Chemical accident emergency response method and system under real-time monitoring

Abstract

The invention provides a chemical accident emergency response method and system under real-time monitoring, and relates to the technical field of emergency response, and the method comprises the steps: carrying out the time alignment through collecting multi-source heterogeneous data in real time, and generating multi-source synchronous data; extracting multi-modal features of the multi-source synchronous data to carry out chemical accident identification; accident scene parameters are called, accident response analysis is carried out based on an accident identification result, an emergency response scheme is formulated, resource scheduling analysis is carried out, and a resource scheduling network diagram is constructed; and performing accident monitoring and early warning according to the resource scheduling network diagram in combination with the multi-source synchronous data, generating a real-time decision instruction, and pushing the real-time decision instruction and an emergency response scheme to an emergency command center. According to the method and the device, the technical problems of low chemical accident identification accuracy and response flow lag caused by lack of a multi-source heterogeneous data fusion and real-time linkage mechanism in the prior art are solved, and the technical effects of improving the chemical accident identification precision and the emergency response speed are achieved.

Description

Technical Field

[0001] The present application relates to the field of emergency response technology, and specifically to a method and system for emergency response to chemical accidents under real-time monitoring. Background Art

[0002] Establishing an efficient and intelligent chemical accident emergency response mechanism in the process of chemical production, storage, transportation and disposal is of great significance for protecting people's lives and property and the ecological environment. At present, chemical accident emergency response mainly relies on traditional sensing equipment and preset response plans, based on single sensor data or manual alarm signals to make accident judgments, and manually command and deploy resources to perform response tasks. However, in actual applications, this type of method lacks the ability to fully perceive the accident scene due to the lack of coordinated collection and real-time fusion of multiple types of data such as sensor data and video surveillance data, which in turn affects the accurate identification of accident types and risk levels. At the same time, existing methods mostly rely on fixed processes and static plans, lack a mechanism for linkage with real-time scenarios, and the links between each link from accident monitoring to emergency response are not tightly connected, the response process lags, and cannot meet the needs of fast and accurate emergency response. Summary of the Invention

[0003] This application provides a chemical accident emergency response method and system under real-time monitoring, which solves the technical problems of the existing technology such as low chemical accident identification accuracy and delayed response process due to the lack of multi-source heterogeneous data fusion and real-time linkage mechanism, and achieves the technical effect of improving the accuracy of chemical accident identification and the speed of emergency response.

[0004] In view of the above problems, on the one hand, the present application provides a chemical accident emergency response method under real-time monitoring, which includes: real-time collection of multi-source heterogeneous data, time alignment based on the multi-source heterogeneous data, and generation of multi-source synchronous data; multi-modal deep learning based on the multi-source synchronous data, extraction of multi-modal features, chemical accident identification according to the multi-modal features, and acquisition of accident identification results; retrieving accident scene parameters, performing accident response analysis based on the accident identification results, formulating an emergency response plan, executing the emergency response plan to perform resource scheduling analysis, and constructing a resource scheduling network diagram; performing accident monitoring according to the resource scheduling network diagram in combination with the multi-source synchronous data, triggering an early warning signal according to the monitoring results, generating real-time decision instructions, and pushing the real-time decision instructions in combination with the emergency response plan to the emergency command center for emergency response to the chemical accident.

[0005] On the other hand, the present application also provides a chemical accident emergency response system under real-time monitoring, the system including: a data acquisition module for real-time acquisition of multi-source heterogeneous data, time alignment based on the multi-source heterogeneous data, and generation of multi-source synchronous data; an accident identification module for performing multimodal deep learning based on the multi-source synchronous data, extracting multimodal features, identifying chemical accidents according to the multimodal features, and obtaining accident identification results; a response analysis module for retrieving accident scene parameters, performing accident response analysis based on the accident identification results, formulating an emergency response plan, executing the emergency response plan for resource scheduling analysis, and constructing a resource scheduling network diagram; an emergency response module for performing accident monitoring according to the resource scheduling network diagram in combination with the multi-source synchronous data, triggering an early warning signal based on the monitoring results, generating real-time decision instructions, and pushing the real-time decision instructions in combination with the emergency response plan to the emergency command center for emergency response to chemical accidents.

[0006] One or more technical solutions provided in this application have at least the following beneficial effects:

[0007] By collecting heterogeneous data from multiple sources in real time and performing time alignment to generate multi-source synchronized data, the system ensures synchronization and consistency across different sources, providing a comprehensive and accurate data foundation for subsequent accident analysis and decision-making, avoiding the biased judgments caused by a single data source. Multimodal deep learning is performed on this multi-source synchronized data to extract multimodal features for chemical accident identification. This step utilizes deep learning techniques to extract features from multimodal data, enabling more precise identification of key chemical accident information, improving the accuracy of accident identification and providing a basis for subsequent emergency response. By retrieving accident scenario parameters, conducting accident response analysis, developing emergency response plans, and executing resource scheduling analysis, a resource scheduling network diagram is constructed. This facilitates more efficient deployment of emergency resources, ensuring that resources can be quickly and appropriately dispatched to the accident site after an accident occurs, improving the efficiency of the emergency response and minimizing losses. Accident monitoring is performed based on the resource scheduling network diagram and combined with multi-source synchronized data, triggering early warning signals, generating real-time decision-making instructions, and transmitting them to the emergency command center, forming a dynamic closed-loop monitoring and response mechanism. By continuously monitoring the accident situation and combining it with the resource scheduling network diagram, changes and anomalies in the accident can be discovered in a timely manner, triggering early warning signals, enabling decision makers to quickly generate and adjust decision instructions based on real-time information, ensuring that emergency response measures can be implemented in a timely and effective manner, and enhancing the flexibility and real-time nature of responding to accidents.

