A fire prevention and control system for electric vehicle parking lots based on multi-source information fusion

Abstract

The present invention discloses a fire prevention and control system for electric vehicle parking lots based on multi-source information fusion. The system comprises a perception layer that uses a multimodal sensor network to capture the environmental status of electric vehicle parking lots and the operating parameters of vehicle batteries in real time; a network layer that uses an intelligent fusion network architecture to enable communication between the perception layer and the data processing layer; a data processing layer that performs a fusion analysis of the environmental status and operating parameters, determines the fire risk level, and generates dynamic control instructions; a data processing layer that comprises a multi-dimensional analysis engine and uses a hierarchical and progressive intelligent analysis architecture to achieve multi-level assessment of fire risk; and an application layer that receives the dynamic control instructions generated by the data processing layer and implements closed-loop management of the entire fire prevention and control process through multimodal terminal devices. This method not only improves the accuracy of fire prevention and control in electric vehicle parking lots, but also has good interpretability and can be directly applied to fire prevention and control systems for electric vehicle parking lots.

Description

Technical Field

[0001] The present invention relates to the field of fire prevention and control, and in particular to a fire prevention and control system for electric vehicle parking lots based on multi-source information fusion. Background Art

[0002] With the booming global electric vehicle industry and the rapid growth in electric vehicle ownership, fire safety issues in parking lots are becoming increasingly severe. As core energy storage components, electric vehicle batteries are characterized by highly concentrated energy and fragile thermal stability. Once thermal runaway occurs, fires can spread rapidly, releasing large amounts of toxic and hazardous gases. Traditional firefighting methods are difficult to effectively contain.

[0003] Currently, conventional fire alarm systems exhibit numerous shortcomings when responding to fires in electric vehicle parking lots. Their monitoring methods, based on a single or limited number of parameters, are unable to fully capture the complex early characteristics of electric vehicle fires, resulting in frequent false alarms and missed alerts. Furthermore, the system's response is sluggish, with a long delay between the onset of a fire and the triggering of an alarm, delaying the optimal time to extinguish the fire. Fire extinguishing systems lack specificity in extinguishing agent selection and injection control, making it difficult to accurately match the unique needs of electric vehicle fires and effectively extinguish them. These issues pose a serious threat to human and property safety and hinder the sustainable development of the electric vehicle industry. Therefore, the development of a precise, efficient, and intelligent fire prevention and control system for electric vehicle parking lots is urgent and has significant practical significance for safeguarding public safety and promoting the steady development of the electric vehicle industry. Summary of the Invention

[0004] The purpose of the present invention is to provide a fire prevention and control method for electric vehicle parking lots based on multi-source information fusion.

[0005] To achieve the above object, the present invention is implemented according to the following technical solutions:

[0006] The present invention comprises the following steps:

[0007] The perception layer captures the environmental status of electric vehicle parking spaces and the operating parameters of vehicle batteries in real time through a multimodal sensor network;

[0008] The network layer adopts an intelligent converged network architecture to achieve communication between the perception layer and the data processing layer;

[0009] The data processing layer is used to perform a fusion analysis of the environmental status and the operating parameters, determine the fire risk level, and generate dynamic control instructions. The data processing layer is composed of a multi-dimensional analysis engine and adopts a hierarchical and progressive intelligent analysis architecture to achieve multi-level analysis and judgment of fire risks. The data processing layer includes a heterogeneous data preprocessing module, a multi-dimensional feature extraction module, a fire risk assessment module, and a dynamic control strategy generation module. It includes:

[0010] Based on the feature information extracted by the multidimensional feature extraction module, a fire risk assessment model is constructed using a three-level progressive analysis framework to evaluate the fire risk level of electric vehicle parking lots. The fire risk level includes level one, level two, and level three assessments.

[0011] The application layer receives the dynamic control instructions generated by the data processing layer and realizes the full-process closed-loop management of fire prevention and control through multimodal terminal equipment; the application layer includes an audio and visual alarm module, a smoke exhaust system control module, a fire extinguishing system control module, a remote monitoring platform, and an emergency evacuation and fire extinguishing rescue module.

[0012] Furthermore, the first-level assessment includes:

[0013] Use the deep neural network pre-trained based on historical environment states and operating parameters and load its weight parameters;

[0014] Local data from the target parking lot was collected and the sample size was expanded through data augmentation. Transfer learning was used to freeze the underlying feature extraction layer of the model and adjust only the parameters of the top fully connected layer. The loss function was designed to combine cross entropy loss with KL divergence.

[0015] Real-time sensor data is input into the fine-tuned model to output a fire risk probability value in the range of 0-1; a threshold is set, marked as a level 1 high risk, and a level 2 assessment is triggered.

[0016] Furthermore, the secondary assessment includes:

[0017] The Kalman filter is used to fuse sensor data of the same type to eliminate isolated point errors; the time series alignment algorithm is used to synchronize BMS data and environmental data to the millisecond level;

[0018] Construct a Bayesian network topology whose nodes include sensor data, equipment health, and fire status; calculate the posterior probability based on the prior probability and conditional probability table;

[0019] The improved DS evidence theory is used to perform uncertainty reasoning on multi-source sensor data. Combined with the adaptive weighting algorithm, the trust weights of different sensors are dynamically assigned. The expression is:

[0020]

[0021] The health of the i-th device is The environmental adaptation coefficient is β, the dynamic correlation adjustment coefficient is μ, the global data set is U, and the i-th sensor data is D i , sensor data D i The historical correlation with the global dataset U is C orr , the correlation change rate is

[0022] Set a threshold, mark it as "secondary confirmation risk", and trigger a third-level assessment; use a sliding window mechanism to statistically analyze the distribution of sensor data, analyze the timeliness of data through information entropy, and update the Bayesian network conditional probability table in real time.

