Monitoring, early warning and analysis method and server based on artificial intelligence

Abstract

The present invention provides an artificial intelligence-based monitoring and early warning analysis method and server. First, a multi-source monitoring data stream of a target monitoring area is collected, which contains at least two types of real-time monitoring data and spatiotemporal label information to clarify the collection location and timestamp. Then, spatiotemporal feature extraction is performed to generate a spatiotemporal feature matrix with a hierarchical correlation relationship. The matrix dynamic pattern is then analyzed through a preset anomaly recognition network to determine potential abnormal events and abnormal propagation paths. Then, based on the path topology structure and event attribute parameters, an adaptive dynamic early warning strategy is generated, including differentiated trigger conditions and response instructions. Finally, historical early warning feedback data is used to optimize and adjust the strategy in real time and output it to a terminal device cluster. In this way, the monitoring data can be comprehensively and accurately analyzed to achieve efficient and intelligent monitoring and early warning.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based monitoring, early warning and analysis method and server. Background Art

[0002] Effective monitoring and early warning analysis of various target monitoring areas are crucial in the operation and management of various fields today, involving many areas such as industrial production, urban security, and environmental monitoring. However, existing monitoring and early warning analysis methods have many shortcomings and cannot meet the growing complex monitoring needs.

[0003] For example, related technologies fail to fully consider the temporal and spatial information inherent in the data. Furthermore, they rely primarily on pre-set fixed thresholds, which are too rigid in the face of complex and ever-changing real-world situations. They struggle to adapt to dynamic environments and lack the ability to effectively predict potential anomalies that haven't yet manifested themselves.

[0004] In addition, related technologies usually adopt unified and static warning rules, which do not take into account the differences between different locations and different time stages within the target monitoring area, and cannot provide targeted warnings based on actual conditions. Summary of the Invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a monitoring and early warning analysis method based on artificial intelligence, the method comprising:

[0006] Collecting a multi-source monitoring data stream within a target monitoring area, the multi-source monitoring data stream including at least two types of real-time monitoring data and corresponding spatiotemporal tag information, the spatiotemporal tag information being used to describe the collection location and timestamp of the real-time monitoring data within the target monitoring area;

[0007] Extracting spatiotemporal features from the multi-source monitoring data stream to generate a spatiotemporal feature matrix with a hierarchical association relationship, wherein each element in the spatiotemporal feature matrix corresponds to a monitoring indicator feature of a target position in the target monitoring area within a target time window;

[0008] Performing dynamic pattern analysis on the spatiotemporal feature matrix based on a preset anomaly recognition network to determine potential abnormal events and their corresponding abnormal propagation paths within the target monitoring area, wherein the abnormal propagation paths are used to indicate the diffusion direction and impact range of the potential abnormal events within the target monitoring area;

[0009] Generate a dynamic early warning strategy adapted to the target monitoring area based on the topological structure of the abnormal propagation path and the attribute parameters of the potential abnormal event, wherein the dynamic early warning strategy includes differentiated early warning trigger conditions and response instruction sets for different locations and different time stages;

[0010] The dynamic warning strategy is optimized and adjusted in real time based on historical warning feedback data, and the optimized dynamic warning strategy is output to the terminal device cluster corresponding to the target monitoring area.

[0011] On the other hand, an embodiment of the present invention also provides a server, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0012] Based on the above aspects, the embodiment of the present application collects a multi-source monitoring data stream containing at least two types of real-time monitoring data and corresponding spatiotemporal tag information, extracts spatiotemporal features from the multi-source monitoring data stream and generates a spatiotemporal feature matrix with hierarchical correlation relationships, thereby mining the potential spatiotemporal connections between the data. Each element in the spatiotemporal feature matrix corresponds to the monitoring indicator feature within a specific location and time window, so that the multi-source data is presented in a structured, hierarchical and easy-to-analyze form, thereby improving the efficiency and depth of data processing. Furthermore, based on a preset anomaly recognition network, a dynamic pattern analysis of the spatiotemporal feature matrix is ​​performed, which can not only accurately determine the potential abnormal events in the target monitoring area, but also depict the abnormal propagation path, clearly indicating the diffusion direction and impact range of the potential abnormal events in the area. Then, based on the topological structure of the abnormal propagation path and the attribute parameters of the potential abnormal events, a dynamic early warning strategy adapted to the target monitoring area is generated. The dynamic early warning strategy takes into account the differences between different locations and different time stages. The differentiated early warning trigger conditions and response instruction sets included can be accurately implemented according to the actual situation, significantly improving the accuracy and effectiveness of the early warning, and ensuring that early warnings can be issued and actions taken in a timely and appropriate manner in different scenarios. Finally, the dynamic warning strategy is optimized and adjusted in real time based on historical warning feedback data. By continuously absorbing historical feedback information, it can dynamically adapt to the complex and changing situations in the target monitoring area, continuously optimize the warning strategy, and ensure that the warning strategy output to the terminal device cluster always maintains optimal performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a schematic diagram of the execution flow of the artificial intelligence-based monitoring, early warning and analysis method provided by an embodiment of the present invention.

[0014] Figure 2 Schematic diagram of the hardware architecture of the server provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of an artificial intelligence-based monitoring, early warning and analysis method provided by an embodiment of the present invention. The artificial intelligence-based monitoring, early warning and analysis method is introduced in detail below.

[0016] Step S110, collecting multi-source monitoring data streams within the target monitoring area, wherein the multi-source monitoring data streams include at least two types of real-time monitoring data and corresponding spatiotemporal tag information, and the spatiotemporal tag information is used to describe the collection location and timestamp of the real-time monitoring data within the target monitoring area.

[0017] In this example, an industrial park is used as the target monitoring area. Within this industrial park, there are multiple sources of real-time monitoring data. First, there are environmental monitoring sensors, distributed throughout the park. These sensors include air quality sensors, temperature sensors, and humidity sensors. Air quality sensors monitor the concentrations of various air pollutants, such as sulfur dioxide, nitrogen oxides, and particulate matter, in real time. Temperature sensors and humidity sensors measure the ambient temperature and humidity, respectively. These sensors collect data every 10 minutes and simultaneously record the location (e.g., using the sensor's own positioning module or a pre-set location marker, such as near Building 3, Area A) and a timestamp (accurate to the second).

[0018] In addition, monitoring devices related to production equipment can be configured. For example, sensors are installed throughout each factory floor to monitor the operating status of production equipment. These sensors can monitor key operating parameters such as vibration frequency, current intensity, and oil temperature. For large power equipment, data is collected every 5 minutes, while for smaller auxiliary equipment, data is collected every 15 minutes. Similarly, this equipment monitoring data is accompanied by the collection location (e.g., Production Line 5 in Workshop B) and timestamp information.

[0019] In addition, a security monitoring system can be deployed, with cameras located at the park's entrances and exits, main corridors, and key areas. These cameras not only capture real-time video image data but also provide information about the image acquisition time and camera location. For example, a camera at the park's main gate captures one frame of image data every second. These different types of monitoring data, including those from the environment, production equipment, and security monitoring, along with their corresponding spatiotemporal labels, together constitute a multi-source monitoring data stream within the target monitoring area.

[0020] Step S120 , extracting spatiotemporal features from the multi-source monitoring data stream to generate a spatiotemporal feature matrix with a hierarchical association relationship, wherein each element in the spatiotemporal feature matrix corresponds to a monitoring index feature of a target position in the target monitoring area within a target time window.

[0021] Specifically, the multi-source monitoring data stream collected from the industrial park is first divided into multiple spatiotemporal data blocks based on spatiotemporal tag information. For example, with a time interval of one hour, all sensor data in Area A is divided into a spatiotemporal data block. This spatiotemporal data block corresponds to the monitoring data set for Area A within that one hour.

[0022] Feature extraction is performed for each spatiotemporal data block. Taking the spatiotemporal data block for environmental monitoring in Area A as an example, the three different types of monitoring data, air quality, temperature, and humidity, are coupled in the same spatiotemporal dimension. For example, under conditions of high temperature and high humidity, the chemical reaction rate of certain pollutants may accelerate, thereby affecting air quality. By analyzing a large amount of historical data and real-time data, this cross-modal correlation feature is extracted. At the same time, the time series fluctuation characteristics of the monitoring data within this spatiotemporal data block are considered, such as a sudden increase in temperature or a sharp deterioration in air quality within a certain period of time. The cross-modal correlation features are fused with the time series fluctuation features to generate a local spatiotemporal feature vector corresponding to the spatiotemporal data block.

[0023] Based on the physical connections between various areas within the industrial park (such as road and pipeline connections) and historical event propagation records (for example, fires spreading along ventilation ducts), a global spatiotemporal correlation map of the entire industrial park is constructed. Nodes in the global spatiotemporal correlation map represent predefined sub-areas (such as workshops, warehouses, and office buildings), while edges represent the probability of event propagation and the strength of correlation between sub-areas. For example, workshops A and B are connected by a material transportation channel. Based on historical statistical data, there is a 30% probability that a fire in workshop A will spread to workshop B through this channel. This 30% represents the event propagation probability, while the strength of correlation may also be affected by factors such as the width of the channel and fire prevention facilities.

[0024] Based on the global spatiotemporal correlation map, spatial weights are assigned to each local spatiotemporal feature vector and time series are aligned. For example, for areas close to the center of the park, due to their high density of personnel and equipment, higher spatial weights may be assigned; and for time series alignment, the temporal consistency of different spatiotemporal data blocks is ensured in order to generate a hierarchical structure of the spatiotemporal feature matrix, which includes a basic feature layer, a spatiotemporal correlation layer, and an event prediction layer. In the basic feature layer, the original monitoring indicator statistics of each sub-area are stored, such as the average temperature and air quality index of area A; the spatiotemporal correlation layer stores the spatiotemporal dependency coefficients between sub-areas, such as the correlation coefficient between workshop A and adjacent workshops; the event prediction layer stores the predicted values ​​of the probability of abnormal events based on training of historical event data, such as predicting the probability of future fires in area A based on past fire records.

[0025] Step S130, based on a preset anomaly recognition network, dynamic pattern analysis is performed on the spatiotemporal feature matrix to determine potential abnormal events and their corresponding abnormal propagation paths in the target monitoring area, where the abnormal propagation paths are used to indicate the diffusion direction and impact range of the potential abnormal events in the target monitoring area.

[0026] The spatiotemporal feature matrix of the industrial park is input into the preset anomaly recognition network, which includes a spatiotemporal convolution module, a long short-term memory module and a graph neural network module.

