Rail transit fire early warning system and method based on edge intelligence multi-source heterogeneous data fusion

Abstract

The application provides a rail transit fire warning system and method based on edge intelligent multi-source heterogeneous data fusion, relates to the technical field of rail transit fire fighting and alarm, and edge intelligent computing nodes synchronously collect multi-source heterogeneous sensor data such as fire detection and environment monitoring and the like at the edges of each fireproof partition, perform local fusion reasoning, output local fire probability evaluation results, a station-level fire warning host receives inference results of each edge node, adopts a weighted evidence theory fusion method to comprehensively vote, outputs a station-level fire judgment conclusion and generates a linkage control instruction, a line-level fire situation center collects whole line situations and realizes cross-station coordination, and meanwhile, the edge inference model is incrementally trained, updated and iteratively optimized through a federal learning framework. The application realizes local rapid identification of early fire characteristics and multi-level fusion decision, and effectively improves the accuracy and real-time performance of rail transit fire warning.

Description

Technical Field

[0001] The present invention relates to the technical field of rail transit fire alarm, and specifically relates to a rail transit fire early warning system and method based on edge intelligence multi-source heterogeneous data fusion. Background Art

[0002] As a densely populated public place, the fire safety guarantee of urban rail transit stations is of vital importance. FAS (Fire Automatic Alarm System) is the core component of the rail transit weak current system, responsible for early fire detection, alarm and linkage control. At present, the design and implementation of FAS systems for rail transit stations are mainly based on the national standard GB 50116 "Code for Design of Automatic Fire Alarm Systems" and industry specifications, and the system architecture usually adopts the mode of centralized or distributed detectors plus station-level Fire Alarm Control Panel (FACP).

[0003] In the actual operation of rail transit stations, fire detection faces the following special challenges: (1) Complex station space structure: Rail transit stations usually include multi-layer spaces such as concourse floor, platform floor and equipment floor, and connected passages, ventilation shafts and other areas form an open space structure, which makes the smoke diffusion path difficult to predict, and traditional point-type detectors are prone to false alarms or missing alarms due to the interference of spatial airflow.

[0004] (2) Numerous environmental interference factors: Dust generated by train operation, airflow disturbance of the ventilation system, and temperature and humidity changes caused by a large number of people's activities during peak hours may all generate interference signals to fire detection sensors, leading to a high false alarm rate.

[0005] (3) Isolated multi-system data: In the existing rail transit weak current systems, data of the FAS, BAS and ISCS subsystems are collected and processed independently, lacking an effective cross-system data linkage mechanism. The FAS only relies on its own sensor data to make fire judgments, and cannot use the environmental monitoring data (temperature, humidity, CO₂ concentration, wind pressure, etc.) of the BAS and the real-time passenger flow density data of the ISCS (Integrated Supervision and Control System) to assist judgment, resulting in insufficient confidence of fire early warning.

[0006] Existing rail transit FAS systems based on traditional threshold judgment mainly adopt a fire detection method based on threshold judgment. Devices such as point-type smoke detectors, point-type temperature-sensitive detectors, and line-type temperature-sensitive detectors (such as line-type fixed-temperature cables) are arranged in each fire compartment of the station, and each detector transmits collected physical quantity signals such as smoke concentration and temperature to the station-level Fire Alarm Control Panel (FACP) through hard wires or buses. The FACP makes judgments according to preset threshold rules: when the detector signal exceeds the set threshold, a fire alarm signal is triggered, and fire-fighting equipment is linked and controlled (such as starting exhaust fans, closing fire dampers, turning on fire broadcasts, etc.). Its disadvantages include: (1) Single threshold judgment cannot adapt to dynamic environmental changes: The environmental parameters (temperature, humidity, airflow speed, etc.) of rail transit stations change dynamically with factors such as operating hours, passenger flow density, and ventilation mode. Fixed thresholds are difficult to balance detection sensitivity and false alarm suppression requirements. Environmental noise signals during peak hours may resemble the characteristics of fire signals, causing false alarms; setting higher thresholds during off-peak hours may lead to missed alarms.

[0007] (2) Insufficient information dimensions of single sensor types: Traditional FAS relies on only a single physical quantity such as smoke concentration or temperature for judgment, and cannot distinguish between real fire signals and environmental interference signals. For example, the sudden change in airflow caused when the tunnel ventilation system switches operating modes may be misjudged by smoke detectors as smoke diffusion, when it is not actually a fire event.

[0008] (3) Centralized data processing leads to delayed response: All detector data needs to be uploaded to FACP for centralized processing. The data transmission link is long. When the station is large and there are many detectors, the data processing cycle is extended and the fire alarm response time increases.