[0023] Furthermore, the three-level evaluation method further includes:

[0024] Define entities, including battery type, fire cause, environmental conditions, and extinguishing agent characteristics, and establish causal chains based on historical cases and domain knowledge;

[0025] Input the current risk characteristics, traverse the knowledge graph through the graph neural network, match similar subgraphs; calculate the subgraph confidence and identify potential risk paths;

[0026] Combining the first-level probability output and the second-level confidence level, the cause of the fire is deduced through the logic rule engine; if the knowledge graph reasoning result is consistent with the first-level and second-level assessments, a high-risk level is determined; if there is a contradiction, manual review is triggered.

[0027] Furthermore, the method for screening the multidimensional features further includes:

[0028] Calculate the correlation between multidimensional features and fire, sort the multidimensional features in descending order according to the correlation, and select the multidimensional features with a correlation greater than 0.469 as the candidate feature set;

[0029] Combine candidate features according to the univariate perturbation principle according to their dimensions to obtain a multidimensional feature set, introduce a particle swarm, and use the multidimensional feature set as particles;

[0030] The prediction error of different multidimensional feature sets on the training set is calculated using the base model, and the fitness value of the particle is calculated based on the prediction error:

[0031]

[0032] The fitness function of the cth particle is The number of dimensions is N1, the multidimensional feature set is Q, and the minimum prediction error is y min , the prediction error of the cth multidimensional feature set is y c , the a-th multidimensional feature of the s-th dimension is u a (s), the z-th multidimensional feature of the k-th dimension is u z (k), multidimensional feature u a (s) and multidimensional features u z (k) Joint probability distribution, multidimensional feature u a The marginal probability distribution of (s) is p(u a (s)), multidimensional features u zThe marginal probability distribution of (k) is p(u z (k)), the number of multidimensional features in the multidimensional feature set is N1;

[0033] The particle with the largest fitness value is used as the attacker, and the average Euclidean distance between the particle and the attacker is calculated;

[0034] Select particles whose Euclidean distance to the attacker is greater than the average Euclidean distance and update the particle position. The expression is:

[0035]

[0036] The position of the cth particle in the t+1th iteration is The upper bound of the search space is K max , the lower bound of the search space is K min , the position of the cth particle in the tth iteration is The current iteration number is t, the random number from 0 to 1 is τ1, and the position of the attacker is

[0037] After updating the particles, the fitness value is calculated. The attackers are divided into two groups according to the fitness value. The position of the particles is updated according to the search step length of the particles to obtain the mutation position. The expression is:

[0038]

[0039] The random numbers from 0 to 1 are τ4 and τ5, and the search step is The mutation position of the cth particle in the t+1th iteration is The maximum number of iterations is t max ;

[0040] Until the fitness reaches the maximum, the corresponding multidimensional feature set is used as the output result.

[0041] Furthermore, the heterogeneous data preprocessing module and the multidimensional feature extraction module include:

[0042] The heterogeneous data preprocessing module uses a spectral-spatial joint noise reduction algorithm, establishes a noise model based on sensor characteristics, and employs a sliding window mechanism to perform adaptive wavelet denoising on time-series data such as temperature and smoke, achieving dynamic threshold noise reduction. It also applies a Kalman filter to fuse data from similar sensors, eliminating isolated point errors and performing multi-source calibration. It also uses a convolutional neural network to identify interference sources such as device shadows and water vapor in the video stream, and corrects false trigger signals for pattern recognition correction.

[0043] The multidimensional feature extraction module is used to extract fire-related feature information from the preprocessed data and build a three-dimensional spatiotemporal feature extraction system, which includes time domain features, frequency domain features, spatial correlation features, and cross-modal features;

[0044] The time domain features are used to calculate the dynamic indicators of battery temperature gradient and second-order derivative of VOCs concentration; the pressure sensor spectrum energy distribution is analyzed through FFT, and the characteristic frequency of micro-detonation inside the battery is detected as a spatial correlation feature; a sensor topology model is constructed to analyze the spatial propagation path of abnormal signals and determine the spatial correlation features; a twin network is used to align the feature space of thermal imaging and voltage fluctuation data as cross-modal features.

[0045] Furthermore, the dynamic control strategy generation module includes:

[0046] Based on the risk level and feature analysis results output by the fire risk assessment module, dynamic control instructions are generated through a multimodal decision engine; the dynamic control strategy generation module includes hierarchical response strategies and instruction generation logic and a dynamic strategy optimization mechanism;

[0047] The hierarchical response strategy and instruction generation logic generates differentiated control instructions based on the fire risk level output by the three-level assessment and the causal reasoning results of the knowledge graph: the risk levels include low risk, medium risk, and high risk;

[0048] When the risk is low, the ventilation system in the target area is activated to reduce the concentration of combustible gases; a battery health check instruction is sent to the BMS to forcibly limit the charging current to a safe threshold; a yellow warning message is pushed to the management personnel through the remote monitoring platform, prompting manual review; when the risk is medium, the full-area sound and light alarm is activated, and the charging pile power is cut off for protection; the smoke exhaust system pre-mode is activated, the smoke diffusion path is calculated in real time, and the direction of the smoke exhaust valve is pre-adjusted; when the risk is high, the fire source is located based on thermal imaging data, and fine water mist is released to suppress the initial fire; if the temperature gradient continues to rise, the fire is fully extinguished with perfluorohexanone gas; the main power supply of the parking lot is simultaneously cut off, and the emergency lighting and evacuation guidance system are activated; a three-dimensional thermal map, vehicle location and battery type data are sent to the city fire command center through a blockchain encrypted channel;

[0049] A dynamic strategy optimization mechanism is used to establish a control instruction execution effect evaluation model. The fire extinguishing efficiency is calculated through real-time data from the fire extinguishing system pressure sensor and camera flame recognition results. If the efficiency is lower than the preset threshold, the strategy adjustment is automatically triggered: the historical case library in the knowledge graph is called to match the optimal strategy for similar scenarios. Conversely, the Q-Learning reinforcement learning algorithm is used to dynamically update the control strategy weights to achieve adaptive optimization.