[0008] In summary, this application achieves accurate identification and rapid judgment of chemical accidents through the real-time collection and fusion of multi-source heterogeneous data, combined with multimodal deep learning technology. At the same time, the resource scheduling network diagram constructed based on accident scenarios and resource scheduling analysis optimizes the resource allocation and scheduling efficiency of emergency response. Finally, through real-time monitoring, early warning, and rapid push of decision-making instructions, a closed-loop emergency response mechanism is formed, enabling emergency response to have continuous linkage and intelligent adjustment capabilities. Overall, this solution effectively improves the identification accuracy, response speed, and handling coordination capabilities of chemical accidents, providing technical support and intelligent guarantees for emergency management in high-risk scenarios.

[0009] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 This is a flow chart of a method for emergency response to chemical accidents under real-time monitoring provided in an embodiment of the present application.

[0011] Figure 2 This is a schematic diagram of the structure of a chemical accident emergency response system under real-time monitoring provided in an embodiment of the present application.

[0012] Explanation of the reference numerals: data collection module 10 , accident identification module 20 , response analysis module 30 , emergency response module 40 . DETAILED DESCRIPTION

[0013] The embodiments of the present application provide a chemical accident emergency response method and system under real-time monitoring, thereby solving the technical problems in the prior art caused by the lack of multi-source heterogeneous data fusion and real-time linkage mechanism, resulting in low chemical accident identification accuracy and delayed response process, thereby achieving the technical effect of improving the accuracy of chemical accident identification and the speed of emergency response.

[0014] Example 1, as Figure 1 As shown, the embodiment of the present application provides a chemical accident emergency response method under real-time monitoring, the method comprising:

[0015] Step S100: collecting multi-source heterogeneous data in real time, performing time alignment based on the multi-source heterogeneous data, and generating multi-source synchronized data.

[0016] Specifically, multi-source heterogeneous data includes data from the hazardous chemical risk monitoring and early warning system, basic chemical information in the enterprise basic information database, and video surveillance network data. Among them, the hazardous chemical risk monitoring and early warning system data includes real-time monitoring data uploaded by devices such as toxic gas sensors, environmental parameter collectors (temperature, humidity, wind speed, etc.), and hazardous material storage tank level and pressure sensors. These data are usually in structured or semi-structured formats and have strong temporal characteristics. The enterprise basic information database contains static structured data such as the registration information of hazardous chemical users, storage locations, types of hazardous materials, emergency contacts, emergency material allocation, and historical safety records. Video surveillance network data includes on-site video image streams, covering key areas inside and outside the factory, such as storage tank areas, loading and unloading areas, and transportation routes. The data is an unstructured video frame stream or image sequence.

[0017] Through data interfaces, the system connects to the hazardous chemical risk monitoring and early warning system, the enterprise basic information database, and the video surveillance network, collecting these multi-source heterogeneous data in real time. The structured data is then standardized into a unified format (e.g., JSON, CSV). The video data is frame-processed, key frames are extracted (e.g., 1 frame per second), and the corresponding timestamps are recorded. A sliding time window (e.g., 5 seconds) is introduced to map the multi-source heterogeneous data to a unified time anchor point based on the timestamp. If any data type is missing, linear interpolation or historical averages are used to compensate for it, aligning all data in time. After all data are time-series aligned, they form a multi-source synchronized data structure. Each time point in the multi-source synchronized data contains corresponding data from different data sources, facilitating subsequent comprehensive analysis and processing.

[0018] This step collects multi-source heterogeneous data in real time from hazardous chemical risk monitoring and early warning systems, enterprise basic information databases, and video surveillance networks, and performs time alignment to generate multi-source synchronous data. This provides a comprehensive, accurate, and time-consistent data foundation for subsequent chemical accident identification and emergency response, enabling the subsequent emergency response process to comprehensively consider various information at the accident site, thereby more accurately judging the accident situation and improving the efficiency and accuracy of emergency response.

[0019] Step S200: performing multimodal deep learning based on the multi-source synchronous data, extracting multimodal features, and performing chemical accident identification according to the multimodal features to obtain an accident identification result.

[0020] Specifically, multimodal deep learning is a method that uses deep learning algorithms to jointly analyze and learn data from multiple modalities (such as images, time series signals, and text). It can automatically mine the correlations and deep features between data from different modalities to more comprehensively understand complex scenarios.

[0021] Build a multi-channel deep learning model, such as a fusion architecture that uses convolutional neural networks to process image and video data, recurrent neural networks or long-short-term memory networks to process time-series sensor data, and natural language processing models to process text data. This fusion architecture is trained using a large amount of labeled and unlabeled chemical accident-related data to learn feature representations for data in different modalities. These features are then fused using methods such as attention mechanisms or weighted averaging to output multimodal features. For example, a multimodal deep learning model can be constructed using a deep learning framework such as TensorFlow or PyTorch. Three-dimensional convolution analysis can be performed on video stream data to extract spatiotemporal features, long-short-term memory network analysis can be performed on sensor time-series signals to extract time series features, and natural language processing based on word embeddings and attention mechanisms can be performed on structured text data to extract semantic keyword features. These features are then fused through weighted fusion (weights can be automatically learned by the model or manually set based on feature importance) to construct multimodal features.

[0022] The multi-source synchronous data generated in step S100 is input into a multi-channel deep learning model to extract representative and discriminative feature vectors to obtain multimodal features. These features can comprehensively reflect various key information of the current accident, such as flame color, size, and shape features extracted from the video, gas leakage rate and concentration peak features extracted from sensor data, accident type description and urgency keyword features extracted from the text, etc. Based on the constructed multimodal features, a classification algorithm (such as a support vector machine, a decision tree, etc.) is used or a fully connected classification layer is added to the last layer of the multi-channel deep learning model to identify chemical accidents, and output an accident identification result including the current accident type (such as fire, leakage, explosion, etc.) and the severity of the accident (such as minor, general, major, etc.).