[0027] First, the spatiotemporal feature matrix is ​​segmented into multiple segments along the time dimension and fed into the spatiotemporal convolution module for feature dimensionality reduction and pattern extraction. For example, a spatiotemporal feature matrix containing monitoring data for a single day is segmented into hourly data segments and then fed into the spatiotemporal convolution module. The spatiotemporal convolution module can extract local spatiotemporal patterns in the monitoring data. For example, if the air quality index in area A exhibits a specific fluctuation pattern during a certain time period, this may be related to abnormal production emissions from a nearby factory.

[0028] The output of the spatiotemporal convolution module is reorganized according to the spatial dimension and then fed into the graph neural network module to generate spatial correlation feature vectors for each subregion. For example, the graph neural network module can analyze the spatial correlation features between workshop A and warehouse B because of the material transportation relationship between them, which may affect the propagation of abnormal events between them.

[0029] The spatially correlated feature vectors are fed into the long-short-term memory module in chronological order to generate a global spatiotemporal state representation of the industrial park. The long-short-term memory module is capable of capturing long-term dependencies in time series. For example, it can detect a gradual increase in power consumption within the industrial park over several consecutive days, which may be due to the continued abnormal operation of certain large equipment.

[0030] The output features of the spatiotemporal convolution module, the graph neural network module, and the long-short-term memory module are integrated to generate a comprehensive discrimination index for abnormal event identification. For example, if this comprehensive discrimination index reveals that the oil temperature of a large piece of production equipment in Area A is too high and has persisted for some time, and the temperature in the surrounding area is also rising, this is identified as a potential abnormal event.

[0031] Next, the anomaly propagation path is determined. The attention allocation module of the anomaly recognition network dynamically weights anomaly pattern features, determining the anomaly confidence level for each sub-area within the industrial park and the strength of correlation between anomaly events. For example, the confidence level for the high oil temperature anomaly in Area A is 0.8, and a high correlation strength is found with the adjacent Area B, as Area B's ventilation system is connected to Area A.

[0032] An abnormal event propagation network is constructed based on anomaly confidence and correlation strength. Nodes represent subregions where the detected anomaly confidence exceeds a set confidence level (e.g., 0.6), and edges represent the likelihood of an abnormal event propagating between subregions. For example, an edge between area A and area B indicates that an abnormal event indicating excessive oil temperature is likely to propagate from area A to area B.

[0033] Based on the node degree distribution and edge weight distribution of the abnormal event propagation network, the core location of potential abnormal events (such as the location of the large production equipment in area A) and the key nodes of the propagation path (such as the entrance of the ventilation duct connecting areas A and B) are identified.

[0034] Finally, the path connectivity of key nodes along the propagation path is optimized based on the temporal continuity of the spatiotemporal tag information. For example, the timestamp sequence corresponding to the spatiotemporal tag information of the key nodes is extracted. For example, suppose the timestamp of the abnormal oil temperature of the equipment in area A is 10:00 AM, and the timestamp of the temperature rise in area B is 10:15 AM. These key nodes are arranged in timestamp order. The timestamp intervals between adjacent nodes are calculated. If the time interval between areas A and B is 15 minutes, which is less than a preset temporal continuity threshold (e.g., 30 minutes), these nodes are grouped into the same temporal continuity group. The spatial coordinate information of the nodes within this temporal continuity group is traversed, and the spatial distance matrix between the nodes within the group is calculated. If the spatial distance between the ventilation duct inlet connecting areas A and B and the temperature sensor in area B is less than a preset spatial proximity threshold (e.g., 10 meters), these nodes are marked as spatially adjacent nodes. A spatial proximity relationship graph is then generated based on the spatial distance matrix. Based on this spatial proximity relationship graph, the equipment in area A, the ventilation duct inlet, and the temperature sensor in area B are linearly connected in timestamp order to form the initial propagation path segment. Detect spatiotemporal conflicts between initial propagation path segments generated by different time-continuous groups. For example, if another propagation path segment is found to have overlapping timestamps with the current path segment in Area B, the path segments with later timestamps are prioritized and removed according to the rules. After a series of adjustments, a complete topological structure of the abnormal propagation path is generated. This structure includes path branching information (for example, the impact of equipment in Area A on Area C may also occur through power lines), node connection sequence (for example, first passing through the ventilation duct entrance and then to the temperature sensor in Area B), and path propagation direction weight parameters (for example, the propagation weight from Area A to Area B is higher).

[0035] Step S140, based on the topological structure of the abnormal propagation path and the attribute parameters of the potential abnormal event, a dynamic early warning strategy adapted to the target monitoring area is generated, and the dynamic early warning strategy includes differentiated early warning trigger conditions and response instruction sets for different locations and different time stages.

[0036] Specifically, the priority index of each sub-area in the potential abnormal event propagation chain is calculated based on the topological structure of the abnormal propagation path of the industrial park. For example, for an abnormal event of excessive oil temperature in equipment in Area A, the sub-areas directly connected to the equipment (such as Area B) have a higher priority index because they are more urgently affected.

[0037] Determine the trigger threshold range for dynamic warning strategies based on the attribute parameters of potential abnormal events. For example, assume that the abnormal intensity of an equipment oil temperature overheating event is 30% above the normal range, the propagation rate is 10 meters per hour, and the impact is expected to last for 2 hours. Based on these attribute parameters, set the warning trigger threshold range. For example, when the oil temperature exceeds the normal range by 20%, the warning preparation phase begins.

[0038] Multi-level warning triggering rules are generated based on the priority index and trigger threshold range. For the high-priority zone B, when the oil temperature exceeds the normal range by 20%, a level 1 warning is triggered, with a high level and a response delay of 0 minutes. For the more distant and lower-priority zone C, when the oil temperature exceeds the normal range by 25%, a level 2 warning is triggered, with a medium level and a response delay of 5 minutes.

[0039] Build a set of response instructions that match the multi-level warning trigger rules. For a level 1 warning, the equipment control instruction might be to immediately stop the operation of equipment in Area A, the personnel dispatch instruction might be to dispatch maintenance and security personnel to Areas A and B, and the data reporting protocol requires reporting equipment status and environmental data every 30 seconds. For a level 2 warning, the equipment control instruction might be to reduce the operating power of equipment in Area C, the personnel dispatch instruction might be to notify staff in Area C to observe equipment status, and the data reporting protocol requires data reporting every 5 minutes.

[0040] Dynamically bind multi-level warning triggering rules and response instruction sets according to the chronological order of the anomaly propagation path, forming a logical chain for executing dynamic warning strategies. For example, the instruction to stop equipment in area A is executed first, followed by subsequent operations based on the situation in area B, and then the corresponding instructions are executed in sequence based on the warning status of area C.

[0041] Step S150 , optimizing and adjusting the dynamic warning strategy in real time based on historical warning feedback data, and outputting the optimized dynamic warning strategy to the terminal device cluster corresponding to the target monitoring area.

[0042] Specifically, we can obtain feedback log data generated by the execution of historical early warning strategies by the industrial park's terminal equipment cluster. For example, during the last abnormal oil temperature overheating event in Area A, the early warning response delay in Area B was actually 2 minutes (while the default was 0 minutes). The equipment execution status showed that the equipment shutdown in Area A had a certain impact on the entire production line (such as some semi-finished products being affected). The actual impact of the abnormal event was slightly larger than predicted, affecting a small warehouse adjacent to Area B.

[0043] Extract strategy execution effect indicators from feedback log data, such as warning accuracy (the actual warning situation was found to deviate from expectations, and the accuracy was reduced), response efficiency deviation (the response in area B was delayed by 2 minutes), and resource consumption rate (equipment shutdown resulted in some additional resource waste, such as waste of raw materials).

[0044] Compare the policy execution performance indicators with the preset optimization objective function. Assume that the optimization objective function includes a warning coverage efficiency term (positively correlated with warning accuracy), a resource consumption penalty term (negatively correlated with the energy consumption and computing resource usage of the device when executing warning instructions), and a response time reward term (positively correlated with warning response speed). If warning accuracy decreases, resource consumption increases, and response speed slows, determine the trigger condition parameters and response instruction priorities that need to be adjusted in the dynamic warning strategy.

[0045] The trigger threshold range and response delay of the dynamic warning strategy are iteratively modified based on the gradient direction of the optimization objective function. For example, the trigger condition for the first-level warning in zone B is adjusted to the oil temperature exceeding the normal range by 22%, and the response delay is adjusted to 1 minute.

[0046] Based on the revised parameters, the coverage and resource allocation weights for each level of warning in the multi-level warning trigger rules are recalculated to generate an optimized dynamic warning strategy. For example, the warning level and resource allocation of different regions under the new trigger conditions are reassessed to ensure that the warning strategy can be executed more effectively and accurately the next time a similar abnormal event occurs. The optimized dynamic warning strategy is then output to the relevant terminal device clusters within the industrial park (such as production equipment controllers and security monitoring equipment).

[0047] Based on the above steps, the embodiment of the present application collects a multi-source monitoring data stream containing at least two types of real-time monitoring data and corresponding spatiotemporal tag information, extracts spatiotemporal features from the multi-source monitoring data stream and generates a spatiotemporal feature matrix with hierarchical correlation relationships, thereby mining the potential spatiotemporal connections between the data. Each element in the spatiotemporal feature matrix corresponds to the monitoring indicator feature within a specific location and time window, so that the multi-source data is presented in a structured, hierarchical and easy-to-analyze form, thereby improving the efficiency and depth of data processing. Furthermore, based on a preset anomaly recognition network, a dynamic pattern analysis of the spatiotemporal feature matrix is ​​performed, which can not only accurately determine the potential abnormal events in the target monitoring area, but also depict the abnormal propagation path, clearly indicating the diffusion direction and impact range of the potential abnormal events in the area. Then, based on the topological structure of the abnormal propagation path and the attribute parameters of the potential abnormal events, a dynamic early warning strategy adapted to the target monitoring area is generated. The dynamic early warning strategy takes into account the differences between different locations and different time stages. The differentiated early warning trigger conditions and response instruction sets included can be accurately implemented according to the actual situation, significantly improving the accuracy and effectiveness of the early warning, and ensuring that early warnings can be issued and actions taken in a timely and appropriate manner in different scenarios. Finally, the dynamic warning strategy is optimized and adjusted in real time based on historical warning feedback data. By continuously absorbing historical feedback information, it can dynamically adapt to the complex and changing situations in the target monitoring area, continuously optimize the warning strategy, and ensure that the warning strategy output to the terminal device cluster always maintains optimal performance.

[0048] In a possible implementation, step S120 includes:

[0049] Step S121 : dividing the multi-source monitoring data stream into a plurality of spatiotemporal data blocks according to the spatiotemporal tag information, each of the spatiotemporal data blocks corresponding to a set of monitoring data of a preset sub-area in the target monitoring area within a continuous time series.

