Intelligent modular prefabricated cabin type tunnel fire station terminal dynamic early warning system based on edge computing and multi-source perception

Abstract

The application relates to the technical field of intelligent fire station, in particular to an intelligent modular prefabricated cabin type tunnel fire station terminal dynamic early warning system based on edge computing and multi-source sensing. The system comprises a multi-source data acquisition unit, an edge computing processing unit, an intelligent early warning analysis unit and a data management unit. The edge computing processing unit is used for real-time fusion and dynamic modeling of the collected data, multi-source data are fused based on Kalman filtering and LSTM, and a risk evolution model with fast and slow dynamic components is constructed by adopting singular perturbation theory. The improved long short-term memory network adopts a parallel GRU branch to construct a multi-scale feature extraction architecture: a short-term branch captures second-level fast features such as flame flickering and smoke concentration sudden change, a long-term branch captures minute-level slow change trends such as temperature gradient rise, and an adaptive attention mechanism is combined to give dynamic weights to different time scale features. The design improves the pertinence and accuracy of feature extraction in complex scenes.

Description

Technical Field

[0001] This invention relates to the field of intelligent fire station technology, and more specifically, to a dynamic early warning system for intelligent modular prefabricated tunnel fire station terminals based on edge computing and multi-source sensing. Background Technology

[0002] Under the policy guidance of intelligent transportation and digital transformation of tunnel safety management, the traditional fire early warning mode for highway tunnels faces significant technical bottlenecks. Traditional building-type substations and single-sensor monitoring solutions suffer from pain points such as limited site selection, high construction costs, and low efficiency of manual operation and maintenance, especially in complex mountainous terrain where it is difficult to achieve real-time response to emergencies such as fires.

[0003] For example, Chinese patent CN202023079185.X discloses a tunnel fire monitoring system, including a server, an integrated monitoring and management module, and a cloud data platform; a front-line device end, which is signal-connected to the monitoring end, includes a high-definition real-time camera, a water level gauge, a wireless smart pressure water meter, a wireless pressure transmitter, a fire hydrant monitoring device, a wireless liquid level transmitter, a water pump monitoring device, and a GPS positioning module. The water level gauge is paired with the high-definition real-time camera. This utility model provides a tunnel fire monitoring system that achieves accurate, real-time, multi-point concurrent monitoring of the water pressure of the fire water source network. Through dynamic analysis of the water pressure value, it ensures that the water level of the fire water tank and fire pool is within the normal range, ensuring the smooth operation of the fire pipeline network system. When the water level or pipeline system is abnormal, it can quickly issue alarm information and promptly investigate potential fire water source hazards. For example, Chinese patent CN202510273256.8 discloses a fire equipment management method and system based on a railway tunnel fire monitoring system. This includes a fire monitoring module for real-time acquisition of multi-dimensional status data of fire equipment; a redundant data processing module employing a hot standby redundancy mechanism for real-time analysis and processing of the multi-dimensional status data of fire equipment, generating a fire risk assessment report; and an emergency dispatch plan based on the fire risk assessment report, generating execution instructions and sending them to the execution module; an execution module for controlling the operation of fire equipment according to the execution instructions; and a data display module for real-time display of the multi-dimensional status data of fire equipment, the fire risk assessment report, and the emergency dispatch plan. This invention can detect potential safety hazards earlier, achieving early warning of fire risks, significantly improving the accuracy and timeliness of fire warnings, thereby gaining valuable time for taking effective preventative measures.

[0004] While the aforementioned technical solutions each possess their own design advantages, they also suffer from the following shortcomings: First, a lack of hierarchical response architecture: Neither Chinese patent CN202023079185.X (pure cloud architecture) nor Chinese patent CN202510273256.8 (centralized server architecture) constructs a layered collaborative mechanism of "edge local emergency decision-making + cloud global analysis." Tunnel fires require millisecond-level response, but the logic of cloud backhaul delays and passive edge forwarding cannot prioritize triggering local alarms in the early stages of a fire, nor can it reconcile the conflict between real-time early warning and global data synchronization, leading to an imbalance between alarm timeliness and decision rationality in emergency situations. Second, insufficient depth of data fusion: Chinese patent CN202023079185.X focuses on the single dimension of fire-fighting water pressure, while Chinese patent CN202510273256.8 extends to the multi-dimensional status of fire-fighting equipment, but both remain at the basic level of "data collection-summarization," failing to construct a multimodal correlation model of "temperature + smoke + equipment current + environmental dust." Due to the lack of multi-source data spatiotemporal fusion and dynamic simulation capabilities, the coupling characteristics of slow temperature rise and rapid smoke diffusion in the early stages of a fire cannot be distinguished, resulting in a high false alarm rate under complex conditions and limiting the accuracy of early warning. Therefore, we propose an intelligent modular prefabricated tunnel fire station terminal dynamic early warning system based on edge computing and multi-source sensing. Summary of the Invention

[0005] The purpose of this invention is to provide a dynamic early warning system for intelligent modular prefabricated tunnel fire station terminals based on edge computing and multi-source sensing, so as to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, the present invention aims to provide a dynamic early warning system for intelligent modular prefabricated tunnel fire station terminals based on edge computing and multi-source sensing, comprising:

[0007] A multi-source data acquisition unit is used to collect environmental parameters, equipment status, and personnel and material information in the fire protection area in real time, and to achieve three-dimensional perception using a heterogeneous sensor network.

[0008] An edge computing processing unit is used to perform real-time fusion and dynamic modeling of the collected data. It fuses multi-source data based on Kalman filtering and LSTM, and uses singular perturbation theory to construct a risk evolution model with fast and slow dynamic components.

[0009] The intelligent early warning analysis unit is used to generate graded early warning signals and link fire-fighting equipment. It determines the early warning level through fuzzy logic algorithm and links fire pumps, perfluorohexanone fire extinguishing devices and emergency lighting devices based on a preset strategy library.

[0010] The data management unit uses blockchain technology to store early warning data and equipment operation records in a chain, and uses a federated learning algorithm to collaboratively optimize the model among edge nodes.

[0011] The human-computer interaction terminal unit displays a risk heat map and escape routes overlaid by an AR device, and supports voice command input and synchronization of multi-site dispatch information.

[0012] As a further improvement to this technical solution, the multi-source data acquisition unit includes an environmental sensing module, an equipment status sensing module, and a personnel sensing module, wherein:

[0013] The environmental sensing module is used to collect temperature and humidity, smoke concentration and combustible gas concentration in the fire-fighting area, and uses infrared thermal imaging sensor and electrochemical gas sensor to realize multi-parameter synchronous sensing.

[0014] The equipment status sensing module is used to monitor the water pressure of fire hydrants and the location of perfluorohexanone fire extinguishing devices, and to obtain status data of fire hydrants and perfluorohexanone fire extinguishing devices through pressure sensors and RFID anti-metal tags.

[0015] The personnel sensing module is used to acquire firefighter location information and uses UWB positioning technology to achieve dynamic tracking in three-dimensional space.

[0016] As a further improvement to this technical solution, the personnel sensing module includes a UWB positioning base station array submodule, a UWB positioning tag submodule, and a data fusion processing submodule, wherein:

[0017] The UWB positioning base station array submodule is deployed at the structural nodes in the fire protection area to form a three-dimensional coordinate grid covering the monitoring area.

[0018] The UWB positioning tag submodule is integrated into the firefighter's helmet and achieves sub-meter positioning accuracy through a two-way time-of-flight ranging algorithm with the UWB positioning base station array submodule.

[0019] The data fusion processing submodule performs multi-sensor fusion processing on the raw UWB data based on a Kalman filter, and compensates for positioning deviations using motion sensor data in non-line-of-sight scenarios.