[0009] Existing video image-based fire detection systems: In recent years, some rail transit stations have introduced fire detection technology based on video image analysis. This approach deploys cameras in key areas of the station and uses flame recognition algorithms (based on color models, flicker frequency, morphological features, etc.) and smoke recognition algorithms (based on motion detection, transmittance estimation, etc.) to analyze the video stream in real time to determine the presence of flame or smoke spread. This system can be linked with traditional fire detection systems (FAS) as an auxiliary verification method. Its disadvantages include: (1) Sensitive to lighting conditions and obstruction: Some areas of rail transit stations (such as equipment rooms and cable channels) have poor lighting conditions, which limits the effectiveness of video fire detection. During peak hours, dense passenger flow may obstruct the camera's field of view, making it impossible to effectively monitor the target area.

[0010] (2) High computational resource requirements and high deployment costs: Video fire detection requires continuous image acquisition and algorithm inference calculations, which places high demands on computational resources. It usually requires the deployment of dedicated servers or GPU acceleration equipment, and the cost of local deployment at the station is high. If cloud processing is used, it is constrained by network bandwidth and latency.

[0011] (3) Inability to integrate physical sensor data: Existing video fire detection systems operate independently and lack a fusion mechanism with FAS physical sensor data. Video analysis results and sensor detection results are output independently, and cannot form a complementary comprehensive judgment.

[0012] Existing fire detection methods based on simple logic combinations of multiple sensors: Some solutions propose integrating multiple types of sensors, such as smoke, heat, and gas (CO / CO2) sensors, into the fire alarm system (FAS), using simple logical combination rules (e.g., "simultaneous smoke and heat alarms trigger a composite alarm") to improve the accuracy of fire detection. This solution reduces the false alarm rate by increasing the number of sensor types, but still relies on centralized data processing and fixed logical rules for judgment. Its disadvantages include: (1) Simple logic combinations cannot capture deep correlations between sensor data: Fixed logic rules (such as "AND / OR" combinations) can only express simple Boolean relationships between sensor signals and cannot model temporal correlations, spatial correlations and nonlinear coupling relationships between sensor data. For example, a slow rise in temperature accompanied by a slight increase in CO concentration may be an early sign of cable overheating, but simple threshold logic cannot capture this gradual abnormal pattern.

[0013] (2) Lack of cross-system data linkage capability: The scheme is still limited to the range of FAS sensor data and cannot introduce auxiliary information dimensions such as BAS environmental monitoring data and ISCS passenger flow data. The information basis for fire judgment is still incomplete.

[0014] (3) Distributed intelligent processing is not realized: Data is still uploaded and processed in a centralized manner, and edge nodes are only responsible for signal acquisition and forwarding. Edge computing capabilities are not used for local preliminary inference, and the overall response speed and resource utilization efficiency of the system are limited. Summary of the Invention

[0015] In response to the shortcomings of existing technologies, this invention proposes a rail transit fire early warning system and method based on edge intelligent multi-source heterogeneous data fusion. The system architecture adopts a three-level hierarchical structure of "edge node - station host - line center". Edge intelligent computing nodes are deployed in each fire protection zone of the station to realize local fusion reasoning and early warning of multi-source heterogeneous sensor data. The station-level host is responsible for the comprehensive voting and linkage control decision of the reasoning results of multiple nodes. The line-level center is responsible for the fire situation perception and cross-station coordination of the entire line.

[0016] This invention is achieved through the following technical solution: A rail transit fire early warning system based on edge intelligence multi-source heterogeneous data fusion: The system includes edge intelligent computing nodes, station-level fire early warning host, and line-level fire situation center: The edge intelligent computing nodes are deployed in each fire protection zone of the station to collect data in real time and preprocess it. After performing local fusion inference, they output the local fire probability assessment results. The station-level fire early warning host is deployed in the station's fire control room. It is used to receive the local inference results of each edge intelligent computing node, combine them with cross-system auxiliary data for comprehensive voting, generate linkage control decisions, and realize situation visualization. The line-level fire situation center is deployed in the line control center to collect fire judgment data from the fire early warning hosts of each station, construct a fire situation map of the entire line, and realize cross-station coordination and centralized updating of edge models. The edge intelligent computing node is communicatively connected to the station-level fire early warning host, which is communicatively connected to the integrated monitoring system and the environmental and equipment monitoring system, respectively. The line-level fire situation center is communicatively connected to the station-level fire early warning host.

