Industrial fire early prevention and control early warning system based on multi-sensor fusion

Abstract

This invention discloses an early warning and prevention system for industrial fires based on multi-sensor fusion, belonging to the field of industrial safety monitoring technology. It includes: a data acquisition module, a false alarm discrimination module, a risk zoning module, a fire identification module, and a prevention and control optimization module. The data acquisition module acquires fire data through a hierarchical distributed multi-sensor network. The false alarm discrimination module filters valid fire data based on dynamic false alarm discrimination rules. The risk zoning module generates a risk heatmap and completes sub-region division and risk label allocation. The fire identification module predicts the probability of fire occurrence and identifies the fire development stage. The prevention and control optimization module matches prevention and control strategies based on the fire development stage and risk labels, combines a dynamic cost map to plan the optimal path for AGVs to execute prevention and control tasks, and feeds the execution results back to the data acquisition module for dynamic correction. This application reduces the false alarm rate and improves the reliability and efficiency of early warning and prevention of industrial fires.

Description

Technical Field

[0001] This invention belongs to the field of industrial safety monitoring technology, specifically an early warning and prevention system for industrial fires based on multi-sensor fusion. Background Technology

[0002] Industrial production areas contain a large number of flammable and explosive materials, electrical equipment, and complex processes, resulting in a high risk of fire accidents. Traditional industrial fire early warning systems often rely on single-type sensors, leading to limited data collection dimensions and high false alarm rates. Furthermore, existing systems lack precise risk zoning capabilities, making it impossible to develop targeted prevention and control strategies based on regional fire risk differences. In addition, delayed fire development stage identification and AGV path planning failing to consider dynamic risk changes result in untimely response, low execution efficiency, and difficulty in effectively curbing the spread of early-stage fires. Therefore, there is an urgent need for a multi-dimensional perception, low false alarm rate, precise zoning, and intelligent early warning system for industrial fire prevention and control. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes an early warning and prevention system for industrial fires based on multi-sensor fusion. The system includes a data acquisition module, a false alarm discrimination module, a risk zoning module, a fire identification module, and a prevention and control optimization module. The data acquisition module acquires fire data through a hierarchical distributed multi-sensor network. The false alarm discrimination module filters valid fire data based on dynamic false alarm discrimination rules. The risk zoning module generates a risk heatmap and completes sub-region division and risk label assignment. The fire identification module predicts the probability of fire occurrence and identifies the fire development stage. The prevention and control optimization module matches prevention and control strategies based on the fire development stage and risk labels, plans the optimal path for AGVs to execute prevention and control tasks using a dynamic cost map, and feeds the execution results back to the data acquisition module for dynamic correction. This application reduces the false alarm rate and improves the reliability and efficiency of early warning and prevention systems for industrial fires.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] An industrial fire early prevention and warning system based on multi-sensor fusion includes: a data acquisition module, a false alarm discrimination module, a risk zoning module, a fire identification module, and a prevention and control optimization module;

[0006] The data acquisition module acquires fire data from the industrial area based on a multi-sensor network; the fire data includes temperature, smoke concentration, flame spectrum, and gas concentration data.

[0007] The false alarm discrimination module uses dynamic false alarm discrimination rules to filter valid fire data based on fire data.

[0008] The risk zoning module identifies the fire risk level of different areas based on valid fire data and assigns a risk label to each sub-area;

[0009] The fire identification module identifies the fire development stage based on valid fire data and the probability of fire occurrence predicted for sub-regions.

[0010] The prevention and control optimization module formulates prevention and control strategies based on the fire development stage, risk labels, and effective fire data. At the same time, the AGV's path planning algorithm plans paths based on sub-regions, executes the prevention and control strategies, and feeds back the execution results to the data acquisition module.

[0011] Specifically, the multi-sensor network adopts a layered distributed architecture, including: a bottom sensor node layer, an intermediate wireless transmission layer, and a top data processing layer; the bottom sensor node layer is deployed in key locations in the industrial area, and each sensor node integrates at least one type of sensor for real-time collection of fire data; the intermediate wireless transmission layer transmits the fire data collected by the bottom sensor node layer to the aggregation node through ZigBee or LoRa communication protocols; the top data processing layer is used to receive the fire data uploaded by the aggregation node and perform time synchronization and format standardization processing on the fire data.

[0012] Specifically, the method for generating the dynamic false alarm discrimination rule includes: establishing a historical fire data sample library, which includes normal data, false alarm data, and real fire data under different operating conditions; training the historical fire data sample library with a machine learning algorithm to generate a dynamic false alarm discrimination model, wherein the input of the dynamic false alarm discrimination model is the real-time collected fire data, and the output is the data validity judgment result.

[0013] Specifically, the dynamic false alarm discrimination rules also include the correlation discrimination between the rate of increase of temperature data and the trend of change of smoke concentration, and the correlation discrimination between the duration of flame spectral signal and the duration of gas concentration exceeding the standard;

[0014] The rate of temperature rise is correlated with the trend of smoke concentration change. When the rate of temperature rise exceeds a preset temperature threshold and the smoke concentration increases exponentially, it is determined to be valid fire data.

[0015] The duration of the flame spectral signal is correlated with the duration of gas concentration exceeding the limit. When the duration of the flame spectral signal exceeds... t When the carbon monoxide concentration exceeds the preset lower explosion limit threshold, it is considered valid fire data.

[0016] Specifically, the risk zoning module identifies the fire hazard level of different areas based on valid fire data and assigns a risk label to each sub-area, including:

[0017] The risk zoning module receives valid fire data and assigns a spatial weight to each valid fire data point, forming valid fire data with spatial weights; the spatial weights are calculated based on the geographic location of the sensors.

[0018] Valid fire data with spatial weights are input into the risk heat map generation algorithm. The geographical location of each valid fire data point is used as the coordinate, and its spatial weight and sensor readings are used as the weighting values. Through interpolation and smoothing, the risk intensity distribution surface is output, and a risk heat map is formed after visualization rendering.

[0019] The risk heatmap is divided into regions based on a preset risk level threshold, including: automatically identifying continuous closed regions with risk intensity within the same risk level threshold range, and dividing the continuous risk heatmap into multiple discrete sub-regions;

[0020] Based on the risk level threshold used when dividing the sub-regions, a corresponding risk label is automatically assigned to each generated sub-region to form a dynamic partition map; the risk label includes first risk, second risk, and third risk.