[0020] As a further improvement to this technical solution, the edge computing processing unit includes a data fusion module, a risk modeling module, and a risk assessment module, wherein:

[0021] The data fusion module performs spatiotemporal alignment and state estimation on real-time data from the environmental perception module, equipment status perception module, and personnel perception module based on the Kalman filter algorithm.

[0022] The risk modeling module uses a Long Short-Term Memory (LSTM) network to construct a multi-source data association model, which is used to extract the temporal features of changes in the state of the fire-fighting area.

[0023] The risk assessment module decomposes the risk evolution process into fast dynamic components (such as smoke diffusion speed) and slow dynamic components (such as temperature rise trend) based on singular perturbation theory, and uses the Lyapunov stability criterion to verify the convergence of the model.

[0024] As a further improvement to this technical solution, the risk modeling module uses an improved Long Short-Term Memory (LSTM) network to construct a multi-source data association model, including the following steps:

[0025] S202.1 Spatiotemporal Feature Fusion Preprocessing:

[0026] Spatiotemporal alignment of multi-source sensor data to construct a tensor ,in For batch size, For time step, For the number of sensors, For feature dimensions; It is the set of real numbers;

[0027] Time series data is generated using a sliding window mechanism, with a window length of [missing information]. satisfy sliding step size satisfy ;

[0028] Z-score normalization is performed on the data from each sensor: ,in and The first Historical data mean and standard deviation of each sensor; For the first The first sensor The standardized values ​​of each sample; For the first The first sensor One original data value;

[0029] S202.2 Multi-scale feature extraction:

[0030] Constructing parallel GRU branches to extract features at different time scales:

[0031] The hidden layer dimension of short-term branches is adapted to the extraction requirements of rapidly changing features within 1-5 seconds;

[0032] The hidden layer dimension of long-term branches is adapted to the extraction requirements of slowly changing features within 30-300 seconds;

[0033] Output feature tensor and ;

[0034] S202.3, Adaptive Attention Fusion:

[0035] Calculate the spatiotemporal attention weight matrix :

[0036] ;

[0037] in, This represents the output feature dimension of the short-run GRU. This represents the output feature dimension of the long-term GRU; , For learnable weight matrix, , It is the bias vector; For the first The first batch, the first Short-term GRU output characteristics at each time step; For the first The first batch, the first Long-term GRU output characteristics at each time step; For the first The first batch, the first Attention weight vectors at each time step;

[0038] Generate weighted feature representation : ; ; For batch indexing, For time step index;

[0039] Fusing features from multiple sensors through a gating mechanism:

[0040]

[0041] ;

[0042] in, , These are learnable parameters; Features of fusion; The gating coefficient; For the multi-scale features of the current time step;

[0043] S202.4 Risk State Prediction and Evolution Modeling:

[0044] Input fusion features LSTM is used to compute the risk state vector. ;

[0045] Specifically, ;

[0046] in, This is the input gate activation value; , Here are the learnable weight matrix and bias of the input gate; Current fusion features Hidden state from the previous moment splicing; It is the sigmoid function;

[0047] ;

[0048] in, The activation value for the forget gate; , Here are the learnable weight matrix and bias for the forget gate;

[0049] ;

[0050] in, This is the output gate activation value; , The learnable weight matrix and bias of the output gate;

[0051] ;

[0052] in, Candidate cell state; , The learnable weight matrix and bias for candidate cells; It is the hyperbolic tangent function;

[0053] ;

[0054] in, This represents the current cell state.

[0055] ;

[0056] in, Currently hidden;

[0057] ;

[0058] Where is the risk state vector; , The learnable weight matrix and bias of the output layer;

[0059] Define a multi-objective loss function :

[0060] ;

[0061] in For mean square error loss, For the average absolute error loss, For attention sparsity constraints, , and 0.5; Indicates the first The true value of the risk status at each time step;

[0062] The model is trained iteratively using the Adam optimizer, and the learning rate is dynamically adjusted using a cosine annealing strategy. :

[0063] ;

[0064] in, This is the current training round; Total number of training rounds; This is the initial learning rate.

[0065] As a further improvement to this technical solution, the intelligent early warning analysis unit includes a fuzzy inference module, a strategy scheduling module, and an execution control module, wherein:

[0066] The fuzzy inference module uses a fuzzy logic algorithm to perform fuzzification transformation, rule matching, and inference on the risk status output by the edge computing processing unit to generate an early warning level.

[0067] The strategy scheduling module calls the preset strategy library based on the early warning level and analyzes the linkage logic of fire pumps, perfluorohexanone fire extinguishing devices and emergency lighting devices.

[0068] The execution control module is used to output timing control signals to drive the corresponding fire-fighting equipment to perform linkage actions according to the analysis results.

[0069] As a further improvement to this technical solution, the fuzzy inference module uses a fuzzy logic algorithm to perform fuzzification transformation, rule matching, and inference on the risk status output by the edge computing processing unit to generate an early warning level, including the following steps:

[0070] S310.1. Convert the multidimensional risk parameters (i.e., raw sensing data, such as temperature, CO concentration, smoke diffusion rate, personnel movement speed, etc.) output by the edge computing processing unit into membership values ​​of preset fuzzy sets respectively; calculate the membership degree of the multidimensional risk parameters to the {low risk, medium risk, high risk} fuzzy sets respectively through membership functions; the membership functions include but are not limited to triangular functions and trapezoidal functions, and the fuzzy set boundaries of each parameter are dynamically adjusted according to the characteristics of the fire scene.

[0071] S310.2. Based on a predefined fuzzy rule library, perform parallel logical matching on the fuzzified parameters to generate rule activation degrees; the fuzzy rule library contains at least fifteen fuzzy rules.

[0072] Furthermore, the fuzzy rules include, but are not limited to: if temperature ∈ high risk and CO concentration ∈ high risk, then an emergency warning is triggered; if temperature ∈ medium risk and smoke diffusion rate ∈ high risk, then a warning is triggered; if personnel movement speed ∈ high risk and any two parameters ∈ medium risk, then a caution warning is triggered.

[0073] S310.3. The conclusions of all activation rules are weighted and synthesized, and the weights are dynamically adjusted based on the accuracy of the rules in historical fire scenarios. The synthesized fuzzy results are converted into specific warning level values ​​by the centroid method and mapped to at least three warning threshold ranges. The warning threshold ranges include safe state (0-30), attention state (31-60), alert state (61-80), and emergency state (81-100).

[0074] S310.4. Based on the real-time status of fire-fighting equipment (such as fire pump pressure and extinguishing agent reserves), the generated warning level is dynamically calibrated; if equipment failure leads to a decrease in linkage capability, the warning level is upgraded by one level (such as the original level three warning is automatically upgraded to level four).

[0075] As a further improvement to this technical solution, the data management unit includes a blockchain evidence storage module and a model collaborative optimization module, wherein:

[0076] The blockchain notarization module is used to perform chain-based notarization of multi-dimensional risk parameters (temperature, CO concentration, smoke diffusion rate) output by the edge computing processing unit and hierarchical early warning instructions generated by the intelligent early warning analysis unit. The chain-based notarization includes a timestamp, edge node ID, and data hash value to ensure that the data cannot be tampered with. The notarization period is synchronized with the early warning data update frequency, and adjacent fire station terminals (such as within a radius of 5km) serve as consensus nodes to verify the validity of the notarization.

[0077] The model collaborative optimization module updates the parameters of the multi-source data association model based on multi-source sensor data (temperature, CO concentration, equipment status); it periodically (e.g., every 24 hours) synchronizes the updated multi-source data association model parameters with adjacent fire station terminals (e.g., within a 5km radius). During the synchronization process, a data desensitization mechanism (masking sensitive location information) is used to ensure the model optimization effect and protect data privacy.

