A fire protection data integration method and system based on distributed heterogeneous protocol adaptation

Abstract

This invention relates to the field of fire protection data integration technology, and discloses a fire protection data integration method and system based on distributed heterogeneous protocol adaptation. This method achieves the acquisition and standardized conversion of multi-source fire protection data through a dual-library collaborative mechanism of a protocol template library and a plugin library. Enhanced fire event feature data is obtained by using sensor spatial topology modeling and multi-scale temporal feature collaborative extraction. This enhanced feature data is then subjected to temporal and spatial dual-path feature extraction and fusion analysis through a multi-scale temporal convolutional network and a spatiotemporal graph attention network. Combined with physical law verification and confidence assessment, intelligent judgment results of fire events are obtained. Hierarchical linkage control and multi-platform message push are executed through an event-driven rule-based process engine. This invention can solve the problems of difficulty in integrating heterogeneous fire protection equipment data, high false alarm rate of traditional fixed threshold methods, and insufficient early warning capabilities.

Description

Technical Field

[0001] This invention relates to the field of fire protection data integration technology, and more specifically, to a fire protection data integration method and system based on distributed heterogeneous protocol adaptation. Background Technology

[0002] With the expansion of new energy power plants, fire safety management faces challenges due to the diversity of equipment types and heterogeneous communication protocols. In existing fire protection systems, sensors and controllers from different manufacturers use their own proprietary protocols, leading to difficulties in data acquisition and integration, poor system scalability, and requiring extensive manual configuration for the integration of new equipment, resulting in high operation and maintenance costs.

[0003] Traditional fire monitoring methods mainly rely on fixed threshold alarms, which are triggered when monitored parameters exceed preset thresholds. This method has two problems: first, it cannot capture early signs of anomalies, making it difficult to provide timely warnings for gradually developing fires such as battery thermal runaway during the slow parameter rise phase; second, it has a high false alarm rate, as parameter fluctuations caused by environmental factors can easily trigger false alarms, affecting system reliability.

[0004] Furthermore, existing methods lack the ability to comprehensively analyze spatiotemporal correlation information from multiple sensors. They rely solely on isolated judgments from a single sensor and cannot effectively utilize the collaborative response patterns between sensors and the propagation patterns of fire in space, resulting in insufficient accuracy in judgment. Summary of the Invention

[0005] This invention provides a fire protection data integration method and system based on distributed heterogeneous protocol adaptation, which solves the technical problems of difficulty in integrating heterogeneous fire protection equipment data, high false alarm rate of traditional fixed threshold method and insufficient early warning capability in related technologies.

[0006] This invention provides a fire protection data integration method based on distributed heterogeneous protocol adaptation, comprising the following steps: S1, through the collaborative mechanism of the protocol template library and the plug-in library, completes the adaptive identification of the protocol and the loading of the plug-in, realizes the collection and standardized conversion of multi-source fire protection data, and obtains structured equipment status data and raw event data; S2 receives structured equipment status data and raw event data, and uses sensor spatial topology modeling and multi-scale temporal feature collaborative extraction to obtain enhanced fire event feature data containing spatiotemporal context information. S3 enhances the fire incident feature data by performing temporal and spatial dual-path feature extraction and fusion analysis through a multi-scale temporal convolutional network and a spatiotemporal graph attention network. Combined with physical law verification and confidence assessment, the intelligent judgment results of fire incidents are obtained. S4 acquires intelligent analysis results of fire incidents, executes hierarchical linkage control and multi-platform message push through an event-driven rule process engine, and outputs the execution status of linkage control commands and delivery confirmation of event push messages. S5, based on the operating status of each gateway node during the fire data collection process, performs automatic fault transfer and device connection migration through comprehensive health assessment and dynamic load balancing strategy, and outputs the updated gateway cluster topology and load distribution status.

[0007] In a preferred embodiment, S1 includes: Distributed gateway nodes discover access to fire protection equipment through network scanning or device registration mechanisms, establish a device connection table, and record the identifier, access time, connection status, and gateway node to which each device belongs. The gateway node queries and matches the protocol template library based on the device type and manufacturer information recorded in the device connection table. If a match is found, the corresponding protocol processing plugin is loaded from the protocol adaptation plugin library. For new devices that cannot be matched, the gateway node collects the raw data stream sent by the device, extracts the protocol feature vector, and calculates the similarity with the feature vectors of known protocols in the protocol template library. If the highest similarity exceeds the preset matching threshold, the protocol type is confirmed and the corresponding plugin is loaded. If it is lower than the preset matching threshold, the protocol self-learning process is started, new parsing rules are generated, and the protocol template library and protocol adaptation plugin library are updated at the same time.

[0008] In a preferred embodiment, S2 includes: A sensor spatial topology map is constructed based on the equipment layout and configuration of the site. The spatial topology map uses sensor devices as nodes and the spatial proximity relationships between devices as edges. For any two sensor nodes, calculate their spatial Euclidean distance. When the spatial distance is less than a preset neighbor distance threshold, establish an edge connection. The weight of the edge is assigned using the reciprocal of the distance. For sensor nodes located within the same fire compartment, an edge connection is established even if the spatial distance exceeds a preset proximity threshold, and the edge weight is set to the reciprocal of the maximum sensor spacing within the fire compartment. For different fire compartments, edge connections are established only when there is a physical connection path, and the edge weight is determined based on the length and cross-sectional area of ​​the connection path.

[0009] In a preferred embodiment, S2 further includes: For the device that triggers the original event, the system maintains a sliding time window. The time window length is set to the first time window length for gradually developing events and the time window length is set to the second time window length for rapidly developing events. The second time window length is less than the first time window length. Collect monitoring data of the triggering device and its neighboring devices in the spatial topology map within the time window, and construct a time-series data matrix; For each sensor in the time series data matrix, the instantaneous rate of change and acceleration of the parameters are calculated on a short time scale, the cumulative change and trend of the parameters are calculated on a medium time scale, and the mean, variance and extreme values ​​of the parameters are calculated on a long time scale. For adjacent sensor nodes in the spatial topology map, calculate the temporal correlation of their monitoring parameters within the time window, find the time offset that maximizes the correlation through the cross-correlation function, and verify the rationality of the time delay based on the spatial distance between the sensors and the physical speed of fire propagation.

[0010] In a preferred embodiment, S3 includes: The network employs a dilated causal convolutional structure, which contains three parallel convolutional branches; The first branch uses a convolution kernel with the first dilation rate to capture the most recent time steps within the receptive field coverage time window, which is used to capture the instantaneous abrupt changes in temperature and gas concentration. The second branch uses a convolution kernel with a second dilation rate to capture the continuous upward or downward trend of parameters within the receptive field coverage time window. The second dilation rate is greater than the first dilation rate. The third branch uses a convolution kernel with a third dilation rate, and its receptive field covers the entire time window to capture the overall distribution characteristics of the parameters and baseline shift. The third dilation rate is greater than the second dilation rate. The output features of the three branches are fused through a concatenation operation to form a multi-scale temporal feature representation.

