A multi-source perception fusion electric vehicle canopy fire intelligent disposal method and system

Abstract

The application relates to the technical field of fire safety, and discloses a multi-source perception fusion electric vehicle shed fire intelligent disposal method and system, wherein the multi-source perception fusion electric vehicle shed fire intelligent disposal method comprises the following steps: performing three-dimensional space structure division on an electric vehicle shed; deploying multi-source sensors based on the three-dimensional space division result and collecting multi-source original sensor data; performing pretreatment; performing multi-modal feature fusion; performing fire detection; performing risk grade discrimination and positioning; reconstructing the three-dimensional space of the electric vehicle shed; simulating fire spread and quantitatively evaluating the risk in different zones; generating an optimal disposal action sequence and issuing an execution instruction; and combining the fire risk grade and the fire source position estimation result updated in real time when the execution instruction is executed to quantitatively evaluate the disposal effect and dynamically adjust; through three-dimensional temperature field reconstruction and fire spread dynamics prediction, full coverage situation awareness of the three-dimensional space of a multi-layer stereoscopic electric vehicle shed is realized.

Description

Technical Field

[0001] This invention relates to the field of fire safety technology, and more specifically, to a method and system for intelligent handling of electric vehicle shed fires using multi-source sensing fusion. Background Technology

[0002] Electric bicycles have become a primary mode of transportation for short-distance travel among urban residents due to their convenience, environmental friendliness, and affordability. With the rapid increase in the number of electric bicycles, centralized charging sheds have been widely constructed in residential communities and commercial complexes. However, these sheds, characterized by concentrated lithium batteries, dense charging equipment, and relatively enclosed spaces, pose an increasingly prominent fire risk. Electric bicycle fires are primarily caused by the thermal runaway of lithium batteries, which burn rapidly, release large amounts of toxic gases, and spread quickly, easily leading to significant casualties and property damage.

[0003] Existing electric vehicle shed fire detection technologies mainly rely on smoke detectors, heat detectors, or flame recognition from video surveillance. These technologies have significant shortcomings: single sensors are easily affected by environmental factors, resulting in false alarms or missed alarms; anomalies can only be detected after a fire has formed (when obvious smoke or flames are produced), making it impossible to provide early warnings in the early stages of lithium battery thermal runaway; and they are designed for two-dimensional planar scenes, making it difficult to meet the three-dimensional spatial monitoring needs of multi-layered, three-dimensional charging sheds.

[0004] Existing fire suppression technologies for electric vehicle sheds typically employ pre-set, fixed procedures, lacking the ability to dynamically generate differentiated response strategies based on the actual fire situation. Activating full-area fire suppression systems in the initial stages of a small-scale fire results in wasted resources, while addressing only the initial fire source area when the fire spreads rapidly leads to untimely response.

[0005] Therefore, there is an urgent need for a fire prevention and control technology for electric vehicle sheds that can achieve ultra-early warning, three-dimensional situational awareness, and intelligent response. Summary of the Invention

[0006] This invention provides a method and system for intelligent handling of electric vehicle shed fires based on multi-source perception fusion, which solves the technical problems of difficulty in early warning of lithium battery thermal runaway, lack of three-dimensional spatial situational awareness, and lack of intelligent handling strategies in related technologies.

[0007] This invention provides an intelligent fire suppression method for electric vehicle sheds based on multi-source sensing fusion, comprising:

[0008] The electric vehicle shed is divided into three-dimensional spatial structures; based on the three-dimensional spatial division results, multi-source sensors are deployed and multi-source raw sensor data are collected to obtain a hierarchical and partitioned multi-source raw data stream;

[0009] Preprocessing of the hierarchical and partitioned multi-source raw data stream yields a spatiotemporally aligned multi-source feature dataset.

[0010] Multimodal feature fusion is performed on a spatiotemporally aligned multi-source feature dataset to obtain a multimodal fused feature vector; fire detection is performed based on the multimodal fused feature vector to obtain fire detection results;

[0011] Based on the fire detection results, risk level is determined and located to obtain the fire risk level and fire source location estimation results;

[0012] The three-dimensional space of the electric vehicle shed is reconstructed based on the spatiotemporally aligned multi-source feature dataset. Combining the reconstruction results, fire risk level and fire source location estimation results, the fire spread is simulated and the risk is quantitatively assessed by region, resulting in a three-dimensional fire situation map and spatial region risk score.

[0013] The optimal sequence of response actions is generated based on the three-dimensional fire situation map and spatial zoning risk score, and execution instructions are issued to obtain execution feedback information.

[0014] Based on the execution feedback information, combined with the real-time updated fire risk level and fire source location estimation results during the execution of instructions, the effectiveness of the response is quantitatively evaluated and dynamically adjusted to obtain the response effectiveness evaluation results and updated strategies.

[0015] In a preferred embodiment, the three-dimensional spatial structural division of the electric vehicle shed includes:

[0016] A three-dimensional mesh spatial partitioning method is adopted to decompose the electric vehicle shed into a regular three-dimensional mesh structure. Each mesh unit corresponds to an independent spatial partition. A triplet coordinate system is used to identify each partition to obtain the spatial partition index.

[0017] The racking system is divided vertically according to the number of rack layers, with each rack layer corresponding to a vertical partition.

[0018] Within each vertical layer, the space is divided into several rectangular areas according to the preset partition size based on the horizontal spatial dimensions, forming a regular grid layout.

[0019] In a preferred embodiment, the deployment of multi-source sensors based on the three-dimensional spatial partitioning results includes:

[0020] Based on the spatial partition index, a partition center deployment strategy is adopted to deploy a combination module of temperature sensor, smoke sensor and gas sensor at the geometric center of each spatial partition to obtain the distribution of partition environmental sensing nodes.

[0021] A layer-top coverage deployment strategy is adopted, infrared thermal imagers are deployed at the top of each layer according to a preset coverage interval to obtain the distribution of layer-level thermal imaging sensing nodes;

[0022] A key point coverage deployment strategy was adopted, and high-definition visible light cameras were deployed at each entrance and exit and at the intersection of the passage to obtain the distribution of visual perception nodes in the passage.

[0023] An active gas sampling enhancement strategy is adopted, in which miniature axial fans are deployed in each spatial partition to form a directional airflow from the periphery of the partition toward the sensor, thereby accelerating the delivery of gas from the partition to the sensor.

[0024] In a preferred embodiment, the preprocessing of the hierarchical, partitioned, multi-source raw data stream includes:

[0025] Based on time-series sensor data from hierarchical and partitioned multi-source raw data streams, Kalman filtering is used to denoise the data and obtain smoothed time-series data.

[0026] Based on the infrared thermal imaging image sequence, a non-uniformity correction and bilateral filtering method are used to obtain the corrected thermal imaging image sequence.

[0027] Based on the visible light video image sequence, an enhanced video image sequence is obtained by using adaptive histogram equalization and median filtering methods.

[0028] Based on the original timestamps of each sensor's data, a linear interpolation time synchronization method is used to obtain a time-aligned multi-source data sequence;

[0029] Based on the sensor installation location information and spatial partition index, a coordinate mapping method is used to map the data of each sensor to a unified three-dimensional spatial coordinate system, resulting in spatially aligned multi-source data.

[0030] In a preferred embodiment, the multimodal feature fusion of the spatiotemporally aligned multi-source feature dataset includes:

[0031] Based on the temporal feature data in the spatiotemporally aligned multi-source feature dataset, a temporal convolutional network is used for feature encoding to obtain temporal modality encoded features.

[0032] Based on infrared thermal imaging feature maps, a two-dimensional convolutional neural network is used for feature encoding to obtain infrared image modality coding features;

[0033] Based on the features of visible light images, a visual Transformer network is used for feature encoding to obtain the modality coding features of visible light images;

[0034] Based on temporal modal coding features, infrared image modal coding features, and visible light image modal coding features, a cross-modal attention mechanism is used to perform feature interaction fusion to obtain a multimodal fusion feature vector;

[0035] Based on the multimodal fusion feature vector, a classification head network is used to output the fire detection confidence score, and a regression head network is used to output the estimated three-dimensional location coordinates of the fire source, thus obtaining the fire detection result.

[0036] In a preferred embodiment, the risk level determination and location based on fire detection results includes:

[0037] A characteristic gas fingerprint library for lithium battery thermal runaway was constructed to obtain standard characteristic gas release patterns;

[0038] Based on gas concentration characteristics, a dynamic time warping algorithm is used to match the feature gas fingerprint database to obtain the similarity score of the thermal runaway stage.

[0039] Based on the principle of gas diffusion dynamics, a gas diffusion time compensation method is used to correct the time lag of gas concentration detection and obtain the compensated thermal runaway stage similarity score.

[0040] Based on the compensated thermal runaway stage similarity score, temperature change rate characteristics, and fire detection results, a multi-threshold cascaded discrimination rule is adopted to obtain the fire risk level.

[0041] Based on the fire risk level and the location of the sensor node that triggered that level, the associated spatial zoning identifier and fire source location estimate are determined.

[0042] In a preferred embodiment, reconstructing the three-dimensional space of the electric vehicle shed based on a spatiotemporally aligned multi-source feature dataset includes:

[0043] Based on temperature sensor data and infrared thermal imaging data, the continuous temperature field distribution in the three-dimensional space of the electric vehicle shed is reconstructed using the Kriging space interpolation method, resulting in a three-dimensional temperature field model.