[0078] As a further improvement to this technical solution, the human-computer interaction terminal unit includes an AR scene overlay module, a voice control module, and a multi-site collaboration module, wherein:

[0079] The AR scene overlay module displays risk heat maps and escape guidance information in real time via AR devices;

[0080] The voice control module integrates a microphone array and noise reduction circuit, enabling firefighters to input voice commands in high-noise environments. The voice commands are processed by a local voice recognition engine and converted into equipment control signals or data query requests, and are compatible with fire protection terminology (including equipment names, operation instructions, and area identifiers).

[0081] The multi-site collaboration module synchronizes early warning levels, equipment status, and personnel location data with adjacent fire station terminals through a dedicated communication protocol; and sets up an independent information dashboard on the AR interface to dynamically display the status of adjacent sites with icons (such as emergency early warning sites being identified by flashing red icons), and supports touch-based access to real-time monitoring video streams.

[0082] As a further improvement to this technical solution, the AR scene overlay module includes a risk heatmap submodule and an escape route planning submodule, wherein:

[0083] The risk heatmap submodule is based on the multi-dimensional risk parameters output by the edge computing processing unit, and uses hierarchical color coding (such as red / orange / yellow / green corresponding to four-color warning levels), and matches them with the spatial coordinates of the three-dimensional building model.

[0084] The escape route planning submodule dynamically generates routes based on preset safety exit locations and real-time risk areas, and continuously displays them as highlighted lines in the AR view.

[0085] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0086] 1. This invention uses the Kalman filter algorithm to perform spatiotemporal alignment and state estimation on multi-source data such as environmental perception, equipment status, and personnel activities. This effectively eliminates sensor sampling bias and time sequence misalignment, improves the consistency and reliability of data fusion, provides a high-precision data base for risk assessment, and solves the problems of asynchronous multi-source data and large noise interference in traditional solutions.

[0087] 2. This invention employs an improved Long Short-Term Memory (LSTM) network with parallel GRU branches to construct a multi-scale feature extraction architecture: short-term branches capture rapid features on the order of seconds, such as flame flickering and sudden changes in smoke concentration, while long-term branches capture slow trends on the order of minutes, such as rising temperature gradients. An adaptive attention mechanism is used to assign dynamic weights to features at different time scales. This design can identify the composite features of "fast smoke and slow temperature" in the early stages of a fire, effectively distinguishing false anomalies such as dust interference and equipment aging, thus improving the targeting and accuracy of feature extraction in complex scenarios.

[0088] 3. This invention decomposes the risk evolution process into fast dynamic components such as smoke diffusion rate and slow dynamic components such as temperature rise trend based on singular perturbation theory. It then verifies the model's convergence using the Lyapunov stability criterion, constructing a risk evolution model supported by dynamic system theory. This method overcomes the limitations of traditional threshold judgment, achieving a systematic characterization of the fire risk evolution process. Through multi-objective loss function optimization, it improves the model's generalization ability and assessment reliability under different operating conditions.

[0089] 4. The edge computing unit of this invention achieves end-to-end low-latency response from data acquisition to risk assessment through localized data processing and dynamic modeling, meeting the second-level early warning requirements of fire scenarios. Compared with traditional centralized architectures, this solution can still autonomously complete risk assessment and emergency decision-making based on edge nodes in abnormal scenarios such as network interruptions, solving the response lag problem caused by communication delays. At the same time, the layered collaborative architecture reduces the computing pressure on the cloud, improving the overall robustness and environmental adaptability of the system. Attached Figure Description

[0090] Figure 1 This is a system framework diagram of the present invention;

[0091] The meanings of the labels in the diagram are as follows:

[0092] 100. Multi-source data acquisition unit; 110. Environmental perception module; 120. Equipment status perception module; 130. Personnel perception module; 131. UWB positioning base station array sub-module; 132. UWB positioning tag sub-module; 133. Data fusion processing sub-module;

[0093] 200. Edge computing processing unit; 210. Data fusion module; 220. Risk modeling module; 230. Risk assessment module;

[0094] 300. Intelligent Early Warning Analysis Unit; 310. Fuzzy Inference Module; 320. Strategy Scheduling Module; 330. Execution Control Module;

[0095] 400. Data Management Unit; 410. Blockchain Evidence Storage Module; 420. Model Collaborative Optimization Module;

[0096] 500. Human-computer interaction terminal unit; 510. AR scene overlay module; 511. Risk heat map sub-module; 512. Escape route planning sub-module; 520. Voice control module; 530. Multi-site collaboration module. Detailed Implementation

[0097] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0098] like Figure 1 As shown, this embodiment provides a dynamic early warning system for intelligent modular prefabricated tunnel fire station terminals based on edge computing and multi-source sensing, including:

[0099] Multi-source data acquisition unit 100 is used to collect environmental parameters, equipment status and personnel and material information in the fire protection area in real time, and adopts a heterogeneous sensor network to realize three-dimensional perception.

[0100] In this embodiment, the multi-source data acquisition unit 100 includes an environmental sensing module 110, an equipment status sensing module 120, and a personnel sensing module 130, wherein:

[0101] The environmental sensing module 110 is used to collect temperature and humidity, smoke concentration and combustible gas concentration in the fire-fighting area, and uses an infrared thermal imaging sensor and an electrochemical gas sensor to achieve synchronous sensing of multiple parameters.

[0102] As a further explanation of this embodiment, the environmental sensing module 110 in this embodiment achieves multi-parameter synchronous acquisition through an infrared thermal imaging sensor and an electrochemical gas sensor. The infrared thermal imaging sensor can be the FLIRA315 model, with a temperature measurement range of -20℃ to 150℃, supporting real-time imaging at 30 frames per second, and transmitting data via a USB 3.0 interface. The electrochemical gas sensor can be either Figaro TGS2610 or TGS2620, with detection ranges of 100-10000ppm and 10-1000ppm respectively, paired with a 16-bit ADC (ADS1115) for synchronous sampling at a frequency of 10Hz, and the sampled data is transmitted via the Modbus RTU protocol.

[0103] The equipment status sensing module 120 is used to monitor the water pressure of fire hydrants and the location of perfluorohexanone fire extinguishing devices, and to obtain status data of fire hydrants and perfluorohexanone fire extinguishing devices through pressure sensors and RFID anti-metal tags.

[0104] As a further explanation of this embodiment, the pressure sensor in this embodiment can be the TEConnectivity P23X series, with a range of 0-1.6MPa, which is compatible with the standard working pressure of fire hydrants; the RFID anti-metal tag can use the Impinj M730 chip, with a working frequency of 860-960MHz, which can be attached to the shell of the perfluorohexanone fire extinguishing device and the location code can be read by a UHF RFID reader (ThingMagicMercury6e).

[0105] The personnel perception module 130 is used to acquire firefighter location information and uses UWB positioning technology to achieve dynamic tracking in three-dimensional space.

[0106] In this embodiment, the personnel sensing module 130 includes a UWB positioning base station array submodule 131, a UWB positioning tag submodule 132, and a data fusion processing submodule 133, wherein:

[0107] UWB positioning base station array submodule 131 is deployed in structural nodes (such as ceilings) in the fire-fighting area to form a three-dimensional coordinate grid covering the monitoring area;

[0108] As a further explanation of this embodiment, the Decawave DW1000 chip can be used, and the chips can be deployed on the ceiling at a density of no less than 4 per 100㎡ to form a triangular array. The base station is constructed with the lower left corner of the fire protection area as the origin (0,0,0) to build a three-dimensional coordinate grid. The horizontal distance between adjacent base stations is ≤15m and the vertical height difference is ≤2m. The position can be calibrated by a laser rangefinder.