[0050] Specifically, various parts of the industrial park are designated as predefined sub-areas, such as workshops, warehouses, office buildings, as well as roads and green areas within the park. Spatiotemporal tag information details the location and timestamp of monitoring data collection. For example, in Workshop A, data collected by various sensors within the workshop (such as equipment operating status sensors and environmental sensors) over a period of time (e.g., eight hours per workday) can be divided into a spatiotemporal data block based on spatiotemporal tags. This spatiotemporal data block contains all monitoring data from Workshop A during that eight-hour period, including operational parameters such as equipment vibration frequency, current intensity, and oil temperature, as well as environmental data such as temperature, humidity, and air quality within the workshop. Similarly, other sub-areas, such as Warehouse B and Office Building C, each have corresponding spatiotemporal data blocks, divided based on the spatiotemporal tags of the sensor data collected within their respective areas.

[0051] Step S122: for each of the spatiotemporal data blocks, extract cross-modal correlation features between the monitoring data contained in the spatiotemporal data block, where the cross-modal correlation features are used to describe the coupling relationship between different types of monitoring data in the same spatiotemporal dimension.

[0052] Taking the spatiotemporal data block from Workshop A as an example, cross-modal correlations exist between the equipment operating status data and environmental data within the workshop. During equipment operation, rising oil temperatures can cause the workshop temperature to rise, demonstrating a coupling relationship between the equipment's operating status and ambient temperature. Increased vibration frequency can cause changes in the surrounding air flow, affecting air quality, demonstrating the correlation between equipment operating status and air quality. Furthermore, when humidity is high within the workshop, condensation may form on the equipment surface, affecting its electrical performance and causing fluctuations in current intensity. This also demonstrates the coupling relationship between different types of monitoring data (humidity and current intensity) within the same spatiotemporal dimension.

[0053] Step S123 , fusing the cross-modal correlation feature with the time series fluctuation feature of the spatiotemporal data block to generate a local spatiotemporal feature vector corresponding to the spatiotemporal data block.

[0054] Specifically, in the spatiotemporal data block for Workshop A, time series fluctuations are very evident. For example, between 9:00 AM and 10:00 AM, due to the concentrated scheduling of production tasks, equipment operation intensity increases, leading to an upward trend in parameters such as oil temperature and vibration frequency. This, in turn, increases the temperature within the workshop, and air quality may slightly decrease due to increased equipment emissions. This time series fluctuation feature is integrated with the previously extracted cross-modal correlation features. For example, the cross-modal correlation feature (temperature increase due to increased oil temperature) is more pronounced between 9:00 AM and 10:00 AM. By combining the strength and trend of this cross-modal correlation feature within this time period with the time series fluctuation feature, a local spatiotemporal feature vector corresponding to Workshop A's spatiotemporal data block is generated. This vector comprehensively describes the comprehensive characteristics of Workshop A during this specific time period, encompassing both the inherent relationships between different types of monitoring data and how these relationships change over time.

[0055] Step S124: Based on the physical connection relationship between the preset sub-areas and the historical event propagation records, a global spatiotemporal correlation map of the target monitoring area is constructed. The nodes in the global spatiotemporal correlation map represent the preset sub-areas, and the edges represent the event propagation probability and correlation strength between the preset sub-areas.

[0056] Specifically, in an industrial park, sub-areas have various physical connections. For example, Workshop A and Warehouse B have a material transport corridor, forming a physical connection. Historical event propagation records show that during a small fire in Workshop A, smoke spread from the material transport corridor to Warehouse B because the corridor was not closed promptly. Based on these physical connections and historical event propagation records, a global spatiotemporal correlation graph is constructed. In this global spatiotemporal correlation graph, nodes represent predefined sub-areas such as Workshop A, Warehouse B, and Office Building C. Edges represent the event propagation probability and correlation strength between predefined sub-areas. For the edge between Workshop A and Warehouse B, the event propagation probability might be hypothetically set at 20% based on statistical data from historical fires and other similar events. The correlation strength might be affected by factors such as the fire prevention facilities in the material transport corridor, as well as its length and width. If the corridor has good fire isolation facilities, the correlation strength will be relatively low; if the corridor is narrow and lacks fire prevention facilities, the correlation strength will be high. For Workshop A and Office Building C, since there is no direct material transportation channel between them, but there is a power line connection, according to the historical records of power fault propagation, the probability of event propagation is 10%, and the correlation strength is affected by factors such as the insulation performance and load conditions of the power line.

[0057] Step S125 , performing spatial weight allocation and time series alignment on the local spatiotemporal feature vectors according to the global spatiotemporal correlation map, to generate a hierarchical structure of the spatiotemporal feature matrix.

[0058] Specifically, based on the global spatiotemporal correlation map of the industrial park, the local spatiotemporal feature vectors of each sub-area (such as Workshop A and Warehouse B) are processed. Regarding spatial weight assignment, Workshop A is the primary production area, with dense equipment and frequent personnel activities. Therefore, any abnormal event would have a wide impact, so it is assigned a higher spatial weight. Office Building C, on the other hand, has a relatively low density of personnel and equipment and is therefore assigned a lower spatial weight. Time series alignment ensures temporal consistency across the spatiotemporal data blocks of each sub-area. For example, the spatiotemporal data blocks for Workshop A, Warehouse B, and Office Building C are all divided into hourly units, allowing monitoring data from different sub-areas to be analyzed on the same time scale. Through this spatial weight allocation and time series alignment, a hierarchical structure of spatiotemporal feature matrices is generated. The basic feature layer in this hierarchy stores the raw monitoring indicator statistics for each sub-area, such as the average oil temperature and average vibration frequency of equipment in Workshop A, and the average temperature and average humidity in Warehouse B. The spatiotemporal correlation layer stores spatiotemporal dependency coefficients between sub-areas, such as the coefficient between Workshop A and Warehouse B, determined based on previously calculated correlation strength and propagation probability. The event prediction layer stores predicted probabilities of abnormal events trained based on historical event data. For example, based on the frequency and propagation of abnormal events such as equipment failures and fires in Workshop A in the past, the probability of similar abnormal events occurring in Workshop A in the future is predicted. This hierarchical structure comprehensively and systematically describes the spatiotemporal characteristics of an industrial park, providing a strong data foundation for subsequent anomaly identification and early warning strategy development.

[0059] In a possible implementation, step S130 includes:

[0060] Step S131: input the spatiotemporal feature matrix into the multi-scale feature extraction module of the anomaly recognition network to obtain the anomaly pattern features of the target monitoring area at different time granularities and spatial resolutions.

[0061] Specifically, for an industrial park, the temporal granularity can be hours, minutes, or even seconds, and the spatial resolution can be at the workshop level, equipment level, or even smaller regional units. For example, at the hourly granularity of Workshop A, it might be discovered that the overall operating parameters of the equipment within the workshop experienced abnormal fluctuations within a certain hour. This is a characteristic of an abnormal pattern at a coarser time granularity. At the minute-level granularity, it might be discovered that the oil temperature of a key piece of equipment rose sharply within a certain minute. This is a characteristic of an abnormal pattern at a finer time granularity. From a spatial resolution perspective, at the workshop level, it might be discovered that the power consumption of the entire workshop A exceeded the normal range for a period of time. This is a characteristic of an abnormal pattern at the workshop-level spatial resolution. At the equipment-level spatial resolution, it might be determined that the abnormal current of a large piece of production equipment caused the abnormal power consumption at the workshop level. Furthermore, between different workshops, warehouses, and other sub-areas, such as between Workshop A and Warehouse B, it is possible to discover abnormal patterns in material transportation volume, environmental parameters, and other aspects at different temporal and spatial resolutions. These abnormal pattern characteristics encompass information at all levels of the industrial park, from macro to micro, providing a comprehensive data foundation for subsequent analysis.

[0062] In step S132, the attention allocation module of the anomaly recognition network dynamically allocates weights to the anomaly pattern features to determine the anomaly confidence of each sub-area in the target monitoring area and the correlation strength between abnormal events.

[0063] Specifically, in an industrial park, for an abnormal pattern characteristic, such as an abnormal increase in oil temperature on a piece of equipment in Workshop A, the attention allocation module assigns weights based on various factors. If this equipment is critical production equipment, the oil temperature increase is significant, and the surrounding ambient temperature is also significantly affected, then the anomaly confidence level for Workshop A sub-area will be assigned a higher value, such as 0.8. For Warehouse B, if only a slight humidity fluctuation is observed and there is little correlation with other sub-areas, the anomaly confidence level might be 0.2. Regarding the strength of the association between abnormal events, such as the association between the abnormal oil temperature of equipment in Workshop A and Warehouse B, if Workshops A and B are connected by a material transportation channel and ventilation system, the association strength between them may be higher, set at 0.6. However, if Office Building C has only a weak electrical connection to Workshop A, the association strength may be only 0.1.

[0064] Step S133: construct an abnormal event propagation network based on the abnormal confidence and the correlation strength. The nodes in the abnormal event propagation network represent sub-regions where the abnormal confidence exceeds the set confidence, and the edges represent the possibility of abnormal events propagating between sub-regions.

[0065] Specifically, in the abnormal event propagation network, a node represents a subregion where the confidence level of an abnormality detected exceeds a set confidence level (e.g., 0.5). Within the industrial park, Workshop A becomes a node because its equipment oil temperature abnormality confidence level is 0.8, exceeding the set value. Warehouse B also becomes a node if its combined abnormality confidence level exceeds 0.5 due to its connection with Workshop A. Edges represent the probability of an abnormal event propagating between subregions. The edge between Workshop A and Warehouse B represents the probability of an abnormal oil temperature event propagating from Workshop A to Warehouse B. This probability is determined based on factors such as the previously calculated connection strength and historical event propagation data, such as 0.6. For Workshop A and Office Building C, due to the lower connection strength, the probability of an abnormal event propagating from Workshop A to Office Building C may be only 0.1, which is also represented by the edge between them.

[0066] Step S134 : Based on the node degree distribution and edge weight distribution of the abnormal event propagation network, identify the core occurrence location of the potential abnormal event and the key nodes of the propagation path.

[0067] Specifically, in the abnormal event propagation network of an industrial park, node degree represents the number of edges connected to that node. If Workshop A has a high node degree, it indicates that it is associated with multiple sub-areas. This may mean that Workshop A is the core location of potential abnormal events, as abnormalities can propagate in multiple directions from here. Key nodes in the propagation path, such as the entrance to the material transportation channel between Workshop A and Warehouse B and the connection points of the ventilation system, may become key nodes because these locations play a key connecting and conducting role in the propagation of abnormal events. Edge weight distribution also helps identify key nodes. If the weight of an edge is very high, such as the edge weight of 0.6 from Workshop A to Warehouse B, then the connection point on this edge (such as the entrance to the material transportation channel) is more likely to be a key node.