[0050] Secondly, a fire prevention and control system for electric vehicle parking lots based on multi-source information fusion includes:

[0051] Data acquisition module: used in the perception layer to capture the environmental status of electric vehicle parking spaces and the operating parameters of vehicle batteries in real time through a multimodal sensor network;

[0052] Communication module: used for the network layer to adopt intelligent fusion network architecture to achieve communication between the perception layer and the data processing layer;

[0053] Evaluation strategy generation module: used in the data processing layer to perform integrated analysis on the environmental status and the operating parameters, determine the fire risk level, and generate dynamic control instructions; the data processing layer is composed of a multi-dimensional analysis engine, and adopts a hierarchical and progressive intelligent analysis architecture to achieve multi-level analysis and judgment of fire risks; the data processing layer includes a heterogeneous data preprocessing module, a multi-dimensional feature extraction module, a fire risk assessment module, and a dynamic control strategy generation module; including:

[0054] Based on the feature information extracted by the multidimensional feature extraction module, a fire risk assessment model is constructed using a three-level progressive analysis framework to evaluate the fire risk level of electric vehicle parking lots. The fire risk level includes level one, level two, and level three assessments.

[0055] Positioning and rescue module: used for the application layer to receive the dynamic control instructions generated by the data processing layer, and to achieve full-process closed-loop management of fire prevention and control through multimodal terminal equipment; the application layer includes sound and light alarm module, smoke exhaust system control module, fire extinguishing system control module, remote monitoring platform, emergency evacuation and fire extinguishing rescue module

[0056] The beneficial effects of the present invention are:

[0057] The present invention is a fire prevention and control system for electric vehicle parking lots based on multi-source information fusion. Compared with the existing technology, the present invention has the following technical effects:

[0058] The present invention can improve the accuracy of fire prevention and control in electric vehicle parking lots through preprocessing, fusion analysis, generation of dynamic control instructions, multi-level analysis and judgment of fire risks, assessment of fire risk levels and full-process closed-loop management steps, thereby improving the accuracy of fire prevention and control in electric vehicle parking lots, optimizing fire prevention and control in electric vehicle parking lots, greatly saving resources and improving work efficiency, realizing intelligent prevention and control of fires in electric vehicle parking lots, and performing fusion analysis and assessment of fire risk levels in electric vehicle parking lots in real time, which is of great significance to fire prevention and control in electric vehicle parking lots, and can adapt to fire prevention and control in electric vehicle parking lots of different standards and different fire prevention and control needs of electric vehicle parking lots, and has a certain universality. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a flowchart of the steps of a fire prevention and control method for electric vehicle parking lots based on multi-source information fusion according to the present invention. DETAILED DESCRIPTION

[0060] The present invention will be further described below through specific examples. The illustrative examples and descriptions of the present invention are used to explain the present invention but are not intended to limit the present invention.

[0061] The present invention provides a fire prevention and control system for electric vehicle parking lots based on multi-source information fusion, comprising the following steps:

[0062] like Figure 1 As shown, in this embodiment, the following steps are included:

[0063] The perception layer captures the environmental status of electric vehicle parking spaces and the operating parameters of vehicle batteries in real time through a multimodal sensor network;

[0064] The network layer adopts an intelligent converged network architecture to achieve communication between the perception layer and the data processing layer;

[0065] The data processing layer is used to perform a fusion analysis of the environmental status and the operating parameters, determine the fire risk level, and generate dynamic control instructions. The data processing layer is composed of a multi-dimensional analysis engine and adopts a hierarchical and progressive intelligent analysis architecture to achieve multi-level analysis and judgment of fire risks. The data processing layer includes a heterogeneous data preprocessing module, a multi-dimensional feature extraction module, a fire risk assessment module, and a dynamic control strategy generation module. It includes:

[0066] Based on the feature information extracted by the multidimensional feature extraction module, a fire risk assessment model is constructed using a three-level progressive analysis framework to evaluate the fire risk level of electric vehicle parking lots. The fire risk level includes level one, level two, and level three assessments.

[0067] The application layer receives the dynamic control instructions generated by the data processing layer and implements full-process closed-loop management of fire prevention and control through multimodal terminal devices; the application layer includes an audio and visual alarm module, a smoke exhaust system control module, a fire extinguishing system control module, a remote monitoring platform, and an emergency evacuation and fire extinguishing rescue module;

[0068] In actual evaluation, β is the environmental adaptability coefficient, which represents the adjustment factor of the equipment health to the credibility weight. Its value is dynamically adjusted by environmental conditions and can be dynamically fitted through the regression model of the historical failure rate of environmental parameters.

[0069] μ is the dynamic correlation adjustment coefficient, which represents the contribution ratio of the control historical correlation change rate to the credibility and reflects the system's sensitivity to data timeliness. It is based on the variance analysis of the correlation change within the sliding window and is optimized in combination with the information entropy theory.

[0070] The health of the i-th device quantifies the operating status of the sensor itself, including calibration accuracy, fault history, signal stability, etc., and is normalized to [0,1], where 1 indicates complete health and 0 indicates failure;

[0071] C orr (D i ,U) is the historical correlation, representing the i-th sensor data D i The statistical correlation with the global data set U reflects the consistency between the sensor data and the overall state of the system;

[0072] is the correlation change rate, the first-order derivative of the historical correlation over time, which characterizes the dynamic change trend of the consistency between sensor data and the global state;

[0073] Temperature and humidity monitoring submodule: ground temperature 25°C, air temperature 26°C, ambient humidity 48% RH; gas monitoring submodule: CO concentration 25ppm, H2 concentration 53ppm, VOCs concentration 20ppm, HF concentration not detected;

[0074] Flame and smoke detection submodule: no flame spectrum characteristics, smoke concentration 6.73mg / m 3 ;