[0021] Specifically, the fire identification module identifies the fire development stage based on valid fire data and the predicted probability of fire occurrence in sub-regions, including:

[0022] The fire identification module receives a dynamic partition map from the risk partitioning module, obtains the geographical boundary of a specified sub-region based on the dynamic partition map, and filters out all sensor data located within the geographical boundary from the valid fire data based on the geographical boundary to form the feature dataset of the corresponding partition.

[0023] The feature dataset and the risk label are input into the time series prediction model; the time series prediction model is an attention-enhanced recurrent neural network model, which dynamically allocates different attention weights according to the importance of different sensor data in the feature dataset to fire prediction;

[0024] The time series prediction model outputs the probability curve of fire occurrence in the corresponding partition within a preset time period, and identifies the fire development stage based on the slope and key inflection points of the probability curve; the fire development stage includes smoldering stage, initial open flame stage and spread stage.

[0025] Specifically, the prevention and control optimization module includes:

[0026] The prevention and control optimization module receives the probability curve of fire occurrence and the fire development stage from the fire identification module, and simultaneously obtains the dynamic zoning map from the risk zoning module, and reads the risk label corresponding to the sub-region where the fire occurred from the dynamic zoning map.

[0027] The currently identified fire development stage and the corresponding risk label of the sub-area are used as query keys to retrieve the preset prevention and control strategy knowledge base and obtain the corresponding set of prevention and control actions. The prevention and control strategy knowledge base stores a strategy mapping table, which defines the mapping relationship between the input condition combination of fire development stage and risk label and the output predefined set of prevention and control actions.

[0028] The matched set of prevention and control actions is parsed into a structured prevention and control strategy; the elements of the prevention and control strategy include prevention and control actions, resource allocation, and execution order.

[0029] Based on the prevention and control strategy and the dynamic partition map, generate fire extinguishing or isolation task instructions for one or more AGVs;

[0030] For each AGV that receives a fire extinguishing or isolation task instruction, the AGV path planning algorithm uses the dynamic partition map and the AGV's real-time position as input to calculate the optimal path that avoids the first risk sub-area and leads to the task target location. The optimal path is then bound to the fire extinguishing or isolation task instruction to form an AGV action instruction set, which is then sent to the AGV actuator to drive its execution.

[0031] Specifically, the prevention and control strategies in the prevention and control strategy knowledge base include:

[0032] For sub-areas that are in the smoldering stage and are classified as risk level 3, ventilation and smoke extraction and early warning notifications will be activated.

[0033] For sub-areas in the initial stage of open flame and with a risk label of second risk, dispatch AGVs to carry fire extinguishers and activate the sprinkler system;

[0034] For sub-areas that are in the spread stage and are labeled as the first risk level, multiple AGVs are dispatched to work together to perform tasks such as setting up firebreaks and extinguishing fires.

[0035] Specifically, the path planning algorithm for the AGV is as follows:

[0036] The dynamic partition map is mapped to a dynamic cost map, wherein each sub-region is regarded as an independent navigation unit;

[0037] Based on the risk label of each sub-region in the dynamic zoning map, the passage cost is dynamically calculated for each sub-region in the dynamic cost map; the passage cost increases exponentially with the risk level represented by the risk label.

[0038] Simultaneously, the current position of the AGV and the target position specified by the control strategy are obtained. Using the current position and the target position as the start and end points of the path, a graph search algorithm is used to search on the generated dynamic cost map. The goal of the graph search algorithm is to find a path from the current position to the target position such that the sum of the travel costs of all sub-regions traversed by the path, i.e., the cumulative travel cost, is the minimum among all paths. The path with the lowest cumulative travel cost is then output as the optimal path for the AGV to execute.

[0039] Specifically, the process of feeding back the execution results to the data acquisition module includes:

[0040] During the execution of AGV action command sets, the AGV collects environmental data after execution through its onboard auxiliary sensors;

[0041] The environmental data after execution is compared with the valid fire data that triggered the execution task to generate a strategy effectiveness evaluation signal;

[0042] The effectiveness evaluation signal of the strategy is sent to the data acquisition module as a dynamic correction parameter for data acquisition and discrimination.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] This invention proposes an early prevention and warning system for industrial fires based on multi-sensor fusion, and optimizes and improves its architecture, operation steps and processes. The system has the advantages of simple process, low investment and operating costs and low production and working costs.

[0045] This invention proposes an early warning and prevention system for industrial fires based on multi-sensor fusion. This system achieves real-time acquisition and standardized processing of multi-dimensional fire data such as temperature, smoke concentration, flame spectrum, and gas concentration through multi-sensor fusion and a hierarchical distributed network architecture, solving the problem of single data dimensions in traditional systems. Combined with dynamic false alarm discrimination rules trained on historical data, it filters effective fire data through dual correlation discrimination, reducing the false alarm rate. At the same time, relying on spatial weights and risk heat maps, it achieves accurate zoning and dynamic updating of fire risk levels.

[0046] This invention proposes an early warning and prevention system for industrial fires based on multi-sensor fusion. Utilizing an attention-enhanced recurrent neural network model, it can accurately predict the probability of fire occurrence and identify development stages such as smoldering, initial open flame, and spread. Combined with a strategy mapping mechanism that integrates fire development stages with risk labels, it achieves targeted matching of prevention and control strategies. AGV dynamic optimal path planning avoids sub-regions with the highest risk, improving execution efficiency. Furthermore, the feedback of execution results forms a dynamic correction closed loop, comprehensively optimizing the response speed, resource allocation rationality, and execution effectiveness of early fire prevention and control, effectively curbing the spread of fire. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the principle of the industrial fire early prevention and warning system based on multi-sensor fusion according to the present invention.

[0048] Figure 2 This is a flowchart of the risk label allocation process for the industrial fire early prevention and warning system based on multi-sensor fusion, as described in this invention.

[0049] Figure 3 This is a flowchart illustrating the fire development stage identification process of the industrial fire early prevention and warning system based on multi-sensor fusion, as described in this invention. Detailed Implementation

[0050] Example 1:

[0051] Please see Figure 1 One embodiment of the present invention provides an industrial fire early prevention and warning system based on multi-sensor fusion, comprising:

[0052] Data acquisition module, false alarm detection module, risk zoning module, fire identification module, and prevention and control optimization module;

[0053] The data acquisition module acquires fire data from the industrial area based on a multi-sensor network; the fire data includes temperature, smoke concentration, flame spectrum, and gas concentration data.