[0068] Step S135 , performing path connection optimization on key nodes of the propagation path in combination with the time continuity of the spatiotemporal tag information, and generating a complete topological structure of the abnormal propagation path.

[0069] In a possible implementation, step S135 includes:

[0070] Step S1351, extract the timestamp sequence in the spatiotemporal tag information corresponding to the key nodes, and arrange the key nodes into a time-ordered node queue according to the chronological order of the timestamp sequence, wherein each node in the time-ordered node queue carries spatial coordinate information and a corresponding timestamp.

[0071] Specifically, assume that the timestamp of the abnormal oil temperature of the equipment in Workshop A is 10:00 AM, the timestamp of the abnormal airflow change detected at the ventilation system connection point between the equipment and Warehouse B is 10:10 AM, and the timestamp of the abnormal temperature increase in Warehouse B is 10:15 AM. These key nodes are arranged into a time-ordered node queue according to the chronological order of the timestamp sequence. Each node carries spatial coordinate information and a corresponding timestamp, such as the spatial coordinates (x1, y1, z1) of the equipment in Workshop A and the timestamp 10:00 AM, the spatial coordinates (x2, y2, z2) of the ventilation system connection point and the timestamp 10:10 AM, and the spatial coordinates (x3, y3, z3) of the temperature sensor in Warehouse B and the timestamp 10:15 AM.

[0072] Step S1352, calculate the timestamp intervals of adjacent nodes in the time-ordered node queue, divide the adjacent nodes whose timestamp intervals are less than a preset time continuity threshold into the same time continuity group, and record the start timestamp and end timestamp of each time continuity group.

[0073] Assuming the time continuity threshold is 15 minutes, in the above example, the timestamp intervals for the equipment in workshop A, the ventilation system connection point, and the temperature sensor in warehouse B are 10 minutes and 5 minutes, respectively, all less than 15 minutes. Therefore, they can be classified into the same time continuity group with a starting timestamp of 10:00 AM and an ending timestamp of 10:15 AM.

[0074] Step S1353, traverse the spatial coordinate information of the nodes in each time-continuous group, calculate the spatial distance matrix between the nodes in the group, mark the node pairs whose spatial distance is less than the preset spatial proximity threshold as spatially adjacent node pairs, and generate a spatial proximity relationship map based on the spatial distance matrix.

[0075] Specifically, the spatial distance between the equipment in Workshop A and the ventilation system connection point is d1, and the spatial distance between the ventilation system connection point and the temperature sensor in Warehouse B is d2. Based on these spatial distances, a spatial distance matrix is ​​calculated. Node pairs whose spatial distance is less than a preset spatial proximity threshold (e.g., 10 meters) are marked as spatially adjacent node pairs, and a spatial proximity relationship map is generated based on the spatial distance matrix. If the distance d1 between the equipment in Workshop A and the ventilation system connection point is less than 10 meters, and the distance d2 between the ventilation system connection point and the temperature sensor in Warehouse B is also less than 10 meters, both pairs of nodes are marked as spatially adjacent node pairs, and a spatial proximity relationship map is generated based on this information.

[0076] Step S1354: Based on the spatial proximity relationship graph, the nodes in the temporal continuous group that are in spatial proximity are linearly connected in timestamp order to form an initial propagation path segment, which includes path direction information and path connection strength parameters.

[0077] For example, the equipment in workshop A, the ventilation system connection point, and the temperature sensor in warehouse B are linearly connected in timestamp order. The path direction is from the equipment in workshop A to the temperature sensor in warehouse B. The path connection strength parameter can be determined based on factors such as the previously calculated association strength and spatial distance.

[0078] Step S1355 , detecting spatiotemporal conflicts between initial propagation path segments generated by different time-continuous groups, wherein the spatiotemporal conflicts include timestamp overlap conflicts of path segments in the same spatial region and topological conflicts of mismatch between path directions and spatial proximity relationship maps.

[0079] For example, suppose there's another propagation path segment involving the power supply system in Workshop A, running from the distribution room in Workshop A to the power equipment in Office Building C. This path segment may experience a spatiotemporal conflict with the previous path segment from the equipment in Workshop A to Warehouse B. Spatiotemporal conflicts include overlapping timestamps between path segments in the same spatial region and topological inconsistencies between the path directions and the spatial proximity graph. If events occur at a certain point in time on both path segments near Warehouse B, there's a timestamp overlap conflict. If the spatial proximity graph indicates that the connection direction of a path segment doesn't match the actual physical layout or logical relationship, there's a topological inconsistency.

[0080] Step S1356: Dynamically adjust the conflicting path segments based on the detected spatiotemporal conflict type to generate adjusted initial propagation path segments. Specifically, for timestamp overlap conflicts, prioritize the path segments with later timestamps and remove the overlapping segments. For topological conflicts, recalculate the connection directions of the path segments based on the spatial proximity graph.

[0081] For example, suppose the timestamp for the power equipment path segment from the power distribution room to Office Building C near Warehouse B is 10:20 AM, while the timestamp for the path segment from Workshop A to the temperature sensor in Warehouse B near Warehouse B is between 10:15 AM and 10:25 AM. Prioritize retaining the portion of the power equipment path segment from the power distribution room to Office Building C near Warehouse B, and remove the overlapping portion of the path segment from Workshop A to the temperature sensor in Warehouse B after 10:20 AM. For topological inconsistencies, the path segment connection direction is recalculated based on the spatial proximity graph. If the original connection direction of a path segment is opposite to the reasonable direction indicated by the spatial proximity graph, the correct connection direction is re-determined based on the graph.

[0082] Step S1357: Globally sort the adjusted initial propagation path segments according to the start timestamps of the time continuity groups, splice the path segments end to end based on the sorting results, and verify during the splicing process whether the timestamp intervals and spatial jump distances of adjacent path segments meet the temporal continuity threshold and spatial proximity threshold.

[0083] Step S1358: When it is detected that the timestamp interval or spatial jump distance of adjacent path segments exceeds a threshold, a transition node is inserted and the spatial proximity relationship map is updated. The spatial coordinates of the transition node are generated by interpolation based on the spatial coordinates of the end nodes of the adjacent path segments, and the timestamps are linearly allocated based on the timestamp interval of the adjacent path segments.

[0084] Assume that after adjustment, there are three path segments, starting at 10:00 AM, 10:30 AM, and 11:00 AM respectively. After global sorting by the starting timestamp, the path segments at 10:00 AM and 10:30 AM are first spliced ​​end to end to verify whether the timestamp interval between them is less than 15 minutes and whether the spatial jump distance is less than 10 meters. If the conditions are met, continue splicing the next path segment; if not, when it is detected that the timestamp interval or spatial jump distance of adjacent path segments exceeds the threshold, a transition node is inserted and the spatial proximity relationship map is updated. For example, if the timestamp interval between the path segment at 10:30 AM and the path segment at 11 AM is more than 15 minutes than 30 minutes, and the spatial jump distance is more than 10 meters than 15 meters, a transition node is inserted at this time. The spatial coordinates of the transition node are generated by interpolating the spatial coordinates of the end nodes of the adjacent path segments. Assume that the spatial coordinates of the end node of the path segment at 10:30 am are (x4, y4, z4), and the spatial coordinates of the starting node of the path segment at 11 am are (x5, y5, z5). The spatial coordinates of the transition node (x6, y6, z6) are calculated according to the interpolation algorithm, and the timestamp is linearly assigned to 10:45 am based on the timestamp interval of the adjacent path segments.

[0085] Step S1359: Based on the final set of spliced ​​path segments and the inserted transition nodes, a complete topological structure of the abnormal propagation path is generated. The complete topological structure includes path branch information, node connection order, and path propagation direction weight parameters.

[0086] For example, a path branch runs from the equipment in Workshop A through the ventilation system connection point to the temperature sensor in Warehouse B. Another path branch runs from the power distribution room in Workshop A to the power equipment in Office Building C. The node connection order follows the previous connection order, such as first the equipment in Workshop A, then the ventilation system connection point, and finally the temperature sensor in Warehouse B. The path propagation direction weight parameter is determined based on previously calculated factors such as correlation strength and spatial distance. For example, the propagation direction weight from the equipment in Workshop A to the ventilation system connection point is 0.8, and the propagation direction weight from the ventilation system connection point to the temperature sensor in Warehouse B is 0.6. This complete topology accurately describes the propagation path of potential abnormal events within the industrial park.

[0087] In a possible implementation, step S140 includes:

[0088] Step S141 , calculating the priority index of each sub-region in the potential abnormal event propagation chain according to the topological structure of the abnormal propagation path, wherein the priority index is used to quantify the urgency of the sub-region being affected by the abnormal event.

[0089] For example, in an industrial park, let's continue with the aforementioned example of abnormal oil temperature in Workshop A. If the abnormality propagates from Workshop A through a ventilation system connection point to Warehouse B, and then to some small equipment storage areas adjacent to Warehouse B, Workshop A, as the source of the abnormality, is directly affected by the abnormal oil temperature, potentially leading to serious consequences such as equipment damage, production halts, or even fires, thus giving Workshop A the highest priority index, set at 0.9. Warehouse B, directly connected to Workshop A and containing materials potentially affected by fire or equipment failure, has the next highest priority index, set at 0.7. Small equipment storage areas, while slightly further away and less likely to be affected, still present a risk and may receive a priority index of 0.5. This priority index calculation takes into account the sub-area's distance from the abnormality source, the closeness of the propagation path, and the sub-area's importance (such as the value of the stored materials and the criticality of the equipment).

[0090] Step S142: determining a trigger threshold range of the dynamic warning strategy based on attribute parameters of the potential abnormal event, wherein the attribute parameters include abnormal intensity, propagation speed, and impact duration.

[0091] Specifically, for a potential abnormal event involving abnormal oil temperature in equipment in Workshop A, the severity of the abnormality is measured by the percentage by which the oil temperature exceeds the normal range. For example, a 30% excess is considered a high-severity abnormality. The propagation rate is estimated based on historical and current monitoring data. Based on factors such as ventilation system airflow velocity and material transport speed, the abnormal event is expected to affect an area within a 10-meter radius per hour. The duration of the impact is estimated to be two hours, based on the equipment's maximum operating time at abnormal oil temperatures and a model for the attenuation of the abnormal event during propagation. Based on these attributes, such as the abnormality's severity, propagation rate, and duration, the warning trigger threshold range is set. For example, when the oil temperature exceeds the normal range by 15%, the warning preparation phase begins; when the oil temperature exceeds the normal range by 20%, a formal warning is triggered.