[0011] In a preferred embodiment, S3 further includes: The association weight between each pair of adjacent sensor nodes is dynamically calculated using a graph attention mechanism. The weight is calculated based on the temporal feature representation and spatial positional relationship of the two nodes. For each edge in the spatial topology graph, multi-scale temporal feature representations of the nodes at both ends of the edge are extracted. The two feature vectors are concatenated and then input into the attention computation layer. The attention computation layer maps the concatenated feature vectors into scalar attention scores. The attention scores are then processed by exponential normalization to obtain the association weights. Each node updates its own feature representation by aggregating the features of its neighbors through attention weighting. The update method is that the new feature representation of the current node is equal to the weighted sum of the original feature representation of the current node and the feature representations of all its neighbors. The network progressively expands the receptive field of each node by stacking multiple graph attention layers, outputting an enhanced feature representation for each sensor node.

[0012] In a preferred embodiment, S3 further includes: The physical law verification first examines the temporal relationship between temperature and smoke. If the rise time of smoke concentration is earlier than the rise time of temperature and exceeds the preset temporal deviation threshold, it is determined that the temporal relationship is abnormal. Secondly, the spatial continuity of heat propagation is verified. The theoretical time for heat to propagate to surrounding devices is estimated based on the spatial location of the triggering device, and the theoretical propagation time is compared with the actual observed triggering time of surrounding devices. The correlation between gas concentration and temperature was verified again. The correlation coefficient between the rate of change of gas concentration and the rate of change of temperature within the time window was calculated. If the correlation coefficient was lower than the preset correlation threshold, it was likely a gas leak rather than a fire. The confidence score is calculated using a weighted fusion method, which takes the highest class probability value output by the deep learning classifier as the base confidence score, multiplies it by the consistency coefficient of physical law verification, and then multiplies it by the deterministic coefficient output by uncertainty quantification to obtain the final confidence score.

[0013] In a preferred embodiment, S4 includes: The rule process engine maintains a rule base, and triggering conditions are matched based on event type, event level, site identifier, equipment type, and confidence score; The process for warning-level events includes generating an event log and sending a notification message to the mobile terminal of the operations and maintenance personnel; The process for warning-level events has been enhanced with a video review step. The video analysis component is used to detect the presence of visual features such as open flames or smoke. If fire features are confirmed, the event level is upgraded to the danger level. The process for hazardous events includes generating linkage control commands, which are then sent to on-site execution equipment via the interface of the fire alarm linkage controller. The message push component encapsulates fire incident data into push messages in a standard format and sends them to the upper-level management platform through a message queue middleware.

[0014] In a preferred embodiment, S5 includes: The gateway node establishes a heartbeat connection with the central management service and periodically reports the node's health status data; The central management service conducts a comprehensive health assessment of each node, compares each indicator with the preset normal range, assigns different weights to each indicator according to its impact on system availability, and performs a weighted summation to obtain the comprehensive health score of the node. When the health score of a node is lower than the preset sub-health threshold, the load migration strategy is triggered; when the heartbeat of a node times out or the health score drops to zero, the failover strategy is triggered. When migrating device connections, the target node loads the corresponding plugin from the protocol adaptation plugin library and establishes a new communication connection with the device. After the new connection is successfully established, the device connection table is updated.

[0015] In a preferred embodiment, a fire data integration system based on distributed heterogeneous protocol adaptation is used to perform the steps in the aforementioned fire data integration method based on distributed heterogeneous protocol adaptation, including: The protocol adaptation module is used to complete the adaptive identification of protocols and the loading of plugins through the collaborative mechanism of the protocol template library and the plugin library, so as to realize the collection and standardized conversion of multi-source fire protection data and obtain structured equipment status data and raw event data. The spatiotemporal feature extraction module is used to receive structured equipment status data and raw event data, and adopts sensor spatial topology modeling and multi-scale temporal feature collaborative extraction to obtain enhanced fire event feature data containing spatiotemporal context information. The intelligent assessment module is used to extract and analyze the enhanced fire incident feature data through a multi-scale temporal convolutional network and a spatiotemporal graph attention network, combining temporal and spatial dual-path feature extraction and fusion analysis. Combined with physical law verification and confidence assessment, the intelligent assessment results of the fire incident are obtained. The linkage control module is used to obtain the intelligent analysis results of fire incidents, execute hierarchical linkage control and multi-platform message push through the event-driven rule process engine, and output the execution status of linkage control commands and the delivery confirmation of event push messages. The high availability assurance module is used to automatically transfer faults and migrate device connections based on the operating status of each gateway node during the fire data collection process, through comprehensive health assessment and dynamic load balancing strategies, and output the updated gateway cluster topology and load distribution status.

[0016] The beneficial effects of this invention are as follows: Through the collaborative mechanism of the protocol template library and the plug-in library, the system achieves adaptive access and protocol self-learning extension of heterogeneous fire protection equipment, solves the problem of data integration difficulties caused by the inconsistency of communication protocols of equipment from different manufacturers, reduces the manual configuration cost of accessing new equipment, and improves the scalability and operation and maintenance efficiency of the system. By combining dual-path feature extraction and fusion analysis of multi-scale temporal convolutional networks and spatiotemporal graph attention networks, and integrating physical law verification and confidence assessment, intelligent judgment of fire incidents is achieved. It can capture early abnormal signs before the alarm is triggered by the traditional fixed threshold method, while effectively reducing the false alarm rate caused by environmental factors and improving the accuracy and timeliness of fire identification. Attached Figure Description

[0017] Figure 1 This is a flowchart of a fire data integration method based on distributed heterogeneous protocol adaptation according to the present invention; Figure 2 This is a flowchart of a fire data integration method based on distributed heterogeneous protocol adaptation according to the present invention. Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses a fire protection data integration method based on distributed heterogeneous protocol adaptation, such as... Figures 1 to 2 As shown, it includes the following steps: S1, through the collaborative mechanism of the protocol template library and the plug-in library, completes the adaptive identification of the protocol and the loading of the plug-in, realizes the collection and standardized conversion of multi-source fire protection data, and obtains structured equipment status data and raw event data; During the system deployment phase, maintenance personnel input the basic configuration of the site through the management interface. This configuration includes the 3D spatial coordinates of each sensor device's installation location, its fire compartment identifier, device type, and manufacturer information. This configuration information is stored in the site configuration database, serving as the foundational data for subsequent spatiotemporal feature extraction and spatial topology modeling. The system maintains two independent but collaborative data structures: a protocol template library and a protocol adapter plugin library. The protocol template library stores feature vector descriptions of various known protocols for protocol type identification and matching. The protocol adapter plugin library stores executable parsing plugins corresponding to various protocols. Each plugin encapsulates the encoding / decoding logic and communication mechanism of a specific protocol for actual data acquisition and parsing execution. The relationship between the two is as follows: first, the protocol type used by the device is identified through the protocol template library; then, the corresponding plugin is loaded from the protocol adapter plugin library to execute data acquisition. For new or unknown protocols, both libraries are updated simultaneously through a self-learning mechanism.