[0017] Furthermore, the edge intelligent computing node includes a multi-source heterogeneous sensor module, an edge AI module, an edge data processing module, and a communication module; The multi-source heterogeneous sensor module is connected to fire detection sensors, environmental monitoring sensors and video monitoring equipment in this zone, and simultaneously collects physical quantity signals and environmental parameter signals; The edge AI module embeds a lightweight fire feature recognition model. The edge data processing module is used to preprocess the raw data; The communication module communicates with the station-level host via industrial Ethernet or fieldbus.

[0018] Furthermore, the station-level fire early warning host includes a multi-node fusion computing module, a cross-system data fusion module, a linkage control decision module, and a situation visualization module; The multi-node fusion computing module uses a weighted evidence theory fusion method to comprehensively vote on the reasoning results of multiple edge nodes and outputs a station-level fire judgment conclusion. The cross-system data fusion module obtains real-time passenger flow density data from the integrated monitoring system and ventilation and air conditioning system operation status data from environmental sensors, which serve as auxiliary information to participate in fire judgment and decision-making. The linkage control decision module generates fire linkage control commands based on the fusion voting results; The situation visualization module displays the fire probability distribution, sensor data trends, and status of linked equipment in real time on the graphic display device.

[0019] Furthermore, the line-level fire situation center includes a full-line situation awareness module, a cross-station coordination module, a central AI computing module, and a fire event database module; The full-line situational awareness module aggregates data from each station to construct a full-line fire situation map. When a fire alarm occurs at a station, the cross-station coordination module automatically triggers the early warning upgrade and ventilation system linkage adjustment of adjacent stations. The central AI computing module is used to optimize and update the edge inference model; The fire event database module records historical fire events and response data and uses them for model training.

[0020] A method for early warning of rail transit fires based on edge intelligence multi-source heterogeneous data fusion; The method specifically includes the following steps: Step 1: Data acquisition and preprocessing of multi-source heterogeneous sensors. The edge intelligent computing nodes of each fire compartment synchronously acquire data from fire detection sensors and environmental monitoring sensors, and preprocess the raw sensor data. Step 2: Edge local multi-source data fusion inference. Input the preprocessed multi-source heterogeneous sensor data sequence into the edge inference model. Use the lightweight fire feature recognition model to perform local fusion inference on the multi-source data, extract early fire feature patterns, and output the local fire probability assessment results. Step 3: Station-level multi-node fusion voting. The station-level fire early warning host receives the local fire probability assessment results reported by the intelligent computing nodes at the edge of each fire protection zone, and at the same time obtains the operating status data of the integrated monitoring system and the environmental and equipment monitoring system. The weighted DS evidence theory fusion method is used for comprehensive voting, and the station-level fire judgment conclusion is output. Step 4: Linkage control and situation reporting. Based on the station-level fire judgment conclusion, generate fire linkage control instructions and report the fire judgment conclusion, key sensor data and linkage equipment status to the line-level fire situation center to update the fire situation of the entire line and trigger early warning coordination and ventilation system adjustment of adjacent stations. Step 5: Online model update and iterative optimization. The edge inference model is incrementally trained and updated using accumulated fire event data, and the updated model parameters are distributed to each edge intelligent computing node to achieve continuous optimization of the edge inference model.

[0021] Furthermore, in step 1, the preprocessing operations include outlier filtering, missing value imputation, timestamp synchronization, and feature normalization.

[0022] Furthermore, in step 2, the lightweight fire feature recognition model adopts a structure that combines a temporal convolutional network with a channel attention mechanism; Temporal convolutional networks expand the receptive field through multi-layer causal convolution and dilated convolution to capture fire feature patterns at different time scales. The channel attention mechanism dynamically adjusts the feature weights based on the contribution of each sensor dimension to the current fire judgment. When passenger flow interference is strong, it automatically reduces the weight of environmental parameter dimensions and increases the weight of smoke and gas dimensions. When the ventilation system mode is switched, it dynamically adjusts the suppression weight of wind pressure dimension.

[0023] Furthermore, in step 3, the local fire probability of each edge node is transformed into a basic probability allocation function, and three identification framework elements are defined: fire, normal, and uncertain. Passenger flow density is obtained through an integrated monitoring system, and the operation mode of the ventilation system is obtained through an environmental and equipment monitoring system. The weight coefficients of the evidence from each edge node are dynamically adjusted. When the passenger flow density is high, the weight of nodes in densely populated areas is reduced and the weight of nodes in equipment areas is increased. When the ventilation system is in a special operation mode, the weight of nodes in areas affected by airflow interference is adjusted. All node evidence is successively fused according to the weighted synthesis rules to obtain the station-level comprehensive fire probability.