[0054] The multi-sensor network adopts a layered distributed architecture, including: a bottom sensor node layer, an intermediate wireless transmission layer, and a top data processing layer. The bottom sensor node layer is deployed in key locations in the industrial area, and each sensor node integrates at least one type of sensor for real-time fire data collection. The intermediate wireless transmission layer transmits the fire data collected by the bottom sensor node layer to the aggregation node via ZigBee or LoRa communication protocols. The top data processing layer receives the fire data uploaded by the aggregation node and performs time synchronization and format standardization processing on the fire data. The ZigBee or LoRa communication protocols are existing technologies in this field and are not inventive solutions for this application; therefore, they will not be elaborated upon here.

[0055] Furthermore, key locations within an industrial area refer to areas with high fire risk or requiring close monitoring, including but not limited to high-risk fire areas such as flammable and explosive material storage areas, production operation areas, and equipment rooms; areas prone to hidden fires such as electrical control rooms, cable trenches, and ventilation ducts; and core areas with a large fire impact, such as warehousing areas and densely populated operation areas.

[0056] Furthermore, the sensor types deployed in the bottom sensor node layer include temperature sensors, smoke sensors, flame spectrum sensors, and gas sensors. Among them, temperature sensors are used to collect ambient temperature and temperature rise rate data, smoke sensors are used to monitor smoke concentration and its changing trend, flame spectrum sensors are used to capture flame characteristic spectral signals and duration, and gas sensors are used to detect the concentration of combustible or toxic gases such as carbon monoxide and determine whether it exceeds the standard.

[0057] The false alarm discrimination module uses dynamic false alarm discrimination rules to filter valid fire data based on fire data.

[0058] The risk zoning module identifies the fire risk level of different areas based on valid fire data and assigns a risk label to each sub-area;

[0059] The fire identification module identifies the fire development stage based on valid fire data and the probability of fire occurrence predicted for sub-regions.

[0060] The prevention and control optimization module formulates prevention and control strategies based on the fire development stage, risk labels, and effective fire data. At the same time, the AGV's path planning algorithm plans paths based on sub-regions, executes the prevention and control strategies, and feeds back the execution results to the data acquisition module.

[0061] The method for generating the dynamic false alarm discrimination rule includes: establishing a historical fire data sample library, which includes normal data, false alarm data, and real fire data under different operating conditions; training the historical fire data sample with a machine learning algorithm to generate a dynamic false alarm discrimination model, wherein the input of the dynamic false alarm discrimination model is real-time collected fire data, and the output is the data validity judgment result. The machine learning algorithm is existing technology in this field and is not an inventive solution of this application, and will not be described in detail here.

[0062] The dynamic false alarm discrimination rules also include the correlation between the rate of increase of temperature data and the trend of smoke concentration change, and the correlation between the duration of flame spectral signal and the duration of gas concentration exceeding the standard.

[0063] The rate of temperature rise is correlated with the trend of smoke concentration change. When the rate of temperature rise exceeds a preset temperature threshold and the smoke concentration increases exponentially, it is determined to be valid fire data.

[0064] Furthermore, the preset temperature threshold refers to the critical value of abnormal changes in ambient temperature per unit time in an industrial setting. For example, the preset temperature threshold for a typical production workshop is 5℃ / min, while the preset temperature threshold for a hazardous chemical storage area is 3℃ / min.

[0065] The duration of the flame spectral signal is correlated with the duration of gas concentration exceeding the limit. When the duration of the flame spectral signal exceeds... t When the carbon monoxide concentration exceeds the preset lower explosion limit threshold, it is considered valid fire data.

[0066] Furthermore, the lower explosion limit refers to the lowest volume percentage concentration of a combustible gas, such as carbon monoxide, that can cause an explosion in the air. In this invention, the lower explosion limit for carbon monoxide is set to 12.5%, meaning that when the concentration of carbon monoxide in the air is greater than or equal to 12.5%, it is considered an explosion hazard.

[0067] Example 2:

[0068] Please see Figure 2 In this embodiment, the risk zoning module identifies the fire hazard level of different areas based on valid fire data and assigns a risk label to each sub-area, including:

[0069] A1: The risk zoning module receives valid fire data and assigns a spatial weight to each valid fire data point, forming valid fire data with spatial weights; the spatial weights are calculated based on the geographic location of the sensors and are used to quantify the risk impact of the valid fire data point on its surrounding area;

[0070] Furthermore, the specific steps of A1 include:

[0071] (1) The risk zoning module receives valid fire data from the false alarm discrimination module at the front end. Each valid fire data point carries the precise geographical location information of its source sensor, thereby constructing a set of discrete data points with spatial coordinates within the risk zoning module.

[0072] (2) Based on a preset spatial influence radius, scan the entire discrete data point set, identify all valid fire data points that are adjacent to each other within this spatial influence radius, and classify these points into a group of data points that have spatial correlation;

[0073] (3) The system reads the confidence label from the false alarm discrimination module for each valid fire data point, and calculates a basic weight value for each valid fire data point based on the data reliability represented by the confidence label. The higher the confidence, the higher the basic weight value.

[0074] (4) Examine the spatial context of each valid fire data point. For points located within a group of data points, their basic weight values ​​will be enhanced with higher synergy, because the simultaneous anomalies of multiple sensors can better characterize regional risks. Conversely, for isolated anomalies of valid fire data points that are not related to each other within the spatial influence radius, their basic weight values ​​will be systematically reduced to suppress possible local interference.

[0075] (5) After the above adjustment of the basic weight values, each valid fire data point obtains its final spatial weight.

[0076] A2: Input the valid fire data with spatial weights into the risk heat map generation algorithm. Using the geographical location of each valid fire data point as coordinates and its spatial weight and sensor readings as weighting values, the algorithm outputs a risk intensity distribution surface through interpolation and smoothing. After visualization rendering, a risk heat map is formed to express the current potential fire risk intensity at each location in the area.