[0044] Based on the three-dimensional temperature field model and fire source location estimation, the fire spread dynamics simulation method is used to predict the spatial range of fire spread within a preset time window and obtain the fire spread prediction results.

[0045] Based on the parking space distribution information and charging pile status information of the electric vehicle shed, the combustible load calculation method is used to calculate the combustible load of each spatial zone and obtain the combustible load distribution of the spatial zone.

[0046] Based on the location of the fire source, the predicted results of fire spread, and the distribution of combustible load in the spatial zones, a risk quantification assessment model is used to calculate the fire risk quantification score for each spatial zone.

[0047] In a preferred embodiment, generating the optimal response sequence based on the three-dimensional fire situation map and spatial zoning risk score includes:

[0048] Based on the fire protection facility deployment information of the electric vehicle shed, a fire protection facility topology model is constructed to obtain a complete set of executable response actions;

[0049] Based on the three-dimensional fire situation map, spatial zoning risk scoring and fire protection facility topology model, the state space and action space of the reinforcement learning environment are constructed to obtain the Markov decision process model.

[0050] Based on the Markov decision process model, a reward function is designed, a reinforcement learning policy network is trained, and an optimized disposition policy network is obtained.

[0051] Based on the trained response strategy network and the current fire situation, a strategy reasoning method is used to generate a sequence of response actions, thereby obtaining an intelligent response strategy.

[0052] Based on the sequence of actions in the intelligent response strategy, control commands are issued to various fire protection facilities using an Internet of Things (IoT) control protocol.

[0053] In a preferred embodiment, the quantitative evaluation of the treatment effect includes:

[0054] Based on the multi-source sensor data continuously collected during the execution of the response strategy, the fire situation assessment is updated in real time to obtain the fire situation evolution sequence during the response process.

[0055] Based on the fire situation evolution sequence, a quantitative evaluation method for handling effectiveness is adopted to calculate the handling effectiveness evaluation indicators and obtain the handling effectiveness evaluation results. The handling effectiveness evaluation indicators include fire control rate, temperature reduction rate, smoke dissipation rate, and handling completion time.

[0056] In a preferred embodiment, a multi-source sensing fusion-based intelligent fire response system for electric vehicle sheds is used to execute the aforementioned multi-source sensing fusion-based intelligent fire response method for electric vehicle sheds, including:

[0057] The spatial partitioning and sensor deployment module divides the electric vehicle shed into three-dimensional spatial structures; based on the three-dimensional spatial partitioning results, it deploys multi-source sensors and collects multi-source raw sensor data to obtain a hierarchical and partitioned multi-source raw data stream.

[0058] The data processing module preprocesses the hierarchical and partitioned multi-source raw data stream to obtain a spatiotemporally aligned multi-source feature dataset.

[0059] The multimodal fusion detection module performs multimodal feature fusion on a spatiotemporally aligned multi-source feature dataset to obtain a multimodal fusion feature vector; fire detection is then performed based on the multimodal fusion feature vector to obtain the fire detection result.

[0060] The risk assessment and location module assesses and locates the risk level based on fire detection results, and obtains the fire risk level and fire source location estimation results.

[0061] The risk simulation and prediction module reconstructs the three-dimensional space of the electric vehicle shed based on the spatiotemporally aligned multi-source feature dataset. Combining the reconstruction results, fire risk level and fire source location estimation results, it simulates the fire spread and performs risk quantification assessment by region, resulting in a three-dimensional fire situation map and spatial region risk score.

[0062] The intelligent response decision-making module generates the optimal sequence of response actions based on the three-dimensional fire situation map and spatial zoning risk score, issues execution instructions, and obtains execution feedback information;

[0063] The evaluation and optimization module, based on execution feedback information and combined with the real-time updated fire risk level and fire source location estimation results during the execution of instructions, performs quantitative evaluation and dynamic adjustment of the response effectiveness, and obtains the response effectiveness evaluation results and updated strategies.

[0064] The beneficial effects of this invention are as follows: By constructing a multi-source sensor network with a hierarchical and partitioned architecture, and fusing heterogeneous data from multiple sources such as temperature, smoke, characteristic gases, infrared thermal imaging, and visible light video, and employing a cross-modal attention mechanism to achieve deep data fusion, this invention effectively solves the technical problems of insufficient reliability of single sensors and high false alarm and false alarm rates in existing technologies. In particular, through the characteristic gas fingerprint matching technology for lithium battery thermal runaway, accurate early warning can be achieved in the ultra-early stage (internal short circuit induction stage) before obvious smoke and flames are generated by lithium battery thermal runaway. Compared with traditional smoke and temperature detection methods, the warning time is earlier, providing a valuable window of opportunity for fire prevention and control. At the same time, through three-dimensional temperature field reconstruction and fire spread dynamics prediction, full-coverage situational awareness of the three-dimensional space of multi-layered electric vehicle sheds is achieved, solving the technical problem that existing technologies cannot accurately locate the three-dimensional fire source and predict the direction of fire spread.

[0065] This invention employs a reinforcement learning strategy network to generate differentiated intelligent response strategies, dynamically adjusting the sequence of response actions based on the real-time fire situation, thus achieving a technological leap from fixed response procedures to intelligent adaptive response. Through a closed-loop feedback mechanism, the system continuously evaluates the response effectiveness and optimizes the strategy network parameters, enabling it to have self-learning and continuous evolution capabilities, with the response effectiveness continuously improving over time. Attached Figure Description

[0066] Figure 1 This is a flowchart of the main process of a multi-source sensing fusion intelligent fire handling method for electric vehicle sheds in this invention.

[0067] Figure 2 This is a detailed flowchart of a multi-source sensing fusion-based intelligent fire handling method for electric vehicle sheds in this invention;

[0068] Figure 3 This is a module diagram of an intelligent fire response system for electric vehicle sheds based on multi-source sensing fusion, as described in this invention. Detailed Implementation

[0069] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0070] At least one embodiment of the present invention discloses an intelligent fire-fighting method for electric vehicle sheds based on multi-source sensing fusion, such as... Figures 1 to 2 As shown, it includes:

[0071] Step 1: Divide the electric vehicle shed into three-dimensional spatial structures; based on the three-dimensional spatial division results, deploy multi-source sensors and collect multi-source raw sensor data to obtain a hierarchical and partitioned multi-source raw data stream;

[0072] Step 1.1, 3D mesh space generation;

[0073] Based on the physical structure information of the electric vehicle shed, a three-dimensional mesh spatial partitioning method is adopted to decompose the shed into a regular three-dimensional mesh structure. Each mesh unit corresponds to an independent spatial partition, and a triplet coordinate system is used to identify each partition, resulting in a spatial partition index. The specific implementation process of the three-dimensional mesh spatial partitioning is as follows: Detailed dimensional information of the electric vehicle shed is obtained, including length, width, height, and the number and height of the internal rack-type parking structure; vertically, the shed is divided according to the number of rack layers, with each rack layer corresponding to a vertical partition. For example, a three-layer rack structure is divided into three vertical layers; within each vertical layer, the shed is divided into several rectangular areas according to the preset partition sizes based on the horizontal spatial dimensions, forming a regular mesh layout. The principle for determining the partition size is that the coverage area of ​​a single partition should match the effective detection radius of the sensor, while also considering the parking density of electric bicycles and the layout of charging stations.

[0074] Step 1.2, Deployment of zoned environment awareness;

[0075] Based on spatial partitioning indexing, a partition-centric deployment strategy is adopted. Temperature sensors, smoke sensors, and gas sensors are deployed at the geometric center of each spatial partition, resulting in a partitioned environmental sensing node distribution. The temperature sensors utilize high-precision negative temperature coefficient thermistors, covering a measurement range from room temperature to high temperatures, with a response time in the second range, enabling rapid detection of temperature changes. The smoke sensors employ a scattering photoelectric smoke detector, determining smoke concentration by detecting the intensity of infrared light scattering by smoke particles, exhibiting high sensitivity and a low false alarm rate. The gas sensors utilize a multi-channel electrochemical sensor array, capable of simultaneously detecting the concentrations of multiple gases, including carbon monoxide, hydrogen, methane, and volatile organic compounds. Carbon monoxide and hydrogen are typical characteristic gases released in the early stages of lithium battery thermal runaway. All sensors read data through a unified acquisition board with a built-in clock synchronization module, ensuring that all sensor data has a unified timestamp.

[0076] Step 1.3, Hierarchical thermal imaging deployment;

[0077] Based on spatial partitioning indexing, a top-layer coverage deployment strategy is adopted. Infrared thermal imagers are deployed at the top of each layer according to a preset coverage interval, resulting in a hierarchical distribution of thermal imaging sensing nodes. The infrared thermal imagers utilize uncooled focal plane array detectors, whose thermal sensitivity is superior to a preset temperature resolution threshold, enabling them to detect minute temperature differences. The field of view and installation height of the infrared thermal imagers are calculated and determined based on the coverage area requirements, ensuring a certain overlap between the fields of view of adjacent thermal imagers and avoiding detection blind spots. The infrared thermal imagers continuously acquire thermal distribution images at a preset frame rate, with each frame containing the radiation temperature value of each pixel within that field of view.