[0092] Step S143 : generating a multi-level warning triggering rule according to the priority index and the triggering threshold range, wherein the multi-level warning triggering rule defines warning levels and response delay times for sub-areas of different priorities under different abnormal intensities.

[0093] Specifically, for Workshop A, with a priority index of 0.9, a Level 1 alert is triggered when the oil temperature exceeds the normal range by 20%, with a high alert level and a response delay of 0 minutes. This means that once this threshold is reached, alert-related actions are immediately executed. For Warehouse B, with a priority index of 0.7, a Level 2 alert is triggered when the oil temperature exceeds the normal range by 20%, with a medium alert level and a response delay of 3 minutes. Because the impact on Warehouse B is slightly less urgent than that on Workshop A, a 3-minute response delay is given to allow for preliminary assessment and preparation. For the small equipment storage area, with a priority index of 0.5, a Level 3 alert is triggered when the oil temperature exceeds the normal range by 25%, with a low alert level and a response delay of 5 minutes. Because the likelihood and urgency of the impact are relatively low, the alert trigger threshold is relatively high, and the response delay is also relatively long.

[0094] Step S144: construct a response instruction set that matches the multi-level warning triggering rule, wherein the response instruction set includes equipment control instructions, personnel dispatch instructions, and data reporting protocols for each warning level.

[0095] Specifically, for the Level 1 alert in Workshop A, the equipment control command is to immediately halt the operation of the relevant equipment to prevent further damage and the escalation of the abnormal event. The personnel dispatch command is to dispatch maintenance personnel, safety personnel, and technical experts to Workshop A. The maintenance personnel are responsible for inspecting and repairing the equipment, the safety personnel are responsible for on-site safety management, and the technical experts are responsible for conducting a comprehensive assessment of the abnormal event. The data reporting protocol requires reporting of equipment status (including oil temperature, vibration frequency, current intensity, etc.), environmental data (temperature, humidity, air quality, etc.), and personnel operation records every 30 seconds. For the Level 2 alert in Warehouse B, the equipment control command is to conduct a preliminary inspection of potentially affected equipment, such as ventilation equipment and fire prevention facilities. The personnel dispatch command is to notify warehouse management and some security personnel to conduct inspections in the warehouse to ensure the safety of materials. The data reporting protocol requires reporting of warehouse temperature, humidity, smoke concentration, and other data every 2 minutes. For the third-level warning in the small equipment storage area, the equipment control instruction is to conduct a simple status observation of the equipment, such as checking whether the equipment has any abnormal sounds or odors; the personnel dispatch instruction is to notify nearby staff to pay attention to the equipment status; the data reporting protocol requires reporting the basic operating status data of the equipment every 5 minutes.

[0096] Step S145 , dynamically binding the multi-level warning triggering rules and the response instruction set according to the time sequence of the abnormality propagation path to form an execution logic chain of the dynamic warning strategy.

[0097] Specifically, as the anomaly propagates from Workshop A to Warehouse B and then to the small equipment storage area, the equipment control instructions, personnel scheduling instructions, and data reporting protocols associated with Workshop A's Level 1 warning are executed first. Then, actions are taken based on Warehouse B's Level 2 warning status. Finally, actions are taken based on the Small Equipment Storage Area's Level 3 warning status. This way, starting from the source of the anomaly and following its propagation, warning and response operations are carried out in an orderly fashion according to pre-set rules and instructions, forming a complete logical chain for the execution of dynamic warning strategies.

[0098] In a possible implementation, step S150 includes:

[0099] Step S151 : obtaining feedback log data generated when the terminal device cluster executes the historical warning strategy, wherein the feedback log data includes the warning response delay time, the device execution status, and the actual impact range of the abnormal event.

[0100] For example, in the recent abnormal oil temperature event at Workshop A, the early warning response delay for Workshop A was recorded as 1 minute (although it was set to 0 minutes, actual factors such as communication and personnel response resulted in a 1-minute delay). After the equipment's execution status indicated that the equipment had stopped operating, some associated equipment experienced a certain degree of impact due to the sudden stop, such as a brief pressure fluctuation in the cooling system connected to the equipment. The actual impact of the abnormal event was slightly larger than previously predicted, affecting not only Warehouse B but also a small tool room adjacent to Warehouse B, where some tools were damaged by the high temperature.

[0101] Step S152: extracting policy execution effect indicators from the feedback log data, wherein the policy execution effect indicators include warning accuracy, response efficiency deviation, and resource consumption rate.

[0102] For example, the accuracy of early warnings decreased because the actual impact range exceeded the predicted range. Originally predicted to affect only Workshop A and Warehouse B, the impact actually affected the small tool room. Response efficiency deviations were manifested as a 1-minute response delay in Workshop A, deviating from the planned 0-minute response time. Regarding resource consumption, the sudden equipment shutdown required the cooling system to be readjusted, consuming additional electricity and manpower to restore normal operation. The emergency dispatch of maintenance and safety personnel also consumed a considerable amount of human resources.

[0103] Step S153 : comparing the strategy execution effect index with a preset optimization objective function to determine the trigger condition parameters and response instruction priority that need to be adjusted in the dynamic early warning strategy.

[0104] For example, suppose the optimization objective function includes a warning coverage efficiency term (positively correlated with warning accuracy), a resource consumption penalty term (negatively correlated with the energy consumption and computing resource usage of the equipment when executing warning instructions), and a response timeliness reward term (positively correlated with warning response speed). Due to decreased warning accuracy and increased resource consumption, the warning trigger conditions need to be adjusted. For example, the first-level warning trigger condition for workshop A can be adjusted from oil temperature exceeding the normal range by 20% to 18% to improve the timeliness and coverage of warnings. At the same time, the priority of the equipment shutdown instruction can be lowered, and a transition instruction for equipment power reduction can be added to reduce the impact of sudden equipment shutdown on associated equipment, thereby reducing resource consumption.

[0105] Step S154 , iteratively correcting the trigger threshold range and response delay time of the dynamic warning strategy according to the gradient direction of the optimization objective function to obtain corrected parameters.

[0106] For example, by analyzing the sensitivity of the optimization objective function to various parameters, we found that the warning trigger threshold significantly impacted warning accuracy and resource consumption, while the response delay significantly affected response efficiency deviation. Based on these analysis results, we adjusted the first-level warning trigger threshold for Workshop A to 18%, and reassessed its impact on warning accuracy and resource consumption. We also adjusted the response delay for Workshop A from 0 minutes to 30 seconds, considering its effect on improving response efficiency deviation. Through multiple iterations of corrections, we ultimately achieved optimal values ​​for each parameter.

[0107] Step S155 , recalculating the coverage and resource allocation weight of each level of warning in the multi-level warning triggering rule based on the revised parameters, and generating an optimized dynamic warning strategy.

[0108] For example, if the new warning trigger threshold for Workshop A is 18%, the coverage of each level of warning at different abnormal intensities under this threshold will be reassessed. For example, when the oil temperature exceeds the normal range by 18%-25%, it is covered by the first-level warning, and 25%-35% is covered by higher-level warnings (if any). At the same time, resource allocation weights are reallocated based on resource consumption. For example, more maintenance resources will be prioritized for equipment inspection and repair in Workshop A during the first-level warning, while the proportion of resources allocated to Warehouse B during the second-level warning will be reduced (because the adjusted strategy may reduce the probability of Warehouse B being affected). Through such adjustments, an optimized dynamic warning strategy is generated, which can play a more efficient and accurate role in responding to subsequent abnormal events.

[0109] In a possible implementation, before step S110, the method further includes:

[0110] Step S210 : dividing a plurality of monitoring data collection units according to the geographical distribution characteristics of the target monitoring area, wherein each of the collection units is configured with at least three heterogeneous sensors and a data preprocessing module.

[0111] Specifically, industrial parks are geographically large and encompass diverse functional areas, such as production workshops, warehouses, office buildings, and park roads. Collection units are divided based on the characteristics of these distinct areas. For example, Workshop A is designated as a collection unit. Due to the large number of devices and complex production processes within the workshop, the heterogeneous sensors installed here include equipment operating status sensors (such as those monitoring equipment vibration frequency, oil temperature, and current intensity), environmental sensors (such as those monitoring temperature, humidity, and air quality), and sensors for monitoring human activity (such as human location sensors and human traffic counters). Each collection unit is equipped with a data preprocessing module. For Workshop A, this module is responsible for performing preliminary processing on the data collected by the sensors.

[0112] Step S220 : setting the data sampling frequency and transmission protocol of the heterogeneous sensors so that different types of sensor data in the same acquisition unit are transmitted through independent channels after their timestamps are aligned.

[0113] For example, within the collection unit in Workshop A, the equipment status sensor is set to collect data every minute due to its rapidly changing operating status. Environmental sensors, which change more slowly, are set to collect data every five minutes. The activity monitoring sensor, based on the characteristics of personnel flow, is set to collect data every 30 seconds. Furthermore, to ensure that different sensor data types within the same collection unit are transmitted over independent channels after timestamp alignment, a unified time synchronization protocol is implemented. For example, all sensors are time-calibrated using a high-precision clock source within the campus to ensure accurate and consistent timestamps on the collected data. Data from the equipment status sensor is transmitted over an industrial Ethernet channel, data from the environmental sensor is transmitted over a ZigBee channel, and data from the activity monitoring sensor is transmitted over a Wi-Fi channel. This ensures that different sensor data types are transmitted over independent channels, avoiding interference and ensuring orderly transmission after timestamp alignment.

[0114] Step S230: configuring a noise filtering algorithm and a data integrity check rule for the data preprocessing module. The noise filtering algorithm uses a wavelet transform method with an adaptive threshold to eliminate environmental interference components in the sensor signal.

[0115] For example, in the acquisition unit of Workshop A, the data preprocessing module uses an adaptive threshold wavelet transform to eliminate environmental interference from sensor signals. For equipment status sensors, electromagnetic interference and vibration interference may be generated by electrical equipment and mechanical operation within the workshop. This adaptive threshold wavelet transform automatically adjusts the threshold based on the characteristics of the sensor signal, filtering out these interference components as noise. For example, an equipment vibration frequency sensor may be affected by low-frequency vibrations from other equipment within the workshop. The wavelet transform algorithm accurately identifies and filters this interference. Furthermore, data integrity verification rules are configured to perform integrity checks on the collected data. For example, these checks include ensuring that the data length meets requirements and that there are no missing values. If data is found to be incomplete, the data preprocessing module will flag the data and attempt to repair it through methods such as interpolation or notify relevant personnel for inspection.

[0116] Step S240: establishing a bidirectional communication link between the acquisition unit and the central server, wherein the bidirectional communication link supports real-time uploading of the multi-source monitoring data stream and synchronous issuance of control instructions.