[0024] Furthermore, the station-level fire judgment conclusion in step 3 includes two levels: alarm status and early warning status. When the overall fire probability reaches the alarm threshold, a fire alarm is triggered. When the overall fire probability reaches the early warning threshold but does not reach the alarm threshold, a fire early warning is triggered and the system enters an enhanced monitoring state. In step 4, when a fire alarm is triggered, the linkage control decision module generates fire linkage control commands, including starting and stopping the smoke exhaust fan, switching the fire damper, triggering the fire broadcast, and releasing the gate.

[0025] Furthermore, in step 5, the incremental training and update of the model adopts a federated learning framework, where gradient calculation is performed locally on each edge intelligent computing node. Only the updated model parameters are uploaded to the line-level fire situation center for aggregation, avoiding the uploading of raw sensitive data. The updated model parameters are then verified and distributed to each edge node.

[0026] Compared to existing fire detection methods using FAS (Fire Alarm System) in rail transit, the advantages of this invention are at least as follows: This invention is specifically applied to data linkage and intelligent fire early warning scenarios in urban rail transit stations using FAS (Automatic Fire Alarm System), BAS (Environmental and Equipment Monitoring System), and ISCS (Integrated Monitoring System).

[0027] (1) Fire detection accuracy is significantly improved: Through the multi-source heterogeneous data fusion and channel attention mechanism of the edge inference engine, fire feature identification is carried out by comprehensively utilizing multi-dimensional information such as FAS detector data and BAS environmental data, effectively distinguishing real fire signals from environmental interference signals. Compared with the traditional single threshold judgment scheme, the false alarm rate can be reduced by more than 60%, and the false alarm rate can be reduced by more than 40%.

[0028] (2) Enhanced early warning capability for fire: The dilated causal convolutional structure of the TCN temporal convolutional network can capture the historical data features of 32 time steps and has a sensitive ability to identify early fire gradual change signals (such as cable overheating signs with slow temperature rise accompanied by slight increase in CO concentration). Compared with the traditional threshold scheme, it can issue early warning signals 2-5 minutes earlier, providing more time window for emergency response.

[0029] (3) Improved confidence in fire warning: The weighted DS evidence theory fusion voting method is adopted to integrate the local inference results of multiple EICNs and dynamically adjust the evidence weights in combination with ISCS passenger flow density and BAS ventilation mode, so that the confidence of the fire judgment conclusion is improved from the local assessment of a single EICN to the station-level comprehensive assessment, effectively eliminating the impact of misjudgment of a single node on the system decision.

[0030] (4) Faster system response speed: The edge intelligent computing node completes multi-source data fusion inference locally and reports the fire probability assessment results (rather than the original sensor data) to SFHP, which greatly reduces the amount of data transmission and the central processing load. The edge inference cycle is 100ms, which is 5-10 times faster than the 500ms-1s cycle of data centralized uploading and processing in the traditional solution.

[0031] (5) Enhanced system adaptability: The channel attention mechanism dynamically adjusts the weight coefficients of each sensor dimension based on environmental factors such as passenger flow density and ventilation mode, and the weighted DS evidence theory dynamically adjusts the weights of each EICN evidence based on cross-system data, enabling the system to automatically adapt to changes in the operating environment without the need for manual adjustment of threshold parameters.

[0032] (6) Continuous model optimization capability: The edge inference engine model is updated and iteratively optimized online through the federated learning framework, so that the system can continuously accumulate experience and improve detection accuracy during long-term operation and has the ability to continuously evolve. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the system composition of the present invention; Figure 2 This is a station-level fusion voting reasoning diagram. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] Unless otherwise specified, the experimental methods used in the following examples are conventional methods. Unless otherwise specified, the materials, reagents, methods, and instruments used are all conventional materials, reagents, methods, and instruments in the art, and can be obtained commercially by those skilled in the art.

[0036] Edge Intelligent Computing Node (EICN): This invention defines an embedded computing unit that is deployed in various fire protection zones of the station and has the ability to acquire data from multiple sources of sensors and perform local AI inference. It is the core innovative component of this system.

[0037] Weighted DS Evidence Theory Fusion: Dempster-Shafer evidence theory is an uncertainty reasoning theory that expresses the degree of support of evidence for a proposition through a basic probability allocation (BPA) function and fuses multiple pieces of evidence through a synthesis rule. This invention introduces adaptive weighting coefficients based on cross-system data (passenger flow density, ventilation mode) into the DS synthesis rule, which is called weighted DS evidence fusion.

[0038] Temporal Convolutional Network (TCN): A network structure for processing temporal data based on causal convolution and dilated convolution. It expands the temporal receptive field by progressively increasing the dilation factor layer by layer, and is suitable for capturing multi-scale feature patterns in temporal data.