[0078] Step 1.4, Deployment of channel visual perception;

[0079] Based on the locations of the entrances and main passageways of the electric vehicle shed, a key-point coverage deployment strategy was adopted. High-definition visible light cameras were deployed at each entrance and passageway intersection to obtain the distribution of visual perception nodes in the passageways. The visible light cameras selected were network cameras with a wide dynamic range, supporting clear imaging in low-light environments, and their resolution met the requirements for extracting detailed features of flames and smoke. The cameras continuously acquired video images at a preset frame rate, and after local compression encoding, transmitted them to the edge computing nodes.

[0080] Step 1.5, Directional airflow sampling system;

[0081] Based on the distribution of environmental sensing nodes in different zones and the distribution of hierarchical thermal imaging sensing nodes, an active gas sampling enhancement strategy is adopted. Miniature axial flow fans are deployed in each spatial zone to form a directional airflow sampling system. The miniature axial flow fans are installed on the inlet side of the gas sensor and operate at low speed under normal monitoring conditions, creating a directional airflow from the periphery of the zone towards the sensor, accelerating the transport of gas from the zone to the sensor. When an abnormal increase in gas concentration is detected, the fan speed automatically increases, further accelerating the gas sampling rate. The directional airflow sampling system effectively solves the problem of slow gas diffusion in underground semi-enclosed spaces, which leads to delays in ultra-early warnings.

[0082] Step 1.6, Wireless Network Data Transmission;

[0083] Based on various sensor nodes, a wireless mesh network transmission architecture is adopted to transmit the collected data to edge computing nodes in real time, resulting in a hierarchical and partitioned multi-source raw data stream. Each sensor node is equipped with a low-power wireless communication module and is networked using a mesh network topology, ensuring that the failure of a single node does not affect the overall network communication. The edge computing nodes are deployed inside the electric vehicle shed and are responsible for receiving data from each sensor node, performing preliminary data aggregation and preprocessing, and connecting to the cloud data processing center via a wired network.

[0084] Furthermore, due to potential wireless signal attenuation and multipath effects leading to communication instability in underground electric vehicle sheds, a hybrid wired and wireless transmission architecture can be adopted instead of a purely wireless one. Specifically, for fixed-deployment infrared thermal imagers and visible light cameras, given their large data volume and fixed locations, Power over Ethernet (PoE) technology is used to integrate data transmission and device power supply. For zoned environmental sensing nodes, due to their large number and wide distribution, Low Power Wide Area Network (LPWAN) technology is used for wireless transmission. This hybrid transmission architecture ensures communication reliability while reducing cabling costs and construction complexity.

[0085] Step 2: Preprocess the hierarchical and partitioned multi-source raw data stream to obtain a spatiotemporally aligned multi-source feature dataset;

[0086] Step 2.1, time series data filtering and denoising;

[0087] Based on time-series sensor data from hierarchical, partitioned, multi-source raw data streams, a Kalman filter denoising method is employed to obtain smoothed time-series data. The specific implementation process of Kalman filtering is as follows: a mathematical model of physical quantity changes is established based on historical data to predict the value at the next moment; the predicted value is compared with the actual sensor measurement value, and a weighted average is performed based on the difference between the two and their respective reliability to obtain a corrected estimate; the corrected estimate is used as the basis for the next prediction, forming a recursive processing process. For time-series data such as temperature, smoke concentration, and gas concentration, corresponding state transition models and observation models are established respectively. The state transition model describes the change law of physical quantities over time; the observation model describes the relationship between sensor measurements and actual values. After Kalman filtering, random noise and abnormal spikes in the time-series data are effectively suppressed, while preserving the true trend of physical quantity changes.

[0088] Step 2.2, thermal imaging image correction;

[0089] Based on the infrared thermal imaging image sequence from the hierarchical and partitioned multi-source raw data stream, a non-uniformity correction and bilateral filtering method are used to obtain the corrected thermal imaging image sequence. Non-uniformity correction employs a two-point correction method, specifically implemented as follows: Corrected images are acquired in front of two blackbody radiation sources with known temperatures, one a low-temperature blackbody and the other a high-temperature blackbody; the response value of each pixel at the two temperature points is calculated, and the gain correction coefficient and bias correction coefficient for each pixel are determined based on a linear relationship; during actual imaging, the corresponding correction coefficient is applied to the original output value of each pixel to eliminate response inconsistencies. Bilateral filtering, when calculating the filter weights, considers not only the spatial distance between pixels but also the similarity of pixel values. Pixels with large differences in grayscale values ​​are given smaller weights, thus maintaining the edge sharpness of high-temperature areas while removing image noise, ensuring that the temperature boundaries of the fire source area are not blurred.

[0090] Step 2.3, visible light image enhancement;

[0091] Based on visible light video image sequences from a hierarchical, multi-source raw data stream, an enhanced video image sequence is obtained using adaptive histogram equalization and median filtering. Adaptive histogram equalization divides the image into several sub-regions, calculates the histogram for each sub-region, and performs equalization. Bilinear interpolation eliminates discontinuities at sub-region boundaries, effectively enhancing image contrast in low-light and high-light regions. Median filtering effectively removes salt-and-pepper noise by replacing each pixel value with the median of its neighboring pixels, while preserving image edges relatively well.

[0092] Step 2.4, Multi-source data time synchronization;

[0093] Based on the original timestamps of data from each sensor, a linear interpolation time synchronization method is used to obtain a time-aligned multi-source data sequence. Since the sampling frequencies of different types of sensors vary, the data from all sensors are unified to the same time reference. The time reference frequency is selected as an integer multiple of the greatest common divisor of the sampling frequencies of all sensors. Data with sampling frequencies higher than the reference frequency is downsampled, while data with sampling frequencies lower than the reference frequency is upsampled using linear interpolation. Linear interpolation calculates the estimated value of the intermediate time using the values ​​of two adjacent sampling points; this method is simple to calculate and can better preserve the data's trend.

[0094] Step 2.5, Spatial alignment of multi-source data;

[0095] Based on the sensor installation location information and spatial partition index, a coordinate mapping method is used to map the data from each sensor to a unified three-dimensional spatial coordinate system, resulting in spatially aligned multi-source data. A three-dimensional Cartesian coordinate system is established with a corner point of the electric vehicle shed as the origin. The three-dimensional coordinates of each sensor in this coordinate system are calculated according to its actual installation location. For point sensors, their data are directly associated with the installation location coordinates; for area sensors, image pixels need to be mapped to three-dimensional spatial coordinates based on their field of view parameters. Through spatial alignment, data collected by different sensors can be compared and fused within a unified spatial reference frame.

[0096] Step 2.6, Multidimensional Feature Extraction Processing;

[0097] Based on multi-source data aligned temporally and spatially, a multi-dimensional feature extraction method was employed to obtain a spatiotemporally aligned multi-source feature dataset. Features extracted from temperature data include: instantaneous temperature value, mean of the temperature sliding window, rate of temperature change, and second-order rate of temperature change. Features extracted from smoke data include: instantaneous concentration value, mean of the concentration sliding window, rate of concentration change, and duration of concentration peak. Features extracted from gas data include: instantaneous concentration values ​​of various gases, rate of concentration change of various gases, carbon monoxide to hydrogen concentration ratio, and concentration anomaly index of volatile organic compounds. Features extracted from infrared thermal imaging data include: pixel value of the highest temperature, pixel area of ​​the high-temperature region, centroid coordinates of the high-temperature region, standard deviation of the temperature distribution, and entropy of the temperature distribution. Features extracted from visible light image data include: color histogram features of flame candidate regions, shape descriptors of flame candidate regions, texture features of smoke candidate regions, and optical flow motion features between adjacent frames.

[0098] Furthermore, due to the complex lighting conditions and potential reflective surfaces in underground spaces, flame detection in visible light images is easily affected by changes in illumination and reflective interference. An infrared and visible light image registration and fusion method can be used instead of processing the visible light image separately. The aim is to utilize the temperature information from the infrared image to assist in the accurate segmentation of the flame region in the visible light image. Specifically, based on the field-of-view parameters of the infrared thermal imaging image and the visible light image, an image registration method based on feature point matching is used to calculate the geometric transformation matrix between the two types of images. Based on the geometric transformation matrix, the infrared image is transformed to the pixel coordinate system of the visible light image, achieving pixel-level alignment. Based on the aligned infrared image, a mask of the high-temperature region is extracted as a constraint condition for the flame candidate region. In the visible light image, flame color and texture features are extracted only within the constraint range of the high-temperature region, effectively eliminating interference from non-flame bright areas.

[0099] Step 3: Perform multimodal feature fusion on the spatiotemporally aligned multi-source feature dataset to obtain a multimodal fusion feature vector; perform fire detection based on the multimodal fusion feature vector to obtain the fire detection result;

[0100] Step 3.1, temporal feature encoding processing;

[0101] Based on temporal feature data from a spatiotemporally aligned multi-source feature dataset, a temporal convolutional network (TCNN) is used for feature encoding to obtain temporal modality encoded features. The TCNN ensures that the model output depends only on historical inputs through causal convolution; it expands the receptive field through dilated convolution to capture long-term dependencies, capturing dependencies over a longer time range without increasing the number of parameters. The input to the TCNN is a numerical sequence of multiple temporal feature channels within a time window, such as temperature change sequences or gas concentration change sequences, and the output is a fixed-length temporal encoded vector. The TCNN contains multiple convolutional blocks, each consisting of dilated convolutional layers, batch normalization layers, activation function layers, and residual connections. Dilated convolutional layers are responsible for feature extraction, and the dilation coefficient of dilated convolution increases exponentially; batch normalization layers stabilize the training process and accelerate convergence; activation function layers introduce nonlinear transformation capabilities; and residual connections allow information to be directly passed to subsequent layers, avoiding the gradient vanishing problem.