[0109] The UWB positioning tag submodule 132 is integrated into the firefighter's helmet and achieves sub-meter positioning accuracy through a two-way time-of-flight ranging algorithm with the UWB positioning base station array submodule 131.

[0110] The data fusion processing submodule 133 performs multi-sensor fusion processing on the raw UWB data based on the Kalman filter, and compensates for positioning deviations through motion sensor data in non-line-of-sight scenarios.

[0111] As a further explanation of this embodiment, the data fusion processing submodule 133 in this embodiment uses a two-way time-of-flight (TWR) algorithm for ranging, the formula of which is:

[0112] ;

[0113] in The distance between the tag and the base station, The speed of electromagnetic wave propagation. The timestamp for the request signal sent by the base station. For the timestamp of the response signal received by the base station;

[0114] The ranging data is fused using a Kalman filter, and the state equation is:

[0115] ;

[0116] Wherein, the state vector ,for The 6-dimensional state vector at time t. Represents three-dimensional spatial coordinates (m). Indicates the velocity in the corresponding direction. Represents the transpose of a matrix;

[0117] State transition matrix ( =0.05s is the sampling interval); where, It is a 3×3 identity matrix. It is a 3×3 zero matrix;

[0118] This is the process noise gain matrix; Zero-mean Gaussian white noise;

[0119] The observation equation is: ;

[0120] in, ,for Ranging observations from four base stations at any given time;

[0121] The Jacobian matrix is ​​calculated based on the base station coordinates. The observation noise vector; the observation noise covariance matrix. ;

[0122] In non-line-of-sight scenarios, positioning errors are compensated using acceleration and angular velocity data from motion sensors (such as ICM-20948). The compensation formula is as follows:

[0123] ;

[0124] in, For acceleration, Angular velocity, This refers to the position at the previous moment; This is the position compensation vector; for The acceleration vector at time t.

[0125] Edge computing processing unit 200 is used to perform real-time fusion and dynamic modeling of collected data. It fuses multi-source data based on Kalman filtering and LSTM, and uses singular perturbation theory to construct a risk evolution model with fast and slow dynamic components.

[0126] In this embodiment, the edge computing processing unit 200 includes a data fusion module 210, a risk modeling module 220, and a risk assessment module 230, wherein:

[0127] The data fusion module 210 performs spatiotemporal alignment and state estimation on the real-time data of the environmental perception module 110, the equipment status perception module 120 and the personnel perception module 130 based on the Kalman filter algorithm.

[0128] As a further explanation of this embodiment, the data fusion module 210 in this embodiment is built on the NVIDIA Jetson Orin edge computing platform, integrating an 8-core ARM CPU, a 2048-core Ampere architecture GPU, and configured with 16GB LPDDR5 memory and 256GB NVMe storage. It accesses multi-source sensor data in parallel through Gigabit Ethernet and SPI bus, meeting the computing power requirements for real-time fusion of multi-dimensional data.

[0129] As a further explanation of this embodiment, this embodiment uses the Kalman filter algorithm to achieve multi-source heterogeneous data fusion, and its core logic is as follows:

[0130] Prediction phase: Based on the physical model of the sensor (such as the temperature change over time, the kinematic model of UWB positioning), derive the predicted state value for the next moment (such as the theoretical estimate of environmental parameters, equipment status, and personnel position).

[0131] Update phase: Combine real-time observation data (actual sensor sampling values) to correct prediction bias and output the fused optimal state estimate;

[0132] The core function of the Kalman filter algorithm is to suppress two types of errors, specifically including:

[0133] Inherent noise of sensors (such as drift error of electrochemical gas sensors and measurement noise of pressure sensors).

[0134] Scene interference (such as non-line-of-sight occlusion errors in UWB positioning and environmental radiation interference in infrared thermal imaging) ultimately improves data accuracy and reliability.

[0135] Furthermore, before performing the Kalman filter calculation, the original data needs to undergo third-order preprocessing, specifically including:

[0136] Time synchronization: Timestamps of multi-source data are aligned through a circular buffer, and the sampling period is unified to 100ms (matching the task scheduling frequency of the edge computing unit), which solves the problem of different sampling rates of different sensors (e.g., environmental sensor sampling rate of 10Hz, UWB positioning sampling rate of 50Hz).

[0137] Outlier removal: Outliers are filtered using the 3σ criterion; the historical mean of sensor data is calculated. and standard deviation If the current sample value satisfy (For example, a sudden change in smoke concentration) is considered an outlier and removed to avoid interfering with the fusion results;

[0138] Range normalization: Mapping the output values ​​of different sensors to the [0,1] range (e.g., temperature 0-100℃ corresponds to 0-1, water pressure 0-1.6MPa corresponds to 0-1), which is used to eliminate dimensional differences and ensure that the algorithm processes multi-source data uniformly.

[0139] The risk modeling module 220 uses a long short-term memory network (LSTM) to construct a multi-source data association model to extract the temporal features of changes in the state of fire-fighting areas.

[0140] The risk assessment module 230 decomposes the risk evolution process into fast dynamic components (such as smoke diffusion speed) and slow dynamic components (such as temperature rise trend) based on singular perturbation theory, and uses the Lyapunov stability criterion to verify the convergence of the model.

[0141] As a further explanation of this embodiment, the state equation of the risk evolution model in this embodiment is expressed as follows: ;in For fast dynamic components (such as smoke concentration s, diffusion velocity v). For slow dynamic components (such as ambient temperature T, structural stability) ), =0.01 is a small parameter. The characteristic timescale of the fast dynamic subsystem. =10s, characteristic timescale of slow dynamic subsystem =300s, the system is decomposed into a boundary layer system and a simplified system by using singular perturbation theory and solved separately.

[0142] Furthermore, in fire scenarios, data such as smoke concentration and temperature exhibit nonlinear abrupt changes (e.g., smoke concentration surges by 50% per second during the open flame stage), making traditional LSTM prone to model divergence due to data non-stationarity. This embodiment ensures the robustness of the algorithm through the Lyapunov stability criterion:

[0143] First, the state equation of the improved LSTM is defined as a dynamic change model of fire time series data:

[0144] + ;

[0145] in, for The system state vector at time step contains core variables such as cell state and hidden layer output; The time derivative of the system state;

[0146] Subsequently, to verify the convergence of the improved LSTM, a positive definite energy function was constructed to describe the "bias degree" of the system, and its time derivative was derived:

[0147] Constructing Lyapunov functions ;in It is a positive definite matrix; The state vector is the fast dynamic component; for Transpose of ; calculate its derivative:

[0148] ;

[0149] in, This is the time derivative of the Lyapunov function;

[0150] Derivation of derivatives (chain rule): ;in, Energy pairs are state vectors The partial derivatives (row vectors) of . Including smoke concentration, LSTM cell status, etc.; To improve the state equation of LSTM;

[0151] Then, through constraints Satisfying Lipschitz continuity, prove the system's energy decay (model convergence):

[0152] Lipschitz constant calibration:

[0153] The characteristic rate of change of statistical fire data during abrupt changes (such as a sudden increase in smoke concentration from 5% / second to 50% / second) is used to calculate the maximum Lipschitz constant L. 1.0, with a safety factor of 0.8, constraint L 0.8;

[0154] Stability proof:

[0155] Positive definiteness: because Positive definite, for any non-zero , (The system's energy is bounded);

[0156] Negative qualitative: combined Lipschitz constraints ( ), derived <0 (System energy decays, model converges);

[0157] Finally, the improved LSTM attention fusion residual term directly serves the Lipschitz constraint: by reducing the feature weights of non-fire source regions (such as equipment heat-generating areas), the constraint is limited. The rate of change, ensuring 0.8.

