Intelligent identification method for fire safety hazards based on multimodal perception

Abstract

This invention discloses an intelligent identification method for fire safety hazards based on multimodal perception. To accurately identify and efficiently handle battery safety hazards, a six-dimensional multimodal intelligent prevention and control system is constructed. First, a distributed sensor array collects six-dimensional data, including temperature and pressure, covering key indicators such as electrolyte characteristic gases and electrical parameters. After denoising and standardization using the 3σ criterion and sliding window, time alignment with a time deviation ≤10ms is achieved. A CNN-LSTM hybrid architecture combined with a cross-modal attention module is used to fuse features. Through a ResNet-Attention hybrid architecture model, and after transfer and incremental training inference, three hazard types and three hazard levels are output. Based on the level, a graded closed-loop response is triggered, executing strategies such as alarms, cooling, and power-off, and collecting feedback samples. The model is incrementally trained and optimized every 1-3 months or when the cumulative sample size is ≥1000 groups. This system achieves accurate identification, graded handling, and dynamic model optimization of battery hazards, providing technical support for battery safety prevention and control and ensuring operational safety.

Description

Technical Field

[0001] This invention relates to the field of fire safety, and in particular to an intelligent method for identifying fire safety hazards based on multimodal perception. Background Technology

[0002] With the rapid development of the new energy industry, fire safety issues related to electric vehicles and electrochemical energy storage equipment have become a key focus for the industry. The "Safety Requirements for Power Batteries for Electric Vehicles" (GB38031-2025) clearly stipulates that power batteries must possess capabilities for early thermal runaway alarm, fire prevention, explosion prevention, and smoke neutralization. Existing technologies have already been explored by relevant enterprises and research institutions. Geely Automobile (CN115830247A) disclosed a thermal runaway early warning method that integrates temperature, voltage, gas, and smoke data, but it lacks vibration sensing and electrolyte characteristic gas (such as hydrogen fluoride) monitoring, and cannot capture the physical precursors of cell expansion and rupture. Moreover, the handling only focuses on cooling and lacks graded linkage. Qianye Technology (CN116011238B) proposed a multi-sensor fire identification system for energy storage compartments, but it did not cover pressure, electrical parameters, and vibration data. The algorithm used a single improved CNN, which had weak anti-interference ability. The disposal strategy was only fire extinguishing agent spraying, and it lacked closed-loop optimization. The University of Science and Technology of China (CN114823579A) used the Attention-LSTM model to warn of thermal runaway, but the perception dimensions were insufficient, with no pressure or smoke monitoring, and no coordinated response or multi-scenario adaptation involved. Existing technologies generally use traditional dry powder, heptafluoropropane and other fire extinguishing agents, which have problems such as being toxic, corrosive and non-degradable, and are not suitable for occupied places or enclosed spaces.

[0003] In summary, the existing technology has the following core defects: 1. Incomplete perception dimensions, unable to fully capture the characteristics of all stages of potential hazards, and prone to false alarms and missed alarms; 2. Single algorithm architecture, lacking the ability to adapt to multiple scenarios and multiple battery types; 3. Fragmented handling strategies, lacking hierarchical closed-loop logic, and failing to form a full-link optimization of "identification-handling-feedback"; 4. Insufficient environmental friendliness, fire extinguishing agents are prone to causing secondary damage.

[0004] Therefore, a method for intelligent identification of fire safety hazards based on multimodal perception is proposed to solve the above problems. Summary of the Invention

