooming in and out of the video and pan-tilt control.

[0094] The emergency command and dispatch module calls the pre-built emergency plan library, matches it with the risk warning information, generates the optimal emergency plan, and assigns it to the corresponding operation and maintenance personnel; The risk report generation module is used to establish a standardized report template library. Based on risk warning information, it automatically fills in the standardized report templates to generate risk reports. The risk report content includes warning data: number of warnings, level distribution, type distribution, regional distribution, and handling status; hidden danger data: number of hidden dangers reported, number of rectifications, rectification rate, and overdue hidden dangers; equipment data: equipment online rate, number of failures, runtime, and maintenance records; and personnel data: number of personnel violations, training status, and inspection records.

[0095] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. An intelligent monitoring and early warning system for safety production risks in industrial parks, characterized by: It includes a perception layer, a network layer, a platform layer, and an application layer. The perception layer deploys perception devices at major hazard sources and key locations in the industrial park to collect various types of perception data and generate multi-source datasets. The network layer constructs communication and transmission between the perception layer, platform layer, and application layer through a network architecture deployed in the industrial park. The platform layer receives multi-source datasets from the perception layer, performs preprocessing and data analysis to obtain risk warning information for the industrial park, and then transmits it to the application layer. The application layer uses the digital twin of the industrial park to display risk warning information and conduct emergency command and dispatch, and generates risk reports; among which: The platform layer includes a data center and a computing platform. The data center receives multi-source datasets transmitted from the perception layer, stores and preprocesses them, and generates preprocessed data. The computing platform includes a spatiotemporal grid coding module, a cross-modal causal attention fusion module, a spatiotemporal coupled risk modeling module, and a dynamic risk propagation and tracing module. The spatiotemporal grid coding module divides the industrial park plane into spatiotemporal grid units and encodes them. The preprocessed data is then mapped to the spatiotemporal grid units. The cross-modal causal attention fusion module independently mines the intermodal direct causal relationships with time delay in each spatiotemporal grid cell through a preset cross-modal feature mining mechanism, calculates the average direct causal effect and constructs a spatiotemporal direct causal matrix that dynamically changes in spatial location and time, and fuses multimodal features through an end-to-end causal attention mechanism that embeds the spatiotemporal direct causal matrix to generate a cross-modal fusion feature vector. The spatiotemporal coupled risk modeling module constructs a real-time updated spatial risk map of the industrial park. It models the bidirectional causal interaction between spatial risk distribution and temporal evolution trend through a two-way spatiotemporal causal interaction layer, and introduces a risk accumulation dynamics model to calculate the final risk level value of each spatiotemporal grid unit. The dynamic risk propagation and tracing module constructs an adaptive causal dynamic Bayesian network, and calculates primary and secondary risks by combining a multi-dimensional root cause ranking algorithm based on posterior probability, causal betweenness centrality, and propagation time, generating a hierarchical risk propagation path diagram and risk warning information based on root cause tracing results.

2. The intelligent monitoring and early warning system for safety production risks in industrial parks according to claim 1, characterized in that: The cross-modal causal attention fusion module includes a single-modal feature extraction unit, a spatiotemporal heterogeneous causal discovery unit, a direct causal effect calculation unit, and a causal attention fusion unit, wherein: The single-modal feature extraction unit is used to obtain grid-level feature vectors from each spatiotemporal grid cell. The feature vectors of m modalities are extracted and concatenated to obtain the initial feature vector of each spatiotemporal grid cell. ; The spatiotemporal heterogeneous causality discovery unit introduces a time delay parameter. The PC algorithm is used to mine causal pairs with time delay. ,in As a causal mode, The resulting mode; and a local causal graph for each spatiotemporal grid cell is constructed. Where V is the modality set, The set of causal edges is the union of all time delays, and each set of causal edges consists of causal edges. composition, The preset time delay amount; Indicates grid i, time t, and delay Below, mode k is paired with mode The existence of causal relationships; The direct causal effect calculation unit is used to calculate each causal edge. The backdoor adjustment formula is used to calculate mode k for modes. In the spacetime grid Placement, delay Average direct causal effect The expression is: in, For causal budgeting, Adjust the variable set for the backdoor; For mathematical expectation; The direct causal effect calculation unit is also used to construct the spatiotemporal direct causality matrix and the time delay weight matrix. The expression is: Time Delay Weight Matrix The expression is: The time delay to generate the maximum ADE value; The causal attention fusion unit is used to fuse multimodal features through an end-to-end causal attention mechanism that embeds a spatiotemporal direct causal matrix to generate a cross-modal fused feature vector. Specifically: Initial eigenvectors for each spatiotemporal grid cell Perform a linear transformation to generate the query matrix. Key matrix Sum matrix , , , The weight matrix is ​​a learnable weight matrix; To calculate the time-delay-weighted causal attention score, the spatiotemporal direct causality matrix and the time-delay weight matrix are simultaneously embedded into the attention mechanism, expressed as: in, The dimension of the key matrix is... Hadamard product, causal matrix Filtering spurious associations, time delay weighting Strengthen the causal relationships that are time-sensitive; Calculate attention output: Introducing a causal regularization term and performing end-to-end joint optimization, the total loss function is: in, To predict cross-entropy loss for risk, To discover loss in causal discovery, A true causal matrix constructed for domain knowledge. For the sparseness loss of the causal matrix, , The regularization coefficient is used. Output cross-modal causal fusion feature vector: in, This is a cross-modal causal fusion feature vector.