[0023] This step realizes multimodal information fusion intelligent perception, effectively improves the recognition accuracy and judgment speed of accident scenes, provides a key basis for the subsequent formulation of reasonable emergency response plans, and improves the pertinence and timeliness of emergency response.

[0024] Step S300: Retrieve accident scenario parameters, perform accident response analysis based on the accident identification result, formulate an emergency response plan, execute the emergency response plan to perform resource scheduling analysis, and construct a resource scheduling network diagram.

[0025] Specifically, accident scenario parameters describe the various states and conditions at the accident scene, including the location and time of the accident, the type and quantity of chemicals, the on-site meteorological conditions (wind direction, wind speed, temperature, etc.), and the surrounding environment (buildings, population density, traffic conditions, etc.). For example, an accident occurred in a large chemical warehouse storing a variety of hazardous chemicals. The wind direction was northeast at a speed of 5m / s at the time of the accident, and there were residential areas within 500 meters.

[0026] The current accident scenario parameters are retrieved from data sources such as the enterprise's safety management information system, geographic information system (GIS), and meteorological department data interface. Based on the accident identification results obtained in step S200, a pre-established accident response model (such as one based on reinforcement learning modeling) is used to perform accident response analysis, and an optimal emergency response plan is formulated in the state-action space, including task allocation, personnel evacuation paths, resource calls, etc. Subsequently, rescue resource distribution, traffic information, and team locations are extracted from the geographic information system, and a resource scheduling network diagram is constructed using graph theory tools (such as Dijkstra or A-STAR algorithms), with nodes representing sites or resources, edges representing travel paths, and edge weights representing travel time. The resource scheduling network diagram graphically displays the scheduling paths, nodes, and priorities of emergency rescue resources, intuitively presents the flow process and time schedule of resources from the supply point to the accident site, and provides decision support for emergency commanders to schedule resources.

[0027] Step S400: Accident monitoring is performed according to the resource scheduling network diagram in combination with the multi-source synchronous data, an early warning signal is triggered according to the monitoring results, a real-time decision instruction is generated, and the real-time decision instruction is pushed to the emergency command center in combination with the emergency response plan to perform an emergency response to the chemical accident.

[0028] Specifically, during the emergency response process, combined with the resource scheduling network diagram constructed in step S300 and the multi-source synchronous data constructed in step S100, the accident site and surrounding environment are continuously monitored in real time to track the development of the accident, including changes in the scope of chemical leakage and diffusion, fire spread trends, changes in on-site meteorological conditions, casualties, etc. The real-time monitoring results are compared with preset monitoring indicator thresholds (such as gas concentration thresholds, flame height thresholds, etc.). When the real-time monitoring results exceed the thresholds, the early warning rule engine is used to trigger the corresponding early warning signal, and the strategy engine is used to generate real-time decision instructions for guiding the adjustment and optimization of on-site rescue operations. The real-time decision instructions are integrated with the emergency response plan and pushed to the emergency command center through the emergency command communication system (such as satellite communication, cluster intercom, emergency command and dispatch software, etc.) for manual confirmation or automatic execution.

[0029] This step realizes dynamic closed-loop control during the accident process, provides real-time decision support for emergency commanders, ensures that emergency response measures can be flexibly adjusted according to the development of the accident, and enhances the flexibility and adaptability of emergency response.

[0030] Furthermore, step S100 includes:

[0031] Step S110: Real-time data collection is performed through a multi-source data interface to obtain multi-source heterogeneous data.

[0032] Step S120: dynamically setting the time window length to construct a time window, and using the time window to align the multi-source heterogeneous data according to timestamps to generate an aligned data set.

[0033] Step S130: performing timestamp offset analysis on the aligned data set to generate multiple offsets.

[0034] Step S140: Determine whether the multiple offsets exceed the time window length. If any offset among the multiple offsets exceeds the time window length, trigger a data compensation instruction.

[0035] Step S150: performing missing compensation on the aligned data set through the data compensation instruction to obtain the multi-source heterogeneous data.

[0036] Specifically, by configuring the industrial bus interface, network API interface, database connector or message queue mechanism, heterogeneous data such as sensor monitoring data, enterprise static information and video images from multiple data sources such as hazardous chemical risk monitoring and early warning systems, enterprise basic information databases and video surveillance networks are accessed and collected. During the collection process, a unified timestamp is added to each data item to achieve real-time and synchronous access to multi-source data.

[0037] According to the upload delay characteristics and frequency distribution of various data sources, the time window length is dynamically adjusted, for example, set to 5 seconds or 10 seconds, and multi-source data is time-matched within each time window to form a combined structure of video frames, sensor data and enterprise attribute information in the same window, and construct an aligned data set.

[0038] For each data source in the aligned dataset, the time difference between the actual timestamp of each data item and the center anchor point of its time window is calculated. This time difference is the timestamp offset, which is used to measure the degree of temporal deviation of data from different data sources. For example, if a sensor data item is recorded at 10:31:03 and its corresponding window anchor time is 10:31:00, the offset is +3 seconds.

[0039] The timestamp offset of each data source is compared with the time window length to determine whether the offset exceeds the time window length. If any of the multiple offsets exceeds the time window length, the data item is considered invalid or lost, triggering a data compensation instruction. This instruction contains information such as the data source identifier that needs to be compensated and the time range of the missing data, which is used to guide subsequent data compensation operations.