[0020] S1.1, the distributed gateway node discovers the accessed fire protection equipment through network scanning or device registration mechanisms, obtains basic information such as the network address, port number, and connection method of the equipment, establishes a device connection table, and records the identifier, access time, connection status, and gateway node to which each device belongs, providing a device list and connection parameters for subsequent data collection and protocol adaptation.

[0021] S1.2, the gateway node, based on the device type and manufacturer information recorded in the device connection table, first performs a query and matching in the protocol template library. The protocol template library pre-contains standard protocol descriptions such as Modbus TCP, OPC UA, MQTT, and BACnet / IP, as well as standard feature vector descriptions of proprietary protocols from various domestic and international fire equipment manufacturers. If a matching protocol type is found in the protocol template library, the corresponding protocol processing plugin is loaded from the protocol adaptation plugin library based on the protocol type identifier. The plugin is developed using a unified interface specification, including standard methods for connection establishment, data reception, data parsing, and data transmission, enabling devices using different protocols to be uniformly scheduled and managed by the operation control engine.

[0022] For new devices whose device type and manufacturer information cannot be directly matched in the protocol template library, the gateway node collects the raw data stream sent by the device, analyzes the data frame structure to identify components such as the frame header, length field, data field, and checksum, extracts the protocol feature vector, and calculates the similarity between this feature vector and the feature vectors of all known protocols in the protocol template library one by one. The protocol with the highest similarity is selected as the candidate matching result. If the highest similarity exceeds the matching threshold, the protocol type is confirmed and the corresponding plugin is loaded from the plugin library. If the highest similarity is lower than the matching threshold, the protocol self-learning process is initiated. New parsing rules are generated through iterative analysis of the raw data stream, the feature vector of the new rule is stored in the protocol template library, and the corresponding parsing plugin is encapsulated and stored in the protocol adaptation plugin library to achieve adaptive recognition and expansion of the protocol, so that it can be directly reused when similar devices are connected in the future.

[0023] S1.3 After establishing a communication connection with the fire protection equipment, the protocol adapter plugin continuously collects real-time data reported by the equipment, including temperature values ​​from temperature sensors, smoke concentration from smoke detectors, gas concentration from combustible gas detectors, flame detection status from flame detectors, alarm flags from video surveillance equipment, and battery temperature and voltage from the battery management system. The collected raw data undergoes field splitting and type conversion by the decoder component. Based on the field mapping relationships defined in the protocol description file, the device-specific field names and enumerated values ​​are converted to corresponding fields in the standard data model, completing unit conversion, range mapping, status code conversion, and other data standardization processing. The converted data undergoes rationality verification by the validator component, including numerical range verification, timestamp validity verification, and required field integrity verification. Data that fails verification is processed according to the failure type: data with values ​​exceeding the physical range is marked as abnormal and logged; data with abnormal timestamps is supplemented using the gateway's local time; and data packets with missing required fields are discarded.

[0024] In step S1.4, the validated data is assembled into a structured device status data object, represented in JSON format, containing fields such as device identifier, device type, acquisition time, monitoring parameter list, alarm status, and data source. The gateway node categorizes the device status data object according to the device identifier, updates the device status cache table, and generates status change events by comparing the current status with historical statuses. When a change in the device's alarm status field is detected, or when a monitored parameter exceeds a preset trigger range, raw event data is generated, recording information such as event trigger time, event type, triggering device, trigger conditions, and parameter snapshot. Structured device status data and raw event data are output, providing standardized data input for subsequent spatiotemporal feature extraction and intelligent analysis.

[0025] S2 receives structured equipment status data and raw event data, and uses sensor spatial topology modeling and multi-scale temporal feature collaborative extraction to obtain enhanced fire event feature data containing spatiotemporal context information. S2.1 After the raw event data enters the spatiotemporal feature extraction module, a sensor spatial topology map is constructed based on the equipment layout configuration of the site. This map uses sensor devices as nodes and spatial proximity relationships between devices as edges. The system reads the three-dimensional spatial coordinates, fire compartment identifier, equipment type, and other attribute information of each sensor device from the site configuration database. For any two sensor nodes, their spatial Euclidean distance is calculated. When the spatial distance between two sensors is less than the proximity distance threshold, an edge connection is established between the two nodes. This proximity distance threshold is determined based on the spatial scale of the site and the fire propagation characteristics. By statistically analyzing the typical installation spacing of various types of equipment within the site and the spatial impact range of historical fires, a distance value that can cover all sensors within the same functional unit is selected as the threshold. Preferably, for the electrochemical energy storage power station scenario, the proximity distance threshold is set as the first distance threshold.

[0026] Edge weights are assigned using the reciprocal of the distance, meaning the edge weight between two sensor nodes is equal to the reciprocal of their spatial distance. The closer the sensor nodes are, the greater the edge weight, reflecting the physical principle that closer locations are more easily affected when a fire spreads through space. For sensor nodes located within the same fire compartment, an edge connection is established even if the spatial distance exceeds the proximity threshold. The edge weight is set to the reciprocal of the maximum sensor spacing within that fire compartment. This is because the space within the same fire compartment is considered a single unit in fire management, and a fire can affect the entire area through smoke diffusion, heat radiation, and other means.

[0027] For different fire compartments, edge connections are established only where a physical connection path exists. Physical connection paths include ventilation ducts, cable trenches, corridors, etc. The system reads connection path information from the building structure configuration of the site. For boundary sensor nodes of different fire compartments connected by a connection path, an edge connection is established. The edge weight is determined based on the length and cross-sectional area of ​​the connection path; the shorter the path and the larger the cross-sectional area, the faster the smoke and heat propagate, and the greater the edge weight. The spatial topology map constructed using these rules accurately reflects the spatial relationships between sensors, providing structured spatial information for subsequent collaborative analysis.

[0028] S2.2 For devices that trigger the original event, the system maintains a sliding time window, the length of which is determined based on the typical development time of the fire event. The system analyzes the parameter evolution curves of various confirmed fire events in the historical fire event database, statistically analyzing the time interval distribution from the appearance of abnormal symptoms to the triggering of a traditional fixed threshold alarm. The goal is to capture early symptoms before the traditional method triggers an alarm, while also considering the increased computational complexity caused by excessively long windows and the introduction of irrelevant historical fluctuations, thus comprehensively determining the time window length for various events.

[0029] For events that develop gradually, such as battery thermal runaway, the time window length is set to the first time window length; for events that develop rapidly, such as electrical fires, the time window length is set to the second time window length, which is shorter than the first time window length to ensure timely response. Within the time window, the system collects all monitoring data of the triggering device and its neighboring devices in the spatial topology graph. The selection of neighboring devices is achieved through a graph traversal algorithm. Starting from the triggering device node, a breadth-first search is performed along the edges of the topology graph to collect data from all neighboring nodes within a preset hop count range from the triggering device. This hop count range can cover the area where the triggering device is located and the sensors in its adjacent areas, ensuring sufficient spatial coverage while avoiding the introduction of too much data from irrelevant devices.