[0117] For example, a two-way communication link is established between the collection unit in Workshop A and the central server. This link supports the real-time upload of multi-source monitoring data streams and the simultaneous issuance of control instructions. During normal operation, the data collected by the sensors in the collection unit in Workshop A is processed by the data preprocessing module and then uploaded to the central server in real time via the two-way communication link. For example, equipment oil temperature data collected by the equipment operation status sensor, temperature data collected by the environmental sensor, and personnel location data collected by the personnel activity monitoring sensor are all uploaded to the central server according to the set time interval and transmission protocol. At the same time, the central server can also issue control instructions to the collection unit in Workshop A through this two-way communication link. For example, if the central server needs to adjust the data sampling frequency of the equipment operation status sensor in Workshop A based on the overall operation of the park, it can issue the corresponding control instructions to the collection unit through this link. After receiving the instructions, the collection unit will adjust the data sampling frequency of the sensor as required.

[0118] Step S250 , regularly performing self-checking and calibration on the acquisition unit, generating a sensor health status report and dynamically adjusting the data sampling frequency and noise filtering parameters.

[0119] In this embodiment, the acquisition unit in Workshop A performs self-tests and calibrations regularly (e.g., at 2:00 AM daily). During this self-test, the equipment operating status sensor performs a self-test to check performance indicators such as sensor sensitivity and measurement range. Environmental sensors check their calibration parameters, and personnel activity monitoring sensors test signal transmission stability. A sensor health status report is generated based on the self-test results. If the sensitivity of the equipment operating status sensor decreases, the collected data may be inaccurate. In this case, the data sampling frequency is dynamically adjusted based on the sensor health status report, for example, by appropriately increasing the sampling frequency to obtain more data for analysis and compensation. Furthermore, if changes in environmental interference within the workshop are detected (e.g., the addition of a new large piece of equipment increases electromagnetic interference), the noise filtering parameters are dynamically adjusted to ensure data accuracy.

[0120] In one possible implementation, the hierarchical structure of the spatiotemporal feature matrix includes a basic feature layer, a spatiotemporal correlation layer, and an event prediction layer. The basic feature layer stores the original monitoring indicator statistics of each sub-region, the spatiotemporal correlation layer stores the spatiotemporal dependency coefficients between sub-regions, and the event prediction layer stores the predicted values ​​of the probability of abnormal events trained based on historical event data.

[0121] For example, in a large industrial park scenario, the basic feature layer stores the raw monitoring indicator statistics for each sub-area. For example, in Workshop A, the basic feature layer stores raw monitoring indicator statistics such as the average oil temperature, average vibration frequency, maximum temperature, and minimum humidity of the equipment within Workshop A. These statistics are calculated based on the large amount of real-time monitoring data uploaded by Workshop A's collection units. For example, if 100 samples of equipment oil temperature data are collected within an hour, the average, maximum, and minimum values ​​of these samples are calculated and stored in the basic feature layer.

[0122] The spatiotemporal correlation layer stores the spatiotemporal dependency coefficients between sub-areas. In an industrial park, Workshop A and Warehouse B are different sub-areas, but they share a spatiotemporal dependency relationship. For example, production activities in Workshop A may affect the storage environment (such as temperature and humidity) in Warehouse B. By analyzing and calculating a large amount of historical and real-time data, the spatiotemporal dependency coefficient between Workshop A and Warehouse B is calculated. This coefficient reflects the extent and likelihood of changes in Workshop A affecting Warehouse B. If the production equipment in Workshop A increases output over a period of time, causing the temperature inside the workshop to rise, and Warehouse B is connected to Workshop A through a ventilation system, a spatiotemporal dependency coefficient is calculated based on factors such as the correlation of temperature changes and the ventilation system's flow rate and stored in the spatiotemporal correlation layer.

[0123] The event prediction layer stores predicted values ​​for the occurrence probability of abnormal events, trained based on historical event data. For industrial parks, this value is trained and stored in the event prediction layer based on historical event data, such as fires and equipment failures. For example, if Workshop A has experienced multiple equipment failures due to excessive oil temperatures, the relevant data (such as oil temperature trends, equipment operating time, and ambient temperature) can be analyzed and model trained to derive a predicted probability of equipment failure (excessive oil temperature) in Workshop A, given the current monitoring data.

[0124] The method further comprises:

[0125] Step S310: updating the statistical values ​​in the basic feature layer according to the real-time monitoring data, and triggering the dynamic recalculation of the dependency coefficients in the spatiotemporal correlation layer.

[0126] During the actual operation of Workshop A, real-time monitoring data is continuously uploaded. For example, as production progresses, equipment operating status sensors continuously collect equipment oil temperature data. Each time a new set of oil temperature data is collected, the average equipment oil temperature statistics in the basic feature layer are updated. When equipment oil temperature data undergoes significant changes (such as a sudden increase in oil temperature), this change may affect the spatiotemporal dependency relationship between Workshop A and other sub-areas (such as Warehouse B). This triggers a dynamic recalculation of the dependency coefficient between Workshop A and Warehouse B in the spatiotemporal correlation layer. Because the increase in oil temperature may affect the environment of Warehouse B through the ventilation system or other channels, the relationship coefficient between them needs to be reassessed.

[0127] Step S320: When the predicted value of the probability of occurrence of an abnormal event in the event prediction layer exceeds a preset threshold, the incremental learning module of the abnormality recognition network is activated to update the calculation model of the predicted value of the probability of occurrence of the abnormal event based on the latest monitoring data.

[0128] Assume that in Workshop A, the preset probability threshold for equipment failure (due to excessive oil temperature) is 30%. When the predicted probability of equipment failure in the event prediction layer exceeds 30%, the incremental learning module of the anomaly recognition network is activated. This incremental learning module retrains and updates the calculation model for the abnormal event probability prediction value based on the latest data collected from Workshop A, such as equipment oil temperature, equipment operating time, and ambient temperature. For example, it may adjust model parameters, consider new factors, or re-evaluate the weights of existing factors to more accurately predict the probability of equipment failure.

[0129] Step S330 : Dynamically maintain the hierarchical structure of the spatiotemporal feature matrix through a sliding time window mechanism, retain feature data within a preset time length, and remove expired data blocks.

[0130] In the industrial park's monitoring system, a sliding time window is set to one day. For monitoring data from Workshop A, only feature data from the most recent day, including the basic feature layer statistics, the spatiotemporal correlation layer dependency coefficients, and the event prediction layer's predicted probability of abnormal events, is retained in the spatiotemporal feature matrix. As time passes, when a new day's data collection begins, data blocks from one day ago (including data corresponding to the basic feature layer, spatiotemporal correlation layer, and event prediction layer) are deemed expired and removed. This ensures that the data in the spatiotemporal feature matrix is ​​always up-to-date and most relevant to the current situation, while also avoiding the waste of computing and storage resources caused by excessive data volumes.

[0131] In one possible embodiment, the anomaly recognition network includes a spatiotemporal convolution module, a long short-term memory module, and a graph neural network module. The spatiotemporal convolution module is used to extract local spatiotemporal patterns of monitoring data, the long short-term memory module is used to capture long-term dependencies in time series, and the graph neural network module is used to model spatial correlations between sub-regions.

[0132] The method further comprises:

[0133] Step S410: dividing the spatiotemporal feature matrix into multiple segments according to the time dimension, and inputting each segment into the spatiotemporal convolution module for feature dimensionality reduction and pattern extraction.

[0134] In the context of an industrial park, the spatiotemporal feature matrix contains features related to monitoring data from various sub-areas (such as workshops and warehouses). Assume the spatiotemporal feature matrix covers a full day of monitoring data, segmented into hourly time segments. For Workshop A, the feature components, consisting of data such as equipment operating status (oil temperature, vibration frequency, etc.) and environmental parameters (temperature, humidity, etc.), are segmented into 24-hour segments. These segments are then fed into the spatiotemporal convolution module. The spatiotemporal convolution module uses a specific convolution kernel to perform convolution operations in both time and space to extract local spatiotemporal patterns in the monitoring data. For example, using the oil temperature data from equipment in Workshop A, the spatiotemporal convolution module can identify oil temperature fluctuation patterns during specific time periods (such as the morning peak production period). These fluctuations may be related to factors such as equipment load changes and ambient temperature fluctuations. Furthermore, the spatiotemporal convolution module can also extract local spatial patterns from oil temperature sensor data at different locations within the workshop. For example, oil temperature fluctuations at sensors near core equipment components may be more dramatic than those at edge locations. Therefore, the spatiotemporal convolution module performs feature dimensionality reduction and pattern extraction on the input data of each time segment, converting the high-dimensional raw data into more representative low-dimensional features.

[0135] Step S420: The output of the spatiotemporal convolution module is reorganized according to the spatial dimension and input into the graph neural network module to generate a spatial correlation feature vector for each sub-region.

[0136] Specifically, the output of the spatiotemporal convolution module for each time segment contains local spatiotemporal features, but these local spatiotemporal features need to be further integrated to reflect the spatial connections between sub-regions. Taking Workshop A and Warehouse B as an example, the operating status of the equipment in Workshop A may be correlated with the environment and the status of stored items in Warehouse B through factors such as the ventilation system and material transportation channels. The data related to Workshop A and Warehouse B output by the spatiotemporal convolution module is reorganized according to the spatial dimension and input into the graph neural network module. The graph neural network module constructs a graph structure based on the physical layout of the industrial park and the connectivity between sub-regions. Nodes represent sub-regions such as Workshop A and Warehouse B, and edges represent the connections between them (such as channels and lines). The graph neural network module generates a spatial correlation feature vector for each sub-region. For Workshop A, its spatial correlation feature vector may include information such as the strength of spatial connections with adjacent workshops and warehouses and the likelihood of mutual influence. For example, if the ventilation system connectivity between workshop A and warehouse B is good, and historical data shows that environmental changes in workshop A have a greater impact on warehouse B, then in the spatial correlation feature vector of workshop A, the elements related to warehouse B will reflect this strong correlation.

[0137] Step S430: Input the spatial correlation feature vector into the long short-term memory module in chronological order to generate a global spatiotemporal state representation of the target monitoring area.

[0138] For example, the states of sub-regions within an industrial park change continuously over time. The LSTM module can capture long-term dependencies within these time series. The previously obtained spatial correlation feature vectors for each sub-region are input into the LSTM module in chronological order. For example, over a period of time, information such as the operating status of equipment in Workshop A and its spatial correlations with other sub-regions forms a time series. The LSTM module can learn the dependencies between state changes in Workshop A at different points in time. For example, equipment maintenance performed the previous day may affect today's operating efficiency, which in turn affects its correlations with other sub-regions. By processing the spatial correlation feature vectors of all sub-regions across the entire industrial park in a time series, the LSTM module generates a global spatiotemporal state representation of the target monitored area (the industrial park). This global spatiotemporal state representation integrates the spatial correlations and state change information of each sub-region at different points in time, comprehensively reflecting the operational status of the industrial park.