[0040] Based on data compensation instructions, differentiated compensation strategies are applied to different data types, including linear interpolation, adjacent-time substitution, historical data infill, or model-based predicted value substitution, to complete missing or invalid data. For example, for hazardous chemical sensor data, linear interpolation can be used to estimate the data value at the missing time point based on existing sensor data points within the adjacent time window. For missing video frames, the previous frame can be used as a replacement. For missing enterprise database fields, the default or latest value can be used. Alternatively, if the data source supports data retransmission, the data source can be requested to resend data within a specified time range. The compensated data is then replaced or appended to the original aligned dataset, forming a structurally complete and temporally consistent multi-source synchronized dataset.

[0041] The above steps enable efficient access, dynamic time-series coordination, and robust compensation fusion of multi-source heterogeneous data, building a highly consistent, low-latency, multi-source synchronized dataset. This dataset, serving as the foundation for subsequent incident identification and response analysis, effectively addresses issues such as data silos, inaccurate synchronization, and significant latency in traditional systems, fundamentally improving the timeliness and accuracy of intelligent chemical incident perception.

[0042] Furthermore, step S120 includes:

[0043] Step S121: Retrieve historical transmission delay statistics of multiple data sources in the multi-source data interface, and initialize a reference length of a time window according to the historical transmission delay statistics.

[0044] Step S122: monitoring the transmission jitter rates of the plurality of data sources in real time, and generating an extended parameter when the transmission jitter rate exceeds a preset threshold.

[0045] Step S123: Dynamically extend the reference length of the time window according to the extension parameter to set the time window length.

[0046] Specifically, by calling the historical data interface, the average transmission delay and fluctuation range of various data sources (such as video streams, sensor sampling signals, and database records) over a period of time are obtained. The initial benchmark length of the time window is determined based on the historical transmission delay statistics of each data source. The maximum value, average value, or other representative statistics of the historical transmission delay of each data source are usually selected as a reference for the benchmark length of the time window. For example, if the average delay of sensor data is 1.8 seconds and the average delay of video frame data is 2.5 seconds, the initial length of the benchmark window can be set to 3 seconds to accommodate the normal delay range of the data.

[0047] The transmission jitter rate refers to the degree of fluctuation in transmission delay during data transmission from a data source. It reflects the stability of data transmission. The transmission jitter rate is calculated using a sliding window statistical method, using the standard deviation or coefficient of variation of the transmission jitter. If the transmission jitter rate exceeds a preset threshold, it is considered that the current data transmission fluctuation is severe and the time window needs to be adjusted. At this time, an extension parameter is automatically generated. This extension parameter is a proportional value that indicates the percentage by which the time window length should be increased from the initial value. For example, a setting of 20% means that the time window length should be increased by 20% from the base length.

[0048] The time window length is increased by the ratio specified by the expansion parameter. For example, when the base length is 3 seconds and the expansion parameter is 1.5, the final time window length is set to 4.5 seconds. This expansion method can automatically tolerate a wider range of data delays during abnormal data fluctuations, improving the robustness of time alignment.

[0049] Through the above steps, adaptive analysis and window adjustment of multi-source heterogeneous data transmission characteristics can be achieved, ensuring that the time tolerance in the data alignment process is reasonable and dynamic, and effectively enhancing the accuracy of data fusion.

[0050] Furthermore, step S200 includes:

[0051] Step S210: Input the multi-source synchronized data set into a multimodal deep learning network for multimodal analysis to obtain multimodal data, where the multimodal data includes video stream data, sensor timing signals, and structured text data.

[0052] Step S220: performing three-dimensional convolution analysis based on the video stream data to extract spatiotemporal features.

[0053] Step S230: performing a bidirectional analysis based on the sensor time series signal to extract time series features.

[0054] Step S240: performing natural language processing based on the structured text data to extract semantic keyword features.

[0055] Step S250: performing weighted fusion on the spatiotemporal features, the time series features, and the semantic keyword features to construct the multimodal features.

[0056] Specifically, multimodal data sets are fed into a multimodal deep learning network for analysis, yielding multimodal data. This data includes image sequences (video stream data) from video surveillance, time series signals (sensor timing signals) from hazardous chemical risk monitoring and early warning systems, and structured text data from enterprise basic information databases. After time alignment, these data are synchronized and can be fed into the multimodal learning pipeline in parallel.

[0057] In the video stream data processing pipeline, a 3D convolutional neural network performs spatiotemporal image analysis, extracting the spatial location and dynamic characteristics of accident precursors (such as flashes of flame and smoke spread) within continuous footage. For example, if a camera detects the rapid spread of smoke in a storage tank area, the 3D convolutional neural network can determine its speed and direction.

[0058] In the sensor data processing pipeline, a bidirectional long short-term memory (LSTM) network is used to analyze sensor time series signals, extracting temporal dependencies and trend characteristics between variables, such as abnormal mutations or regular fluctuations in sensor signals like temperature, pressure, and concentration. For example, for sensor data from a gas leak in a storage tank, a bidirectional LSTM network can simultaneously consider past and future data points to capture temporal patterns in gas concentration, such as the rapid concentration increase at the beginning of a leak, concentration fluctuations during the stable leak period, and possible concentration mutations. This helps predict potential development trends of an accident.

[0059] In the text data processing channel, semantic analysis of structured text is performed based on natural language processing models such as BERT to extract representative keywords and semantic tags, such as "corrosive liquid", "excess storage", "explosion critical value" and other descriptions closely related to the accident.

[0060] The spatiotemporal features, time series features, and semantic keyword features output by the three channels are weighted and fused to construct a unified multimodal feature vector. The weighting strategy can be adjusted based on the performance differences between the data sources during the training phase, using an attention mechanism or projection fusion algorithm to form a holistic perception of the chemical accident status.