[0075] Visual perception submodule: Battery pack surface temperature 30°C, no charging plug overheating and battery bulging; Battery body monitoring submodule: Single cell temperature 32°C, battery deformation pressure 50kPa, battery internal resistance 0.01Ω;

[0076] Electrical parameter monitoring submodule: battery charge and discharge current 7.2A, battery pack total voltage 400V, single cell voltage 3.3V;

[0077] BMS data fusion interface: SOC estimation value 58%, no fault code, thermal management system operating normally;

[0078] Hybrid network architecture module: Assuming a 5G network signal strength of -80dBm, a Wi-Fi 6E network rate of 800Mbps, a ZigBee network node transmission success rate of 98%, a LoRa network node transmission success rate of 95%, and no packet loss in the fiber ring network;

[0079] Adaptive network dynamic adjustment module: Sets initial bandwidth allocation to 800Mbps for video streams, 100Mbps for battery status data, and 50Mbps for environmental monitoring data; channel quality indicators are 0.01% bit error rate for video streams, 0.001% bit error rate for battery status data, and 0.05% bit error rate for environmental monitoring data;

[0080] Heterogeneous data preprocessing module: The sensor noise level is set to ±0.3°C for temperature sensor noise, ±2ppm for gas sensor noise (taking CO as an example), and ±1mg / m 3 ;

[0081] Multidimensional feature extraction module: Calculates the battery temperature gradient of 0.1°C / min and the second-order derivative of VOCs concentration of 0.01ppm / min 2 ;

[0082] Fire risk assessment module: In the first-level assessment, the model accuracy on local data before fine-tuning based on transfer learning was 70%. In the second-level assessment, the initial sensor credibility weights were: 0.3 for temperature sensor, 0.3 for gas sensor, 0.2 for smoke sensor, and 0.2 for vision sensor.

[0083] Sound and light alarm module: alarm volume 80dB(A), alarm light brightness 500cd / m 2 ;

[0084] Smoke exhaust system control module: the initial speed of the smoke exhaust fan is 800r / min, and the initial opening of the smoke exhaust valve is 30%;

[0085] Fire extinguishing system control module: water mist nozzle initial pressure 2MPa, perfluorohexanone gas storage pressure 4MPa.

[0086] In this embodiment, the perception layer includes:

[0087] The perception layer includes an environmental information acquisition module and an electric vehicle status information acquisition module;

[0088] The environmental information acquisition module includes a temperature and humidity monitoring submodule and a gas monitoring submodule;

[0089] The temperature and humidity monitoring submodule uses distributed fiber optic temperature sensors, which are laid along the parking space layout to monitor ground and air temperature in real time with an accuracy of ±0.5°C. Wavelength demodulation technology is used to eliminate electromagnetic interference. An integrated capacitive humidity sensor detects ambient humidity with a range of 0-100% RH and an accuracy of ±2%.

[0090] The gas monitoring submodule is equipped with a multi-spectral gas sensor array to simultaneously detect CO, H2 (range 0-5000ppm), VOCs and HF with a resolution of ppb level;

[0091] The flame and smoke detection submodule uses a dual-band infrared flame detector, 3-5μm & 8-14μm, to achieve multiple verification of flame spectral characteristics. It also deploys a photoelectric smoke detector, which analyzes the particle size distribution through scattering principles and distinguishes between fire smoke and dust / water mist interference.

[0092] The visual perception submodule is equipped with an infrared thermal imaging camera with a resolution of 640×480 and a temperature measurement range of -20°C to 1200°C, which can capture abnormal temperature fields on the battery pack surface in real time. It also integrates an AI visual analysis unit to identify overheating of the charging plug and battery swelling.

[0093] The electric vehicle status information acquisition module includes a battery body monitoring submodule, an electrical parameter monitoring submodule, and a BMS data fusion interface;

[0094] The battery monitoring submodule uses a multi-channel thermocouple array embedded within the battery module to monitor the temperature of individual cells with a sampling rate of 10Hz and an accuracy of ±0.2°C, constructing a three-dimensional temperature field model. A fiber Bragg grating strain sensor is attached to the battery casing to measure deformation pressure with a range of 0-500kPa and a resolution of 0.1% FS, providing an early warning of battery expansion risks. An AC impedance spectrum analyzer uses a high-frequency excitation signal (1kHz-10MHz) to detect changes in the battery's internal resistance and quantify the degree of aging, with a SOH error of less than 2%.

[0095] The electrical parameter monitoring submodule uses a wideband Hall effect sensor to collect battery charge and discharge current in real time, with a range of ±500A and a bandwidth of DC-1MHz, capturing micro-short-circuit transient characteristics. A high-precision voltage acquisition unit is used to synchronously monitor the total battery pack voltage and single cell voltage with an accuracy of ±0.05%, identifying voltage consistency degradation trends.

[0096] The BMS data fusion interface uses the CAN bus protocol to obtain SOC estimation values, fault codes, and thermal management status in real time. It uses a time series alignment algorithm to synchronize BMS data with environmental sensor data to the millisecond level.

[0097] In actual evaluation, the measurement range for CO is 0-1000ppm; for H2, it is 0-5000ppm; for VOCs, it is 0-1000ppm; and for HF, it is 0-50ppm.

[0098] In this embodiment, the network layer includes:

[0099] The network layer includes a hybrid network architecture module, an adaptive network dynamic adjustment module, an edge-fog-cloud collaborative computing module, and blockchain-enabled network security;

[0100] The hybrid network architecture module includes multimodal communication protocol integration and network slicing and QoS classification;

[0101] The multimodal communication protocol integration includes high-speed data transmission channels, low-power wide-area coverage, and wired redundant backup. The high-speed data transmission channels utilize 5G and Wi-Fi 6E technologies to provide transmission rates exceeding 10Gbps for high-bandwidth data such as video streaming and 3D heat maps, with end-to-end latency less than 5ms. Low-power wide-area coverage deploys ZigBee and LoRa dual-mode base stations, supporting wide-area coverage of low-frequency sensor data such as smoke concentration and temperature, while reducing power consumption to the μW level. Wired redundant backup utilizes a fiber ring network based on the GPON standard, ensuring 100% packet loss-free transmission in areas with severe electromagnetic interference.