[0158] In this embodiment, the risk modeling module 220 uses an improved Long Short-Term Memory (LSTM) network to construct a multi-source data association model, including the following steps:

[0159] S202.1 Spatiotemporal Feature Fusion Preprocessing:

[0160] Spatiotemporal alignment of multi-source sensor data to construct a tensor ,in For batch size, For time step, For the number of sensors, For feature dimensions; It is the set of real numbers;

[0161] Time series data is generated using a sliding window mechanism, with a window length of [missing information]. satisfy sliding step size satisfy ;

[0162] Z-score normalization is performed on the data from each sensor: ,in and The first Historical data mean and standard deviation of each sensor; For the first The first sensor The standardized values ​​of each sample; For the first The first sensor One original data value;

[0163] S202.2 Multi-scale feature extraction:

[0164] Constructing parallel GRU branches to extract features at different time scales:

[0165] The hidden layer dimension of short-term branches is adapted to the extraction requirements of rapidly changing features within 1-5 seconds;

[0166] The hidden layer dimension of long-term branches is adapted to the extraction requirements of slowly changing features within 30-300 seconds;

[0167] Furthermore, the hidden layer dimension of the short-term GRU branch is selected as 64-96 (preferably 80): this range is sufficient to capture rapid changes within 1-5 seconds (such as flame flickering, sudden changes in smoke concentration), while avoiding computational redundancy caused by excessively high dimensions.

[0168] The hidden layer dimension of the long-term GRU branch is selected as 96-128 (preferably 120): this range can accommodate slow-changing dependencies within 30-300 seconds (such as continuous temperature increase, smoke diffusion trend), ensuring the effective extraction of long-term features.

[0169] The aforementioned dimensional range is determined by the adaptability of "time scale complexity - model dimensional complexity", and the specific value can be adjusted under this logic according to the actual scenario (such as sensor sampling rate, data noise).

[0170] Output feature tensor and ;

[0171] S202.3, Adaptive Attention Fusion:

[0172] Calculate the spatiotemporal attention weight matrix :

[0173] ;

[0174] in, This represents the output feature dimension of the short-run GRU. This represents the output feature dimension of the long-term GRU; , For learnable weight matrix, , It is the bias vector; For the first The first batch, the first Short-term GRU output characteristics at each time step; For the first The first batch, the first Long-term GRU output characteristics at each time step; For the first The first batch, the first Attention weight vectors at each time step;

[0175] Generate weighted feature representation : ; ; For batch indexing, For time step index;

[0176] Fusing features from multiple sensors through a gating mechanism:

[0177]

[0178] ;

[0179] in, , These are learnable parameters; Features of fusion; The gating factor; For the multi-scale features of the current time step;

[0180] S202.4 Risk State Prediction and Evolution Modeling:

[0181] Input fusion features LSTM is used to compute the risk state vector. ;

[0182] Specifically, ;

[0183] in, The input gate activation value; , Here are the learnable weight matrix and bias of the input gate; Current fusion features Hidden state from the previous moment splicing; It is the sigmoid function;

[0184] ;

[0185] in, The activation value for the forget gate; , Here are the learnable weight matrix and bias for the forget gate;

[0186] ;

[0187] in, This is the output gate activation value; , The learnable weight matrix and bias of the output gate;

[0188] ;

[0189] in, Candidate cell state; , For candidate cells, the learnable weight matrix and bias are defined. It is the hyperbolic tangent function;

[0190] ;

[0191] in, This represents the current cell state.

[0192] ;

[0193] in, Currently hidden;

[0194] ;

[0195] Where is the risk state vector; , The learnable weight matrix and bias of the output layer;

[0196] Define a multi-objective loss function :

[0197] ;

[0198] in For mean square error loss, For the average absolute error loss, For attention sparsity constraints, , and 0.5; Indicates the first The true value of the risk status at each time step;

[0199] The model is trained iteratively using the Adam optimizer, and the learning rate is dynamically adjusted using a cosine annealing strategy. :

[0200] ;

[0201] in, This is the current training round; Total number of training rounds; This is the initial learning rate.

[0202] Furthermore, during the model training phase, the Adam optimizer is used to iteratively update the network parameters, with an initial learning rate set to... =0.001. To balance the training convergence speed with the risk of parameter oscillation, this embodiment performs a cosine annealing learning rate adjustment every 5 epochs. The adjustment formula is: ;

[0203] in, This refers to the current training epoch. The preset period threshold is used (T=5 in this embodiment). This setting has been verified using a fire risk dataset, which improves the model convergence speed and avoids gradient oscillations.

[0204] Understandably, the epoch interval (e.g., 3 or 10 epochs) can be adjusted based on the data scale and model complexity (e.g., number of network layers or sample size).

[0205] The intelligent early warning analysis unit 300 is used to generate graded early warning signals and link fire-fighting equipment. It determines the early warning level through fuzzy logic algorithm and links fire pumps, perfluorohexanone fire extinguishing devices and emergency lighting devices based on a preset strategy library.

[0206] In this embodiment, the intelligent early warning analysis unit 300 includes a fuzzy inference module 310, a strategy scheduling module 320, and an execution control module 330, wherein:

[0207] The fuzzy inference module 310 uses a fuzzy logic algorithm to perform fuzzification transformation, rule matching, and inference on the risk status output by the edge computing processing unit 200, and generate an early warning level.

[0208] The strategy scheduling module 320 calls the preset strategy library based on the warning level and analyzes the linkage logic of fire pumps, perfluorohexanone fire extinguishing devices and emergency lighting devices.

[0209] The execution control module 330 is used to output timing control signals to drive the corresponding fire-fighting equipment to perform linkage actions according to the analysis results.

[0210] In this embodiment, the fuzzy inference module 310 uses a fuzzy logic algorithm to perform fuzzification transformation, rule matching, and inference on the risk status output by the edge computing processing unit 200 to generate a warning level, including the following steps:

[0211] S310.1 The multidimensional risk parameters (i.e., raw sensing data, such as temperature, CO concentration, smoke diffusion rate, personnel movement speed, etc.) output by the edge computing processing unit 200 are converted into membership values ​​of preset fuzzy sets respectively; the membership degree of the multidimensional risk parameters to the {low risk, medium risk, high risk} fuzzy sets is calculated respectively through membership functions; the membership functions include but are not limited to triangular functions and trapezoidal functions, and the fuzzy set boundaries of each parameter are dynamically adjusted according to the characteristics of the fire scene.

[0212] S310.2 Based on a predefined fuzzy rule base, perform parallel logical matching on the fuzzified parameters to generate rule activation degrees; the fuzzy rule base contains at least fifteen fuzzy rules;

[0213] Furthermore, the fuzzy rules include, but are not limited to: if temperature ∈ high risk and CO concentration ∈ high risk, then an emergency warning is triggered; if temperature ∈ medium risk and smoke diffusion rate ∈ high risk, then a warning is triggered; if personnel movement speed ∈ high risk and any two parameters ∈ medium risk, then a caution warning is triggered.

[0214] S310.3. The conclusions of all activation rules are weighted and synthesized, with the weights dynamically adjusted based on the accuracy of the rules in historical fire scenarios. The synthesized fuzzy results are converted into specific warning level values ​​using the centroid method and mapped to at least three warning threshold ranges. The warning threshold ranges include safe state (0-30), attention state (31-60), alert state (61-80), and emergency state (81-100).

[0215] S310.4. Based on the real-time status of fire-fighting equipment (such as fire pump pressure and extinguishing agent storage), the generated warning level is dynamically calibrated, with a calibration range of 5%-15% (preferably 10%). If equipment failure leads to a decrease in linkage capability, the warning level is upgraded by one level (e.g., the original level three warning is automatically upgraded to level four).