[0061] Through the above-mentioned multimodal deep learning processing, not only the spatiotemporal and semantic information of multi-source data is effectively integrated, but also a highly reliable input basis is provided for accident identification and response strategy formulation in subsequent steps, significantly improving the intelligent recognition capability of complex and hidden accident situations.

[0062] Furthermore, step S300 includes:

[0063] Step S310: performing vectorization based on the accident scene parameters to define a state space, performing emergency action matching based on the accident identification result to define an action space.

[0064] Step S320: performing reinforcement learning on the state space in combination with the action space to formulate an emergency response plan.

[0065] Step S330: Perform resource scheduling analysis according to the emergency response plan to obtain rescue resource distribution data, traffic network information, and real-time location of the rescue team.

[0066] Step S340: construct multiple nodes based on the rescue resource distribution data, construct multiple edges of the multiple nodes based on the traffic network information, assign weights to the multiple edges based on the real-time location of the rescue team, and determine multiple travel time weights.

[0067] Step S350: Associating and integrating the multiple nodes, the multiple edges, and the multiple travel time weights to construct the resource scheduling network graph.

[0068] Specifically, accident scenario parameters related to the accident, including key information such as chemical type, location, meteorological conditions, and surrounding environment, are retrieved and vectorized to form a state space representing the state of the accident environment. Simultaneously, based on the accident identification results in step S200, a set of pre-defined emergency response actions (such as lockdown, evacuation, cooling, fire extinguishing, and chemical neutralization) is matched from the emergency action library to construct a set of optional response measures, thereby defining the action space.

[0069] Based on the established state and action spaces, a reinforcement learning model (such as DQN and PPO) is constructed. With the goal of minimizing incident response time and optimizing resource allocation, a set of optimal action sequences for the current state is trained to develop an emergency response plan. For example, the model can evaluate the expected benefits of different response action combinations in terms of reducing the spread and shortening response time, and select the solution with the highest benefit for output.

[0070] After executing the emergency response plan, it is necessary to obtain spatial resource information to support decision-making. Through the platform interface, rescue resource distribution data (such as fire stations, chemical material warehouses, and medical team locations), traffic network information (such as road accessibility and construction sections), and the real-time location of rescue teams are extracted in real time as the basis for subsequent scheduling diagram modeling.

[0071] Based on this resource data, the graph first maps key facilities and team locations as multiple nodes. Then, combined with traffic network information, multiple edges (representing road paths) are established between these nodes. These edges are then weighted based on the rescue team's real-time location and traffic status. Weights can represent travel time, traffic risk, or priority. For example, an edge representing the route from a supply warehouse to an incident site could have a weight determined by a combination of current traffic speed and the urgency of the rescue effort.

[0072] The previously constructed node set, edge set, and travel time weights are unified into a model to generate a resource scheduling network diagram. This network diagram can be used for route planning, optimal resource scheduling, and subsequent dynamic monitoring. The graph structure is highly scalable and can be dynamically adjusted with the addition of new resources or road conditions. It supports scheduling optimization and risk prediction using technologies such as graph neural networks based on the graph structure.

[0073] By implementing the above steps, we can quickly build an emergency strategy generation mechanism based on state-action pairs and reinforcement learning according to the actual accident situation, and use spatial graph networks to complete multi-dimensional resource scheduling preparation, providing a foundation for the next stage of dynamic accident monitoring and intelligent decision-making.

[0074] Furthermore, step S400 includes:

[0075] Step S410: Dynamically analyzing the resource scheduling network diagram and the multi-source synchronization data to obtain monitoring results, which include multiple monitoring indicators.

[0076] Step S420: When any of the multiple monitoring indicators exceeds a preset threshold, a graded warning signal is triggered and a real-time decision instruction is generated.

[0077] Step S430: performing analysis based on the real-time decision instruction to obtain resource demand fields and evacuation path information.

[0078] Step S440: Match the resource requirement field with the emergency response plan to determine a material reserve list.

[0079] Step S450: adjusting the rescue priority according to the evacuation route and the material reserve list, constructing an emergency execution instruction set, and pushing the emergency execution instruction set to the emergency command center to respond to the chemical accident.

[0080] Specifically, the constructed resource scheduling network diagram is integrated with the multi-source synchronized data obtained in step S100 in real time for analysis. By monitoring video, sensor, and communication information within the coverage area of ​​the monitoring node, several key monitoring indicators are extracted, such as toxic gas concentration, temperature change rate, personnel density, and the operating status of key equipment. Through a graph neural network or dynamic graph aggregation mechanism, the status of each node is dynamically updated to form a monitoring result set reflecting the accident situation.

[0081] Set grading thresholds for different types of indicators (such as a temperature greater than 80°C is a level one warning, and greater than 100°C is a level two warning), continuously perform threshold judgments on monitoring results, and compare multiple indicators in the monitoring results with the corresponding preset thresholds in real time. If any monitoring indicator is found to exceed the preset threshold, the graded warning signal corresponding to the indicator level will be immediately triggered (such as blue, yellow, and red levels). When multiple indicators exceed the threshold, the corresponding level of warning signal is triggered according to the highest priority principle. At the same time, according to the warning level and the current resource map status, the rule decision engine is called to generate real-time decision instructions. The instructions are output in a structured form, and the content covers response actions, resource call suggestions, path adjustments, etc.

[0082] The decision-making instructions are structured and parsed to extract key execution information, including resource requirements (such as the type and quantity of equipment required, the type of professional team, etc.) and evacuation path information (a dynamically calculated set of optimal evacuation paths based on current traffic conditions). This parsing process utilizes a natural language parsing engine and a graph path search algorithm (such as Dijkstra).

[0083] Based on the extracted resource demand field, it is matched and analyzed with the emergency response plan formulated in the aforementioned step S320 to determine whether the current response strategy has covered all requirements. If there is a gap, a material reserve list is automatically generated. The list will be optimized based on factors such as the available location of the materials and the arrival time.