[0139] Step S440: Fusing the output features of the spatiotemporal convolution module, graph neural network module, and long short-term memory module to generate a comprehensive discrimination index for abnormal event identification.

[0140] For example, the spatiotemporal convolution module provides local spatiotemporal pattern features, the graph neural network module identifies spatial correlation features between subregions, and the long-term short-term memory module captures long-term dependencies in time series. The output features of these three modules are fused. In industrial parks, this integrated discriminant metric is crucial for identifying potential abnormal events, such as equipment failure or fire risk. For example, for equipment failure, the spatiotemporal convolution module might identify abnormal operating patterns of a piece of equipment in local time and space (e.g., abnormal oil temperature fluctuations confined to a specific component of the equipment). The graph neural network module reflects the spatial correlation between the equipment and surrounding subregions (e.g., a nearby warehouse may be threatened by a fire caused by equipment failure). The long-term short-term memory module considers changes in the equipment's operating status over longer periods of time (e.g., frequent recent load fluctuations may lead to equipment fatigue). The discriminant metric formed by combining this information can more accurately determine abnormal event scenarios, such as the presence of equipment failure risk and the potential impact of the failure.

[0141] For example, in one possible implementation, the optimization objective function includes a warning coverage efficiency item, a resource consumption penalty item, and a response time reward item. The warning coverage efficiency item is positively correlated with the warning accuracy of the dynamic warning strategy, the resource consumption penalty item is negatively correlated with the energy consumption and computing resource occupancy when the device executes the warning instruction, and the response time reward item is positively correlated with the warning response speed.

[0142] In industrial parks, early warning accuracy is directly related to the timely detection and accurate identification of abnormal events. For example, for equipment failure early warning, high early warning coverage efficiency means a high probability of accurately detecting equipment failures in different sub-areas (e.g., equipment in each workshop). If the early warning system can issue a warning signal promptly before a critical piece of equipment in Workshop A experiences an oil temperature overheating fault, the early warning accuracy is high, and accordingly, the value of the early warning coverage efficiency term will also be high. This requires comprehensive consideration of factors such as the accuracy of the monitoring data, the performance of the anomaly recognition network, and the rationality of the early warning strategy.

[0143] When executing warning instructions, devices may consume additional energy and computing resources. For example, when an oil temperature warning is triggered in workshop A, the equipment may need to adjust its operating mode or activate the emergency cooling system, which consumes electricity. Furthermore, servers and other devices in the monitoring system consume computing resources when processing warning-related data and instructions. If these energy and resource usage are excessive, the resource consumption penalty will increase. To reduce this penalty, it is necessary to optimize device control instructions and improve the efficiency of computing resource utilization. For example, more energy-efficient device control algorithms and the rational allocation of server computing tasks are necessary.

[0144] Rapid response to abnormal events is crucial. When equipment in Workshop A experiences an abnormality, if the early warning system can quickly issue a warning signal, enabling maintenance personnel and safety personnel to arrive quickly and take action, the reward for timely response will be higher. This depends on factors such as the speed of the warning signal transmission, the personnel's response mechanism, and the operability of the on-site equipment.

[0145] The method further comprises:

[0146] Step S510: Dynamically adjust the weight coefficients of each item in the optimization objective function according to the real-time resource status of the target monitoring area.

[0147] Step S520: When it is detected that the computing resource occupancy rate of the terminal device cluster exceeds a preset threshold, the weight of the resource consumption penalty item is increased and the weight of the response time reward item is reduced.

[0148] Step S530: When the propagation speed of the potential abnormal event exceeds a preset safety threshold, the weights of the early warning coverage efficiency item and the response time reward item are increased.

[0149] Step S540: generating a weight adjustment strategy for the optimization objective function based on a reinforcement learning algorithm, so that the dynamic early warning strategy can achieve a comprehensive optimization effect under resource constraints.

[0150] The weight coefficients of each item in the optimization objective function are dynamically adjusted based on the real-time resource status of the target monitoring area. The resource status of an industrial park is constantly changing. For example, during peak production periods, equipment energy consumption is high and computing resources are relatively limited. If the computing resource utilization of a terminal device cluster exceeds a preset threshold (e.g., 80%), the weight of the resource consumption penalty item is increased and the weight of the response time reward item is decreased to prevent excessive resource consumption. This means that in this situation, reducing resource consumption is prioritized, potentially sacrificing response speed. However, when the propagation speed of a potential abnormal event exceeds a preset safety threshold, such as a fire spreading rapidly within a workshop, the weights of the warning coverage efficiency item and the response time reward item are increased. This is because in such emergencies, quickly and accurately issuing warnings and taking measures to prevent the spread of the fire are paramount, even if this may consume more resources. A weight adjustment strategy for the optimization objective function generated based on a reinforcement learning algorithm enables the dynamic warning strategy to achieve comprehensive optimization results within resource constraints. The reinforcement learning algorithm continuously interacts with the actual operating environment of the industrial park to learn the optimal weight adjustment strategy for different situations. For example, by repeatedly simulating the early warning process under different equipment failure scenarios and different resource states, we can gradually determine how to adjust the weight coefficients in various situations to achieve a comprehensive balance between early warning accuracy, response speed and resource consumption while meeting resource constraints.

[0151] For example, in one possible implementation, the method further includes:

[0152] Step S610: deploying an edge computing node in the terminal device cluster, wherein the edge computing node stores the simplified execution logic of the dynamic warning strategy and local emergency response rules.

[0153] Within the industrial park, edge computing nodes are distributed near various sub-areas, such as Workshop A and Warehouse B. The simplified execution logic of the dynamic early warning policies stored in these edge computing nodes is simplified based on the complete policy issued by the central server, allowing for rapid local execution. Local emergency response rules are specific rules formulated for abnormal events that may occur locally. For example, the local emergency response rules of the edge computing node in Workshop A might include immediately activating the local ventilation system and alarm devices within the workshop when the smoke concentration reaches a certain threshold, without waiting for instructions from the central server.

[0154] Step S620: When the communication link between the central server and the terminal device cluster is interrupted, the autonomous warning mode of the edge computing node is activated, and an emergency warning instruction is generated based on the most recently received complete dynamic warning strategy and local monitoring data.

[0155] Suppose a network failure disrupts the communication link between the central server and the terminal device clusters in sub-areas such as Workshop A and Warehouse B. The edge computing node in Workshop A generates an emergency warning command based on the most recently received complete dynamic warning strategy (including warning trigger conditions and response instructions for different abnormal events) and local monitoring data (such as oil temperature, temperature, and smoke concentration data for equipment within the workshop). If local monitoring data indicates that the oil temperature of a piece of equipment within the workshop is too high and approaching the warning threshold, the edge computing node, based on simplified execution logic and local emergency response rules, issues a command to reduce the equipment's load and notifies workshop staff to conduct an inspection.

[0156] Step S630: After the communication link is restored, the execution log and data difference record generated by the edge computing node in the autonomous warning mode are synchronized to the central server.

[0157] When the communication link returns to normal, the edge computing node in workshop A will synchronize the execution log in the autonomous early warning mode (including the instructions issued, execution time, equipment response, etc.) and the difference records between the local monitoring data and the central server's expected data (such as the difference between the actual change in oil temperature and the central server's predicted value) to the central server.

[0158] Step S640: performing consistency check on the local execution logic of the dynamic early warning strategy and updating the strategy version according to the execution log and data difference record.

[0159] Specifically, after receiving the execution logs and data discrepancy records from the edge computing nodes in Workshop A, the central server performs a consistency check on the local execution logic of Workshop A's dynamic early warning strategy. For example, it checks whether the instructions issued by the edge computing nodes comply with the overall early warning strategy principles and whether there are any misoperations or unreasonable instructions. Simultaneously, the dynamic early warning strategy is updated based on the data discrepancy records. If significant discrepancies are detected between the local monitoring data and the central server's predicted data, it may be necessary to adjust certain parameters in the early warning strategy (such as warning thresholds and response instructions) to improve the accuracy and adaptability of the early warning strategy and ensure better response to similar situations in the future.

[0160] Figure 2 The schematic diagram shows exemplary hardware and software components of a server 100 that can implement the concepts of the present application according to some embodiments of the present application. For example, a processor 120 can be used on the server 100 to perform the functions of the present application.

[0161] The server 100 can be a general-purpose server or a special-purpose server, both of which can be used to implement the artificial intelligence-based monitoring, early warning and analysis method of the present application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0162] For example, the server 100 may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the server 100 may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application may be implemented based on these program instructions. The server 100 also includes an input / output (I / O) interface 150 between the computer and other input / output devices.

[0163] For ease of explanation, only one processor is described in the server 100. However, it should be noted that the server 100 in the present application may also include multiple processors, so the steps performed by one processor described in the present application may also be performed jointly or individually by multiple processors. For example, if the processor of the server 100 performs step A and step B, it should be understood that step A and step B may also be performed jointly by two different processors or individually in one processor. For example, the first processor performs step A and the second processor performs step B, or the first processor and the second processor perform steps A and B together.

[0164] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the above-mentioned artificial intelligence-based monitoring, early warning and analysis method is implemented.