[0030] The collected data is constructed into a time-series data matrix. The rows of the matrix correspond to different sensor devices, the columns correspond to the sampling times within the time window, the sampling frequency is determined according to the device type, and the matrix elements are the sensor monitoring parameter values ​​at the corresponding time, including various physical quantities such as temperature, smoke concentration, carbon monoxide concentration, and hydrogen concentration.

[0031] S2.3, for each sensor in the time-series data matrix, extracts multi-scale temporal features to capture change patterns at different time granularities. The system employs a sliding statistical method with different window lengths to calculate the instantaneous rate of change and acceleration of parameters at short time scales, reflecting the rapid fluctuation characteristics of parameters; at medium time scales, it calculates the cumulative change and trend of parameters, reflecting the continuous evolution characteristics of parameters; and at long time scales, it calculates the overall statistical characteristics of parameters, including mean, variance, and extreme values, reflecting the overall distribution characteristics of parameters. For temperature sensor data, the short-time-scale temperature rise rate can identify sudden heat sources, the medium-time-scale cumulative temperature rise can identify gradual thermal runaway processes, and the long-time-scale temperature baseline can distinguish between abnormal temperature rises and normal fluctuations in ambient temperature.

[0032] For gas sensor data, short-timescale concentration abrupt changes can identify gas leaks or rapid release of combustion products, medium-timescale concentration integrals can assess the cumulative hazard level of gases, and long-timescale concentration baselines can eliminate interference from ambient background gases. After completing multi-scale feature extraction for single sensors, the system further analyzes the cooperative change patterns between different sensors. For adjacent sensor nodes in the spatial topology map, the temporal correlation of their monitoring parameters within a time window is calculated, and the covariance calculation method within a sliding window is used to quantify the degree of synchronous change of the two sensor data. When the data from two sensors show a cooperative upward or downward trend within the time window, the time offset that maximizes the correlation is further found through the cross-correlation function. This time offset reflects the time required for a fire or abnormal state to propagate from one sensor location to another.

[0033] The system verifies the rationality of time delays based on the spatial distance between sensors and the physical speed of fire propagation. Time delays that conform to physical laws are marked as causally related coordinated triggering; abnormal time delays are marked as possible independent events or sensor malfunctions. Simultaneously, the system performs a breadth-first search along the edges of the spatial topology, starting from the triggering device, to check the triggering status and parameter changes of adjacent devices. If the fire propagates along a continuous spatial path, and adjacent devices trigger alarms sequentially or monitoring parameters sequentially exceed abnormal thresholds, it is marked as a spatially continuous event, and the propagation path is recorded. If only an isolated single device triggers without response from surrounding devices, or if the triggering devices exhibit a disjointed spatial distribution, it is marked as a suspected false alarm or localized malfunction, reducing the initial confidence level of the event.

[0034] S2.4 integrates multi-scale temporal features from a single sensor, collaborative change patterns from multiple sensors, and spatial propagation continuity checks to construct enhanced fire event feature data. This feature data adds multiple dimensions to the original event data, including temporal feature vectors, spatial topology feature vectors, and collaborative triggering feature vectors. The temporal feature vectors include characteristic components such as the rate of change, cumulative change, and statistical characteristics of each sensor at different time scales. The spatial topology feature vectors include characteristic components such as the spatial location of triggering devices, the response of neighboring devices, and the continuity of the propagation path. The collaborative triggering feature vectors include characteristic components such as the temporal correlation coefficient between sensors, time delay, and the number of collaboratively triggering devices. Furthermore, the system calculates the event complexity score, which comprehensively considers factors such as the number of sensors involved, spatial range, development speed, and parameter change amplitude. A higher complexity score indicates a wider impact range and a more severe development trend. The output is enhanced fire event feature data containing rich spatiotemporal context information, providing high-quality feature input for subsequent deep learning models.

[0035] S3 enhances the fire incident feature data by performing temporal and spatial dual-path feature extraction and fusion analysis through a multi-scale temporal convolutional network and a spatiotemporal graph attention network. Combined with physical law verification and confidence assessment, the intelligent judgment results of fire incidents are obtained. S3.1, the deep learning-based analysis engine initializes its model parameters through offline training before system deployment. During the data preparation phase, the system collects station operation data from a historical fire database covering a sufficiently long historical period with different seasons and operating conditions. This includes sensor monitoring data during normal operation and confirmed fire incident data. Normal operation data is obtained by filtering periods with no alarm records and stable parameters. Fire incident data includes confirmed battery thermal runaway events, electrical fires, and combustible material combustion events. The number of samples for each type of event is increased using data augmentation techniques. These methods include adding random noise to the time-series data that conforms to the sensor measurement error distribution, randomly cropping and transforming the time scale of the time window, and randomly deleting edges from the spatial topology graph to simulate sensor failures. Data augmentation ensures that the number of training samples reaches the scale required for model training.

[0036] The network input is a 3D tensor consisting of the number of sensors multiplied by the number of time steps multiplied by the feature dimension. The network output is the event category probability distribution and confidence score. During training, the cross-entropy loss function is used to measure the difference between the model's predictions and the true labels. The network parameters are updated using the backpropagation algorithm. A validation set is used to monitor model performance during training; training stops when the accuracy on the validation set no longer improves to avoid overfitting. Finally, an independent test set is used to evaluate the model's generalization performance. After ensuring the model meets the performance requirements of practical applications, the model parameters are saved to a model library. During system runtime, the pre-trained model parameters are loaded from the model library to perform inference analysis on the real-time collected enhanced fire event feature data.

[0037] S3.2, the multi-scale temporal convolutional network extracts the temporal feature vector from the enhanced fire event feature data as input. This temporal feature vector contains feature components such as the rate of change, cumulative change, and statistical characteristics of each sensor at different time scales. The network adopts a dilated causal convolutional structure. By setting convolutional layers with different dilation rates, the network can simultaneously capture short-term instantaneous fluctuations, medium-term continuous evolution, and long-term overall trends. The network contains three parallel convolutional branches, each responsible for extracting features at different time scales.

[0038] The first branch uses a convolution kernel with a smaller dilation rate to cover the most recent time steps within the receptive field time window, in order to capture the instantaneous abrupt changes in temperature and gas concentration. This is crucial for identifying rapidly developing events such as electrical fires, as electrical fires are often accompanied by a sharp rise in temperature and a rapid release of flammable gases.

[0039] The second branch uses a convolution kernel with a moderate expansion rate to capture the continuous upward or downward trend of parameters within the receptive field coverage time window. This is crucial for identifying the early, gradual evolution of battery thermal runaway, as early-stage thermal runaway often manifests as a slow but continuous rise in temperature, which is difficult to identify using simple instantaneous threshold judgment. However, by analyzing the cumulative temperature rise and change trend over a moderate time scale, anomalies can be detected before traditional methods trigger alarms.

[0040] The third branch uses a convolution kernel with a large expansion rate, and the receptive field covers the entire time window to capture the overall distribution characteristics of parameters and baseline shift. This helps to distinguish between anomalous evolution and normal fluctuations in ambient temperature. For example, in hot summer weather, the overall temperature baseline of the battery compartment will rise, but this rise is slow and uniform, which is significantly different from the local rapid temperature rise caused by battery thermal runaway. Long-term baseline analysis can effectively eliminate the interference of environmental factors.