[0084] Further combining evacuation route information with inventory of supplies, the team dynamically adjusted the priority of existing rescue missions, such as prioritizing the deployment of hazardous materials handling equipment to certain corridors and dispatching additional evacuation guidance personnel to specific nodes. Ultimately, a multi-dimensional dispatch and coordination emergency execution instruction set was constructed, covering key fields such as response content, task division, material dispatch, and personnel routing. These instructions were then pushed to the emergency command center in real time through the emergency platform interface, achieving a precise and efficient closed-loop emergency response to chemical accidents.

[0085] Furthermore, step S410 includes:

[0086] Step S411: spatially mapping the multiple nodes in the resource scheduling network diagram with the collection points of the multi-source synchronous data, and establishing a node-data association index table.

[0087] Step S412: continuously extracting real-time sensor data from sensors, associating the real-time sensor data with multiple nodes in the resource scheduling network graph, and determining multiple node state attributes.

[0088] Step S413: Identify the smoke diffusion trend at the accident point, perform collaborative calculations based on the multiple edges in the resource scheduling network graph, and obtain multiple edge state attributes.

[0089] Step S414: When a new risk point appears, the multiple travel time weights in the resource scheduling network diagram are updated to determine multiple weight coefficients.

[0090] Step S415: Based on the multiple node status attributes, the multiple edge status attributes, and the multiple weight coefficients, the multiple monitoring indicators are constructed based on the node-data association index table, and the multiple monitoring indicators are added to the monitoring results.

[0091] Specifically, the GIS system is used for coordinate projection matching, and the spatial coordinates of multiple nodes in the resource scheduling network diagram (such as warehouses and intersections) are bound to the collection points of multi-source synchronous data (such as sensor deployment points and surveillance camera locations). A mapping relationship is established based on the coordinates of each node in the geographic space and the spatial distribution of the collection points to form a node-data association index table for the rapid binding and retrieval of subsequent monitoring data.

[0092] Real-time sensor data is continuously extracted from various sensors (such as temperature and humidity sensors, smoke detectors, and toxic gas concentration probes). This real-time data is then mapped to the corresponding nodes in the resource scheduling network diagram, dynamically assigning corresponding status attributes to each node, such as hazard level, temperature distribution, and gas concentration. For example, based on chlorine concentration data, the status attribute of the liquid chlorine storage tank node is determined to be "normal" (concentration is below the safety threshold), "warning" (concentration is close to the safety threshold), or "alarm" (concentration exceeds the safety threshold).

[0093] Based on video surveillance systems or air quality detection equipment, the diffusion trend of smoke or harmful gases at chemical accident sites is identified, and the diffusion path is collaboratively calculated in combination with multiple edges (representing path connections) in the resource scheduling network diagram to obtain the status attributes of each edge, such as the traffic risk level, blockade status, and crossing time.

[0094] When new risk points (such as sudden leaks, fire sources, and secondary explosions) are detected through real-time analysis, the travel time weights in the resource scheduling network diagram are immediately updated. This update involves adjusting the weight coefficients based on the risk level, determining the passability and urgency of each path, and ensuring that routing avoids high-risk areas.

[0095] Based on the node-data association index table, the status attributes of each node, the status attributes of multiple edges, and the multiple weight coefficients after the travel time are updated are integrated, and finally a monitoring indicator set with multiple dimensions such as concentration level, temperature index, travel time, and personnel distribution density is constructed. This set is recorded in the current monitoring results to support the next step of early warning decision generation.

[0096] Through the above implementation steps, it is possible to achieve accurate integration of resource scheduling network diagrams and multi-source synchronous data, reflect the development trend of accidents in real time, and generate multi-dimensional, high-precision monitoring indicators, effectively supporting the formulation of subsequent graded warnings and emergency response instructions, and significantly improving the perception ability and intelligent response level of chemical accidents.

[0097] In summary, the chemical accident emergency response method under real-time monitoring provided by the embodiments of the present application has the following beneficial effects:

[0098] The embodiment of the present application achieves accurate identification and rapid judgment of chemical accidents through the real-time collection and fusion of multi-source heterogeneous data, combined with multimodal deep learning technology; at the same time, the resource scheduling network diagram constructed based on accident scenarios and resource scheduling analysis optimizes the resource allocation and scheduling efficiency of emergency response; finally, through real-time monitoring, early warning and rapid push of decision-making instructions, a closed-loop emergency response mechanism is formed, which enables emergency response to have continuous linkage and intelligent adjustment capabilities. Overall, this method effectively improves the identification accuracy, response speed and handling coordination capabilities of chemical accidents, and provides technical support and intelligent guarantee for emergency management in high-risk scenarios.

[0099] Example 2, as Figure 2 As shown, based on the same inventive concept as the aforementioned embodiment 1, this embodiment of the present application provides a chemical accident emergency response system under real-time monitoring, the system comprising:

[0100] The data acquisition module 10 is used to acquire multi-source heterogeneous data in real time, perform time alignment based on the multi-source heterogeneous data, and generate multi-source synchronized data.

[0101] The accident identification module 20 is used to perform multimodal deep learning based on the multi-source synchronous data, extract multimodal features, identify chemical accidents according to the multimodal features, and obtain accident identification results.

[0102] The response analysis module 30 is used to retrieve accident scenario parameters, perform accident response analysis based on the accident identification results, formulate an emergency response plan, execute the emergency response plan to perform resource scheduling analysis, and construct a resource scheduling network diagram.

[0103] The emergency response module 40 is used to monitor accidents according to the resource scheduling network diagram in combination with the multi-source synchronous data, trigger an early warning signal based on the monitoring results, generate real-time decision instructions, and push the real-time decision instructions in combination with the emergency response plan to the emergency command center for emergency response to chemical accidents.