[0102] Network slicing and QoS grading, based on SDN technology, divide the physical network into three logical slices: emergency command slice, real-time monitoring slice, and routine maintenance slice. The emergency command slice exclusively uses bandwidth resources, prioritizing the transmission of critical data such as fire-fighting instructions and evacuation alarms, ensuring latency <10ms. The real-time monitoring slice allocates dynamic bandwidth for video streams and battery status data, supporting adaptive QoS adjustment. The routine maintenance slice is used for non-real-time services such as device status heartbeat packets and log uploads.

[0103] The adaptive network dynamic adjustment module includes reinforcement learning-driven bandwidth allocation and multi-path fault-tolerant transmission;

[0104] Reinforcement learning-driven bandwidth allocation includes defining the state space, action space, and reward function; using a deep deterministic policy gradient algorithm to optimize network configuration in real time;

[0105] Multi-path fault-tolerant transmission adopts a multi-armed bandit model to dynamically select the optimal transmission path and seamlessly switch when the channel is congested or the signal is attenuated, with a packet loss rate of less than 0.01%. It integrates forward error correction and ARQ retransmission protocols to ensure 100% accessibility of key instructions.

[0106] The edge-fog-cloud collaborative computing module includes edge computing nodes, fog computing clusters, and cloud collaboration mechanisms. The edge computing nodes deploy AI acceleration chips within charging piles to perform sensor data preprocessing and feature extraction, including temperature gradients and gas concentration change rates. A lightweight LSTM model is used to initially assess local fire risks with an accuracy rate greater than 85%. Only high-risk data is uploaded to the fog computing layer.

[0107] The fog computing cluster is deployed on servers in the parking lot area, aggregating data from multiple edge nodes and integrating multi-source data based on Kalman filtering to eliminate temporal and spatial biases. It also runs a simplified fire risk assessment model to generate pre-control strategies. Using a time-sensitive network protocol, it ensures that command issuance latency is less than 50ms.

[0108] The cloud-based collaboration mechanism only calls on cloud resources when uploading complex analysis tasks or when cross-regional collaboration is required. It uses a federated learning framework to aggregate data from multiple parking lots to update the global model while protecting local data privacy.

[0109] Blockchain-enabled network security includes data integrity assurance, identity authentication, and access control. Data integrity assurance generates a Merkle tree hash value for each data packet, and hash chain consistency is verified through smart contracts, with a tamper detection rate greater than 99.99%. Sensitive data is encrypted using the national secret SM4 algorithm, and keys are dynamically updated through the quantum key distribution protocol. Identity authentication and access control are based on the issuance of device digital certificates on the blockchain, achieving two-way authentication. An attribute-based encryption model is constructed to dynamically authorize data access rights by role.

[0110] In the actual evaluation, ZigBee, 2.4GHz; LoRa, Sub-1GHz; wide-area coverage radius greater than 1km; QoS adaptive adjustment priority is 1080P video > battery temperature > ambient humidity.

[0111] In this embodiment, the first-level assessment includes:

[0112] Use the deep neural network pre-trained based on historical environment states and operating parameters and load its weight parameters;

[0113] Local data from the target parking lot was collected and the sample size was expanded through data augmentation. Transfer learning was used to freeze the underlying feature extraction layer of the model and adjust only the parameters of the top fully connected layer. The loss function was designed to combine cross entropy loss with KL divergence.

[0114] Real-time sensor data is input into the fine-tuned model to output a fire risk probability value in the range of 0-1; a threshold is set: probability > 0.7, marked as a level 1 high risk, triggering a level 2 assessment.

[0115] In this embodiment, the secondary assessment includes:

[0116] The Kalman filter is used to fuse sensor data of the same type to eliminate isolated point errors; the time series alignment algorithm is used to synchronize BMS data and environmental data to the millisecond level;

[0117] Construct a Bayesian network topology whose nodes include sensor data, equipment health, and fire status; calculate the posterior probability based on the prior probability and conditional probability table;

[0118] The improved DS evidence theory is used to perform uncertainty reasoning on multi-source sensor data. Combined with the adaptive weighting algorithm, the trust weights of different sensors are dynamically assigned. The expression is:

[0119]

[0120] The health of the i-th device is The environmental adaptation coefficient is β, the dynamic correlation adjustment coefficient is μ, the global data set is U, and the i-th sensor data is D i , sensor data D i The historical correlation with the global dataset U is C orr , the correlation change rate is

[0121] Set the threshold: if the evidence support is greater than 0.7, it will be marked as "secondary confirmation risk" and trigger the third-level assessment; use the sliding window mechanism to statistically analyze the distribution of sensor data, analyze the timeliness of the data through information entropy, and update the Bayesian network conditional probability table in real time.

[0122] In this embodiment, the three-level evaluation method further includes:

[0123] Define entities, including battery type, fire cause, environmental conditions, and extinguishing agent characteristics, and establish causal chains based on historical cases and domain knowledge;

[0124] Input the current risk characteristics, traverse the knowledge graph through the graph neural network, match similar subgraphs; calculate the subgraph confidence and identify potential risk paths;

[0125] Combining the first-level probability output and the second-level confidence level, the cause of the fire is deduced through the logic rule engine; if the knowledge graph reasoning result is consistent with the first-level and second-level assessments, a high-risk level is determined; if there is a contradiction, manual review is triggered.