[0039] Channel attention mechanism: a mechanism that applies dynamic weight adjustment to multi-channel feature maps, adaptively assigning weights based on the contribution of each channel feature to the current task objective, thereby enhancing the expression strength of key information dimensions and suppressing the influence of noise dimensions.

[0040] Federated learning: a distributed machine learning framework in which each edge node independently performs model gradient calculations on local data and only uploads parameter updates to the central node for aggregation, avoiding the uploading of raw sensitive data and protecting data privacy.

[0041] like Figure 1 As shown, this system consists of the following components: (1) Edge Intelligent Computing Node (EICN) Deployed in each fire compartment of the station, each EICN contains: ① Multi-source heterogeneous sensor module: Connects to data from FAS detectors (smoke detectors, temperature detectors, linear temperature detectors, gas detectors, etc.), BAS environmental sensors (temperature and humidity, CO2 concentration, wind pressure and wind speed sensors, etc.) and video monitoring equipment in this zone to achieve synchronous acquisition of physical quantity signals and environmental parameter signals.

[0042] ② Edge AI Module: Embedded with a lightweight fire feature recognition model, it performs local fusion inference on multi-source heterogeneous sensor data based on temporal convolutional network (TCN) and attention mechanism, extracts early fire feature patterns, and outputs local fire probability assessment results.

[0043] ③ Edge data processing module: Performs preprocessing operations such as outlier filtering, missing value filling, and timestamp synchronization on the raw sensor data to ensure the quality of the data input to the edge inference engine.

[0044] ④ Communication module: It communicates with the station-level host via industrial Ethernet or fieldbus to transmit local inference results and preprocessed sensor data.

[0045] (2) Station-level Fire Warning Host (SFHP) Deployed in the station's fire control room, each SFHP contains: ① Multi-node fusion calculation module: Receives the local fire probability assessment results of each fire compartment EICN, uses the weighted DS evidence theory fusion method to comprehensively vote on the inference results of multiple EICNs, and outputs the station-level fire judgment conclusion.

[0046] ② Cross-system data fusion module: Obtains real-time passenger flow density data from the ISCS and ventilation and air conditioning system operation status data from the BAS, using this as auxiliary information to participate in fire judgment and decision-making. When passenger flow density is high, it automatically lowers the alarm threshold to enhance early detection sensitivity; when the ventilation system is in a special operating mode, it adjusts environmental interference suppression parameters.

[0047] ③ Linkage Control Decision Module: Generates fire linkage control commands based on the fusion voting results, including starting and stopping smoke exhaust fans, switching fire dampers, triggering fire broadcasts, and releasing turnstiles.

[0048] ④ Situation visualization module: Displays real-time information such as fire probability distribution in each zone, sensor data trends, and status of linked equipment on a graphical display device.

[0049] (3) Line-level Fire Situation Center (LFC) Deployed at the Line Control Center (OCC), including: ① Full-line situational awareness module: Summarizes fire judgment conclusions and key sensor data from each station's SFHP to construct a full-line fire situational map.

[0050] ② Cross-station coordination module: When a fire alarm occurs at a station, it automatically triggers the early warning upgrade and ventilation system linkage adjustment of adjacent stations to prevent smoke from spreading to adjacent stations through tunnels.

[0051] ③ Central AI computing module: Optimize the edge computing model and maintain updates to the edge inference computing module.

[0052] ④ Fire Event Database Module: Records historical fire events and response data, and generates AI-generated English data.

[0053] The connection relationships between edge intelligent computing nodes, station-level fire early warning hosts, and line-level fire situation centers in the system are as follows: EICN and SFHP are connected via industrial Ethernet (redundant dual-ring topology) to transmit local inference results and preprocessed data; the multi-source heterogeneous sensor interface modules within EICN are connected to various sensors via hardwired connections or fieldbuses (such as Modbus and CAN bus) to achieve real-time acquisition of physical quantity signals. SFHP and ISCS are connected via an integrated monitoring network interface to acquire passenger flow density and ventilation system operating status data; SFHP and BAS acquire environmental monitoring data via the BAS network interface. SFHP and LFC are connected via a communication transmission network (such as the OTN backbone network) to achieve the aggregation of fire situation data across the entire line and the issuance of cross-station coordination instructions.

[0054] This invention also proposes a fire early warning method based on edge intelligent multi-source heterogeneous data fusion, the specific steps of which are as follows: Step 1: Data Acquisition and Preprocessing from Multi-Source Heterogeneous Sensors EICN's multi-source heterogeneous sensor interface module synchronously acquires FAS detector signals (smoke concentration) from this fire compartment. ,temperature Gas concentration ) and BAS environmental sensor signals (ambient temperature and humidity) CO2 concentration Wind pressure The local data preprocessing module performs the following operations on the raw data: ① Outlier filtering: A sliding window-based statistical method (mean ± 3σ criterion) is used to mark and remove sensor data that exceeds the reasonable range.