[0041] The output features of the three branches are fused through a concatenation operation to form a multi-scale temporal feature representation. This representation contains both fine-grained instantaneous change information and coarse-grained long-term evolution information, which can comprehensively depict the complex dynamic characteristics of fire events in the time dimension.

[0042] S3.3, the spatiotemporal graph attention network takes the spatial topology feature vector and co-triggered feature vector in the enhanced fire event feature data as input, and combines it with the sensor spatial topology graph constructed in S2. It learns the dynamic association weights between different sensor nodes through a graph attention mechanism. Traditional fixed-weight graph convolution methods cannot adapt to the dynamic changes in sensor relationships in fire scenarios. For example, when the ventilation system is on and off, the propagation path and speed of smoke in space differ significantly, and the association strength between sensors also changes accordingly. A fixed graph structure cannot reflect this dynamism.

[0043] The spatiotemporal graph attention network dynamically calculates the association weights between each pair of adjacent sensor nodes through an attention mechanism. The weights are calculated based on the temporal feature representations and spatial positional relationships of the two nodes. The specific implementation process is as follows: For each edge in the spatial topology graph, the network extracts the multi-scale temporal feature representations of the nodes at both ends of the edge. The two feature vectors are concatenated and then input into the attention calculation layer. The attention calculation layer contains a fully connected neural network, which maps the concatenated feature vectors into a scalar attention score. The attention score is then exponentially normalized to all neighboring nodes of each node to obtain the association weights, ensuring that the sum of the weights of all neighbors of each node is one.

[0044] The magnitude of the association weight reflects the degree of synchronization between the response patterns of two sensors. When the temporal feature representations of two sensors are highly similar, it indicates that their response patterns are highly synchronized, and the association weight is large. Conversely, when the temporal feature representations of two sensors are low similarity, it indicates that their response patterns differ, and the association weight is small. The calculated association weight is used to weighted aggregate the feature information of neighboring nodes. Each node updates its own feature representation by weighted aggregation of the features of its neighbors through attention. The update method is that the new feature representation of the current node is equal to the weighted sum of the original feature representation of the current node and the feature representations of all neighboring nodes. The weight is the association weight calculated by the attention mechanism.

[0045] This process enables the node's feature representation to not only include the temporal evolution information of its own sensor but also the cooperative response information of spatially neighboring sensors. The network progressively expands the receptive field of each node through the stacking of multiple graph attention layers, allowing nodes to perceive information from multi-hop neighbors and thus capture the propagation process and impact range of a fire in space. The output of the spatiotemporal graph attention network is an enhanced feature representation for each sensor node, which integrates temporal evolution features and spatial cooperative features. This representation accurately characterizes the spatiotemporal coupling characteristics of fire events, especially for battery thermal runaway events. By analyzing the cooperative response patterns of the triggering device and its neighboring devices, it can determine whether the hazard is a local anomaly of a single battery or a systemic danger to the entire battery compartment, thereby improving the accuracy of the judgment.

[0046] S3.4, the anomaly detection and event classification module receives the enhanced feature representation output by the spatiotemporal graph attention network and uses a pre-trained deep learning classifier to distinguish between normal and abnormal patterns of the current event and to identify the event type. During the training phase, the classifier learns discriminative features of normal operation modes and various abnormal event patterns, enabling normal samples and different types of abnormal samples to form a distinguishable distribution in the feature space.

[0047] During training, the classifier learns sensor response characteristics during normal operation from historical databases. These characteristics are manifested as random fluctuations in parameters within a certain range, with the amplitude and frequency of fluctuations conforming to the normal changing patterns of environmental factors and equipment operating status. At the same time, it learns the characteristics of confirmed fire incident data, including the gradual temperature rise characteristics of battery thermal runaway events, the rapid temperature rise and sudden gas concentration changes of electrical fire events, and the continuous increase in smoke concentration of combustible combustion events. Different types of events have significant differences in temporal evolution patterns and spatial propagation patterns. The classifier learns these differences to achieve accurate identification of event types.

[0048] During the inference phase, the classifier first maps the enhanced feature representations to the classification space through a fully connected layer. Then, it calculates the probability distribution of each category through the output layer. The categories include normal operation, battery thermal runaway, electrical fire, combustible material combustion, gas leak, equipment failure, and environmental interference. The classifier outputs a probability value for each category, reflecting the likelihood that the current event belongs to that category. The category with the highest probability is used as the initial judgment result for the event. For the normal operation category, if its probability is higher than the normal judgment threshold, the current event is judged as normal fluctuation or environmental interference, and fire linkage is not triggered. For the abnormal event category, if its probability is higher than the abnormal judgment threshold, the current event is judged as a real fire or dangerous state, and the process proceeds to the subsequent physical law verification and confidence assessment process.

[0049] S3.5, the physical law verification and confidence assessment module receives the preliminary judgment results from the deep learning classifier, performs a physical consistency check on them, and comprehensively quantifies and assesses the confidence level. The physical law verification first examines the temporal relationship between temperature and smoke. In typical fire scenarios, temperature rise often precedes or coincides with the increase in smoke concentration. The module compares the start times of parameter changes in the temperature sensor and smoke detector. If the smoke concentration rises earlier than the temperature rise, exceeding a timing deviation threshold, it is considered an abnormal timing relationship. This timing deviation threshold is determined based on the sensor's response speed and installation location. When an abnormal timing relationship occurs, it may be due to environmental interference such as oil fumes or water vapor causing false triggering of the smoke detector. The module then lowers the confidence score for this event.

[0050] Secondly, the spatial continuity of heat propagation is verified. Based on the spatial location of the triggering device and the physical model of heat conduction and heat convection, the theoretical time for heat to propagate to the surrounding devices is estimated. The theoretical propagation time is calculated by dividing the spatial distance by the heat propagation speed. The heat propagation speed is determined according to the ventilation conditions and spatial structure of the site. For enclosed spaces, heat mainly propagates through heat radiation and heat conduction. For spaces with ventilation systems, heat propagates rapidly through heat convection. The module compares the theoretical propagation time with the actual observed triggering time of the surrounding devices. If the time deviation is within a reasonable range, it is considered to conform to the spatial propagation law. If the response time of the surrounding devices deviates significantly from the theoretical time or the triggering devices are spatially distributed in a jump pattern while the devices in the middle path do not respond, the credibility of the event is reduced.

[0051] The correlation between gas concentration and temperature was verified again. For combustible gases such as carbon monoxide and hydrogen, an increase in concentration is usually accompanied by an increase in temperature. The module calculates the Pearson correlation coefficient between the rate of change of gas concentration and the rate of change of temperature within a time window. If the correlation coefficient is higher than the correlation threshold, it indicates that the change in gas concentration matches the thermal effect, supporting the judgment of a real fire. If the gas concentration increases but the temperature does not change significantly, resulting in a correlation coefficient lower than the correlation threshold, it may be a gas leak rather than a fire. The module then adjusts the event type discrimination result. After verifying the physical laws, the module uses the Monte Carlo dropout method to quantify the prediction uncertainty of the deep learning model. During the inference phase, the dropout layer in the network is kept active, and multiple forward propagations are performed on the same input. Each forward propagation produces slightly different prediction results due to the randomness of dropout. The mean of the distribution of multiple prediction results is calculated as the final discrimination result, and the variance of the distribution is calculated as a measure of uncertainty. The higher the uncertainty, the more uncertain the model's judgment of the current input.