[0102] Step 3.2, Infrared image feature encoding;

[0103] Based on infrared thermal imaging feature maps from a spatiotemporally aligned multi-source feature dataset, a two-dimensional convolutional neural network (2D convolutional neural network) is used for feature encoding to obtain infrared image modality coding features. The 2D convolutional neural network uses a pre-trained image feature extraction backbone network as the encoder. The input to the 2D convolutional neural network is a stack of multiple frames of infrared thermal imaging images, that is, consecutive frames of infrared images are superimposed in the channel dimension to form a multi-channel input tensor. Image features are extracted progressively through multiple convolutional operations, from low-level edge and texture features to high-level semantic features. A global average pooling layer compresses the 2D convolutional feature map into a one-dimensional vector. Global average pooling calculates the average value of all pixels in each channel of the feature map to obtain the representative value of that channel. The compressed feature vector is then mapped to the target dimension through a fully connected layer to obtain a fixed-length infrared image coding vector containing the key feature information of the infrared image.

[0104] Step 3.3, visible light image encoding;

[0105] Based on visible light image features from a spatiotemporally aligned multi-source feature dataset, a visual Transformer network is used for feature encoding to obtain visible light image modality encoding features. The visual Transformer network divides the input image into fixed-size image patches. Each patch is linearly projected and transformed into an embedding vector, which, after adding positional encoding, is input into a multi-layer Transformer encoder. The Transformer encoder learns the global dependencies between image patches through a self-attention mechanism, and the output class label vector serves as the visible light image encoding vector. The self-attention mechanism works as follows: for each image patch, its similarity score with all other image patches is calculated. The features of all image patches are then weighted and summed based on the similarity scores to obtain a new feature representation that incorporates global information.

[0106] Step 3.4, cross-modal feature fusion;

[0107] Based on temporal modal coding features, infrared image modal coding features, and visible light image modal coding features, a cross-modal attention mechanism is employed for feature interaction and fusion to obtain a multimodal fused feature vector. The cross-modal attention mechanism allows feature vectors from each modality to serve as queries, while feature vectors from other modalities serve as keys and values, achieving cross-modal information interaction through attention weight calculation. Specifically, for a feature vector from modality A, attention is calculated with a feature vector from modality B to obtain a feature representation of modality A enhanced by modality B; similarly, the feature vector from modality B also obtains a new feature representation after enhancement by modality A. Through multiple rounds of cross-modal attention interaction, the features of each modality fully integrate complementary information from other modalities. Finally, the enhanced feature vectors from each modality are concatenated, and nonlinear transformation and dimensionality compression are performed through a multi-layer fully connected network to output a unified multimodal fused feature vector.

[0108] Step 3.5, Fire Detection and Location;

[0109] Based on the multimodal fusion feature vector, a classification head network outputs the fire detection confidence score, and a regression head network outputs the estimated three-dimensional location coordinates of the fire source, thus obtaining the fire detection result. The classification head network consists of a fully connected layer and a sigmoid activation function. The fully connected layer maps the input multimodal fusion feature vector to a scalar value, and the sigmoid activation function compresses this scalar value to the range of 0 to 1. The output value represents the fire detection confidence score, that is, the probability estimate of the existence of a fire. The closer the value is to one, the greater the probability of the fire existing; the closer the value is to zero, the smaller the probability of the fire existing.

[0110] The regression head network consists of multiple fully connected layers, mapping the input multimodal fusion feature vector to three-dimensional coordinates, and outputting an estimated position coordinate of the fire source in a three-dimensional spatial coordinate system, including x, y, and z coordinates. During model training, a loss function needs to be defined to guide the optimization of network parameters. The classification head uses a binary cross-entropy loss function, calculating the loss value by comparing the difference between the predicted probability and the true label; the closer the prediction is to the actual situation, the smaller the loss value. The regression head uses a smoothed L1 loss function. In actual training, the classification loss and regression loss are combined into a total loss function through a weighted summation, and the gradient descent algorithm is used to optimize the total loss function, thereby improving the accuracy of fire detection and fire source localization.

[0111] Furthermore, due to the scarcity of real fire samples, directly training deep fusion networks using a limited number of real samples can easily lead to overfitting. A simulation data augmentation method based on fire physics models can be used for model pre-training. Specifically, fire physics simulation modeling is performed by establishing a physical simulation model of lithium battery thermal runaway and flame combustion based on fire dynamics theory. This model can simulate the temperature field evolution, smoke diffusion process, and gas release process under different ignition locations and fire intensities. Simulated sensor data generation is achieved by generating simulated multi-source sensor data based on the output of the physical simulation model and combined with a sensor noise model. Flame and smoke image rendering is performed by generating simulated flame and smoke images based on the flame parameters of the physical simulation model using a graphics rendering engine. Model pre-training and fine-tuning are then performed by pre-training the fusion network using a large amount of generated simulated data and fine-tuning it using a small amount of real data, effectively improving the model's generalization ability in real-world scenarios.

[0112] Step 4: Based on the fire detection results, determine and locate the risk level to obtain the fire risk level and fire source location estimation results;

[0113] Step 4.1, Construction of the characteristic gas fingerprint library;

[0114] Based on research findings and experimental data on the thermal runaway mechanism of lithium batteries, a characteristic gas fingerprint database for lithium battery thermal runaway was constructed to obtain standard characteristic gas release patterns. The thermal runaway process of lithium batteries can be divided into several stages: the internal short-circuit induction stage, the self-heating acceleration stage, and the thermal runaway outbreak stage. In the internal short-circuit induction stage, the electrolyte begins to decompose, releasing small amounts of carbon monoxide and hydrocarbons. In the self-heating acceleration stage, the positive electrode material begins to decompose, releasing oxygen, while the negative electrode material reacts with the electrolyte to release hydrogen and methane, increasing the gas release rate. In the thermal runaway outbreak stage, a violent chemical reaction causes the battery casing to rupture, releasing a large amount of gas instantaneously, accompanied by combustion. The characteristic gas fingerprint database records information such as the types of characteristic gases, typical concentration ranges, concentration change rate ranges, and concentration ratios between gases at each stage.

[0115] Step 4.2, Gas fingerprint matching analysis;

[0116] Based on the gas concentration characteristics output in step 2, a dynamic time warping algorithm is used to match the data with a feature gas fingerprint database to obtain a similarity score for the thermal runaway stage. The dynamic time warping algorithm can handle the stretching and compression of time-series data along the time axis. In practical applications, even for the same thermal runaway process, the time progression of gas release may differ due to variations in environmental conditions, battery status, and other factors, manifesting as stretching or compression along the time axis. The dynamic time warping algorithm finds the optimal alignment path between two time series using dynamic programming. This path allows for non-linear mapping of the time axis, maximizing the match between the two time series after alignment. Specifically, a two-dimensional distance matrix is ​​constructed, where the rows and columns correspond to the time points of the two time series, and each element represents the distance between the corresponding time point values. The optimal path from the top left corner to the bottom right corner of the matrix is ​​found using dynamic programming, minimizing the cumulative distance while satisfying path continuity and monotonicity constraints. The cumulative distance of the optimal path is then calculated as a similarity metric between the two time series. The real-time detected gas concentration time series is matched with the standard time series of each thermal runaway stage in the fingerprint database, and the similarity score with each stage is calculated. The similarity score is normalized to the range of 0 to 1, where 0 means completely dissimilar and 1 means completely similar. The higher the value, the more similar the current gas release mode is to the thermal runaway stage.

[0117] Step 4.3, gas diffusion time compensation;

[0118] Based on the principles of gas diffusion dynamics, a gas diffusion time compensation method is employed to correct the time lag in gas concentration detection, resulting in a compensated similarity score for the thermal runaway stage. In practical applications, due to the time required for gas to diffuse from the release source to the sensor location, the detected gas concentration change lags behind the actual thermal runaway process. Based on the volume of the spatial partition, ventilation conditions, and parameters of the directional airflow sampling system, a gas diffusion time estimation model is established to calculate the diffusion time from each possible ignition source location to the sensor.

[0119] When performing gas fingerprint matching analysis, the specific implementation method of time compensation is as follows: the time series of gas concentration detected in real time is shifted forward by the corresponding diffusion time. That is, it is assumed that the gas concentration detected at the current moment is actually the gas released several seconds ago. Through this time axis shifting operation, the gas concentration change is synchronized with the actual thermal runaway process in time, thereby improving the accuracy and timeliness of the judgment of the thermal runaway stage.