[0055] ② Missing value imputation: For data missing due to sensor failure or communication interruption, a method based on linear interpolation or weighted estimation of adjacent partition data is used for imputation.

[0056] ③ Timestamp synchronization: Based on the EICN local clock, timestamps of sensor data with different sampling periods are aligned and unified to a 100ms sampling period.

[0057] ④ Feature standardization: Normalize the sensor data of each dimension to eliminate the differences in the dimensions of different sensors and unify the numerical range of the input data to the [0,1] interval.

[0058] Step 2: Edge Local Multi-Source Data Fusion Inference Preprocessed multi-source heterogeneous sensor data sequence X = { , , , , , The edge inference engine is an input source. It employs a lightweight model structure combining a Temporal Convolutional Network (TCN) with a channel attention mechanism. ① TCN Feature Extraction Layer: Multi-layer causal convolution is used to extract features from the input time-series sensor data sequence. Dilated convolution is used to expand the receptive field and capture feature patterns at different time scales. The dilation factor of TCN increases layer by layer in the order of [1, 2, 4, 8]. Four layers of causal convolution can cover historical data of 32 time steps, effectively capturing early fire characteristics (such as a slow rise in temperature accompanied by a slight increase in CO concentration).

[0059] ② Channel Attention Mechanism Layer: Channel attention weights are applied to the multi-channel feature map output by the TCN, and the feature weights are dynamically adjusted according to the contribution of each sensor dimension to the current fire judgment. The formula for calculating the attention weight is:

[0060] in This represents the global average pooling vector for the features of each channel. and For a learnable parameter matrix, For the first i Attention weights for each sensor dimension.

[0061] Through an attention mechanism, the weight of ambient temperature and humidity is automatically reduced and the weight of smoke and gas dimensions is increased when passenger flow interference is strong during peak hours; the suppression weight of wind pressure dimension is dynamically adjusted when the ventilation system mode is switched.

[0062] ③ Local Fire Probability Output Layer: The attention-weighted feature vector is mapped to a local fire probability value through a fully connected layer. ∈[0,1] represents the local confidence level of EICN's assessment of the fire risk in this fire compartment.

[0063] Step 3: Station-level multi-node fusion voting SFHP receives local fire probability reports from the EICN of each fire compartment. ( i =1,2,...,N), and simultaneously obtain the station's passenger flow density from ISCS. and BAS ventilation system operation mode A weighted DS evidence theory fusion method is used for comprehensive voting: ① Constructing the Basic Probability Assignment (BPA): Assigning each EICN's... Transformed into the basic probability assignment function of evidence theory Define three identification frame elements: {Fire F, Normal N, Uncertain U}. The transformation rule is: (F) = ; (N) = 1 - ; (U) = (Uncertain allocation, default) =0.05); ② Adaptive weight allocation: based on ISCS passenger flow density and BAS ventilation mode Dynamically adjust the weighting coefficients of each EICN piece of evidence. .

[0064] The weighting rule is: the higher the passenger flow density ( > The weight of EICN in high-traffic areas such as station halls and platforms decreases (increased environmental interference), while the weight of EICN in equipment areas and cable tunnels increases (less interference, higher reliability of physical sensors). When the ventilation system is in tunnel smoke exhaust mode, the weight of EICN near the smoke exhaust outlet decreases (airflow interference), while the weight of EICN far from the smoke exhaust outlet increases.

[0065] ③ DS Evidence Fusion: The evidence from N EICNs is successively fused according to the weighted DS synthesis rule to obtain the station-level comprehensive fire probability. : (F) = [ (F)· (F) + (F)· (U) + (U)· (F)] / K Where K is the conflict coefficient, K = 1 - [ (F)·m2(N) + (N)· (F)]; After successively fusing all EICN evidence, = (F) ④ Decision output: When ≥ A fire alarm is triggered when the alarm threshold is reached; when ≥ (Warning threshold, < When a fire alarm is triggered, the system enters an enhanced monitoring state.

[0066] Step 4: Joint Control and Situation Reporting ① When ≥ At that time, the SFHP linkage control decision module generates fire linkage control instructions according to the preset linkage logic table, and executes actions such as smoke exhaust, fire damper control, fire broadcast, and gate release.

[0067] ② SFHP reports the fire assessment conclusions, key sensor data, and status of linked equipment to LFC.

[0068] ③ The LFC full-line situational awareness module updates the fire situation map, and the cross-station coordination module automatically triggers the early warning upgrade and ventilation system adjustment of adjacent stations.