[0077] Furthermore, the specific steps of A2 include:

[0078] (1) Obtain valid fire data with spatial weights;

[0079] (2) Define the physical boundary of the entire industrial area and cover it with a regular grid to discretize the continuous space into multiple computing units. The center point of each computing unit is a target location.

[0080] (3) Define an influence range for each effective fire data point with spatial weight; the influence range represents the geographical area centered on this data point, where its spatial weight and sensor readings can have an effective impact;

[0081] (4) Traverse the target location of each computing unit in the regular grid. For the current target location, the system searches for all valid fire data points with attached spatial weights whose influence range can cover it. These points are considered as contribution points.

[0082] (5) For each contribution point, the system calculates its contribution value to the current target location. This contribution value is determined by the product of the sensor reading of the contribution point and its spatial weight, and is positively correlated with this product value. At the same time, the contribution value is negatively correlated with the geographical distance from the contribution point to the current target location. The greater the distance, the smaller the impact.

[0083] (6) The final risk intensity value of the current target location is obtained by weighted superposition of the contribution values ​​of all contribution points, and a quantitative risk characterization value is assigned to each calculation unit;

[0084] (7) After the risk intensity values ​​of all computational units in the regular grid have been calculated, these discrete numerical points together constitute a continuous risk intensity distribution surface covering the entire industrial area.

[0085] (8) Map the risk intensity distribution surface into a tangible image, namely a risk heat map, which is usually expressed using color gradients. For example, use cool blue to represent low-risk areas, and use warm yellow, orange and red to represent medium and high-risk areas respectively.

[0086] A3: Divide the risk heat map into regions based on the preset risk level threshold, including: automatically identifying continuous closed regions with risk intensity within the same risk level threshold range, and dividing the continuous risk heat map into multiple discrete sub-regions;

[0087] Furthermore, the specific steps in A3 include:

[0088] (1) Obtain the risk heat map and call the preset risk level threshold. The risk level threshold divides the risk intensity into continuous intervals of low, medium and high.

[0089] (2) The continuous risk intensity distribution surface represented by the risk heat map is horizontally cut along each risk level threshold. Each cut will generate one or more closed contour lines, and the risk intensity of any point inside them is greater than or equal to the current cutting threshold.

[0090] (3) The system processes each closed contour line generated by the cutting and initially identifies the internal region enclosed by each contour line as a candidate sub-region;

[0091] (4) Assign a risk label to each candidate sub-region. The rule for determining the risk label is: observe which risk level threshold the candidate sub-region was cut by, and combine the risk intensity of its internal points to determine which threshold range it mainly falls into. For example, a region that is cut by a high risk threshold and whose internal intensity is higher than the medium risk threshold is identified as a sub-region with the first risk.

[0092] (5) Perform spatial relationship processing and regional optimization, including:

[0093] Examine the spatial location of all candidate sub-regions. For multiple regions that are not spatially connected but have the same risk label, each one will be identified as an independent sub-region.

[0094] For multiple candidate sub-regions that are spatially adjacent and have the same risk label, a decision is made on whether to merge them into a larger, continuous sub-region based on preset merging rules, such as upper distance limit and lower area limit.

[0095] (6) After all the contour lines are processed and the spatial relationships are clarified, the originally continuous risk heat map is divided into multiple discrete final sub-regions with homogenized internal risk intensity. Each sub-region has a clear geographical boundary and a unique risk label. All these sub-regions and their attributes together constitute the dynamic partition map required by the system.

[0096] A4: Based on the risk level threshold used when dividing the sub-regions, a corresponding risk label is automatically assigned to each generated sub-region to form a dynamic partition map; the risk label includes first risk, second risk, and third risk.

[0097] Furthermore, the first risk is high risk, the second risk is medium risk, and the third risk is low risk.

[0098] Furthermore, the specific steps for A4 include:

[0099] (1) Obtain the sub-regions and their corresponding risk level thresholds;

[0100] (2) The system automatically determines the risk level of a region based on the specific risk level threshold used when defining the boundary of a sub-region, combined with the internally preset mapping rules. For example, a region whose boundary is defined by a high-risk threshold is determined to be high-risk.

[0101] (3) Based on the determined risk level, the system selects the corresponding text-type risk label from the preset label library, such as low-risk area, medium-risk area or high-risk area, and performs a binding operation to inseparably associate the text-type risk label with all the geometric attributes of the sub-region; the geometric attributes include all the coordinate point sequences that constitute the boundary of the region;

[0102] (4) After all sub-regions have completed the risk label binding, the system integrates these labeled sub-regions, spatial boundary information and their relative positional relationships into a unified, structured data object, namely, a dynamic partition map.

[0103] (5) The newly generated dynamic partition map is encapsulated in a standardized data format. The dynamic partition map accurately answers the core question of where the risk is high and what the risk level is, laying the most critical spatial situational awareness foundation for the intelligent decision-making of the entire system.

[0104] Example 3:

[0105] Please see Figure 3 In this embodiment, the fire identification module identifies the fire development stage based on valid fire data and the predicted probability of fire occurrence in sub-regions, including:

[0106] B1: The fire identification module receives a dynamic partition map from the risk partitioning module, obtains the geographical boundary of the specified sub-region based on the dynamic partition map, and filters out all sensor data located within the geographical boundary from the valid fire data based on the geographical boundary to form the feature dataset of the corresponding partition.

[0107] Furthermore, the specific steps for B1 include:

[0108] (1) The fire identification module receives the latest dynamic partition map from the risk partitioning module at the front end. The dynamic partition map clearly records all the sub-regions that have been divided in the current industrial environment, as well as the precise geographical boundary coordinates of each sub-region and its corresponding risk label.

[0109] (2) Based on the current analysis needs or alarm triggering conditions, a target sub-region that needs to be focused on is determined from the dynamic partition map. This target sub-region may be a newly emerging high-risk area or a region where the risk is changing rapidly. Once the target is locked, the system accurately extracts the polygonal geographical boundary that defines the range of the sub-region from the dynamic partition map. The polygonal geographical boundary is composed of multiple geographical coordinate points connected in sequence, forming a closed polygonal region in the digital space.