[0120] Step 4.4, Multi-threshold risk level determination;

[0121] Based on the compensated thermal runaway stage similarity score, the temperature change rate feature output in step 2, and the fire detection result output in step 3, a multi-threshold cascaded discrimination rule is adopted to obtain the fire risk level. The discrimination rule for the multi-threshold cascaded discrimination is designed as follows:

[0122] When the similarity score of the internal short circuit induction stage exceeds the first warning threshold and the temperature change rate exceeds the first temperature rise threshold at the same time, the fire risk level is determined to be the ultra-early warning level. This level indicates that there may be initial signs of thermal runaway of lithium battery, which requires attention but has not yet posed a direct threat. The first warning threshold is set to 0.6; the first temperature rise threshold is set to two degrees Celsius per minute, which means that the temperature rise rate exceeds the normal fluctuation range.

[0123] When the similarity score of the self-heating acceleration stage exceeds the second warning threshold or the smoke concentration exceeds the smoke warning threshold, the fire risk level is determined to be a fire warning level, which indicates that the fire risk has increased and preventive measures need to be taken; the second warning threshold is set to 0.7 and the smoke warning threshold is set to 10mg per cubic meter.

[0124] When the fire detection confidence level exceeds the fire confirmation threshold or the similarity score of the thermal runaway outbreak stage exceeds the third warning threshold, the fire risk level is determined to be the fire confirmation level. This level indicates that a fire has occurred or is about to occur, and emergency response measures need to be initiated immediately. The fire confirmation threshold is set to 0.8, and the third warning threshold is set to 0.85.

[0125] When none of the above conditions are met, the fire risk level is determined to be safe, indicating that the current state is normal and no special action is required.

[0126] Step 4.5: Early warning signal generation and output;

[0127] Based on the fire risk level and the location of the sensor nodes that trigger that level, the associated spatial zoning identifier and fire source location estimate are determined to obtain an ultra-early warning signal or a fire confirmation signal. When the fire risk level is at the ultra-early warning level, an ultra-early warning signal is output, including the warning time, warning level, associated spatial zoning identifier, abnormal gas type and concentration, and estimated thermal runaway stage. When the fire risk level is at the fire confirmation level, a fire confirmation signal is output, including the confirmation time, three-dimensional coordinate estimate of the fire source location, fire detection confidence level, and associated infrared and visible light images.

[0128] Furthermore, since the composition and proportion of gases released during thermal runaway may vary between different brands and types of lithium batteries, a single characteristic gas fingerprint database may not cover all situations. An adaptive fingerprint database update method can be used to improve the adaptability of the matching. Specifically, a site-specific fingerprint database is established by collecting information on common lithium battery types at the site during the deployment of the electric vehicle shed, creating a site-specific initial fingerprint database. The fingerprint database is dynamically updated: during system operation, after a fire event occurs and is handled, the gas release pattern collected during the event is compared with the existing fingerprint database; if differences exist, they are added as new fingerprint samples to the database. Fingerprint patterns are clustered and grouped: a clustering algorithm is used to group the samples in the fingerprint database, with each group representing a typical thermal runaway gas release pattern. During matching, the samples are compared with multiple standard patterns simultaneously, and the highest similarity is taken as the matching result.

[0129] Step 5: Reconstruct the three-dimensional space of the electric vehicle shed based on the spatiotemporally aligned multi-source feature dataset; combine the reconstruction results, fire risk level and fire source location estimation results to simulate the fire spread and perform risk quantification assessment by region, and obtain a three-dimensional fire situation map and spatial region risk score.

[0130] Step 5.1, Reconstruction of the three-dimensional temperature field;

[0131] Based on temperature sensor data and infrared thermal imaging data from a spatiotemporally aligned multi-source feature dataset, a continuous temperature field distribution in the three-dimensional space of an electric vehicle shed is reconstructed using the Kriging spatial interpolation method, resulting in a three-dimensional temperature field model. Kriging interpolation establishes a spatial correlation model by analyzing the spatial correlation between known sampling points. Based on this model, optimal linear unbiased estimation is performed for unsampled locations. The specific implementation process includes the following steps: collecting the temperature values ​​and precise three-dimensional spatial coordinates of each temperature sensor and infrared thermal imaging pixel to form a known sampling point dataset; calculating the variogram of the temperature field, which is established by analyzing the variance of the temperature values ​​at different distance intervals; calculating the temperature estimate for any location in the three-dimensional space according to the variogram model, considering the spatial distance and correlation between the location and all known sampling points, giving greater weight to sampling points that are closer and more correlated; and calculating the estimation variance. The three-dimensional temperature field model is represented in the form of a voxel mesh, that is, dividing the three-dimensional space into regular cubic units, with each voxel storing the temperature estimate and estimation uncertainty at that location.

[0132] Step 5.2, prediction of fire spread dynamics;

[0133] Based on a three-dimensional temperature field model and fire source location estimation, a fire spread dynamics simulation method is used to predict the spatial range of fire spread within a preset time window, yielding the fire spread prediction results. The fire spread dynamics model is based on the heat conduction equation, heat convection equation, and heat radiation equation, which describe the heat conduction process in solid materials, the convection propagation process in the air, and the radiation propagation process via electromagnetic waves, respectively. The input parameters of the fire spread dynamics model include the three-dimensional coordinates of the fire source location, the current temperature field distribution, the geometric structure information of the electric vehicle shed, the distribution location of various combustibles, and their thermal properties. A system of partial differential equations is solved using numerical calculation methods to output the predicted temperature field distribution at each future time. Based on the predicted temperature field distribution, areas with temperatures exceeding the ignition temperature thresholds of various combustibles are marked as potential spread areas, forming the predicted spatial range of fire spread.

[0134] Step 5.3, Calculation of combustible load;

[0135] Based on the parking space distribution information and charging pile status information of the electric bicycle charging shed, a combustible load calculation method is used to calculate the combustible load of each spatial zone, thus obtaining the combustible load distribution of the spatial zones. Combustible load is defined as the total heat released when all combustible materials within a unit area are completely burned, directly reflecting the potential heat release capacity and fire hazard level of a certain area in the event of a fire. A higher combustible load means that once a fire starts in that area, the fire will develop more rapidly, the temperature will be higher, and the spread speed will be faster. For the electric bicycle charging shed scenario, the main combustible materials include the plastic shell, seat cushion, rubber tires, and lithium battery components of the electric bicycles. The calculation process for combustible load is as follows: count the number of electric bicycles parked in each spatial zone, and obtain basic information for each bicycle, including model, weight, battery capacity, and other parameters; calculate the total calorific value of each electric bicycle based on the calorific value data of different materials; sum the calorific values ​​of all electric bicycles in the zone to obtain the total calorific value of that zone; divide the total calorific value by the ground area of ​​the zone to obtain the combustible load value of that zone. During the calculation process, the charging status of electric bicycles needs to be taken into special consideration. Because lithium batteries in the charging state have a higher risk of thermal runaway due to the active internal chemical reaction, they are given a higher weighting factor in the calculation of combustible load. Typically, the risk factor of the battery in the charging state is set to 1.5 to 2 times that of the battery in the non-charging state.

[0136] Step 5.4, Spatial partition risk assessment;

[0137] Based on the location of the fire source, the fire spread prediction results, and the distribution of combustible load in the spatial zones, a risk quantification assessment model is used to calculate the fire risk quantification score for each spatial zone, thus obtaining the spatial zone risk score. The risk quantification assessment model comprehensively considers the following key factors: the spatial distance between the zone and the fire source location; the closer the zone is to the fire source, the greater the likelihood of being affected by the fire, and the higher the risk score; distance and risk are generally inversely proportional. Whether the zone is within the fire spread prediction range; if the fire spread dynamics predict that the zone may be affected by the fire in the future time window, a higher risk score is given. The combustible load value of the zone; the higher the combustible load, the more rapidly the fire will develop once it ignites or is affected by the fire, thus the higher the risk score. The current temperature level of the zone; the higher the temperature, the closer the zone is to the ignition conditions, and the higher the risk score accordingly. The distance between the zone and the evacuation route; the farther the zone is from the evacuation route, the greater the difficulty of evacuation and the greater the threat to personnel safety once a fire occurs, thus the higher the risk score.

[0138] The calculation process for risk quantification assessment is as follows: The original values ​​of each factor are standardized, mapping them to a range of 0 to 1; a corresponding weight coefficient is assigned to each factor; each standardized factor value is multiplied by its corresponding weight coefficient, and then a weighted sum is obtained to obtain the comprehensive risk score for that zone; the comprehensive risk score is normalized to a range of 0 to 1, where 0 represents no risk and 1 represents extremely high risk, with higher values ​​indicating greater fire risk in that zone. The weight coefficients are determined based on expert knowledge in the field of fire safety and statistical analysis of historical fire cases. A reasonable weight allocation is determined by analyzing the correlation between various factors and the degree of fire loss in a large number of fire accident cases. Furthermore, the system can dynamically adjust and optimize the weight coefficients based on actual response results and expert feedback during operation.

[0139] Step 5.5, visualization of the 3D situation map;

[0140] Based on a 3D temperature field model, fire spread prediction results, and spatial zoning risk scoring, a 3D fire situation map is generated using a 3D visualization rendering method, providing a visualized display of the fire situation. The 3D fire situation map is based on a 3D model of the electric vehicle shed, using color coding to represent the temperature value or risk level of each location. High-temperature areas are marked with warm colors, and low-temperature areas with cool colors; high-risk zones are marked with red borders, medium-risk zones with orange borders, and low-risk zones with green borders. The location of the fire source is highlighted with special markers, and the predicted range of fire spread is represented by semi-transparent boundaries. The 3D fire situation map supports multi-view viewing and zooming, providing commanders with intuitive fire situation information.