[0069] Step 5: Online Model Update and Iterative Optimization ① LFC's fire event database records complete data for each fire warning / alarm event (including raw data sequences from multi-source sensors of each EICN, inference results, final judgment conclusions, and linkage response effects).

[0070] ② Regularly (e.g., monthly) use accumulated fire event data to incrementally train and update the edge inference engine model. Gradient calculations are performed locally in each EICN using a federated learning framework. Only the model parameter update is uploaded to LFC aggregation, avoiding the uploading of raw sensitive data and protecting data privacy.

[0071] ③ The updated model parameters are verified and then distributed to each EICN to achieve continuous iterative optimization of the edge inference engine.

[0072] This invention achieves its goals through the above technical solution: (1) Introduce edge intelligent computing nodes (EICN) in the rail transit FAS to complete the fusion reasoning of multi-source heterogeneous sensor data in each fire protection zone of the station, realize distributed intelligent processing of fire detection, and report the reasoning results instead of the original data to the station host, which greatly reduces the amount of data transmission and the central processing load.

[0073] (2) The edge inference engine adopts a lightweight model that combines temporal convolutional network (TCN) and channel attention mechanism to achieve a 100ms inference cycle on embedded processor. It captures early fire gradual features through dilated causal convolution and dynamically adjusts sensor dimension weights according to environmental factors through channel attention mechanism.

[0074] (3) At the station level, the weighted DS evidence theory fusion voting method is adopted to integrate the local inference results of multiple EICNs and dynamically adjust the evidence weight of each EICN in combination with ISCS passenger flow density and BAS ventilation mode to realize cross-system data linkage to assist fire judgment and decision-making.

[0075] (4) Online incremental updates and iterative optimization of the edge inference engine model are achieved through the federated learning framework. Each EICN performs gradient calculations locally and only uploads the model parameter update amount, thus protecting data privacy while enabling continuous model evolution.

[0076] (5) A three-level hierarchical architecture of “edge node-station host-line center” is adopted to realize the hierarchical processing of fire early warning and cross-station coordination. The edge level is responsible for rapid local reasoning, the station level is responsible for comprehensive voting and linkage decision-making, and the line level is responsible for the overall situational awareness and cross-station coordination.

[0077] A computer device system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described method. A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0078] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the method described above.

[0079] The above provides a detailed description of the rail transit fire early warning system and method based on edge intelligent multi-source heterogeneous data fusion proposed in this invention, and elucidates the principles and implementation methods of this invention. The above description of the embodiments is only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A rail transit fire early warning system based on edge intelligent multi-source heterogeneous data fusion, characterized in that: The system includes edge intelligent computing nodes, station-level fire early warning host, and line-level fire situation center: The edge intelligent computing nodes are deployed in each fire protection zone of the station to collect data in real time and preprocess it. After performing local fusion inference, they output the local fire probability assessment results. The station-level fire early warning host is deployed in the station's fire control room. It is used to receive the local inference results of each edge intelligent computing node, combine them with cross-system auxiliary data for comprehensive voting, generate linkage control decisions, and realize situation visualization. The line-level fire situation center is deployed in the line control center to collect fire judgment data from the fire early warning hosts of each station, construct a fire situation map of the entire line, and realize cross-station coordination and centralized updating of edge models. The edge intelligent computing node is communicatively connected to the station-level fire early warning host, which is communicatively connected to the integrated monitoring system and the environmental and equipment monitoring system, respectively. The line-level fire situation center is communicatively connected to the station-level fire early warning host.

2. The system according to claim 1, characterized in that: The edge intelligent computing node includes a multi-source heterogeneous sensor module, an edge AI module, an edge data processing module, and a communication module; The multi-source heterogeneous sensor module is connected to fire detection sensors, environmental monitoring sensors and video monitoring equipment in this zone, and simultaneously collects physical quantity signals and environmental parameter signals; The edge AI module embeds a lightweight fire feature recognition model. The edge data processing module is used to preprocess the raw data; The communication module communicates with the station-level host via industrial Ethernet or fieldbus.

3. The system according to claim 2, characterized in that: The station-level fire early warning host includes a multi-node fusion computing module, a cross-system data fusion module, a linkage control decision module, and a situation visualization module; The multi-node fusion computing module uses a weighted evidence theory fusion method to comprehensively vote on the reasoning results of multiple edge nodes and outputs a station-level fire judgment conclusion. The cross-system data fusion module obtains real-time passenger flow density data from the integrated monitoring system and ventilation and air conditioning system operation status data from environmental sensors, which serve as auxiliary information to participate in fire judgment and decision-making. The linkage control decision module generates fire linkage control commands based on the fusion voting results; The situation visualization module displays the fire probability distribution, sensor data trends, and status of linked equipment in real time on the graphic display device.