[0126] In this embodiment, the method for screening the multi-dimensional features further includes:

[0127] Calculate the correlation between multidimensional features and fire, sort the multidimensional features in descending order according to the correlation, and select the multidimensional features with a correlation greater than 0.469 as the candidate feature set;

[0128] Combine candidate features according to the univariate perturbation principle according to their dimensions to obtain a multidimensional feature set, introduce a particle swarm, and use the multidimensional feature set as particles;

[0129] The prediction error of different multidimensional feature sets on the training set is calculated using the base model, and the fitness value of the particle is calculated based on the prediction error:

[0130]

[0131] The fitness function of the cth particle is The number of dimensions is N1, the multidimensional feature set is Q, and the minimum prediction error is y min , the prediction error of the cth multidimensional feature set is y c, the a-th multidimensional feature of the s-th dimension is u a (s), the z-th multidimensional feature of the k-th dimension is u z (k), multidimensional feature u a (s) and multidimensional features u z (k) Joint probability distribution, multidimensional feature u a The marginal probability distribution of (s) is p(u a (s)), multidimensional features u z The marginal probability distribution of (k) is p(u z (k)), the number of multidimensional features in the multidimensional feature set is N1;

[0132] The particle with the largest fitness value is used as the attacker, and the average Euclidean distance between the particle and the attacker is calculated;

[0133] Select particles whose Euclidean distance to the attacker is greater than the average Euclidean distance and update the particle position. The expression is:

[0134]

[0135] The position of the cth particle in the t+1th iteration is The upper bound of the search space is K max , the lower bound of the search space is K min , the position of the cth particle in the tth iteration is The current iteration number is t, the random number from 0 to 1 is τ1, and the position of the attacker is

[0136] After updating the particles, the fitness value is calculated. The attackers are divided into two groups according to the fitness value. The position of the particles is updated according to the search step length of the particles to obtain the mutation position. The expression is:

[0137]

[0138] The random numbers from 0 to 1 are τ4 and τ5, and the search step is The mutation position of the cth particle in the t+1th iteration is The maximum number of iterations is t max ;

[0139] Until the fitness reaches the maximum, the corresponding multidimensional feature set is used as the output result.

[0140] In this embodiment, the heterogeneous data preprocessing module and the multidimensional feature extraction module include:

[0141] The heterogeneous data preprocessing module uses a spectral-spatial joint noise reduction algorithm, establishes a noise model based on sensor characteristics, and employs a sliding window mechanism to perform adaptive wavelet denoising on time-series data such as temperature and smoke, achieving dynamic threshold noise reduction. It also applies a Kalman filter to fuse data from similar sensors, eliminating isolated point errors and performing multi-source calibration. It also uses a convolutional neural network to identify interference sources such as device shadows and water vapor in the video stream, and corrects false trigger signals for pattern recognition correction.

[0142] The multidimensional feature extraction module is used to extract fire-related feature information from the preprocessed data and build a three-dimensional spatiotemporal feature extraction system, which includes time domain features, frequency domain features, spatial correlation features, and cross-modal features;

[0143] Time-domain features are used to calculate dynamic indicators such as battery temperature gradient and the second-order derivative of VOCs concentration. The pressure sensor spectrum energy distribution is analyzed through FFT, and the characteristic frequency of micro-detonations within the battery is detected as a spatial correlation feature. A sensor topology model is constructed to analyze the spatial propagation path of abnormal signals and determine spatial correlation features. A twin network is used to align the feature spaces of thermal imaging and voltage fluctuation data as cross-modal features.

[0144] In actual evaluation, characteristic information includes temperature change rate, smoke concentration change rate, combustible gas concentration change rate, flame characteristics, battery temperature abnormality, battery voltage abnormality, battery current abnormality, battery deformation pressure abnormality, and battery internal resistance abnormality.

[0145] In this embodiment, the dynamic control strategy generation module includes:

[0146] Based on the risk level and feature analysis results output by the fire risk assessment module, dynamic control instructions are generated through a multimodal decision engine; the dynamic control strategy generation module includes hierarchical response strategies and instruction generation logic and a dynamic strategy optimization mechanism;

[0147] The hierarchical response strategy and instruction generation logic generates differentiated control instructions based on the fire risk level output by the three-level assessment and the causal reasoning results of the knowledge graph: the risk levels include low risk, medium risk, and high risk;

[0148] When the risk is low, the ventilation system in the target area is activated to reduce the concentration of combustible gas; a battery health check instruction is sent to the BMS to forcibly limit the charging current to a safe threshold; a yellow warning message is pushed to the management personnel through the remote monitoring platform, prompting manual review; when the risk is medium, the full-area sound and light alarm is activated, and the charging pile power is cut off for protection; the smoke exhaust system pre-mode is activated, the smoke diffusion path is calculated in real time, and the direction of the smoke exhaust valve is pre-adjusted; when the risk is high, the fire source is located according to the thermal imaging data, and fine water mist is released first to suppress the initial fire; if the temperature gradient continues to rise and the temperature gradient is less than 5℃ / s, the fire is extinguished with full coverage of perfluorohexanone gas; the main power supply of the parking lot is simultaneously cut off, and the emergency lighting and evacuation guidance system is activated; a three-dimensional thermal map, vehicle location and battery type data are sent to the city fire command center through a blockchain encrypted channel;

[0149] A dynamic strategy optimization mechanism is used to establish a control instruction execution effect evaluation model. The fire extinguishing efficiency is calculated through real-time data from the fire extinguishing system pressure sensor and camera flame recognition results. If the efficiency is lower than the preset threshold, the strategy adjustment is automatically triggered: the historical case library in the knowledge graph is called to match the optimal strategy for similar scenarios. Conversely, the Q-Learning reinforcement learning algorithm is used to dynamically update the control strategy weights to achieve adaptive optimization.