[0141] Furthermore, traditional fire spread dynamics simulation requires solving complex systems of partial differential equations, resulting in a large computational load that may not meet the time requirements for real-time prediction. To address this issue, a fast fire spread prediction method based on graph neural networks can be used to replace the traditional physical simulation method. The implementation of this method includes the following steps:

[0142] The 3D spatial graph is structured to convert the 3D space of the electric vehicle shed into a graph data structure, representing the adjacency and connection relationships in the space. Each spatial partition is treated as a node in the graph, containing attribute information such as current temperature, combustible material load, and geometric dimensions. Edges are established between adjacent spatial partitions, representing a direct physical adjacency between the two partitions, allowing heat transfer between them. Each edge is assigned a weight value, representing the ease with which heat can be transferred between two adjacent partitions. The weight calculation considers factors such as the distance between partitions, the presence of obstacles in between, and the thermal conductivity of the materials.

[0143] The training process of the graph neural network uses historical fire case data and data generated by physical simulation as training samples. Each training sample contains the temperature field distribution at a certain moment and the corresponding temperature field distribution at a future moment. Through training with a large number of samples, the graph neural network learns the mapping relationship between the current temperature field distribution and the fire source location to predict the future temperature field distribution. That is, the current state is input and the prediction result of the future state is output.

[0144] Real-time fire prediction inference, in practical applications, inputs the current fire situation information into a trained graph neural network model, including the temperature values ​​of each zone, the location of the fire source, and other state information. The model quickly outputs the prediction results of the fire spread in the future through forward calculation. The entire inference process only requires matrix operations, and the calculation speed is much faster than traditional physical simulation methods, which can meet the time requirements of real-time prediction.

[0145] Step 6: Generate the optimal sequence of response actions based on the three-dimensional fire situation map and spatial zoning risk score, issue execution instructions, and obtain execution feedback information;

[0146] Step 6.1, Fire protection facility topology modeling;

[0147] Based on the fire protection facility deployment information of the electric vehicle shed, a fire protection facility topology model is constructed to obtain a complete set of executable actions. The fire protection facility topology model records the location, coverage area, control interface, and linkage relationships of various fire protection facilities. Fire sprinklers are grouped according to spatial zones, and each group of sprinklers can be independently controlled to start and stop, with the sprinkler coverage area corresponding to the spatial zone. Smoke exhaust fans are grouped according to smoke exhaust zones, and each group of fans can be independently controlled to start and stop and adjust airflow. Power control switches are grouped according to power supply zones, and each group of switches can be independently controlled to cut off and restore power. Emergency lighting systems, evacuation broadcast systems, and fire notification systems are controlled as a whole facility. The complete set of executable actions includes: starting and stopping fire sprinklers in each zone, starting and stopping smoke exhaust fans in each zone and adjusting airflow, cutting off and restoring power to each power supply zone, activating emergency lighting, playing evacuation broadcasts, and issuing fire department notices.

[0148] Step 6.2, Strengthen the construction of the learning environment;

[0149] Based on a 3D fire situation map, spatial zoning risk scoring, and a fire protection facility topology model, a state space and action space for a reinforcement learning environment are constructed, resulting in a Markov decision process model. In the fire response scenario, the state space defines all possible environmental states, including the current temperature and smoke concentration values ​​for each spatial zone, the risk score calculated using the aforementioned method, the estimated 3D coordinates of the fire source location, and the current operational status of each fire protection facility, such as the on / off status of fire sprinklers, the operating status and airflow of smoke exhaust fans, and the on / off status of power switches. The action space is a combination of all executable actions. Due to the high dimensionality of the action space, a hierarchical action space design is adopted, categorizing actions into fire extinguishing actions, smoke exhaust actions, power outage actions, evacuation actions, and notification actions, with discrete selection within each type of action. The state transition function, based on a fire spread dynamics model and a fire protection facility effectiveness model, describes the probability distribution of transitioning to the next state after performing an action in the current state.

[0150] Step 6.3, Policy Network Training Optimization;

[0151] Based on the Markov decision process model, a reward function is designed, and a reinforcement learning policy network is trained to obtain an optimized response policy network. In fire response scenarios, the reward function needs to comprehensively consider multiple objectives. Specific reward design includes the following aspects: minimizing the fire spread range, measured by the size of the high-temperature zone; a smaller high-temperature zone indicates better fire control and a higher positive reward. Minimizing personnel risk, measured by the number of high-risk zones; fewer high-risk zones indicate better personnel safety and a higher positive reward. Minimizing property loss, measured by the number of electric bicycles affected by the fire; fewer affected vehicles indicate better property protection and a higher positive reward. Minimizing extinguishing agent consumption, measured by the number of activated fire extinguishing nozzles; fewer activated nozzles, while ensuring fire extinguishing effectiveness, indicate more economical resource utilization and an appropriate positive reward. Timeliness of response, measured by the time interval from fire warning to the execution of the first response action; a shorter response time indicates more timely response and a higher positive reward. All rewards are combined into a total reward through a weighted sum, with weight coefficients determined based on actual needs and expert experience.

[0152] The policy network employs a deep neural network architecture, containing multiple fully connected layers and activation functions. The input is a numerical representation of the current state, and the output is the probability distribution of each action selection. Backpropagation is used to learn the mapping relationship from state to action probability, enabling the selection of actions that yield high rewards in similar states. A proximal policy optimization algorithm is used for training the policy network. Limiting the magnitude of each policy update ensures the stability of the training process and avoids a sharp performance drop due to excessively large update steps. The proximal policy optimization algorithm reuses historical experience data through importance sampling, improving sample utilization efficiency. The policy network is trained in a digital twin simulation environment that accurately simulates the physical characteristics and fire development process of a real electric vehicle shed. This allows the agent to conduct extensive trial-and-error learning in a virtual environment, avoiding the potential safety risks and economic losses associated with exploring in a real environment.

[0153] Step 6.4, intelligent handling strategy generation;

[0154] Based on a trained response strategy network and the current fire situation, a strategy reasoning method is used to generate a sequence of response actions, resulting in an intelligent response strategy. The current state is input into the strategy network, which outputs the selection probability of each action. The action with the highest probability is sampled or selected as the response action for the current time step. After executing the action, the state of the next time step is observed, and the state is input into the strategy network again to obtain the next action. This process is repeated until the fire situation decreases to a safe level or the maximum response time is reached. The response action sequence records the actions executed at each time step and their execution time, forming a complete intelligent response strategy.

[0155] Step 6.5, IoT control execution;

[0156] Based on the sequence of actions in the intelligent response strategy, an IoT control protocol is used to issue control commands to each fire protection facility and receive execution feedback information. The IoT control module connects to the control interfaces of each fire protection facility via wired or wireless means, supporting remote control and status monitoring. Control commands include facility identification, action type, execution parameters, and execution time. Upon receiving the command, each facility executes the corresponding action and reports its execution status. Execution status feedback includes facility response delay, action execution result, and current facility status, used to confirm whether the response strategy is executed as expected. If a facility fails to execute the command, the system automatically triggers a backup facility or adjusts the response strategy.

[0157] Furthermore, since reinforcement learning policy networks may make unreasonable decisions when faced with anomalies outside the training data distribution, a rule-constrained hybrid decision-making method can be adopted to improve the safety of the response. Specifically, a set of mandatory fire safety rules is formulated, including mandatory actions and prohibited actions, such as the mandatory activation of fire alarms after fire confirmation and the prohibition of turning off emergency lighting during evacuation. After the policy network outputs the action probabilities, the mandatory rule constraints are checked, and actions that violate the rules are filtered out and their probabilities are set to zero. For mandatory actions, regardless of the policy network output, they are forcibly included in the action sequence. Under the premise of satisfying the rule constraints, the remaining actions are selected according to the output probabilities of the policy network, achieving a balance between safety and intelligence.

[0158] Step 7: Based on the execution feedback information, and combined with the fire risk level and fire source location estimation results updated in real time during the execution of the instructions, conduct a quantitative assessment and dynamic adjustment of the response effectiveness to obtain the response effectiveness assessment results and updated strategies;

[0159] Step 7.1, Real-time monitoring of situation evolution;

[0160] Based on multi-source sensor data continuously collected during the execution of the response strategy, the fire situation assessment is updated in real time using the same method as steps 2 to 5, resulting in a fire situation evolution sequence during the response process. During the response execution, the sensor network continues to operate, collecting data at the normal sampling frequency. After preprocessing, feature extraction, and multi-source fusion, the data outputs real-time fire situation assessment results. The fire situation assessment results at each moment are recorded in chronological order to form a fire situation evolution sequence, describing the changes in the fire situation from the start to the end of the response.