3. The intelligent monitoring and early warning system for safety production risks in industrial parks according to claim 2, characterized in that: The spatiotemporal coupled risk modeling module includes a directed dynamic spatiotemporal causal graph construction unit, a bidirectional spatiotemporal causal interaction unit, and a risk calculation unit, wherein: The directed dynamic spatiotemporal causal graph construction unit is used to construct a directed dynamic spatiotemporal causal graph of an industrial park that is updated in real time, specifically as follows: Define the dynamic spatiotemporal causal graph of the industrial park as follows: , where the set of nodes For all spatiotemporal grid cells, the set of edges This represents the directed risk propagation relationship between spatiotemporal grid cells at time t; Calculate the directed edge weights from grid i to grid j at time t. It also considers geographical distance, real-time causal effects, wind direction influence, and process correlation, with the expression being: in, Let be the Euclidean distance between grid i and grid j. Let i be the spatiotemporal direct causal matrix. and the direct causal matrix of spacetime j The strength of real-time causal association is obtained by calculating the difference between them; Wind direction is an influencing factor. For equipment and process correlation factors, , , , These are the initial learnable parameters; Construct the adjacency matrix at time t ,in If the preset constraint value is exceeded, then And recalculate the edge weights and adjacency matrix every minute to perform real-time dynamic updates of the graph structure; The bidirectional spatiotemporal causal interaction unit is used to model the bidirectional causal interaction between spatial risk distribution and temporal evolution trends, specifically: Cross-modal feature vector fusion Input a two-layer graph attention network (GAT), with each layer containing 8 attention heads, and output a 256-dimensional feature vector. ; The dilation rate of the temporal convolution is dynamically adjusted based on the spatial risk distribution at the current moment. , This is a rounding function; A three-layer multi-scale temporal convolutional network, MSTCN, is used as input, taking spatial feature sequences of 10 consecutive time windows as input, and outputting a 256-dimensional temporal feature vector. ; It is a spatial feature sequence of 10 consecutive time windows; By leveraging historical evolution trends, the weights of spatial graph attention can be dynamically adjusted. : The original spatial attention coefficients, Let i be the rate of change of risk. A gating fusion mechanism is used to fuse spatial and temporal features: For gating weights, and For the spatial and temporal characteristics after interaction, The spatiotemporal coupling characteristics after fusion; The risk calculation unit is used to calculate the final risk level value of each spatiotemporal grid cell through a risk accumulation dynamics model, specifically: The spatiotemporal coupling features are input into two fully connected layers, and the instantaneous risk value is output through the sigmoid activation function. The expression is: in, , This is the weight matrix. , For bias terms, The symbol for the sigmoid activation function; By introducing a forgetting factor, the cumulative effect and decay characteristics of risk are modeled to obtain the cumulative risk amount. The expression is: Forgetting factor, Let $t$ be the cumulative risk of grid $i$ at time $t-1$, with an initial cumulative risk of $0$. The risk outbreak threshold is dynamically determined based on the historical risk distribution of each grid, expressed as: To adapt to the risk outbreak threshold, The historical risk average of grid i over the past 30 days. The standard deviation of grid i over the past 30 days. For safety factor; Output the final risk level value: This is the final risk level value.