[0216] As a further explanation of this embodiment, this embodiment addresses temperature ( ), CO concentration ( ), smoke diffusion rate ( ), personnel movement speed ( The triangular membership function is used to map to {low, medium, and high risk}, and its calculation formula is as follows:

[0217] ;

[0218] in, Real-time values ​​from the sensor (such as temperature and concentration); Vertices of a fuzzy set; The half-width of the membership function;

[0219] Furthermore, the parameter calibration is based on NFPA 72 standards and historical fire data, and is updated every 24 hours. , .

[0220] As a further explanation of this embodiment, the fuzzy inference module 310 in this embodiment uses the centroid method to complete the defuzzification operation. In this embodiment, the core function of the centroid method is to transform the qualitative conclusions (low, medium, and high risk) output by the fuzzy rules into quantitative warning values ​​of 0-100, providing the strategy scheduling module with numerical basis for directly matching linkage strategies. In specific implementation:

[0221] Fuzzy interval discretization: For the four-level warning intervals of "Safety (0-30), Attention (31-60), Alertness (61-80), Emergency (81-100)", each interval is pre-discretized into 10 sampling points (e.g., 81, 85...100 for the Emergency interval). The membership degree of each sampling point is approximately calculated by the trapezoidal membership function (simplifying engineering calculations and keeping the error within an acceptable range).

[0222] Dynamic calibration correlation: The warning value output by the center of gravity method will be corrected in conjunction with the real-time status of fire-fighting equipment (such as fire pump pressure and extinguishing agent reserves).

[0223] If the equipment capacity decreases (e.g., pump pressure < 60% of rated value), the center of gravity value should be increased by 5%-15% (preferably 10%).

[0224] If equipment malfunctions (such as a pump without pressure signal), the warning level will be directly upgraded (e.g., the "attention level" will automatically jump to the "alert level") to ensure that the warning decision and execution capability are matched.

[0225] This embodiment uses the center-of-gravity method to transform the qualitative conclusions of fuzzy reasoning into quantitative instructions that can directly drive equipment linkage. At the same time, combined with dynamic correction of equipment status, it ensures the practicality and reliability of the early warning strategy.

[0226] The data management unit 400 uses blockchain technology to store early warning data and equipment operation records in a chain, and uses a federated learning algorithm to collaboratively optimize the model among edge nodes.

[0227] In this embodiment, the data management unit 400 includes a blockchain evidence storage module 410 and a model collaborative optimization module 420, wherein:

[0228] The blockchain evidence storage module 410 is used to perform chain-based evidence storage of the multi-dimensional risk parameters (temperature, CO concentration, smoke diffusion rate) output by the edge computing processing unit 200 and the hierarchical early warning instructions generated by the intelligent early warning analysis unit 300. The chain-based evidence storage includes timestamps, edge node IDs, and data hash values ​​to ensure that the data cannot be tampered with. The evidence storage period is synchronized with the early warning data update frequency, and adjacent fire station terminals (such as within a radius of 5km) serve as consensus nodes to verify the validity of the evidence storage.

[0229] As a further explanation of this embodiment, the block in this embodiment consists of a block header (128 bytes) and a block body. The block header includes a version number, block hash value, Merkle root, UTC timestamp, consensus node ID, and random number. The block body encapsulates a list of transactions, and each transaction includes a data hash value, edge node ID (such as "FX-001"), data type (risk parameters / warning instructions), and original data length.

[0230] Furthermore, the specific process of blockchain evidence storage module 410 performing chain-based evidence storage in this embodiment is as follows:

[0231] First, the multi-dimensional risk parameters and early warning instructions output by the edge computing processing unit 200 are used to generate JSON data;

[0232] Subsequently, the data hash value is calculated using the SHA-256 algorithm and packaged into a transaction along with the node ID and timestamp;

[0233] Next, the local node collects transactions within 10 seconds to generate Merkle root, and performs PBFT consensus with neighboring nodes (at least 3 nodes confirm).

[0234] Finally, once consensus is reached, a new block is encapsulated and linked to the blockchain, returning the evidence hash for subsequent queries;

[0235] Furthermore, in this embodiment, the verification mechanism for adjacent nodes is as follows: taking the current fire station as the center, other fire stations within a radius of 5km are selected as consensus nodes through GPS coordinates, and the number of nodes must meet the fault tolerance requirements of the PBFT algorithm.

[0236] It should be added that the blockchain evidence storage module 410 in this embodiment also adopts a dual-channel architecture, specifically including:

[0237] Real-time decision-making channel: Directly accesses sensor data buffers through memory mapping technology to ensure early warning analysis latency ≤800ms;

[0238] Asynchronous evidence storage channel: Every 10 seconds, incremental data snapshots (including the previous hash value) are stored through the RAFT consensus engine;

[0239] The model collaborative optimization module 420 updates the parameters of the multi-source data association model based on multi-source sensor data (temperature, CO concentration, equipment status); it periodically (e.g., every 24 hours) synchronizes the updated multi-source data association model parameters with adjacent fire station terminals (e.g., within a radius of 5km). During the synchronization process, a data desensitization mechanism (masking sensitive location information) is used to ensure the model optimization effect and protect data privacy.

[0240] As a further explanation of this embodiment, the updated multi-source data association model parameters are synchronized with the terminals of adjacent fire stations. The collaborative process is as follows:

[0241] First, each edge node constructs a training set based on multi-source data (such as temperature, CO concentration, and equipment status) collected by local sensors, and updates the model parameters using gradient descent. Through formula Calculate the parameters for the current iteration; where For learning rate, The gradient calculated for local data;

[0242] Then, the gradients calculated locally are... Add Laplacian noise to achieve differential privacy protection (avoid inferring sensitive information from gradient data);

[0243] Next, the encrypted gradient data is transmitted to the regional coordination node (such as the central fire station server) via a TLS 1.3 encrypted channel.

[0244] Finally, the coordinating node performs weighted aggregation of the received gradient parameters according to the proportion of effective data volume of each participating node, using the formula... Generate global model parameters; where The number of nodes participating in the collaboration, and ≥3; For the first The weight of each node's data volume is determined by the proportion of valid samples contributed by each node.

[0245] The human-computer interaction terminal unit 500 displays risk heat maps and escape routes overlaid by AR devices, and supports voice command input and synchronization of multi-site dispatch information.

[0246] In this embodiment, the human-computer interaction terminal unit 500 includes an AR scene overlay module 510, a voice control module 520, and a multi-site collaboration module 530, wherein:

[0247] The AR scene overlay module 510 displays risk heat maps and escape guidance information in real time through AR devices (such as AR helmets and smart glasses);

[0248] The 520 voice control module integrates a microphone array and noise reduction circuit, enabling firefighters to input voice commands in high-noise environments. The voice commands are processed by the local voice recognition engine and converted into equipment control signals or data query requests, and are compatible with fire protection terminology (including equipment names, operating instructions, and area identifiers).

[0249] The multi-site collaboration module 530 synchronizes early warning levels, equipment status, and personnel location data with adjacent fire station terminals via a dedicated communication protocol; and sets up an independent information dashboard on the AR interface to dynamically display the status of adjacent sites with icons (such as emergency early warning sites being identified by flashing red icons), and supports touch access to real-time monitoring video streams.

[0250] As a further explanation of this embodiment, this embodiment can use a custom UDP protocol to establish a data synchronization channel with neighboring fire stations within a 5km radius of the current fire station; at the same time, this embodiment can allocate update frequency according to data characteristics, such as:

[0251] Warning level (0-100 quantified value): Updated every second, synchronized with the decision output of the intelligent warning analysis unit 300;

[0252] Equipment status (fire pump pressure, extinguishing agent reserves): updated every 5 seconds to adapt to the dynamic change cycle of equipment status;

[0253] Personnel positioning (UWB 3D coordinates): Updated every 50ms, supporting real-time personnel dispatching needs for rescue operations.