[0150] Secondly, a fire prevention and control system for electric vehicle parking lots based on multi-source information fusion includes:

[0151] Data acquisition module: used in the perception layer to capture the environmental status of electric vehicle parking spaces and the operating parameters of vehicle batteries in real time through a multimodal sensor network;

[0152] Communication module: used for the network layer to adopt intelligent fusion network architecture to achieve communication between the perception layer and the data processing layer;

[0153] Evaluation strategy generation module: used in the data processing layer to perform integrated analysis on the environmental status and the operating parameters, determine the fire risk level, and generate dynamic control instructions; the data processing layer is composed of a multi-dimensional analysis engine, and adopts a hierarchical and progressive intelligent analysis architecture to achieve multi-level analysis and judgment of fire risks; the data processing layer includes a heterogeneous data preprocessing module, a multi-dimensional feature extraction module, a fire risk assessment module, and a dynamic control strategy generation module; including:

[0154] Based on the feature information extracted by the multidimensional feature extraction module, a fire risk assessment model is constructed using a three-level progressive analysis framework to evaluate the fire risk level of electric vehicle parking lots. The fire risk level includes level one, level two, and level three assessments.

[0155] Positioning and rescue module: used for the application layer to receive the dynamic control instructions generated by the data processing layer, and realize the full-process closed-loop management of fire prevention and control through multimodal terminal equipment; the application layer includes an audio and visual alarm module, a smoke exhaust system control module, a fire extinguishing system control module, a remote monitoring platform, and an emergency evacuation and fire extinguishing and rescue module.

[0156] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A fire prevention and control method for electric vehicle parking lots based on multi-source information fusion, characterized in that: The following steps are involved: The perception layer captures the environmental status of electric vehicle parking spaces and the operating parameters of vehicle batteries in real time through a multimodal sensor network; The network layer adopts an intelligent converged network architecture to achieve communication between the perception layer and the data processing layer; The data processing layer is used to perform a fusion analysis of the environmental status and the operating parameters, determine the fire risk level, and generate dynamic control instructions; The data processing layer is composed of a multi-dimensional analysis engine, which adopts a hierarchical and progressive intelligent analysis architecture to achieve multi-level analysis and judgment of fire risks. The data processing layer includes a heterogeneous data preprocessing module, a multi-dimensional feature extraction module, a fire risk assessment module, and a dynamic control strategy generation module. It includes: Based on the feature information extracted by the multidimensional feature extraction module, a fire risk assessment model is constructed using a three-level progressive analysis framework to evaluate the fire risk level of electric vehicle parking lots. The fire risk level includes level one, level two, and level three assessments. The application layer receives the dynamic control instructions generated by the data processing layer and implements full-process closed-loop management of fire prevention and control through multimodal terminal devices; the application layer includes an audio and visual alarm module, a smoke exhaust system control module, a fire extinguishing system control module, a remote monitoring platform, and an emergency evacuation and fire extinguishing rescue module; The first-level assessment includes: Use the deep neural network pre-trained based on historical environment states and operating parameters and load its weight parameters; Local data from the target parking lot was collected and the sample size was expanded through data augmentation. Transfer learning was used to freeze the underlying feature extraction layer of the model and adjust only the parameters of the top fully connected layer. The loss function was designed to combine cross entropy loss with KL divergence. Real-time sensor data is fed into the fine-tuned model, which outputs a fire risk probability value between 0 and 1. A threshold is set, marking it as "Level 1 High Risk" to trigger a Level 2 assessment. The secondary assessment includes: The Kalman filter is used to fuse sensor data of the same type to eliminate isolated point errors; the time series alignment algorithm is used to synchronize BMS data and environmental data to the millisecond level; Construct a Bayesian network topology whose nodes include sensor data, equipment health, and fire status; calculate the posterior probability based on the prior probability and conditional probability table; The improved DS evidence theory is used to perform uncertainty reasoning on multi-source sensor data, and the adaptive weighting algorithm is combined to dynamically assign trust weights to different sensors. A threshold is set and marked as "Level 2 Confirmed Risk" to trigger a Level 3 assessment. The Level 3 assessment method further includes: Define entities, including battery type, fire cause, environmental conditions, and extinguishing agent characteristics, and establish causal chains based on historical cases and domain knowledge; Input the current risk characteristics, traverse the knowledge graph through the graph neural network, match similar subgraphs; calculate the subgraph confidence and identify potential risk paths; Combining the first-level probability output and the second-level confidence level, the cause of the fire is deduced through the logic rule engine; if the knowledge graph reasoning result is consistent with the first-level and second-level assessments, a high-risk level is determined; if there is a contradiction, manual review is triggered.

2. The method for fire prevention and control in electric vehicle parking lots based on multi-source information fusion according to claim 1, characterized in that: The trust weight expression is: The health of the i-th device is , the environmental adaptation coefficient is , the dynamic correlation adjustment coefficient is , the global data set is U, and the i-th sensor data is , sensor data The historical correlation with the global dataset U is , the correlation change rate is ; The sliding window mechanism is used to statistically analyze the distribution of sensor data, the timeliness of data is analyzed through information entropy, and the Bayesian network conditional probability table is updated in real time.

3. The method for fire prevention and control in electric vehicle parking lots based on multi-source information fusion according to claim 1, characterized in that: The method for screening the multidimensional features further includes: Calculate the correlation between multidimensional features and fire, sort the multidimensional features in descending order according to the correlation, and select the multidimensional features with a correlation greater than 0.469 as the candidate feature set; Combine candidate features according to the univariate perturbation principle according to their dimensions to obtain a multidimensional feature set, introduce a particle swarm, and use the multidimensional feature set as particles; The prediction error of different multidimensional feature sets on the training set is calculated using the base model, and the fitness value of the particle is calculated based on the prediction error: The fitness function of the cth particle is , the number of dimensions is , the multidimensional feature set is Q, and the minimum prediction error is , the prediction error of the c-th multidimensional feature set is , the a-th multidimensional feature of the s-th dimension is , the z-th multidimensional feature of the k-th dimension is , multidimensional features and multidimensional features Joint probability distribution, multidimensional features The marginal probability distribution of , multidimensional features The marginal probability distribution of , the number of multidimensional features in the multidimensional feature set is ; The particle with the largest fitness value is used as the attacker, and the average Euclidean distance between the particle and the attacker is calculated; Select particles whose Euclidean distance to the attacker is greater than the average Euclidean distance and update the particle position. The expression is: The position of the cth particle in the t+1th iteration is , the upper bound of the search space is , the lower bound of the search space is , the position of the cth particle in the tth iteration is , the current iteration number is t, and the random number between 0 and 1 is , the attacker's position is ; After updating the particles, the fitness value is calculated. The attackers are divided into two groups according to the fitness value. The position of the particles is updated according to the search step length of the particles to obtain the mutation position. The expression is: The random number between 0 and 1 is 、 , the search step is , the mutation position of the cth particle in the t+1th iteration is , the maximum number of iterations is ; Until the fitness reaches the maximum, the corresponding multidimensional feature set is used as the output result.