[0161] Step 7.2, Quantitative evaluation of the treatment effect;

[0162] Based on the fire situation evolution sequence, a quantitative evaluation method for fire response effectiveness is adopted to calculate evaluation indicators and obtain the evaluation results. Quantitative evaluation of fire response effectiveness is a method that measures the effectiveness of fire response strategies through a series of objective numerical indicators, providing a quantitative basis for system optimization. Specific evaluation indicators include fire control rate, temperature decrease rate, smoke dissipation rate, and response completion time, specifically manifested as follows:

[0163] Fire control rate is a key indicator for measuring the effectiveness of fire spread control. The calculation method is to determine the area of ​​the high-temperature zone at the start of the treatment as the benchmark value. The high-temperature zone is defined as the spatial range where the temperature exceeds a preset threshold. The area of ​​the high-temperature zone at the end of the treatment is measured. The ratio of the area reduction to the benchmark area is calculated. That is, the fire control rate is equal to the difference in the area of ​​the high-temperature zone before and after the treatment divided by the area of ​​the high-temperature zone before the treatment. The larger the ratio, the better the fire control effect.

[0164] Temperature drop rate is an indicator reflecting the effectiveness of temperature control. The calculation method is as follows: select the fire source area as the monitoring focus, record the highest temperature value of the area at the beginning of the treatment as the benchmark; monitor the temperature change of the area during the treatment process, and record the temperature value at the end of the treatment; calculate the temperature drop rate as the amount of temperature reduction divided by the benchmark temperature value. The larger the ratio, the better the temperature control effect.

[0165] Smoke dissipation rate is an indicator for evaluating the effectiveness of smoke removal. It is calculated by using the average smoke concentration at the start of the response as a baseline; monitoring changes in smoke concentration during the response process; and calculating the smoke dissipation rate as the concentration reduction divided by the baseline concentration. Response time is an indicator for measuring the speed of a system's reaction. It is defined as the time interval between the system issuing a fire warning signal and executing the first response action. The shorter this time, the faster the system response.

[0166] The response completion time is an indicator of response efficiency. It is defined as the time interval between the start of response actions and the reduction of the fire situation to a safe level. The criteria for determining the safety level are that the temperature, smoke concentration, and gas concentration in all monitored areas have dropped to the normal range.

[0167] The above indicators are combined into a comprehensive score for the treatment effect by weighted summation. The weight allocation is determined according to the importance of each indicator. The final score is normalized to a range of 0 to 1, where 0 represents extremely poor treatment effect and 1 represents excellent treatment effect. The higher the value, the better the treatment effect.

[0168] Step 7.3, Adaptive strategy adjustment;

[0169] Based on the evaluation results of the response effectiveness, an adaptive strategy adjustment method is adopted to determine whether dynamic adjustment of the response strategy is necessary, resulting in an adjusted strategy or maintaining the current strategy. If the response effectiveness evaluation indicators do not reach the expected improvement threshold within a preset time window, it indicates that the current response strategy may not be effective enough and strategy adjustment is required. The strategy adjustment methods are as follows: re-inputting the current fire situation as a new initial state into the strategy network to generate a new sequence of response actions; or increasing the response intensity based on a rule-based heuristic method, such as expanding the activation range of fire extinguishing nozzles or increasing the air volume of smoke exhaust fans. The adjusted strategy is executed through the IoT control module, forming a closed-loop control.

[0170] Step 7.4, record in the fire case database;

[0171] Based on complete fire event data, a standardized case record format is used to store event information in a fire event case database, resulting in fire event case records. Case records include: basic event information (time of occurrence, location, fire type, etc.), raw sensing data (raw data files from each sensor), intermediate processing results (feature data, fused features, detection results, etc.), fire situation evolution sequence (situation assessment results at each moment), response strategy record (sequence of response actions and execution time), and response effectiveness evaluation results (values ​​of various evaluation indicators). The case database is stored using a relational database, supporting querying and statistical analysis by field.

[0172] Step 7.5, update the policy network parameters;

[0173] Based on historical case data accumulated in a fire case database, an offline reinforcement learning method is used to update the policy network parameters, resulting in an optimized policy network. Offline reinforcement learning does not require real-time interaction with the environment; instead, it learns the optimal policy solely from pre-collected historical data, avoiding potential safety hazards associated with online learning in real-world environments. The workflow of offline reinforcement learning is as follows: Complete data on historical fire events is extracted from the fire case database, including the environmental state at each moment, the actions taken, the reward obtained, and the next state after transition, forming a four-tuple data sequence of state-action-reward-next state. These four-tuple data are then organized into an offline training dataset, containing handling experiences under various fire scenarios. A conservative Q-learning algorithm is used to update the policy network parameters. The conservative Q-learning algorithm introduces conservative constraints in the value function learning, ensuring that the learned policy does not overly rely on insufficiently explored regions in the training data. After validation and testing, the updated policy network is deployed to an online system for real-time decision-making in subsequent fire events. The update cycle of the policy network can be flexibly set according to actual needs. It can be updated regularly at fixed time intervals, or it can be set to automatically trigger an update when a certain number of new cases are accumulated in the case library, ensuring that the policy network can continuously learn and improve.

[0174] Furthermore, due to potential differences in the spatial structure, fire protection facilities, and fire risk characteristics of different electric vehicle sheds, a policy network trained in one location may not be directly applicable to another. Transfer learning methods can be used to accelerate the deployment of policy networks for new locations. Specifically, the process involves: pre-training model initialization, where the policy network trained in the source location is used as the pre-trained model, and its network parameters are used as initialization parameters; target location data collection, where a small amount of actual operational data from the target location is collected after deployment, or simulated data is generated in the digital twin environment of that location; model fine-tuning and adaptation, where the pre-trained model is fine-tuned using data from the target location, updating only the parameters of the later layers of the network while retaining the general feature extraction capabilities learned from the earlier layers; and location-adaptive deployment, where the fine-tuned policy network adapts to the specific conditions of the target location, while leveraging the accumulated experience from the source location, reducing the sample size and time required to train the policy network for the new location from scratch.

[0175] A multi-source sensing fusion intelligent fire response system for electric vehicle sheds, such as Figure 3 As shown, a multi-source sensing fusion-based intelligent fire response method for electric vehicle sheds, as described above, includes:

[0176] The spatial partitioning and sensor deployment module divides the electric vehicle shed into three-dimensional spatial structures; based on the three-dimensional spatial partitioning results, it deploys multi-source sensors and collects multi-source raw sensor data to obtain a hierarchical and partitioned multi-source raw data stream.

[0177] The data processing module preprocesses the hierarchical and partitioned multi-source raw data stream to obtain a spatiotemporally aligned multi-source feature dataset.

[0178] The multimodal fusion detection module performs multimodal feature fusion on a spatiotemporally aligned multi-source feature dataset to obtain a multimodal fusion feature vector; fire detection is then performed based on the multimodal fusion feature vector to obtain the fire detection result.

[0179] The risk assessment and location module assesses and locates the risk level based on fire detection results, and obtains the fire risk level and fire source location estimation results.

[0180] The risk simulation and prediction module reconstructs the three-dimensional space of the electric vehicle shed based on the spatiotemporally aligned multi-source feature dataset. Combining the reconstruction results, fire risk level and fire source location estimation results, it simulates the fire spread and performs risk quantification assessment by region, resulting in a three-dimensional fire situation map and spatial region risk score.

[0181] The intelligent response decision-making module generates the optimal sequence of response actions based on the three-dimensional fire situation map and spatial zoning risk score, issues execution instructions, and obtains execution feedback information;

[0182] The evaluation and optimization module, based on execution feedback information and combined with the real-time updated fire risk level and fire source location estimation results during the execution of instructions, performs quantitative evaluation and dynamic adjustment of the response effectiveness, and obtains the response effectiveness evaluation results and updated strategies.

[0183] In one embodiment of the present invention, a specific example is provided:

[0184] This invention focuses on the application of intelligent fire handling in multi-story underground electric bicycle charging sheds in residential communities. The following is an example of a real-world application scenario.

[0185] A residential community has a centralized charging shed for electric bicycles on its basement level. It employs a three-tiered rack-style parking and charging structure, with a total area of ​​approximately several square meters, capable of accommodating a number of electric bicycles simultaneously. The multi-source sensing fusion intelligent fire response system described in this invention is deployed within the charging shed.

[0186] Table 1 shows an example of multi-source sensor data collected by the system during a very early warning event:

[0187] Table 1: Examples of collected multi-source sensor data;

[0188]

[0189] The fire risk assessment results based on the multi-source data fusion analysis are shown in Table 2:

[0190] Table 2: Fire Risk Assessment Results;

[0191]

[0192] Through the multi-source sensing fusion method described in this invention, the system issues an early warning in the very early stage before the lithium battery thermal runaway produces obvious smoke and flames, providing managers with sufficient response time for inspection and handling, effectively preventing the occurrence of fire or controlling the fire in its infancy.