[0254] In this embodiment, the AR scene overlay module 510 includes a risk heatmap submodule 511 and an escape route planning submodule 512, wherein:

[0255] The risk heat map submodule 511 is based on the multi-dimensional risk parameters output by the edge computing processing unit 200, and adopts hierarchical color coding (such as red / orange / yellow / green corresponding to four-color warning levels), and matches them with the spatial coordinates of the three-dimensional building model.

[0256] As a further explanation of this embodiment, this embodiment uses four colors—red, orange, yellow, and green—to directly map the grading results of the intelligent early warning analysis unit 300. The specific grading is as follows:

[0257] Red (RGB:255,0,0) indicates an emergency warning (81-100).

[0258] Orange (RGB:255,165,0) indicates a warning level (61-80).

[0259] Yellow (RGB:255,255,0) indicates a warning (31-60).

[0260] Green (RGB:0,255,0) indicates a safe state (0-30).

[0261] As a further explanation of this embodiment, when the AR scene overlay module 510 executes the AR overlay rendering process, it first aligns the building BIM model with the real rescue scene using AzureSpatialAnchors spatial anchoring technology, establishing a precise mapping relationship between the virtual 3D model coordinates (such as the BIM coordinates of room corners and passage boundaries) and the real physical space (relying on the spatial topology information of the BIM model to ensure that the spatial matching error between the AR overlay risk heat map and the real building structure is ≤1 meter); then, based on the spatial distribution of risk parameters output by the edge computing processing unit 200, it renders four-color heat maps that match the actual size of the building at the physical coordinate positions corresponding to the BIM model (for example, if a room in the building is 5 meters long and 4 meters wide, the heat map is rendered with a spatial range of 5m×4m, so that the risk visualization area is completely aligned with the real spatial boundary); finally, it updates the heat map in real time at a refresh rate of 30 frames / second, and updates it synchronously with the risk data of the edge computing processing unit 200, ensuring that the dynamic risk distribution observed by firefighters in the AR view is consistent with the actual changes in the fire scene.

[0262] The escape route planning submodule 512 is dynamically generated based on the preset safety exit locations and real-time risk areas, and continuously displayed as highlighted lines in the AR view.

[0263] As a further explanation of this embodiment, the escape route planning submodule 512 in this embodiment implements dynamic escape guidance based on the A* path planning algorithm. The specific process is as follows:

[0264] First, complete the preliminary preparations for route planning:

[0265] Before fire fighting, the three-dimensional coordinates of all safety exits are pre-extracted from the building BIM model as the set of endpoints for path planning; at the same time, the three-dimensional position of the firefighters is obtained in real time based on the UWB positioning tag integrated in the firefighters' helmets as the starting point for path planning.

[0266] Subsequently, the A* algorithm and cost function calculation are performed:

[0267] The A* algorithm is used to search for the optimal escape path, and the cost function is defined as follows: ;

[0268] in, Real-time location of firefighters to path nodes The Euclidean distance; The calculation is performed using the spatial coordinate formula, which is: In the formula, To locate the firefighters' coordinates using UWB. Let n be the coordinates of node n;

[0269] For nodes Distance to the nearest emergency exit in Manhattan The heuristic function for simplifying calculations is as follows: In the formula, Provide safe exit coordinates to accelerate path search efficiency;

[0270] Overlapping risk penalties: If a node If the area is in a red or orange alert zone (corresponding to the emergency or alert level of the intelligent early warning analysis unit 300), then... An additional 10 times distance cost is added (forcing the algorithm to prioritize avoiding high-risk areas).

[0271] Next, dynamic updates and path recalculation are triggered:

[0272] Every 100ms, the latest risk data is obtained from the edge computing processing unit 200. If there is a node in the current planned path that triggers the "10 times distance penalty" (i.e. the path is blocked by a high-risk area), the A* algorithm is immediately re-executed to generate a new path that bypasses the high-risk area.

[0273] Finally, AR path visualization is implemented:

[0274] The optimal path is rendered as a bright blue line that matches the building floor or passageway and is displayed in real time via AR devices to ensure that the guidance route is updated in sync with changes in fire risk.

[0275] Through the above design, escape routes can be quickly generated and updated when the risks in a fire scene change dynamically, providing firefighters with intuitive and real-time spatial guidance.

[0276] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.

[0277] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A dynamic early warning system for intelligent modular prefabricated tunnel fire station terminals based on edge computing and multi-source sensing, characterized in that, include: A multi-source data acquisition unit (100) is used to collect environmental parameters, equipment status and personnel and material information in the fire protection area in real time, and to realize three-dimensional perception using a heterogeneous sensor network. Edge computing processing unit (200) is used to perform real-time fusion and dynamic modeling of collected data, fuse multi-source data based on Kalman filtering and LSTM, and construct a risk evolution model with fast and slow dynamic components using singular perturbation theory; The edge computing processing unit (200) includes a risk modeling module (220), which uses an improved Long Short-Term Memory (LSTM) network to construct a multi-source data association model, including the following steps: S202.1 Spatiotemporal Feature Fusion Preprocessing; S202.2 Multi-scale feature extraction: Constructing parallel GRU branches to extract features at different time scales: The hidden layer dimension of the short-term branch is adapted to the extraction requirements of rapidly changing features within 1–5 seconds. The short-term branch is used to capture second-level rapid fire features, including flame flashing and sudden changes in smoke concentration. The hidden layer dimension of the long-term branch is adapted to the extraction requirements of slowly changing features within 30–300 seconds. The long-term branch is used to capture the slow fire change trend at the minute level, which includes the increase of temperature gradient. S202.3, Adaptive Attention Fusion; S202.4 Risk Status Prediction and Evolution Modeling; The edge computing processing unit (200) also includes a risk assessment module (230). The risk assessment module (230) decomposes the risk evolution process into fast dynamic components and slow dynamic components based on the singular perturbation theory. The fast dynamic components are used to characterize the rapid changes in the instantaneous spread of fire in the early stage, including fire evolution components related to smoke diffusion speed and sudden changes in smoke concentration. The slow dynamic components are used to characterize the slow changing trend of the overall fire development, including the fire evolution components related to the increase of temperature gradient and changes in ambient temperature; and the convergence of the model is verified using the Lyapunov stability criterion. The intelligent early warning analysis unit (300) is used to generate graded early warning signals and link fire-fighting equipment. It determines the early warning level through fuzzy logic algorithm and links fire pumps, perfluorohexanone fire extinguishing devices and emergency lighting devices based on a preset strategy library. The data management unit (400) uses blockchain technology to store early warning data and equipment operation records in a chain, and uses a federated learning algorithm to collaboratively optimize the model among edge nodes. The human-computer interaction terminal unit (500) displays risk heat maps and escape routes overlaid by AR devices, supports voice command input and synchronization of multi-site dispatch information, and realizes synchronous interaction of dispatch information and early warning data of adjacent fire stations and other related sites based on the blockchain chain storage mechanism and federated learning collaborative optimization capability of the data management unit (400).

2. The intelligent modular prefabricated tunnel fire station terminal dynamic early warning system based on edge computing and multi-source sensing according to claim 1, characterized in that, The multi-source data acquisition unit (100) includes an environmental sensing module (110), an equipment status sensing module (120), and a personnel sensing module (130), wherein: The environmental sensing module (110) is used to collect the temperature and humidity, smoke concentration and combustible gas concentration in the fire-fighting area, and uses an infrared thermal imaging sensor and an electrochemical gas sensor to realize the synchronous sensing of multiple parameters. The equipment status sensing module (120) is used to monitor the water pressure of fire hydrants and the location of perfluorohexanone fire extinguishing devices, and to obtain status data of fire hydrants and perfluorohexanone fire extinguishing devices through pressure sensors and RFID anti-metal tags. The personnel perception module (130) is used to obtain the firefighter's location information and to achieve three-dimensional spatial dynamic tracking using UWB positioning technology.