[0110] (3) Real-time acquisition of effective fire data streams, each effective fire data point carries the unique identifier of its source sensor and its precise geographical coordinates;

[0111] (4) For each valid fire data point in the flow, perform spatial location relationship judgment, that is, determine whether the geographical coordinates of the valid fire data point fall within the polygonal geographical boundary of the corresponding target sub-region; if the valid fire data point is determined to be within the polygonal geographical boundary, it will be automatically captured and marked; if the valid fire data point is determined to be outside the polygonal geographical boundary, it will be temporarily ignored in this analysis to obtain a pure data subset.

[0112] (5) Align and integrate the clean data subsets according to the time series to obtain the feature dataset of the specified partition.

[0113] B2: Input the feature dataset and the risk label together into a time series prediction model; the time series prediction model is an attention-enhanced recurrent neural network model, which dynamically allocates different attention weights according to the importance of different sensor data in the feature dataset to fire prediction. The recurrent neural network model is existing technology in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0114] B3: The time series prediction model outputs the probability curve of fire occurrence in the corresponding partition within a preset time period, and identifies the fire development stage based on the slope and key inflection points of the probability curve; the fire development stage includes the smoldering stage, the initial open flame stage, and the spread stage.

[0115] Furthermore, the specific identification logic for fire development stages includes: when the probability curve remains at a low numerical level and its slope is almost zero, showing only minor fluctuations, it indicates that the risk is in a state of slow brewing but not significant escalation, and the system will identify it as the smoldering stage; when the probability curve shows the first significant critical inflection point, and after this inflection point, the slope of the curve turns into a stable and obvious positive value, it means that the probability begins to increase continuously and steadily, which indicates that an open flame has begun to be generated and is burning stably, and the system will identify it as the initial stage of open flame; when the probability curve shows a second, steeper critical inflection point, after this inflection point, the slope of the curve becomes extremely steep, and the probability value soars almost vertically, which indicates that the fire is breaking through the initial burning state and entering a situation of rapid spread and difficulty in control, and the system will identify it as the spread stage.

[0116] The prevention and control optimization module includes:

[0117] C1: The prevention and control optimization module receives the probability curve of fire occurrence and the fire development stage from the fire identification module, and at the same time obtains the dynamic zoning map from the risk zoning module, and reads the risk label corresponding to the sub-region where the fire occurred from the dynamic zoning map.

[0118] C2: Using the currently identified fire development stage and the risk label of the corresponding sub-area as the query key, retrieve the preset prevention and control strategy knowledge base and match the corresponding set of prevention and control actions; the prevention and control strategy knowledge base stores a strategy mapping table, which defines the mapping relationship between the input condition combination of fire development stage and risk label and the output predefined set of prevention and control actions.

[0119] It is important to understand that the core of the prevention and control strategy knowledge base is a strategy mapping table, which can be understood as a huge decision dictionary. In this decision dictionary, every possible legal combination of fire development stage and risk label is an independent entry, that is, a unique input condition.

[0120] Furthermore, the specific steps of C2 include:

[0121] (1) Receive the risk labels of the currently identified fire development stage and corresponding sub-area;

[0122] (2) Combine the received fire development stage and risk label into a standardized query key according to the preset format;

[0123] (3) Access the prevention and control strategy knowledge base, compare the constructed standardized query key with all the preset input condition terms in the strategy mapping table, and find the term that is completely consistent with the current query key. For example, when the query key is the combination of open flame initial stage and high-risk area, it will only match the term in the strategy mapping table that is also defined as open flame initial stage and high-risk area, and will not match smoldering stage and high-risk area or any other different combination.

[0124] (4) Once a completely matching term is found, the matching process is declared successful. The content corresponding to the term in the strategy mapping table is a predefined set of prevention and control actions.

[0125] C3: Parse the matched set of prevention and control actions into a structured prevention and control strategy; the elements of the prevention and control strategy include prevention and control actions, resource allocation, and execution order;

[0126] Furthermore, the specific steps of C3 include:

[0127] (1) Obtain a set of prevention and control actions;

[0128] (2) Process each action item in the set of prevention and control actions one by one. For each action, the system transforms it from an abstract name into a specific, executable task instance. For example, for the action of extinguishing a fire, the system parses it into a specific instruction to use a dry powder fire extinguisher to extinguish the fire. At the same time, the system binds the target location parameter to this task instance. The target location parameter comes from the precise coordinates of the sub-area where the fire is located in the dynamic partition map provided by the risk partitioning module.

[0129] (3) Analyze each instantiated task to determine the type and quantity of resources required for its execution. For example, the task of using a dry powder fire extinguisher to put out a fire requires an AGV equipped with a dry powder fire extinguisher. The system will query the current resource status table and bind a specific, physically existing resource entity to the task based on the urgency of the task and the location of the resource. This process connects the abstract task with the specific physical resource.

[0130] (4) Analyze the logical relationships between all instantiated and resource-allocated tasks. Some tasks must be executed first. For example, building an isolation zone may need to be completed before going into the fire to extinguish the fire in order to prevent the fire from spreading. Based on the preset rules and the inherent logic between tasks, plan a clear and linear execution sequence for all tasks, clearly define which task starts first, which task starts later, and whether there are strict dependencies between them.

[0131] (5) The three core elements of the completed instantiated prevention and control actions, allocated resources, and determined execution order are integrated and encapsulated into a unified data structure to generate the final, structured prevention and control strategy.

[0132] C4: Based on the prevention and control strategy and the dynamic zoning map, generate fire extinguishing or isolation task instructions for one or more AGVs;

[0133] Furthermore, the specific steps of C4 include:

[0134] (1) Obtain structured prevention and control strategies and dynamic zoning maps. The structured prevention and control strategies clearly define the prevention and control actions to be performed, the types of resources allocated to each action, and the execution order of these actions. The dynamic zoning maps depict the geographical boundaries of all sub-regions and their risk labels.

[0135] (2) Analyze the sequential prevention and control actions in the prevention and control strategy, and transform each independent action item with allocated resources into a logical task to be executed. For example, the action of using a dry powder fire extinguisher to put out the fire in the strategy is initialized as a fire extinguishing task. Then, give this task precise spatial attributes, find the specific sub-area where the fire occurred by cross-referencing the dynamic partition map, and obtain the center coordinates or key point coordinates of the sub-area from the dynamic partition map. Determine this coordinate as the target location of the fire extinguishing task.