[0193] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments based on the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for intelligent handling of electric vehicle shed fires using multi-source sensing fusion, characterized in that, include: Divide the electric vehicle shed into three-dimensional spatial structures; Based on the three-dimensional spatial partitioning results, multi-source sensors are deployed and multi-source raw sensor data is collected to obtain a hierarchical and partitioned multi-source raw data stream. Preprocessing of the hierarchical and partitioned multi-source raw data stream yields a spatiotemporally aligned multi-source feature dataset. Multimodal feature fusion is performed on the spatiotemporally aligned multi-source feature dataset to obtain a multimodal fused feature vector; Fire detection is performed based on multimodal fusion feature vectors to obtain fire detection results; Based on the fire detection results, risk level is determined and located to obtain the fire risk level and fire source location estimation results; The risk level assessment and location based on fire detection results includes: A characteristic gas fingerprint library for lithium battery thermal runaway was constructed to obtain standard characteristic gas release patterns; Based on gas concentration characteristics, a dynamic time warping algorithm is used to match the feature gas fingerprint database to obtain the similarity score of the thermal runaway stage. Based on the principle of gas diffusion dynamics, a gas diffusion time compensation method is used to correct the time lag of gas concentration detection and obtain the compensated thermal runaway stage similarity score. Based on the compensated thermal runaway stage similarity score, temperature change rate characteristics, and fire detection results, a multi-threshold cascaded discrimination rule is adopted to obtain the fire risk level. Based on the fire risk level and the location of the sensor node that triggered the level, the associated spatial zoning identifier and fire source location estimate are determined; The three-dimensional space of the electric vehicle shed is reconstructed based on a spatiotemporally aligned multi-source feature dataset; combining the reconstruction results, fire risk level, and fire source location estimation results, the fire spread potential is simulated and the risk is quantitatively assessed by region, resulting in a three-dimensional fire situation map and spatial region risk score; the reconstruction of the three-dimensional space of the electric vehicle shed based on the spatiotemporally aligned multi-source feature dataset includes: Based on temperature sensor data and infrared thermal imaging data, the continuous temperature field distribution in the three-dimensional space of the electric vehicle shed is reconstructed using the Kriging space interpolation method, resulting in a three-dimensional temperature field model. Based on the three-dimensional temperature field model and fire source location estimation, the fire spread dynamics simulation method is used to predict the spatial range of fire spread within a preset time window and obtain the fire spread prediction results. Based on the parking space distribution information and charging pile status information of the electric vehicle shed, the combustible load calculation method is used to calculate the combustible load of each spatial zone and obtain the combustible load distribution of the spatial zone. Based on the location of the fire source, the predicted results of fire spread, and the distribution of combustible load in the spatial zones, a risk quantification assessment model is used to calculate the fire risk quantification score for each spatial zone. The optimal response sequence is generated based on the 3D fire situation map and spatial zoning risk score, and execution instructions are issued to obtain execution feedback information; the generation of the optimal response sequence based on the 3D fire situation map and spatial zoning risk score includes: Based on the fire protection facility deployment information of the electric vehicle shed, a fire protection facility topology model is constructed to obtain a complete set of executable response actions; Based on the three-dimensional fire situation map, spatial zoning risk scoring and fire protection facility topology model, the state space and action space of the reinforcement learning environment are constructed to obtain the Markov decision process model. Based on the Markov decision process model, a reward function is designed, a reinforcement learning policy network is trained, and an optimized disposition policy network is obtained. Based on the trained response strategy network and the current fire situation, a strategy reasoning method is used to generate a sequence of response actions, thereby obtaining an intelligent response strategy. Based on the sequence of actions in the intelligent response strategy, control commands are issued to various fire protection facilities using an Internet of Things control protocol; Based on the execution feedback information, combined with the real-time updated fire risk level and fire source location estimation results during the execution of instructions, the effectiveness of the response is quantitatively evaluated and dynamically adjusted to obtain the response effectiveness evaluation results and updated strategies.

2. The intelligent fire suppression method for electric vehicle sheds based on multi-source sensing fusion according to claim 1, characterized in that, The three-dimensional spatial structure division of the electric vehicle shed includes: A three-dimensional mesh spatial partitioning method is adopted to decompose the electric vehicle shed into a regular three-dimensional mesh structure. Each mesh unit corresponds to an independent spatial partition. A triplet coordinate system is used to identify each partition to obtain the spatial partition index. The vertical structure is divided into layers according to the number of rack layers in the rack-type parking structure, with each rack layer corresponding to a vertical partition. Within each vertical layer, the space is divided into several rectangular areas according to the preset partition size based on the horizontal spatial dimensions, forming a regular grid layout.

3. The intelligent fire suppression method for electric vehicle sheds based on multi-source sensing fusion according to claim 1, characterized in that, The deployment of multi-source sensors based on the three-dimensional spatial partitioning results includes: Based on the spatial partition index, a partition center deployment strategy is adopted to deploy a combination module of temperature sensor, smoke sensor and gas sensor at the geometric center of each spatial partition to obtain the distribution of partition environmental sensing nodes. A layer-top coverage deployment strategy is adopted, infrared thermal imagers are deployed at the top of each layer according to a preset coverage interval to obtain the distribution of layer-level thermal imaging sensing nodes; A key point coverage deployment strategy was adopted, and high-definition visible light cameras were deployed at each entrance and exit and at the intersection of the passage to obtain the distribution of visual perception nodes in the passage. An active gas sampling enhancement strategy is adopted, in which miniature axial fans are deployed in each spatial partition to form a directional airflow from the periphery of the partition toward the sensor, thereby accelerating the delivery of gas from the partition to the sensor.

4. The intelligent fire suppression method for electric vehicle sheds based on multi-source sensing fusion according to claim 1, characterized in that, The preprocessing of the hierarchical and partitioned multi-source raw data stream includes: Based on time-series sensor data from hierarchical and partitioned multi-source raw data streams, Kalman filtering is used to denoise the data and obtain smoothed time-series data. Based on the infrared thermal imaging image sequence, a non-uniformity correction and bilateral filtering method are used to obtain the corrected thermal imaging image sequence. Based on the visible light video image sequence, an enhanced video image sequence is obtained by using adaptive histogram equalization and median filtering methods. Based on the original timestamps of each sensor's data, a linear interpolation time synchronization method is used to obtain a time-aligned multi-source data sequence; Based on the sensor installation location information and spatial partition index, a coordinate mapping method is used to map the data of each sensor to a unified three-dimensional spatial coordinate system, resulting in spatially aligned multi-source data.

5. The intelligent fire suppression method for electric vehicle sheds based on multi-source sensing fusion according to claim 1, characterized in that, The multimodal feature fusion of the spatiotemporally aligned multi-source feature dataset includes: Based on the temporal feature data in the spatiotemporally aligned multi-source feature dataset, a temporal convolutional network is used for feature encoding to obtain temporal modality encoded features. Based on infrared thermal imaging feature maps, a two-dimensional convolutional neural network is used for feature encoding to obtain infrared image modality coding features; Based on the features of visible light images, a visual Transformer network is used for feature encoding to obtain the modality coding features of visible light images; Based on temporal modal coding features, infrared image modal coding features, and visible light image modal coding features, a cross-modal attention mechanism is used to perform feature interaction fusion to obtain a multimodal fusion feature vector; Based on the multimodal fusion feature vector, a classification head network is used to output the fire detection confidence score, and a regression head network is used to output the estimated three-dimensional location coordinates of the fire source, thus obtaining the fire detection result.

6. The intelligent fire suppression method for electric vehicle sheds based on multi-source sensing fusion according to claim 1, characterized in that, The quantitative evaluation of the treatment effect includes: Based on the multi-source sensor data continuously collected during the execution of the response strategy, the fire situation assessment is updated in real time to obtain the fire situation evolution sequence during the response process. Based on the fire situation evolution sequence, a quantitative evaluation method for handling effectiveness is adopted to calculate the handling effectiveness evaluation indicators and obtain the handling effectiveness evaluation results. The handling effectiveness evaluation indicators include fire control rate, temperature reduction rate, smoke dissipation rate, and handling completion time.

7. A multi-source sensing fusion intelligent fire response system for electric vehicle sheds, characterized in that, A method for intelligent fire response in electric vehicle sheds based on multi-source sensing fusion as described in any one of claims 1-6 includes: The spatial partitioning and sensor deployment module divides the electric vehicle shed into three-dimensional spatial structures; based on the three-dimensional spatial partitioning results, it deploys multi-source sensors and collects multi-source raw sensor data to obtain a hierarchical and partitioned multi-source raw data stream. The data processing module preprocesses the hierarchical and partitioned multi-source raw data stream to obtain a spatiotemporally aligned multi-source feature dataset. The multimodal fusion detection module performs multimodal feature fusion on a spatiotemporally aligned multi-source feature dataset to obtain a multimodal fusion feature vector; fire detection is then performed based on the multimodal fusion feature vector to obtain the fire detection result. The risk assessment and location module assesses and locates the risk level based on fire detection results, and obtains the fire risk level and fire source location estimation results. The risk simulation and prediction module reconstructs the three-dimensional space of the electric vehicle shed based on the spatiotemporally aligned multi-source feature dataset. Combining the reconstruction results, fire risk level and fire source location estimation results, it simulates the fire spread and performs risk quantification assessment by region, resulting in a three-dimensional fire situation map and spatial region risk score. The intelligent response decision-making module generates the optimal sequence of response actions based on the three-dimensional fire situation map and spatial zoning risk score, issues execution instructions, and obtains execution feedback information; The evaluation and optimization module, based on execution feedback information and combined with the real-time updated fire risk level and fire source location estimation results during the execution of instructions, performs quantitative evaluation and dynamic adjustment of the response effectiveness, and obtains the response effectiveness evaluation results and updated strategies.

Images