3. The intelligent modular prefabricated tunnel fire station terminal dynamic early warning system based on edge computing and multi-source sensing according to claim 2, characterized in that, The personnel sensing module (130) includes a UWB positioning base station array submodule (131), a UWB positioning tag submodule (132), and a data fusion processing submodule (133), wherein: The UWB positioning base station array submodule (131) is deployed at the structural nodes in the fire protection area to form a three-dimensional coordinate grid covering the monitoring area; The UWB positioning tag submodule (132) is integrated into the firefighter's helmet and achieves sub-meter positioning accuracy with the UWB positioning base station array submodule (131) through a two-way time-of-flight ranging algorithm; The data fusion processing submodule (133) performs multi-sensor fusion processing on the raw UWB data based on the Kalman filter, and compensates for positioning deviations through motion sensor data in non-line-of-sight scenarios.

4. The intelligent modular prefabricated tunnel fire station terminal dynamic early warning system based on edge computing and multi-source sensing according to claim 2, characterized in that, The edge computing processing unit (200) further includes a data fusion module (210), wherein: The data fusion module (210) performs spatiotemporal alignment and state estimation on the real-time data of the environment perception module (110), equipment status perception module (120) and personnel perception module (130) based on the Kalman filter algorithm.

5. The intelligent modular prefabricated tunnel fire station terminal dynamic early warning system based on edge computing and multi-source sensing according to claim 4, characterized in that, The spatiotemporal feature fusion preprocessing in S202.1 specifically includes: Spatiotemporal alignment of multi-source sensor data to construct a tensor ,in For batch size, For time step, For the number of sensors, For feature dimensions; It is the set of real numbers; Time series data is generated using a sliding window mechanism, with a window length of [missing information]. satisfy sliding step size satisfy ; Z-score normalization is performed on the data from each sensor: ,in and The first Historical data mean and standard deviation of each sensor; For the first The first sensor The standardized values ​​of each sample; For the first The first sensor One original data value; The multi-scale feature extraction in S202.2 also includes: outputting feature tensors. and ; The adaptive attention fusion in S202.3 specifically includes: Calculate the spatiotemporal attention weight matrix : ； in, This represents the output feature dimension of the short-run GRU. This represents the output feature dimension of the long-term GRU; , For learnable weight matrix, , It is the bias vector; For the first The first batch, the first Short-term GRU output characteristics at each time step; For the first The first batch, the first Long-term GRU output characteristics at each time step; For the first The first batch, the first Attention weight vectors at each time step; Generate weighted feature representation : ; ; For batch indexing, For time step index; Fusing features from multiple sensors through a gating mechanism: ； ； in, , These are learnable parameters; Features of fusion; The gating coefficient; For the multi-scale features of the current time step; The risk state prediction and evolution modeling in S202.4 specifically includes: Input fusion features LSTM is used to compute the risk state vector. ; Define a multi-objective loss function : ； in, For mean square error loss, For the average absolute error loss, For attention sparsity constraints, ; Indicates the first The true value of the risk status at each time step; The model is trained iteratively using the Adam optimizer, and the learning rate is dynamically adjusted using a cosine annealing strategy. : ； in, This is the current training round; Total number of training rounds; This is the initial learning rate.

6. The intelligent modular prefabricated tunnel fire station terminal dynamic early warning system based on edge computing and multi-source sensing according to claim 4, characterized in that, The intelligent early warning analysis unit (300) includes a fuzzy reasoning module (310), a strategy scheduling module (320), and an execution control module (330), wherein: The fuzzy reasoning module (310) uses a fuzzy logic algorithm to perform fuzzification transformation, rule matching and reasoning on the risk status output by the edge computing processing unit (200) to generate a warning level; The strategy scheduling module (320) calls the preset strategy library based on the early warning level and analyzes the linkage logic of the fire pump, perfluorohexanone fire extinguishing device and emergency lighting device. The execution control module (330) is used to output timing control signals to drive the corresponding fire-fighting equipment to perform linkage actions according to the analysis results.

7. The intelligent modular prefabricated tunnel fire station terminal dynamic early warning system based on edge computing and multi-source sensing according to claim 6, characterized in that, The fuzzy inference module (310) uses a fuzzy logic algorithm to perform fuzzification transformation, rule matching, and inference on the risk status output by the edge computing processing unit (200) to generate an early warning level, including the following steps: S310.

1. Convert the multidimensional risk parameters output by the edge computing processing unit (200) into the membership values ​​of preset fuzzy sets respectively; calculate the membership degree of the multidimensional risk parameters to the fuzzy sets {low risk, medium risk, high risk} respectively through the membership function; S310.

2. Based on a predefined fuzzy rule library, perform parallel logical matching on the fuzzified parameters to generate rule activation degrees; the fuzzy rule library contains at least fifteen fuzzy rules. S310.

3. The conclusions of all activation rules are weighted and synthesized, with the weights dynamically adjusted based on the accuracy of the rules in historical fire scenarios; the synthesized fuzzy results are converted into specific warning level values ​​using the centroid method and mapped to at least a three-level warning threshold range. S310.

4. Dynamically calibrate the generated early warning level based on the real-time status of the fire-fighting equipment; If equipment malfunctions cause a decrease in linkage capabilities, the warning level will be raised to level one.

8. The intelligent modular prefabricated tunnel fire station terminal dynamic early warning system based on edge computing and multi-source sensing according to claim 1, characterized in that, The data management unit (400) includes a blockchain evidence storage module (410) and a model collaborative optimization module (420), wherein: The blockchain evidence storage module (410) is used to perform chain-based evidence storage of the multi-dimensional risk parameters output by the edge computing processing unit (200) and the hierarchical early warning instructions generated by the intelligent early warning analysis unit (300); The model collaborative optimization module (420) updates the parameters of the multi-source data association model based on multi-source sensor data; it periodically synchronizes the updated multi-source data association model parameters with the terminals of adjacent fire stations, and a data desensitization mechanism is adopted during the synchronization process.

9. The intelligent modular prefabricated tunnel fire station terminal dynamic early warning system based on edge computing and multi-source sensing according to claim 1, characterized in that, The human-computer interaction terminal unit (500) includes an AR scene overlay module (510), a voice control module (520), and a multi-site collaboration module (530), wherein: The AR scene overlay module (510) displays the risk heat map and escape guidance information in real time through AR devices; The voice control module (520) integrates a microphone array and noise reduction circuit, enabling firefighters to input voice commands in high-noise environments; the voice commands are processed by the local voice recognition engine and converted into equipment control signals or data query requests. The multi-site collaboration module (530) synchronizes the early warning level, equipment status, and personnel location data with the adjacent fire station terminals through a dedicated communication protocol; and sets up an independent information dashboard on the AR interface to dynamically display the status of adjacent sites with icons.

10. The intelligent modular prefabricated tunnel fire station terminal dynamic early warning system based on edge computing and multi-source sensing according to claim 9, characterized in that, The AR scene overlay module (510) includes a risk heatmap submodule (511) and an escape route planning submodule (512), wherein: The risk heat map submodule (511) is based on the multi-dimensional risk parameters output by the edge computing processing unit (200), uses hierarchical color coding, and matches the spatial coordinates of the three-dimensional building model; The escape route planning submodule (512) is dynamically generated based on the preset safety exit location and real-time risk area, and continuously displayed in the AR field of view as a highlighted line.

Images