[0136] (3) Traverse all spatialized tasks decomposed by the prevention and control strategy, and select the most suitable AGV execution unit for each task. The matching principle is that the equipment carried by the AGV must meet the task requirements. For example, the fire extinguishing task must assign an AGV carrying a fire extinguisher. Under the premise of meeting the conditions, the idle or schedulable AGV closest to the task target location is selected first to minimize the response time. Once the matching is successful, a specific task is officially bound to a specific AGV entity.

[0137] (4) For each task that is successfully bound to an AGV, the system generates a highly refined task instruction. This task instruction contains an executable command with multiple mandatory fields. The mandatory fields mainly include: a clear instruction type, i.e., whether it is a fire extinguishing instruction or an isolation instruction; precise coordinates of the task target location; specific operation details that the AGV needs to perform, such as the dosage and duration of the fire extinguishing agent spraying, or the number of isolation materials placed; and the priority identifier of the task. This priority identifier is derived from the execution order defined in the prevention and control strategy. High-priority tasks will be planned and executed first.

[0138] (5) After all tasks have generated corresponding task instructions, these task instructions are collected and packaged into a task instruction package. The task instruction package clearly indicates the specific responsibilities, destinations, and operating procedures that each AGV needs to undertake, and finally outputs microscopic, drivable robot instructions.

[0139] C5: For each AGV that has received a fire extinguishing or isolation task instruction, the AGV path planning algorithm is used to calculate the optimal path that avoids the first risk sub-area and leads to the task target location using the dynamic partition map and the AGV's real-time position as input. The optimal path is then bound to the fire extinguishing or isolation task instruction to form an AGV action instruction set, which is then sent to the AGV actuator to drive its execution.

[0140] Furthermore, the specific steps of C5 include:

[0141] (1) When the system has assigned a specific fire extinguishing or isolation task instruction to a particular AGV, the path planning algorithm of that AGV is activated;

[0142] (2) Receive the dynamic partition map from the risk partitioning module and convert it into a dynamic cost map specifically for path planning;

[0143] (3) Obtain the AGV’s current precise position in real time from its own positioning system, and use it as the starting point for path planning. At the same time, the algorithm extracts the target position from the fire extinguishing or isolation task instructions that have been issued, and uses it as the endpoint for path planning.

[0144] (4) On the constructed dynamic cost map, with the determined starting point and ending point as endpoints, find a path from the starting point to the ending point where the sum of the passage costs of all navigation units traversed by the trajectory is minimized. This sum is called the cumulative passage cost. Since the sub-regions with the first risk are given extremely high passage costs, the path with the lowest cumulative passage cost found by this algorithm is naturally an optimal path that can effectively avoid all sub-regions with the first risk.

[0145] (5) Bind and encapsulate the calculated optimal path with the fire extinguishing or isolation task instructions initially assigned to the AGV. The optimal path solves the problem of how to arrive safely, while the task instructions define what to do after arrival. The combination of the two forms a self-contained, closed-loop AGV action instruction set. This AGV action instruction set not only contains all waypoint sequences of the path, but also contains the specific operation commands to be executed after arriving at the target point.

[0146] (6) The AGV action instruction set is sent to the on-board actuator of the corresponding AGV through the communication network. The actuator parses the AGV action instruction set, including: first, controlling the AGV to move autonomously to the task target location along the planned optimal path. After confirming arrival, the actuator triggers the task operation specified in the AGV action instruction set, such as turning on the fire extinguishing device to spray or releasing isolation materials.

[0147] The prevention and control strategies in the knowledge base include:

[0148] For sub-areas that are in the smoldering stage and are classified as risk level 3, ventilation and smoke extraction and early warning notifications will be activated.

[0149] For sub-areas in the initial stage of open flame and with a risk label of second risk, dispatch AGVs to carry fire extinguishers and activate the surrounding sprinkler system;

[0150] For sub-areas that are in the spread stage and are labeled as having the highest risk, multiple AGVs are dispatched to work together to set up firebreaks and carry out key firefighting tasks.

[0151] The path planning algorithm for the AGV is as follows:

[0152] D1: Map the dynamic partition map to a dynamic cost map, wherein each sub-region is regarded as an independent navigation unit;

[0153] Furthermore, the process of mapping the dynamic partition map to a dynamic cost map includes:

[0154] (1) Receive dynamic partitioning map from the risk partitioning module;

[0155] (2) Establish the correspondence between navigation units and cost structures, including: defining each independent sub-region in the dynamic partition map as a basic navigation unit in the dynamic cost map used for internal path planning. At the same time, the system has a preset cost mapping rule. The core of this cost mapping rule is to strongly correlate the passage cost of a sub-region with its risk label and stipulate that this correlation is an exponential growth relationship, that is, for every level increase in risk, the passage cost will increase by several times or even tens of times.

[0156] (3) Perform cost quantification calculation and assignment, including: traversing each navigation unit in the dynamic cost map, that is, each sub-region, for the sub-region currently being processed, reading its risk label, and then calculating a specific, quantified passage cost value according to the above-mentioned preset exponential cost mapping rule.

[0157] (4) After the calculation is completed, the calculated specific passage cost value is assigned to this navigation unit in the dynamic cost map;

[0158] (5) Once all navigation units have been successfully assigned a quantified passage cost, a brand-new dynamic cost map specifically for path planning is officially generated. The dynamic cost map is completely consistent with the original dynamic partition map in terms of spatial scope, but its connotation has undergone a fundamental change: it no longer merely indicates where the risks are, but precisely quantifies the risk costs that need to be borne when traversing different areas.

[0159] D2: Based on the risk label of each sub-region in the dynamic partition map, dynamically calculate a passage cost for each sub-region in the dynamic cost map; the passage cost increases exponentially with the risk level represented by the risk label;

[0160] It should be noted that the system has a pre-set cost configuration table. The core function of the cost configuration table is to map text-descriptive risk labels, such as low risk, medium risk, and high risk, to basic cost coefficients for mathematical calculations. These coefficients are not the final cost value, but rather the basic multipliers for calculations, and their values ​​are set in strict accordance with the law of exponential growth.

[0161] Furthermore, the specific steps of D2 include:

[0162] (1) Establish a mapping relationship between risk level and basic cost coefficient. This mapping relationship is recorded in the cost configuration table. It is clearly stipulated that the basic cost coefficient increases exponentially with the risk level. At the same time, the system sets a general benchmark cost value.