4. The method for fire prevention and control in electric vehicle parking lots based on multi-source information fusion according to claim 1, characterized in that: The heterogeneous data preprocessing module and the multidimensional feature extraction module include: The heterogeneous data preprocessing module uses a spectral-spatial joint noise reduction algorithm, establishes a noise model based on sensor characteristics, and employs a sliding window mechanism to perform adaptive wavelet denoising on temperature / smoke time series data, achieving dynamic threshold noise reduction. A Kalman filter is applied to fuse data from similar sensors, eliminating isolated point errors and performing multi-source calibration. A convolutional neural network is used to identify interference sources such as device shadows and water vapor in the video stream, correcting false trigger signals and performing pattern recognition correction. The multidimensional feature extraction module is used to extract fire-related feature information from the preprocessed data and build a three-dimensional spatiotemporal feature extraction system, which includes time domain features, frequency domain features, spatial correlation features, and cross-modal features; The time domain features are used to calculate the dynamic indicators of battery temperature gradient and second-order derivative of VOCs concentration; the pressure sensor spectrum energy distribution is analyzed through FFT, and the characteristic frequency of micro-detonation inside the battery is detected as a spatial correlation feature; a sensor topology model is constructed to analyze the spatial propagation path of abnormal signals and determine the spatial correlation features; a twin network is used to align the feature space of thermal imaging and voltage fluctuation data as cross-modal features.

5. The method for fire prevention and control in electric vehicle parking lots based on multi-source information fusion according to claim 1, characterized in that: The dynamic control strategy generation module includes: Based on the risk level and feature analysis results output by the fire risk assessment module, dynamic control instructions are generated through a multimodal decision engine; the dynamic control strategy generation module includes hierarchical response strategies and instruction generation logic and a dynamic strategy optimization mechanism; The hierarchical response strategy and instruction generation logic generates differentiated control instructions based on the fire risk level output by the three-level assessment and the causal reasoning results of the knowledge graph: the risk levels include low risk, medium risk, and high risk; When the risk is low, the ventilation system in the target area is activated to reduce the concentration of combustible gases; a battery health check instruction is sent to the BMS to forcibly limit the charging current to a safe threshold; a yellow warning message is pushed to the management personnel through the remote monitoring platform, prompting manual review; when the risk is medium, the full-area sound and light alarm is activated, and the charging pile power is cut off for protection; the smoke exhaust system pre-mode is activated, the smoke diffusion path is calculated in real time, and the direction of the smoke exhaust valve is pre-adjusted; when the risk is high, the fire source is located based on thermal imaging data, and fine water mist is released to suppress the initial fire; if the temperature gradient continues to rise, the fire is fully extinguished with perfluorohexanone gas; the main power supply of the parking lot is simultaneously cut off, and the emergency lighting and evacuation guidance system are activated; a three-dimensional thermal map, vehicle location and battery type data are sent to the city fire command center through a blockchain encrypted channel; A dynamic strategy optimization mechanism is used to establish a control instruction execution effect evaluation model. The fire extinguishing efficiency is calculated through real-time data from the fire extinguishing system pressure sensor and camera flame recognition results. If the efficiency is lower than the preset threshold, the strategy adjustment is automatically triggered: the historical case library in the knowledge graph is called to match the optimal strategy for similar scenarios. Conversely, the Q-Learning reinforcement learning algorithm is used to dynamically update the control strategy weights to achieve adaptive optimization.

6. A fire prevention and control system for electric vehicle parking lots based on multi-source information fusion, used to execute the method according to any one of claims 1 to 5, characterized in that: include: Data acquisition module: used in the perception layer to capture the environmental status of electric vehicle parking spaces and the operating parameters of vehicle batteries in real time through a multimodal sensor network; Communication module: used for the network layer to adopt intelligent fusion network architecture to achieve communication between the perception layer and the data processing layer; Evaluation strategy generation module: used in the data processing layer to perform fusion analysis on the environmental status and the operating parameters, determine the fire risk level, and generate dynamic control instructions; The data processing layer is composed of a multi-dimensional analysis engine, which adopts a hierarchical and progressive intelligent analysis architecture to achieve multi-level analysis and judgment of fire risks. The data processing layer includes a heterogeneous data preprocessing module, a multi-dimensional feature extraction module, a fire risk assessment module, and a dynamic control strategy generation module. It includes: Based on the feature information extracted by the multidimensional feature extraction module, a fire risk assessment model is constructed using a three-level progressive analysis framework to evaluate the fire risk level of electric vehicle parking lots. The fire risk level includes level one, level two, and level three assessments. Positioning and rescue module: used for the application layer to receive the dynamic control instructions generated by the data processing layer, and realize the full-process closed-loop management of fire prevention and control through multimodal terminal equipment; the application layer includes an audio and visual alarm module, a smoke exhaust system control module, a fire extinguishing system control module, a remote monitoring platform, and an emergency evacuation and fire extinguishing and rescue module.

Images