[0052] Based on the above results, the confidence score is calculated using a weighted fusion method: the highest class probability value output by the deep learning classifier is used as the base confidence score, multiplied by the consistency coefficient of the physical law verification's various checks, and then multiplied by the certainty coefficient output by uncertainty quantification (i.e., one minus the normalized uncertainty measure). The final confidence score is obtained by multiplying these three values ​​together, with a value ranging from zero to one. A higher confidence score is achieved when the physical law verification result is consistent with the model prediction and has low uncertainty; a lower confidence score is achieved when the physical law verification result contradicts the model prediction or has high uncertainty. For events with a confidence score below the uncertainty threshold, the system triggers a secondary verification process, calling video surveillance for manual review or extending the observation time to collect more evidence.

[0053] After completing the confidence assessment, the module compares the probability values ​​of the anomaly categories with preset grading thresholds, classifying them into three levels: warning, alert, and danger. The warning level corresponds to events with a low probability of anomaly, which may be early signs of anomalies or slight parameter fluctuations, requiring continuous observation. The alert level corresponds to events with a moderate probability of anomaly, with the potential to develop into a fire, requiring manual verification and on-site inspection. The danger level corresponds to events with a high probability of anomaly, highly suspected or confirmed as a fire, requiring immediate activation of emergency response and coordinated control. If physical verification reveals that the event does not conform to the basic laws of fire development, the event level is downgraded or marked as a suspected false alarm.

[0054] The final intelligent assessment result of the fire incident is generated, which includes fields such as incident identifier, incident level, confidence score, incident type, list of triggering devices, list of associated devices, spatiotemporal feature summary, judgment criteria, and suggested handling measures. The incident identifier adopts a globally unique encoding rule that includes site identifier, timestamp, and incident sequence number. The judgment criteria field records the probability distribution of each category output by the deep learning classifier, the inspection results of various physical law verifications, and the confidence interval of uncertainty quantification. The suggested handling measures match the corresponding handling procedures from the emergency plan library according to the incident level and incident type. The assessment result is output in JSON format and transmitted to the subsequent linkage control module and event push module through a standard interface.

[0055] S4 acquires intelligent analysis results of fire incidents, executes hierarchical linkage control and multi-platform message push through an event-driven rule process engine, and outputs the execution status of linkage control commands and delivery confirmation of event push messages. S4.1 The intelligent assessment results of fire incidents serve as the execution of the event source trigger rule flow engine. The rule flow engine maintains a rule base, which stores processing rules for various fire scenarios. Each rule includes elements such as triggering conditions, execution actions, and output results. Triggering conditions are matched using attributes such as event type, event level, site identifier, equipment type, and confidence score. When the assessment result meets the triggering conditions of a rule, that rule is activated and the corresponding business process is executed.

[0056] S4.2 specifies different rule procedures for different levels of fire incidents. The procedure for early warning level incidents includes querying detailed information about the triggering device, generating an event record, pushing it to the log system, sending a notification message to the mobile terminal of maintenance personnel, and outputting the event record identifier and push status. The procedure for warning level incidents adds a video verification step, calling the video monitoring interface to obtain real-time video streams of the incident area, detecting the presence of open flames or smoke visual features through video analysis components, appending the video analysis results to the event record, and if the video analysis confirms the presence of fire characteristics, the event level is upgraded to the danger level and the corresponding procedures are triggered.

[0057] S4.3, the procedure for hazardous events includes emergency response control. Based on the area where the event occurred and the type of fire, response control commands are generated. These commands include starting fire pumps, opening sprinkler valves in the corresponding area, shutting off ventilation systems to prevent smoke spread, cutting off relevant power supplies to prevent the spread of electrical fires, activating audible and visual alarms, and unlocking emergency evacuation routes. Control commands are sent to on-site execution equipment through the fire alarm control panel interface, and the control actions are recorded in the event handling log to provide a basis for post-event analysis and accountability.

[0058] In S4.4, the rules engine matches the corresponding emergency plan to the current event based on the emergency plan library configured at the site. Each emergency plan describes the standard handling procedures and required resources for a specific type of fire, including the emergency response organizational structure, personnel responsibilities, handling steps, required materials and equipment, and external support contact information. The matched emergency plan number, plan name, and plan content are appended to the event data and pushed to emergency command personnel and on-site personnel via mobile terminals, providing decision-making references and action guidelines for emergency command.

[0059] In S4.5, the message push component encapsulates fire incident data into a standard-format push message. The message body uses JSON format and includes fields such as event identifier, event level, confidence score, event type, occurrence time, site name, site location, equipment information, spatiotemporal feature summary, judgment criteria, handling measures, emergency plan, and attachment list. The attachment list includes auxiliary information such as video screenshots, data graphs, equipment location maps, and propagation path diagrams to help maintenance personnel and emergency command personnel quickly understand the full picture of the incident.

[0060] S4.6 uses a message queue middleware to send push messages to the upper-level management platform. The message queue adopts a publish-subscribe model, supporting multiple platforms to subscribe to fire events simultaneously, including the group-level monitoring center, regional operation and maintenance center, and mobile emergency command platform. Before sending messages, data format conversion is performed according to the target platform's interface specifications to adapt to the field requirements and data structures of different platforms. After the message is sent, the system listens for the platform's confirmation response. If no confirmation is received within the timeout period, the message is resent. After the resentment limit is reached, the message is persisted to local storage, and an alarm is triggered to notify the administrator, ensuring that critical fire event information is not lost. The system also outputs the execution status of linkage control commands and the event push message and its delivery confirmation.

[0061] S5, based on the operating status of each gateway node during the fire data collection process, performs automatic failover and device connection migration through comprehensive health assessment and dynamic load balancing strategy, and outputs the updated gateway cluster topology and load distribution status. In S5.1, the system deploys multiple distributed gateway nodes, each responsible for connecting some fire protection equipment and acquiring data. The gateway nodes establish a heartbeat connection with the central management service and periodically report their health status data. This reporting includes metrics such as CPU utilization, memory usage, network bandwidth usage, number of connected devices, data processing rate, and message queue backlog. The reporting cycle is dynamically adjusted based on system load. Under normal conditions, reports are submitted according to the standard cycle; when a node's load exceeds a warning threshold, the reporting cycle is shortened to allow the central management service to promptly detect changes in node status.

[0062] In S5.2, after receiving health status data from each gateway node, the central management service performs a comprehensive health assessment on each node. The assessment method involves comparing various indicators such as CPU utilization, memory usage, and message queue backlog against preset normal ranges, calculating the deviation of each indicator, and then assigning different weights to each indicator according to its impact on system availability. A weighted sum is then calculated to obtain the node's comprehensive health score, ranging from zero to one hundred, with higher scores indicating a healthier node. High CPU and memory usage directly lead to data processing delays and are given higher weights; message queue backlog reflects whether the node's processing capacity matches the current load pressure and is given a medium weight; network bandwidth usage and the number of connected devices are used as auxiliary reference indicators and are given lower weights.