[0163] (2) Obtain a dynamic cost map; each sub-region in the dynamic cost map has been defined as a navigation unit;

[0164] (3) Traverse each navigation unit in the dynamic cost map. For the navigation unit currently being processed, the system accurately obtains the risk label assigned to the navigation unit by querying the dynamic partition map.

[0165] (4) For the navigation unit currently being processed, query the cost configuration table based on the obtained risk label, find the basic cost coefficient that corresponds to it, and then use a preset benchmark cost value as the base and the found basic cost coefficient as the exponent to perform exponentiation to calculate the final passage cost value of the navigation unit.

[0166] (5) Assign the calculated passage cost value to the current navigation unit.

[0167] D3: Simultaneously obtain the current position of the AGV and the target position specified by the control strategy. Using the current position and the target position as the start and end points of the path, perform a graph search algorithm on the generated dynamic cost map. The goal of the graph search algorithm is to find a path from the current position to the target position such that the sum of the travel costs of all sub-regions traversed by the path, i.e., the cumulative travel cost, is the minimum among all possible paths. The path with the lowest cumulative travel cost is then output as the optimal path for the AGV to execute.

[0168] Furthermore, the specific steps of D3 include:

[0169] (1) Load the assigned dynamic cost map and confirm the starting point and ending point of this navigation. Then, model the assigned dynamic cost map as a graph structure, where sub-regions are nodes and the connection relationship between adjacent regions is an edge. At the same time, initialize an open set and a closed set, and maintain the known minimum cumulative passage cost and the estimated total cost to reach the destination for each node. The initialized open set contains only the starting node, and the initialized closed set is empty.

[0170] (2) Select the node with the minimum estimated total cost from the open set as the current node. If the current node is the destination, backtrack from the destination to the starting point. The sequence of nodes traversed by the backtracking path is output as the optimal path with the minimum cumulative travel cost. Otherwise, identify all adjacent nodes of the current node.

[0171] (3) For each adjacent node, calculate the cumulative passage cost of reaching the new path through the current node, that is, the known minimum cumulative passage cost of the current node plus the passage cost of the adjacent node itself. If the new cost is lower than the known minimum cumulative passage cost recorded by the adjacent node, update the known minimum cumulative passage cost of the node to the new lower value, set its predecessor node as the current node, and then put the adjacent node into the open set.

[0172] (4) Move the processed current node into the closed set and repeat (2)-(3) until the endpoint is selected as the current node;

[0173] (5) When the destination is selected, the node sequence traversed by the predecessor node pointer of each node is output as the optimal path with the lowest cumulative passage cost.

[0174] The specific process of feeding back the execution result to the data acquisition module includes:

[0175] E1: During the execution of the AGV action command set, the AGV collects environmental data after execution through its onboard auxiliary sensors;

[0176] E2: Compare the environmental data after execution with the valid fire data that triggered this execution task to generate a strategy effectiveness evaluation signal;

[0177] Furthermore, the specific steps of E2 include:

[0178] (1) The system acquires data at two key time points: one is the environmental data after the AGV actuator completes the task and is transmitted back by the auxiliary sensor it is equipped with; the other is the effective fire data that was initially triggered by the system archived. The system extracts the same set of key fire indicators, such as temperature and specific gas concentration, from the environmental data and the effective fire data that triggered the action, and ensures that the type, monitoring location and statistical caliber of the key fire indicators are completely consistent.

[0179] (2) For each set of aligned key fire indicators, calculate the value of the key fire indicators in the environmental data after execution, and calculate the change and direction of the value of the key fire indicators in the effective fire data, such as decrease, increase, or remain unchanged. The system comprehensively analyzes the change and direction of all key fire indicators to form a comprehensive situation change profile.

[0180] (3) Retrieve the expected goals set by the prevention and control strategies that have been implemented this time, such as extinguishing open flames. Then, logically match the resulting situation change profile with the expected goals to determine whether the actual situation changes meet the expectations. For example, if the goal is to extinguish open flames, and the key fire indicators show that the temperature and smoke have dropped significantly, then the consistency is high.

[0181] (4) Based on the compliance assessment results, generate a structured strategy effectiveness assessment signal. The strategy effectiveness assessment signal includes the overall effectiveness conclusion, such as effective, partially effective or ineffective, and may include detailed information, such as specific indicators that have not been met. Finally, the strategy effectiveness assessment signal is output.

[0182] E3: The effectiveness evaluation signal of the strategy is sent to the data acquisition module as a dynamic correction parameter for data acquisition and discrimination.

[0183] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the present invention. All of these variations are within the protection scope of the present invention.

Claims

1. An industrial fire early prevention and warning system based on multi-sensor fusion, characterized in that, include: Data acquisition module, false alarm detection module, risk zoning module, fire identification module, and prevention and control optimization module; The data acquisition module acquires fire data from the industrial area based on a multi-sensor network; the fire data includes temperature, smoke concentration, flame spectrum, and gas concentration data. The false alarm discrimination module uses dynamic false alarm discrimination rules to filter valid fire data based on fire data. The risk zoning module identifies the fire risk level of different areas based on valid fire data and assigns a risk label to each sub-area; The fire identification module identifies the fire development stage based on valid fire data and the probability of fire occurrence predicted for sub-regions. The prevention and control optimization module formulates prevention and control strategies based on the fire development stage, risk labels, and effective fire data. At the same time, the AGV's path planning algorithm plans paths based on sub-regions, executes the prevention and control strategies, and feeds back the execution results to the data acquisition module. The risk zoning module identifies the fire hazard level of different areas based on valid fire data and assigns a risk label to each sub-area, including: The risk zoning module receives valid fire data and assigns a spatial weight to each valid fire data point, forming valid fire data with spatial weights; the spatial weights are calculated based on the geographic location of the sensors. Valid fire data with spatial weights are input into the risk heat map generation algorithm. The geographical location of each valid fire data point is used as the coordinate, and its spatial weight and sensor readings are used as the weighting values. Through interpolation and smoothing, the risk intensity distribution surface is output, and a risk heat map is formed after visualization rendering. The risk heatmap is divided into regions based on a preset risk level threshold, including: automatically identifying continuous closed regions with risk intensity within the same risk level threshold range, and dividing the continuous risk heatmap into multiple discrete sub-regions; Based on the risk level threshold used when dividing the sub-regions, a corresponding risk label is automatically assigned to each generated sub-region to form a dynamic partition map; the risk label includes first risk, second risk, and third risk.