4. The system according to claim 3, characterized in that: The line-level fire situation center includes a full-line situation awareness module, a cross-station coordination module, a central AI computing module, and a fire event database module; The full-line situational awareness module aggregates data from each station to construct a full-line fire situation map. When a fire alarm occurs at a station, the cross-station coordination module automatically triggers the early warning upgrade and ventilation system linkage adjustment of adjacent stations. The central AI computing module is used to optimize and update the edge inference model; The fire event database module records historical fire events and response data and uses them for model training.

5. A method for early warning of rail transit fires based on edge intelligent multi-source heterogeneous data fusion, characterized in that: The method is used to execute the rail transit fire early warning system based on edge intelligent multi-source heterogeneous data fusion as described in any one of claims 1 to 4; The method specifically includes the following steps: Step 1: Data acquisition and preprocessing of multi-source heterogeneous sensors. The edge intelligent computing nodes of each fire compartment synchronously acquire data from fire detection sensors and environmental monitoring sensors, and preprocess the raw sensor data. Step 2: Edge local multi-source data fusion inference. Input the preprocessed multi-source heterogeneous sensor data sequence into the edge inference model. Use the lightweight fire feature recognition model to perform local fusion inference on the multi-source data, extract early fire feature patterns, and output the local fire probability assessment results. Step 3: Station-level multi-node fusion voting. The station-level fire early warning host receives the local fire probability assessment results reported by the intelligent computing nodes at the edge of each fire protection zone, and at the same time obtains the operating status data of the integrated monitoring system and the environmental and equipment monitoring system. The weighted DS evidence theory fusion method is used for comprehensive voting, and the station-level fire judgment conclusion is output. Step 4: Linkage control and situation reporting. Based on the station-level fire judgment conclusion, generate fire linkage control instructions and report the fire judgment conclusion, key sensor data and linkage equipment status to the line-level fire situation center to update the fire situation of the entire line and trigger early warning coordination and ventilation system adjustment of adjacent stations. Step 5: Online model update and iterative optimization. The edge inference model is incrementally trained and updated using accumulated fire event data, and the updated model parameters are distributed to each edge intelligent computing node to achieve continuous optimization of the edge inference model.

6. The method according to claim 5, characterized in that: In step 1, the preprocessing operations include outlier filtering, missing value imputation, timestamp synchronization, and feature normalization.

7. The method according to claim 6, characterized in that: In step 2, the lightweight fire feature recognition model adopts a structure that combines a temporal convolutional network with a channel attention mechanism; Temporal convolutional networks expand the receptive field through multi-layer causal convolution and dilated convolution to capture fire feature patterns at different time scales. The channel attention mechanism dynamically adjusts the feature weights based on the contribution of each sensor dimension to the current fire judgment. When passenger flow interference is strong, it automatically reduces the weight of environmental parameter dimensions and increases the weight of smoke and gas dimensions. When the ventilation system mode is switched, it dynamically adjusts the suppression weight of wind pressure dimension.

8. The method according to claim 7, characterized in that: In step 3, the local fire probability of each edge node is transformed into a basic probability assignment function, and three identification framework elements are defined: fire, normal, and uncertain. Passenger flow density is obtained through an integrated monitoring system, and the operation mode of the ventilation system is obtained through an environmental and equipment monitoring system. The weight coefficients of the evidence from each edge node are dynamically adjusted. When the passenger flow density is high, the weight of nodes in densely populated areas is reduced and the weight of nodes in equipment areas is increased. When the ventilation system is in a special operation mode, the weight of nodes in areas affected by airflow interference is adjusted. All node evidence is successively fused according to the weighted synthesis rules to obtain the station-level comprehensive fire probability.

9. The method according to claim 8, characterized in that: The station-level fire judgment conclusion in step 3 includes two levels: alarm status and early warning status. When the comprehensive fire probability reaches the alarm threshold, a fire alarm is triggered. When the comprehensive fire probability reaches the early warning threshold but does not reach the alarm threshold, a fire early warning is triggered and the enhanced monitoring state is entered. In step 4, when a fire alarm is triggered, the linkage control decision module generates fire linkage control commands, including starting and stopping the smoke exhaust fan, switching the fire damper, triggering the fire broadcast, and releasing the gate.

10. The method according to claim 9, characterized in that: In step 5, the incremental training and update of the model adopts a federated learning framework, where gradient calculation is performed locally on each edge intelligent computing node. Only the updated model parameters are uploaded to the line-level fire situation center for aggregation, avoiding the uploading of raw sensitive data. The updated model parameters are then verified and distributed to each edge node.

Images