[0165] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, multiple features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. A monitoring and early warning analysis method based on artificial intelligence, characterized in that: The method comprises: Collecting a multi-source monitoring data stream within a target monitoring area, the multi-source monitoring data stream including at least two types of real-time monitoring data and corresponding spatiotemporal tag information, the spatiotemporal tag information being used to describe the collection location and timestamp of the real-time monitoring data within the target monitoring area; Extracting spatiotemporal features from the multi-source monitoring data stream to generate a spatiotemporal feature matrix with a hierarchical association relationship, wherein each element in the spatiotemporal feature matrix corresponds to a monitoring indicator feature of a target position in the target monitoring area within a target time window; Performing dynamic pattern analysis on the spatiotemporal feature matrix based on a preset anomaly recognition network to determine potential abnormal events and their corresponding abnormal propagation paths within the target monitoring area, wherein the abnormal propagation paths are used to indicate the diffusion direction and impact range of the potential abnormal events within the target monitoring area; Generate a dynamic early warning strategy adapted to the target monitoring area based on the topological structure of the abnormal propagation path and the attribute parameters of the potential abnormal event, wherein the dynamic early warning strategy includes differentiated early warning trigger conditions and response instruction sets for different locations and different time stages; Optimize and adjust the dynamic warning strategy in real time based on historical warning feedback data, and output the optimized dynamic warning strategy to the terminal device cluster corresponding to the target monitoring area; The step of extracting spatiotemporal features from the multi-source monitoring data stream to generate a spatiotemporal feature matrix having a hierarchical correlation relationship includes: Dividing the multi-source monitoring data stream into a plurality of spatiotemporal data blocks according to the spatiotemporal tag information, each of the spatiotemporal data blocks corresponding to a set of monitoring data of a preset sub-area in the target monitoring area within a continuous time series; For each of the spatiotemporal data blocks, extracting cross-modal correlation features between the monitoring data contained in the spatiotemporal data block, wherein the cross-modal correlation features are used to describe the coupling relationship between different types of monitoring data in the same spatiotemporal dimension; Fusing the cross-modal correlation feature with the time series fluctuation feature of the spatiotemporal data block to generate a local spatiotemporal feature vector corresponding to the spatiotemporal data block; Based on the physical connection relationship between the preset sub-areas and the historical event propagation records, a global spatiotemporal correlation map of the target monitoring area is constructed, wherein the nodes in the global spatiotemporal correlation map represent the preset sub-areas, and the edges represent the event propagation probability and correlation strength between the preset sub-areas; The local spatiotemporal feature vectors are spatially weighted and time series aligned according to the global spatiotemporal correlation map to generate a hierarchical structure of the spatiotemporal feature matrix.

2. The artificial intelligence-based monitoring and early warning analysis method according to claim 1 is characterized in that: The method of performing dynamic pattern analysis on the spatiotemporal feature matrix based on a preset anomaly recognition network to determine potential abnormal events and their corresponding abnormal propagation paths within the target monitoring area includes: Inputting the spatiotemporal feature matrix into the multi-scale feature extraction module of the anomaly recognition network to obtain the anomaly pattern features of the target monitoring area at different time granularities and spatial resolutions; Dynamically weight the abnormal pattern features through the attention allocation module of the abnormal recognition network to determine the abnormal confidence of each sub-area in the target monitoring area and the correlation strength between abnormal events; Constructing an abnormal event propagation network based on the abnormal confidence and the correlation strength, wherein the nodes in the abnormal event propagation network represent sub-regions where the abnormal confidence exceeds the set confidence, and the edges represent the possibility of the abnormal event propagating between sub-regions; Based on the node degree distribution and edge weight distribution of the abnormal event propagation network, identify the core occurrence location of the potential abnormal event and the key nodes of the propagation path; The path connection optimization is performed on the key nodes of the propagation path in combination with the time continuity of the spatiotemporal label information to generate a complete topological structure of the abnormal propagation path.

3. The artificial intelligence-based monitoring and early warning analysis method according to claim 2 is characterized in that: The step of optimizing the path connections of the key nodes of the propagation path in combination with the temporal continuity of the spatiotemporal tag information to generate a complete topological structure of the abnormal propagation path includes: Extracting a timestamp sequence from the spatiotemporal tag information corresponding to the key nodes, and arranging the key nodes into a time-ordered node queue according to the chronological order of the timestamp sequence, wherein each node in the time-ordered node queue carries spatial coordinate information and a corresponding timestamp; Calculating the timestamp intervals of adjacent nodes in the time-ordered node queue, dividing the adjacent nodes whose timestamp intervals are less than a preset time continuity threshold into the same time continuity group, and recording the start timestamp and end timestamp of each time continuity group; Traverse the spatial coordinate information of the nodes in each time-continuous group, calculate the spatial distance matrix between the nodes in the group, mark the node pairs whose spatial distance is less than a preset spatial proximity threshold as spatially adjacent node pairs, and generate a spatial proximity relationship map based on the spatial distance matrix; Based on the spatial proximity relationship graph, linearly connect the nodes in the temporally continuous group that are spatially adjacent in order of timestamps to form an initial propagation path segment, wherein the initial propagation path segment includes path direction information and a path connection strength parameter; Detecting spatiotemporal conflicts between initial propagation path segments generated by different time-continuous groups, including timestamp overlap conflicts in the same spatial region and topological conflicts where path directions do not match the spatial proximity relationship map; Dynamically adjust the conflicting path segments based on the detected spatiotemporal conflict type to generate adjusted initial propagation path segments. Specifically, for timestamp overlap conflicts, prioritize the path segments with later timestamps and remove the overlapping segments. For topological conflicts, recalculate the connection directions of the path segments based on the spatial proximity relationship graph. Globally sorting the adjusted initial propagation path segments according to the start timestamps of the time-continuous groups, splicing the path segments end to end based on the sorting results, and verifying during the splicing process whether the timestamp intervals and spatial jump distances of adjacent path segments meet the temporal continuity threshold and spatial proximity threshold; When it is detected that the timestamp interval or spatial jump distance of adjacent path segments exceeds a threshold, a transition node is inserted and the spatial proximity relationship map is updated. The spatial coordinates of the transition node are generated by interpolation based on the spatial coordinates of the end nodes of the adjacent path segments, and the timestamps are linearly allocated based on the timestamp intervals of the adjacent path segments. Based on the final set of spliced ​​path segments and the inserted transition nodes, a complete topological structure of the abnormal propagation path is generated, and the complete topological structure includes path branch information, node connection order and path propagation direction weight parameters.

4. The artificial intelligence-based monitoring and early warning analysis method according to claim 2 is characterized in that: Generating a dynamic early warning strategy adapted to the target monitoring area based on the topological structure of the abnormal propagation path and the attribute parameters of the potential abnormal event includes: Calculating a priority index of each sub-region in the potential abnormal event propagation chain according to the topological structure of the abnormal propagation path, wherein the priority index is used to quantify the urgency of the sub-region being affected by the abnormal event; Determining a trigger threshold range of the dynamic early warning strategy based on attribute parameters of the potential abnormal event, wherein the attribute parameters include abnormal intensity, propagation speed, and impact duration; Generate a multi-level warning trigger rule based on the priority index and the trigger threshold range, wherein the multi-level warning trigger rule defines the warning level and response delay time of different priority sub-areas under different abnormal intensities; Constructing a response instruction set that matches the multi-level warning triggering rules, the response instruction set including equipment control instructions, personnel dispatch instructions, and data reporting protocols for each warning level; The multi-level warning triggering rules and the response instruction set are dynamically bound according to the time sequence of the abnormal propagation path to form an execution logic chain of the dynamic warning strategy.

5. The artificial intelligence-based monitoring and early warning analysis method according to claim 4 is characterized in that: The real-time optimization and adjustment of the dynamic warning strategy based on historical warning feedback data includes: Acquire feedback log data generated when the terminal device cluster executes the historical warning strategy, wherein the feedback log data includes the warning response delay time, the device execution status, and the actual impact range of the abnormal event; Extracting policy execution effect indicators from the feedback log data, wherein the policy execution effect indicators include warning accuracy, response efficiency deviation, and resource consumption rate; Comparing the strategy execution effect index with the preset optimization objective function to determine the trigger condition parameters and response instruction priorities that need to be adjusted in the dynamic early warning strategy; Iteratively correcting the trigger threshold range and response delay time of the dynamic warning strategy according to the gradient direction of the optimization objective function to obtain corrected parameters; The coverage and resource allocation weights of each level of warning in the multi-level warning triggering rule are recalculated based on the revised parameters to generate an optimized dynamic warning strategy.

6. The artificial intelligence-based monitoring and early warning analysis method according to claim 1 is characterized in that: Before collecting the multi-source monitoring data stream within the target monitoring area, the method further includes: Divide a plurality of monitoring data acquisition units according to the geographical distribution characteristics of the target monitoring area, each of the acquisition units is configured with at least three heterogeneous sensors and a data preprocessing module; Setting the data sampling frequency and transmission protocol of the heterogeneous sensors so that different types of sensor data in the same acquisition unit are transmitted through independent channels after the timestamps are aligned; The data preprocessing module is configured with a noise filtering algorithm and a data integrity verification rule, wherein the noise filtering algorithm uses a wavelet transform method with an adaptive threshold to eliminate environmental interference components in the sensor signal; Establishing a two-way communication link between the acquisition unit and the central server, wherein the two-way communication link supports real-time uploading of the multi-source monitoring data stream and synchronous issuance of control instructions; The acquisition unit is regularly self-checked and calibrated to generate a sensor health status report and dynamically adjust the data sampling frequency and noise filtering parameters.

7. The artificial intelligence-based monitoring and early warning analysis method according to claim 1 is characterized in that: The hierarchical structure of the spatiotemporal feature matrix includes a basic feature layer, a spatiotemporal correlation layer, and an event prediction layer. The basic feature layer stores the original monitoring indicator statistics of each sub-region, the spatiotemporal correlation layer stores the spatiotemporal dependency coefficients between sub-regions, and the event prediction layer stores the predicted values ​​of the probability of abnormal events trained based on historical event data. The method further comprises: Update the statistical values ​​in the basic feature layer according to the real-time monitoring data, and trigger the dynamic recalculation of the dependency coefficients in the spatiotemporal correlation layer; When the predicted value of the probability of occurrence of an abnormal event in the event prediction layer exceeds a preset threshold, the incremental learning module of the abnormality recognition network is activated to update the calculation model of the predicted value of the probability of occurrence of the abnormal event based on the latest monitoring data; The hierarchical structure of the spatiotemporal feature matrix is ​​dynamically maintained through a sliding time window mechanism, feature data within a preset time length is retained, and expired data blocks are removed.

8. The artificial intelligence-based monitoring and early warning analysis method according to claim 2 is characterized in that: The anomaly recognition network includes a spatiotemporal convolution module, a long short-term memory module, and a graph neural network module. The spatiotemporal convolution module is used to extract local spatiotemporal patterns of monitoring data, the long short-term memory module is used to capture long-term dependencies in time series, and the graph neural network module is used to model spatial correlations between sub-regions. The method further comprises: The spatiotemporal feature matrix is ​​divided into multiple segments according to the time dimension, and the segments are respectively input into the spatiotemporal convolution module for feature dimensionality reduction and pattern extraction; The output of the spatiotemporal convolution module is reorganized according to the spatial dimension and input into the graph neural network module to generate the spatial correlation feature vector of each sub-region; Inputting the spatial correlation feature vector into the long short-term memory module in chronological order to generate a global spatiotemporal state representation of the target monitoring area; The output features of the spatiotemporal convolution module, graph neural network module and long short-term memory module are integrated to generate a comprehensive discrimination index for abnormal event identification.

9. A server, characterized in that: The server includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the artificial intelligence-based monitoring and early warning analysis method described in any one of claims 1 to 8.

Images