[0063] S5.3 The central management service continuously monitors the health score trends of each node. When a node's health score continuously declines and falls below the sub-health threshold, the node is determined to be in a sub-healthy state, triggering a load migration strategy to migrate some of the device connections carried by that node to other healthy nodes. The migration priority is determined based on the importance of the devices, with critical sensors in key fire protection zones being migrated first. When a node's heartbeat times out or its health score drops to zero, the node is determined to have failed, triggering a failover strategy. The central management service queries the device connection table for a list of all devices carried by the failed node. Based on the current load of each healthy node, a load balancing algorithm is used to redistribute the device connection tasks to the healthy nodes. The load balancing algorithm comprehensively considers factors such as the node's remaining processing capacity, network latency with the devices, and the geographical distribution of the devices to calculate the optimal device allocation scheme.

[0064] S5.4 The device connection migration process is as follows: The central management service sends a device access command to the target healthy node. The command includes information such as the identifier, network address, protocol configuration parameters, and priority of the device to be migrated. After receiving the command, the target node loads the corresponding plugin from the protocol adaptation plugin library and establishes a new communication connection with the device. After the new connection is successfully established and the first data collection verification is completed, the target node sends a confirmation message to the central management service. After receiving the confirmation, the central management service notifies the original node to close the connection with the device to avoid duplicate data collection. Finally, the device connection table is updated to modify the mapping relationship between the device and the gateway node to the new node, ensuring the correct data flow path in the future. Throughout the migration process, the system ensures the continuity of data collection and avoids data loss due to migration by setting a brief parallel collection window between the old and new connections.

[0065] S5.5. When a new device requests access, the central management service selects the node with the highest comprehensive health score and the fewest currently connected devices for device allocation based on the real-time load statistics of each node, avoiding performance degradation caused by load concentration. When it is detected that the overall system load remains at a high level and the health scores of all nodes are lower than the expected values, the system generates an expansion recommendation report, including the number of nodes to be added, the recommended deployment locations, the expected load sharing effect, etc. After the administrator completes the deployment of new nodes according to the recommendation, the new nodes automatically register with the central management service and participate in subsequent load balancing scheduling, outputting the updated gateway cluster topology and load distribution status to ensure the continuous high-availability operation of the system in scenarios of large-scale device access.

[0066] In an embodiment of the present invention, an application example is provided. This example is applied to the new energy field fire integration control system of a new energy group company, covering 12 energy storage power stations, 8 wind farm stations, and 5 photovoltaic power stations in the province, with a total of about 1500 fire-fighting devices connected. Taking a certain energy storage power station as an example, this site is equipped with a battery compartment temperature monitoring system, a combustible gas detection system, a smoke alarm system, a video monitoring system, etc. Before applying the method of the present invention, the site adopted the traditional fixed threshold alarm method, which had problems such as a high false alarm rate and difficulty in early warning. The data examples are shown in Tables 1 and 2: Table 1. Example of original heterogeneous fire-fighting device data

[0067] The system automatically identifies the protocol types of the above devices through the protocol adaptation gateway, completes data parsing and standardization conversion, and realizes the unified access of heterogeneous devices without manual configuration.

[0068] Table 2. Example of fire event data after standardization processing

[0069] Through the deep learning intelligent judgment mechanism, the system can accurately identify real fire situations, mark events with low confidence as suspected false alarms or equipment failures, and avoid triggering unnecessary linkage controls. Especially in event EVT20260312001, the system captured the small but continuous upward trend of temperature through a multi-scale temporal convolutional network and issued an early warning before the traditional threshold method triggered an alarm, winning precious time for emergency response. In event EVT20260312003, the system found that the increase in smoke concentration did not conform to the basic law of fire development through physical law verification, and combined with a low confidence score, marked this event as a suspected false alarm, avoiding unnecessary linkage controls triggered by false alarms.

[0070] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A fire protection data integration method based on distributed heterogeneous protocol adaptation, characterized in that, Includes the following steps: S1, through the collaborative mechanism of the protocol template library and the plug-in library, completes the adaptive identification of the protocol and the loading of the plug-in, realizes the collection and standardized conversion of multi-source fire protection data, and obtains structured equipment status data and raw event data; S2 receives structured equipment status data and raw event data, and uses sensor spatial topology modeling and multi-scale temporal feature collaborative extraction to obtain enhanced fire event feature data containing spatiotemporal context information. S3 enhances the fire incident feature data by performing temporal and spatial dual-path feature extraction and fusion analysis through a multi-scale temporal convolutional network and a spatiotemporal graph attention network. Combined with physical law verification and confidence assessment, the intelligent judgment results of fire incidents are obtained. S4 acquires intelligent analysis results of fire incidents, executes hierarchical linkage control and multi-platform message push through an event-driven rule process engine, and outputs the execution status of linkage control commands and delivery confirmation of event push messages. S5, based on the operating status of each gateway node during the fire data collection process, performs automatic fault transfer and device connection migration through comprehensive health assessment and dynamic load balancing strategy, and outputs the updated gateway cluster topology and load distribution status.

2. The fire protection data integration method based on distributed heterogeneous protocol adaptation according to claim 1, characterized in that, S1 includes: Distributed gateway nodes discover access to fire protection equipment through network scanning or device registration mechanisms, establish a device connection table, and record the identifier, access time, connection status, and gateway node to which each device belongs. The gateway node queries and matches the protocol template library based on the device type and manufacturer information recorded in the device connection table. If a match is found, the corresponding protocol processing plugin is loaded from the protocol adaptation plugin library. For new devices that cannot be matched, the gateway node collects the raw data stream sent by the device, extracts the protocol feature vector, and calculates the similarity with the feature vectors of known protocols in the protocol template library. If the highest similarity exceeds the preset matching threshold, the protocol type is confirmed and the corresponding plugin is loaded. If it is lower than the preset matching threshold, the protocol self-learning process is started, new parsing rules are generated, and the protocol template library and protocol adaptation plugin library are updated at the same time.

3. The fire protection data integration method based on distributed heterogeneous protocol adaptation according to claim 1, characterized in that, S2 includes: A sensor spatial topology map is constructed based on the equipment layout and configuration of the site. The spatial topology map uses sensor devices as nodes and the spatial proximity relationships between devices as edges. For any two sensor nodes, calculate their spatial Euclidean distance. When the spatial distance is less than a preset neighbor distance threshold, establish an edge connection. The weight of the edge is assigned using the reciprocal of the distance. For sensor nodes located within the same fire compartment, an edge connection is established even if the spatial distance exceeds a preset proximity threshold, and the edge weight is set to the reciprocal of the maximum sensor spacing within the fire compartment. For different fire compartments, edge connections are established only when there is a physical connection path, and the edge weight is determined based on the length and cross-sectional area of ​​the connection path.