4. The intelligent monitoring and early warning system for safety production risks in industrial parks according to claim 3, characterized in that: The dynamic risk propagation and tracing module includes an adaptive structure causal dynamic Bayesian network construction unit, a risk propagation path dynamic prediction unit, and a multi-dimensional root cause tracing and hierarchical propagation unit, wherein: The adaptive structure causal dynamic Bayesian network building unit constructs a causal dynamic Bayesian network based on knowledge in the field of chemical safety and historical accident data. The causal dynamic Bayesian network consists of a set of risk factor nodes and a set of risk event nodes. The expression of the causal dynamic Bayesian network is as follows: This is the initial Bayesian network structure. This is the initial conditional probability table. This is an initial conditional probability table across time slices, learned from historical time-series data based on the introduced time slices. and Conduct online learning updates; The risk propagation path dynamic prediction unit is used to predict the future propagation path, scope of impact, and arrival time of a risk in real time based on a causal dynamic Bayesian network when an anomaly in a risk factor is detected, and to generate a risk propagation path map, specifically: When the final risk level value of a risk factor node exceeds the preset first risk threshold, it is determined to be an abnormal node, and its observation status is input into the causal dynamic Bayesian network. A joint tree inference algorithm is used for forward filtering and prediction to calculate the future. The posterior probability of risk factor nodes in each time slice; Based on the average propagation time between risk factor nodes in historical accidents, the risk propagation speed and arrival time are calculated. For each predicted time slice, the risk impact range is defined as the set of spatiotemporal grid cells corresponding to risk factor nodes with a posterior probability greater than a preset probability threshold. The geographical area of ​​this set and the number of devices involved are calculated to generate a risk propagation path map in the form of a directed acyclic graph. The multi-dimensional root cause tracing and hierarchical propagation unit is used to locate the root cause of risk factor anomalies based on causal dynamic Bayesian networks when anomalies are detected, distinguish between primary and secondary risks, and generate a hierarchical propagation path diagram, specifically as follows: When the final risk level value of a spatiotemporal grid cell is greater than the second risk threshold, the corresponding risk event node is taken as the target node. The joint tree backward reasoning algorithm is used to calculate the posterior probability of all potential root cause nodes and the causal betweenness centrality of potential root cause nodes. Based on posterior probability, causal betweenness centrality, and propagation time, a comprehensive score for root cause nodes is calculated, and primary and secondary risk nodes are identified to generate a hierarchical propagation path map.

5. The intelligent monitoring and early warning system for safety production risks in industrial parks according to claim 1, characterized in that: The spatiotemporal grid coding module divides the industrial park's plane into spatiotemporal grid cells and encodes them. The preprocessed data is then mapped into these spatiotemporal grid cells as follows: The industrial park's layout is divided into square grid cells of equal size, and a unique spatial code is generated for each grid cell. Divide the continuous time axis into non-overlapping fixed-length time windows and generate a unique time code for each time window; By combining spatial and temporal codes, a globally unique spatiotemporal grid code is generated; The preprocessed multi-source heterogeneous data is mapped to corresponding spatiotemporal grid cells according to its geographical location and collection time, and feature aggregation is performed to generate standardized grid-level feature vectors.

6. The intelligent monitoring and early warning system for safety production risks in industrial parks according to claim 1, characterized in that: The application layer includes a digital twin construction module, an early warning and display module, an emergency command and dispatch module, and a risk report generation module, wherein: The digital twin construction module constructs a three-dimensional geographic base for the industrial park based on high-precision terrain data and satellite imagery data, imports park drawings and BIM models, generates a digital model of the industrial park, and connects to real-time data of all elements to obtain a digital twin of the industrial park. The early warning and display module is used to receive risk warning information pushed by the platform layer and generate a visual warning effect in the digital twin of the industrial park; The emergency command and dispatch module calls the pre-built emergency plan library, matches it with the risk warning information, generates the optimal emergency plan, and assigns it to the corresponding operation and maintenance personnel; The risk report generation module is used to establish a standardized report template library. It automatically fills in the standardized report templates based on risk warning information to generate risk reports.

Images