2. The industrial fire early prevention and warning system based on multi-sensor fusion as described in claim 1, characterized in that, The multi-sensor network adopts a layered distributed architecture, including: a bottom sensor node layer, an intermediate wireless transmission layer, and a top data processing layer. The bottom sensor node layer is deployed in key locations in the industrial area, and each sensor node integrates at least one type of sensor for real-time fire data acquisition. The intermediate wireless transmission layer transmits the fire data acquired by the bottom sensor node layer to the aggregation node via ZigBee or LoRa communication protocols. The top data processing layer receives the fire data uploaded by the aggregation node and performs time synchronization and format standardization processing on the fire data.

3. The industrial fire early prevention and warning system based on multi-sensor fusion as described in claim 2, characterized in that, The method for generating the dynamic false alarm discrimination rule includes: establishing a historical fire data sample library, which includes normal data, false alarm data and real fire data under different operating conditions; training the historical fire data sample library with a machine learning algorithm to generate a dynamic false alarm discrimination model, wherein the input of the dynamic false alarm discrimination model is the real-time collected fire data, and the output is the data validity judgment result.

4. The industrial fire early prevention and warning system based on multi-sensor fusion as described in claim 3, characterized in that, The dynamic false alarm discrimination rules also include the correlation between the rate of increase of temperature data and the trend of smoke concentration change, and the correlation between the duration of flame spectral signal and the duration of gas concentration exceeding the standard. The rate of temperature rise is correlated with the trend of smoke concentration change. When the rate of temperature rise exceeds a preset temperature threshold and the smoke concentration increases exponentially, it is determined to be valid fire data. The duration of the flame spectral signal is correlated with the duration of gas concentration exceeding the limit. When the duration of the flame spectral signal exceeds... t When the carbon monoxide concentration exceeds the preset lower explosion limit threshold, it is considered valid fire data.

5. The industrial fire early prevention and warning system based on multi-sensor fusion as described in claim 4, characterized in that, The fire identification module identifies the fire development stage based on valid fire data and the predicted probability of fire occurrence in sub-regions, including: The fire identification module receives a dynamic partition map from the risk partitioning module, obtains the geographical boundary of a specified sub-region based on the dynamic partition map, and filters out all sensor data located within the geographical boundary from the valid fire data based on the geographical boundary to form the feature dataset of the corresponding partition. The feature dataset and the risk label are input into the time series prediction model; the time series prediction model is an attention-enhanced recurrent neural network model, which dynamically allocates different attention weights according to the importance of different sensor data in the feature dataset to fire prediction; The time series prediction model outputs the probability curve of fire occurrence in the corresponding partition within a preset time period, and identifies the fire development stage based on the slope and key inflection points of the probability curve; the fire development stage includes smoldering stage, initial open flame stage and spread stage.

6. The industrial fire early prevention and warning system based on multi-sensor fusion as described in claim 5, characterized in that, The prevention and control optimization module includes: The prevention and control optimization module receives the probability curve of fire occurrence and the fire development stage from the fire identification module, and simultaneously obtains the dynamic zoning map from the risk zoning module, and reads the risk label corresponding to the sub-region where the fire occurred from the dynamic zoning map. The currently identified fire development stage and the corresponding risk label of the sub-area are used as query keys to retrieve the preset prevention and control strategy knowledge base and obtain the corresponding set of prevention and control actions. The prevention and control strategy knowledge base stores a strategy mapping table, which defines the mapping relationship between the input condition combination of fire development stage and risk label and the output predefined set of prevention and control actions. The matched set of prevention and control actions is parsed into a structured prevention and control strategy; the elements of the prevention and control strategy include prevention and control actions, resource allocation, and execution order. Based on the prevention and control strategy and the dynamic partition map, generate fire extinguishing or isolation task instructions for one or more AGVs; For each AGV that receives a fire extinguishing or isolation task instruction, the AGV path planning algorithm uses the dynamic partition map and the AGV's real-time position as input to calculate the optimal path that avoids the first risk sub-area and leads to the task target location. The optimal path is then bound to the fire extinguishing or isolation task instruction to form an AGV action instruction set, which is then sent to the AGV actuator to drive its execution.

7. The industrial fire early prevention and warning system based on multi-sensor fusion as described in claim 6, characterized in that, The prevention and control strategies in the knowledge base include: For sub-areas that are in the smoldering stage and are classified as risk level 3, ventilation and smoke extraction and early warning notifications will be activated. For sub-areas in the initial stage of open flame and with a risk label of second risk, dispatch AGVs to carry fire extinguishers and activate the sprinkler system; For sub-areas that are in the spread stage and are labeled as the first risk level, multiple AGVs are dispatched to work together to perform tasks such as setting up firebreaks and extinguishing fires.

8. The industrial fire early prevention and warning system based on multi-sensor fusion as described in claim 7, characterized in that, The path planning algorithm for the AGV is as follows: The dynamic partition map is mapped to a dynamic cost map, wherein each sub-region is regarded as an independent navigation unit; Based on the risk label of each sub-region in the dynamic zoning map, the passage cost is dynamically calculated for each sub-region in the dynamic cost map; the passage cost increases exponentially with the risk level represented by the risk label. Simultaneously, the current position of the AGV and the target position specified by the control strategy are obtained. Using the current position and the target position as the start and end points of the path, a graph search algorithm is used to search on the generated dynamic cost map. The goal of the graph search algorithm is to find a path from the current position to the target position such that the sum of the travel costs of all sub-regions traversed by the path, i.e., the cumulative travel cost, is the minimum among all paths. The path with the lowest cumulative travel cost is then output as the optimal path for the AGV to execute.

9. The industrial fire early prevention and warning system based on multi-sensor fusion as described in claim 8, characterized in that, The specific process of feeding back the execution results to the data acquisition module includes: During the execution of AGV action command sets, the AGV collects environmental data after execution through its onboard auxiliary sensors; The environmental data after execution is compared with the valid fire data that triggered the execution task to generate a strategy effectiveness evaluation signal; The effectiveness evaluation signal of the strategy is sent to the data acquisition module as a dynamic correction parameter for data acquisition and discrimination.

Images