[0104] Furthermore, the data acquisition module 10 of the embodiment of the present application is further configured to perform the following steps:

[0105] Real-time data collection is performed through a multi-source data interface to obtain multi-source heterogeneous data; a time window is dynamically set to construct a time window, and the multi-source heterogeneous data is aligned according to timestamps using the time window to generate an aligned data set; timestamp offset analysis is performed based on the aligned data set to generate multiple offsets; it is determined whether the multiple offsets exceed the time window length, and if any offset among the multiple offsets exceeds the time window length, a data compensation instruction is triggered; and missing compensation is performed on the aligned data set using the data compensation instruction to obtain the multi-source heterogeneous data.

[0106] Furthermore, the data acquisition module 10 of the embodiment of the present application is further configured to perform the following steps:

[0107] Retrieve historical transmission delay statistics of multiple data sources in the multi-source data interface, and initialize a baseline length of a time window based on the historical transmission delay statistics; monitor the transmission jitter rates of the multiple data sources in real time, and generate an extension parameter when the transmission jitter rate exceeds a preset threshold; dynamically extend the baseline length of the time window based on the extension parameter, and set the time window length.

[0108] Furthermore, the accident identification module 20 of the embodiment of the present application is further configured to perform the following steps:

[0109] The multi-source synchronous data set is input into a multimodal deep learning network for multimodal analysis to obtain multimodal data, wherein the multimodal data includes video stream data, sensor time series signals, and structured text data; three-dimensional convolution analysis is performed based on the video stream data to extract spatiotemporal features; bidirectional analysis is performed based on the sensor time series signals to extract time series features; natural language processing is performed based on the structured text data to extract semantic keyword features; the spatiotemporal features, the time series features, and the semantic keyword features are weightedly fused to construct the multimodal features.

[0110] Furthermore, the response analysis module 30 of the embodiment of the present application is further configured to perform the following steps:

[0111] Vectorization is performed based on the accident scene parameters to define a state space, and emergency action matching is performed based on the accident identification result to define an action space; reinforcement learning is performed on the state space in combination with the action space to formulate an emergency response plan; resource scheduling analysis is performed according to the emergency response plan to obtain rescue resource distribution data, traffic network information, and the real-time location of the rescue team; multiple nodes are constructed based on the rescue resource distribution data, multiple edges of the multiple nodes are constructed based on the traffic network information, and weights are assigned to the multiple edges based on the real-time location of the rescue team to determine multiple travel time weights; the multiple nodes, the multiple edges, and the multiple travel time weights are associated and integrated to construct the resource scheduling network diagram.

[0112] Furthermore, the emergency response module 40 of the embodiment of the present application is further configured to perform the following steps:

[0113] Dynamic analysis is performed in combination with the resource scheduling network diagram and the multi-source synchronous data to obtain monitoring results, which include multiple monitoring indicators; when any of the multiple monitoring indicators exceeds a preset threshold, a graded warning signal is triggered and a real-time decision instruction is generated; parsing is performed based on the real-time decision instruction to obtain resource demand fields and evacuation path information; the resource demand field is matched with the emergency response plan to determine a material reserve list; the rescue priority is adjusted according to the evacuation path in combination with the material reserve list, an emergency execution instruction set is constructed, and the emergency execution instruction set is pushed to the emergency command center for emergency response to the chemical accident.

[0114] Furthermore, the emergency response module 40 of the embodiment of the present application is further configured to perform the following steps:

[0115] The multiple nodes in the resource scheduling network diagram are spatially mapped with the collection points of the multi-source synchronous data to establish a node-data association index table; the real-time sensing data of the sensor is continuously extracted, the real-time sensing data is associated with the multiple nodes in the resource scheduling network diagram, and multiple node status attributes are determined; the smoke diffusion trend at the accident point is identified, and the multiple edges in the resource scheduling network diagram are combined for collaborative calculation to obtain multiple edge status attributes; when a new risk point appears, the multiple travel time weights in the resource scheduling network diagram are updated to determine multiple weight coefficients; based on the multiple node status attributes, the multiple edge status attributes, and the multiple weight coefficients, they are integrated according to the node-data association index table to construct the multiple monitoring indicators, and the multiple monitoring indicators are added to the monitoring results.

[0116] Through the detailed description of the chemical accident emergency response method under real-time monitoring in the foregoing specification, those skilled in the art can clearly understand the chemical accident emergency response system under real-time monitoring in this embodiment. For the system disclosed in Example 2, since it corresponds to the method disclosed in Example 1 and has corresponding functional modules and beneficial effects, the relevant details can be referred to the description of the method section.

[0117] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A chemical accident emergency response method under real-time monitoring, characterized in that: The method comprises: Collect multi-source heterogeneous data in real time, perform time alignment based on the multi-source heterogeneous data, and generate multi-source synchronized data; performing multimodal deep learning based on the multi-source synchronized data, extracting multimodal features, and identifying chemical accidents based on the multimodal features to obtain accident identification results; Retrieving accident scenario parameters, performing accident response analysis based on the accident identification results, formulating an emergency response plan, executing the emergency response plan to perform resource scheduling analysis, and constructing a resource scheduling network diagram; Accident monitoring is performed according to the resource scheduling network diagram in combination with the multi-source synchronous data, an early warning signal is triggered according to the monitoring results, a real-time decision instruction is generated, and the real-time decision instruction is pushed to the emergency command center in combination with the emergency response plan to respond to the chemical accident.