4. The fire protection data integration method based on distributed heterogeneous protocol adaptation according to claim 1, characterized in that, S2 further includes: For the device that triggers the original event, the system maintains a sliding time window. The time window length is set to the first time window length for gradually developing events and the time window length is set to the second time window length for rapidly developing events. The second time window length is less than the first time window length. Collect monitoring data of the triggering device and its neighboring devices in the spatial topology map within the time window, and construct a time-series data matrix; For each sensor in the time series data matrix, the instantaneous rate of change and acceleration of the parameters are calculated on a short time scale, the cumulative change and trend of the parameters are calculated on a medium time scale, and the mean, variance and extreme values ​​of the parameters are calculated on a long time scale. For adjacent sensor nodes in the spatial topology map, calculate the temporal correlation of their monitoring parameters within the time window, find the time offset that maximizes the correlation through the cross-correlation function, and verify the rationality of the time delay based on the spatial distance between the sensors and the physical speed of fire propagation.

5. The fire protection data integration method based on distributed heterogeneous protocol adaptation according to claim 1, characterized in that, S3 includes: The network employs a dilated causal convolutional structure, which contains three parallel convolutional branches; The first branch uses a convolution kernel with the first dilation rate to capture the most recent time steps within the receptive field coverage time window, which is used to capture the instantaneous abrupt changes in temperature and gas concentration. The second branch uses a convolution kernel with a second dilation rate to capture the continuous upward or downward trend of parameters within the receptive field coverage time window. The second dilation rate is greater than the first dilation rate. The third branch uses a convolution kernel with a third dilation rate, and its receptive field covers the entire time window to capture the overall distribution characteristics of the parameters and baseline shift. The third dilation rate is greater than the second dilation rate. The output features of the three branches are fused through a concatenation operation to form a multi-scale temporal feature representation.

6. The fire protection data integration method based on distributed heterogeneous protocol adaptation according to claim 1, characterized in that, In S3, the spatiotemporal graph attention network takes the spatial topological feature vector and co-triggered feature vector in the fire incident feature data as input, including: The association weight between each pair of adjacent sensor nodes is dynamically calculated using a graph attention mechanism. The weight is calculated based on the temporal feature representation and spatial position relationship of the two nodes. For each edge in the spatial topology graph, multi-scale temporal feature representations of the nodes at both ends of the edge are extracted. The two feature vectors are concatenated and then input into the attention computation layer. The attention computation layer maps the concatenated feature vectors into scalar attention scores. The attention scores are then processed by exponential normalization to obtain the association weights. Each node updates its own feature representation by aggregating the features of its neighbors through attention weighting. The update method is that the new feature representation of the current node is equal to the weighted sum of the original feature representation of the current node and the feature representations of all its neighbors. The network progressively expands the receptive field of each node by stacking multiple graph attention layers, and outputs an enhanced feature representation for each sensor node.

7. The fire protection data integration method based on distributed heterogeneous protocol adaptation according to claim 1, characterized in that, In S3, the combination of physical law verification and confidence assessment includes: The physical law verification first examines the temporal relationship between temperature and smoke. If the rise time of smoke concentration is earlier than the rise time of temperature and exceeds the preset temporal deviation threshold, it is determined that the temporal relationship is abnormal. Secondly, the spatial continuity of heat propagation is verified. The theoretical time for heat to propagate to surrounding devices is estimated based on the spatial location of the triggering device, and the theoretical propagation time is compared with the actual observed triggering time of surrounding devices. The correlation between gas concentration and temperature was verified again. The correlation coefficient between the rate of change of gas concentration and the rate of change of temperature within the time window was calculated. If the correlation coefficient was lower than the preset correlation threshold, it was likely a gas leak rather than a fire. The confidence score is calculated using a weighted fusion method, which takes the highest class probability value output by the deep learning classifier as the base confidence score, multiplies it by the consistency coefficient of physical law verification, and then multiplies it by the deterministic coefficient output by uncertainty quantification to obtain the final confidence score.

8. The fire protection data integration method based on distributed heterogeneous protocol adaptation according to claim 1, characterized in that, In S4, hierarchical linkage control and multi-platform message push are executed through an event-driven rule flow engine, including: The rule process engine maintains a rule base, and triggering conditions are matched based on event type, event level, site identifier, equipment type, and confidence score; The process for warning-level events includes generating an event log and sending a notification message to the mobile terminal of the operations and maintenance personnel; The process for warning-level events has been enhanced with a video review step. The video analysis component is used to detect the presence of visual features such as open flames or smoke. If fire features are confirmed, the event level is upgraded to the danger level. The process for hazardous events includes generating linkage control commands, which are then sent to on-site execution equipment via the interface of the fire alarm linkage controller. The message push component encapsulates fire incident data into push messages in a standard format and sends them to the upper-level management platform through a message queue middleware.

9. A fire data integration method based on distributed heterogeneous protocol adaptation according to claim 1, characterized in that, In step S5, automatic failover and device connection migration are performed through comprehensive health assessment and dynamic load balancing strategies, including: The gateway node establishes a heartbeat connection with the central management service and periodically reports the node's health status data; The central management service conducts a comprehensive health assessment of each node, compares each indicator with the preset normal range, assigns different weights to each indicator according to its impact on system availability, and performs a weighted summation to obtain the comprehensive health score of the node. When the health score of a node is lower than the preset sub-health threshold, the load migration strategy is triggered; when the heartbeat of a node times out or the health score drops to zero, the failover strategy is triggered. When migrating device connections, the target node loads the corresponding plugin from the protocol adaptation plugin library and establishes a new communication connection with the device. After the new connection is successfully established, the device connection table is updated.

10. A fire protection data integration system based on distributed heterogeneous protocol adaptation, used to execute the steps in the fire protection data integration method based on distributed heterogeneous protocol adaptation as described in any one of claims 1-9, characterized in that, include: The protocol adaptation module is used to complete the adaptive identification of protocols and the loading of plugins through the collaborative mechanism of the protocol template library and the plugin library, so as to realize the collection and standardized conversion of multi-source fire protection data and obtain structured equipment status data and raw event data. The spatiotemporal feature extraction module is used to receive structured equipment status data and raw event data, and adopts sensor spatial topology modeling and multi-scale temporal feature collaborative extraction to obtain enhanced fire event feature data containing spatiotemporal context information. The intelligent assessment module is used to extract and analyze the enhanced fire incident feature data through a multi-scale temporal convolutional network and a spatiotemporal graph attention network, combining temporal and spatial dual-path feature extraction and fusion analysis. Combined with physical law verification and confidence assessment, the intelligent assessment results of the fire incident are obtained. The linkage control module is used to obtain the intelligent analysis results of fire incidents, execute hierarchical linkage control and multi-platform message push through the event-driven rule process engine, and output the execution status of linkage control commands and the delivery confirmation of event push messages. The high availability assurance module is used to automatically transfer faults and migrate device connections based on the operating status of each gateway node during the fire data collection process, through comprehensive health assessment and dynamic load balancing strategies, and output the updated gateway cluster topology and load distribution status.

Images