2. The chemical accident emergency response method under real-time monitoring according to claim 1, characterized in that: The method includes: collecting multi-source heterogeneous data in real time, performing time alignment based on the multi-source heterogeneous data, and generating multi-source synchronized data. Real-time collection is performed through multi-source data interfaces to obtain multi-source heterogeneous data; Dynamically setting the time window length to construct a time window, and using the time window to align the multi-source heterogeneous data according to timestamps to generate an aligned data set; Performing timestamp offset analysis on the aligned data set to generate multiple offsets; determining whether the multiple offsets exceed the time window length, and triggering a data compensation instruction if any offset among the multiple offsets exceeds the time window length; The aligned data set is compensated for missing data by using the data compensation instruction to obtain the multi-source heterogeneous data.

3. The chemical accident emergency response method under real-time monitoring according to claim 2, characterized in that: The process of dynamically setting the time window length includes: Retrieving historical transmission delay statistics of multiple data sources in the multi-source data interface, and initializing a reference length of a time window according to the historical transmission delay statistics; monitoring transmission jitter rates of the plurality of data sources in real time, and generating an extended parameter when the transmission jitter rate exceeds a preset threshold; The reference length of the time window is dynamically extended according to the extension parameter to set the time window length.

4. The chemical accident emergency response method under real-time monitoring according to claim 1, characterized in that: Performing multimodal deep learning based on the multi-source synchronized data to extract multimodal features, the method comprising: Inputting the multi-source synchronized data set into a multimodal deep learning network for multimodal analysis to obtain multimodal data, wherein the multimodal data includes video stream data, sensor timing signals, and structured text data; Performing three-dimensional convolution analysis based on the video stream data to extract spatiotemporal features; Performing bidirectional analysis based on the sensor time series signal to extract time series features; Performing natural language processing based on the structured text data to extract semantic keyword features; The spatiotemporal features, the time series features, and the semantic keyword features are weighted and fused to construct the multimodal features.

5. The chemical accident emergency response method under real-time monitoring according to claim 1, characterized in that: Retrieving accident scenario parameters, performing accident response analysis based on the accident identification results, formulating an emergency response plan, executing the emergency response plan to perform resource scheduling analysis, and constructing a resource scheduling network diagram, the method includes: Vectorizing the accident scene parameters to define a state space, and matching emergency actions based on the accident identification results to define an action space; Perform reinforcement learning on the state space and the action space to formulate an emergency response plan; Perform resource scheduling analysis according to the emergency response plan to obtain rescue resource distribution data, traffic network information, and the real-time location of the rescue team; Constructing multiple nodes based on the rescue resource distribution data, constructing multiple edges of the multiple nodes based on the traffic network information, assigning weights to the multiple edges based on the real-time location of the rescue team, and determining multiple travel time weights; The multiple nodes, the multiple edges, and the multiple travel time weights are associated and integrated to construct the resource scheduling network graph.

6. The chemical accident emergency response method under real-time monitoring according to claim 5, characterized in that: Accident monitoring is performed according to the resource scheduling network diagram in combination with multi-source synchronous data, an early warning signal is triggered according to the monitoring results, a real-time decision instruction is generated, and the real-time decision instruction is pushed to the emergency command center in combination with the emergency response plan to perform an emergency response to the chemical accident, the method comprising: Performing dynamic analysis on the resource scheduling network diagram and the multi-source synchronization data to obtain monitoring results, wherein the monitoring results include multiple monitoring indicators; When any of the multiple monitoring indicators exceeds a preset threshold, a graded warning signal is triggered and a real-time decision instruction is generated; Analyze the real-time decision instructions to obtain resource demand fields and evacuation path information; Matching the resource requirement field with the emergency response plan to determine a material reserve list; The rescue priority is adjusted according to the evacuation route in combination with the material reserve list, an emergency execution instruction set is constructed, and the emergency execution instruction set is pushed to the emergency command center for emergency response to the chemical accident.

7. The chemical accident emergency response method under real-time monitoring according to claim 6, characterized in that: Dynamically analyzing the resource scheduling network diagram and the multi-source synchronization data to obtain monitoring results includes: Performing spatial mapping between the multiple nodes in the resource scheduling network diagram and the collection points of the multi-source synchronous data, and establishing a node-data association index table; Continuously extracting real-time sensor data from sensors, associating the real-time sensor data with multiple nodes in the resource scheduling network graph, and determining multiple node state attributes; Identify the smoke diffusion trend at the accident point, perform collaborative calculations based on the multiple edges in the resource scheduling network graph, and obtain multiple edge state attributes; When a new risk point appears, the multiple travel time weights in the resource scheduling network diagram are updated to determine multiple weight coefficients; Based on the multiple node status attributes, the multiple edge status attributes, and the multiple weight coefficients, they are integrated according to the node-data association index table to construct the multiple monitoring indicators, and the multiple monitoring indicators are added to the monitoring results.

8. A chemical accident emergency response system under real-time monitoring, characterized by: The system is used to implement the chemical accident emergency response method under real-time monitoring according to any one of claims 1 to 7, comprising: A data acquisition module is used to collect multi-source heterogeneous data in real time, perform time alignment based on the multi-source heterogeneous data, and generate multi-source synchronized data; An accident identification module, configured to perform multimodal deep learning based on the multi-source synchronized data, extract multimodal features, identify chemical accidents based on the multimodal features, and obtain accident identification results; A response analysis module is used to retrieve accident scenario parameters, perform accident response analysis based on the accident identification results, formulate an emergency response plan, execute the emergency response plan to perform resource scheduling analysis, and construct a resource scheduling network diagram; The emergency response module is used to monitor accidents according to the resource scheduling network diagram in combination with the multi-source synchronous data, trigger early warning signals based on the monitoring results, generate real-time decision instructions, and push the real-time decision instructions in combination with the emergency response plan to the emergency command center for emergency response to chemical accidents.

Images