[0005] This invention overcomes the shortcomings of the prior art and provides an intelligent identification method for fire safety hazards based on multimodal perception.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: an intelligent identification method for fire safety hazards based on multimodal perception, comprising the following steps: S1. Six-dimensional multimodal sensing data acquisition: Real-time acquisition of temperature data, pressure data, electrolyte characteristic gas concentration data, smoke concentration data, electrical parameter data, and vibration data of the monitoring area through a distributed sensor array; the electrolyte characteristic gas includes carbon monoxide, hydrogen fluoride, and hydrocarbon compounds; the electrical parameters include battery cell voltage, total voltage, and charging / discharging current; the sensor array includes distributed fiber optic temperature sensors, electrochemical gas sensors, laser smoke sensors, battery management system integrated sensors, piezoelectric vibration sensors, and high-precision pressure sensors; S2. Time series synchronization preprocessing: Outliers are removed by using the 3σ criterion and sliding window, and the dimensional differences are eliminated by min-max standardization. Six-dimensional data time series alignment is achieved based on timestamp synchronization technology, with a time deviation ≤10ms, resulting in a standardized time series dataset. S3. Hybrid Architecture Feature Fusion: Spatial features of data from each modality are extracted through a convolutional neural network, and temporal dynamic features are extracted through a long short-term memory network. The cross-modal attention module is combined to dynamically allocate the weights of each feature. The weight coefficients are adaptively adjusted according to the development stage of the hidden danger, and a unified dimension fused feature vector is output. S4. Transfer and incremental hybrid training model inference: The fused feature vector is input into the pre-trained hazard identification model. The model is based on the ResNet-Attention-CNN-LSTM hybrid architecture, which is adapted to multiple scenarios and battery types through transfer learning and dynamically optimized through incremental training. Types of potential hazards: Precursors to battery thermal runaway, electrolyte combustion, and metal short circuit fire; Output hazard levels: general hazard, major hazard, and serious hazard; S5. Tiered closed-loop response: Based on the level of hazard, corresponding response strategies are triggered, including tiered alarm, cooling and temperature reduction, high-voltage power cut-off, directional spraying of environmentally friendly fire extinguishing agent, pressure relief and decompression, and emergency transfer. At the same time, multimodal data after the response is collected to form feedback samples for incremental training of the model. S6. Dynamic Model Optimization: Every 1-3 months at preset intervals or when the cumulative number of feedback samples reaches a threshold of ≥1000 sets, incremental training is performed based on the feedback sample set to optimize model parameters and ensure continuous improvement in recognition accuracy.

[0007] In a preferred embodiment of the present invention, the parameters of the sensor array in step S1 satisfy the following: the temperature measurement range of the distributed optical fiber temperature sensor is -40℃ to 1200℃, and the accuracy is ±0.5℃. The electrochemical gas sensor has a detection limit of ≤0.5ppm for hydrogen fluoride and a response time of ≤2s. The laser smoke sensor has a detection concentration range of 0.001~20%obs / m; The piezoelectric vibration sensor has a measurement range of 10~1000Hz and a sensitivity of ≥10mV / g; the high-precision pressure sensor has a range of 0~5MPa and a response time of ≤10ms.

[0008] In a preferred embodiment of the present invention, the weight allocation logic of the cross-modal attention module in step S3 is as follows: in the early stage of potential hazards: the concentration of characteristic gas exceeds the standard and the temperature rises slowly, the weight of gas feature is 0.35, the weight of temperature feature is 0.25, the weight of electrical parameter is 0.2, and the remaining weight is allocated to pressure and smoke modes; The development stage of the hidden danger is characterized by a sudden increase in temperature and pressure. The weights for temperature characteristics are 0.4, pressure characteristics are 0.3, and vibration characteristics are 0.15. The remaining weights are allocated to the other modes. Combustion stage: Smoke generation and open flame appearance. Smoke feature weight 0.4, temperature feature weight 0.3, pressure feature weight 0.2, and the remaining weights are allocated to other modes. The sum of all modal weights in each stage is 1.

[0009] In a preferred embodiment of the present invention, the training process of the hazard identification model in step S4 includes the following steps: S41. Construct a multi-scenario sample library: covering normal operating condition data, full-stage data of potential hazards, and false trigger data for scenarios such as thermal runaway of electric vehicle batteries, fire of electrochemical energy storage chamber, and short circuit of live equipment. The sample labels refer to the "Safety Requirements for Power Batteries for Electric Vehicles" standard and fire hazard classification specifications. S42. Transfer learning adaptation: The ResNet-Attention-CNN-LSTM model is pre-trained based on a public fire dataset, and then fine-tuned using a multi-scene sample library to achieve an initial recognition accuracy of ≥99% in different scenarios. S43. Incremental Training Optimization: Employ mini-batch gradient descent and weight freezing strategies, updating only the fully connected layer parameters of the model to avoid overfitting. The model accuracy improves by ≥0.1% after each incremental training.

[0010] In a preferred embodiment of the present invention, the hierarchical closed-loop processing strategy in step S5 specifically includes: General hazards: Issue audible and visual alarm signals, push early warning information to the management terminal, increase the collection frequency to 10Hz for continuous monitoring, and upgrade the level if there is no relief within 10 minutes. Major hazard: Activate the cooling system, heat pump cooling or coolant circulation, shut down non-essential electrical equipment, issue personnel evacuation warnings, and output a hazard status assessment every 5 seconds; Major hazard: Immediately trigger high-voltage power outage, control the solenoid valve to open the directional spray of silicone nanostructured alcohol-resistant foam extinguishing agent at a spray pressure of 0.3~0.5MPa, release internal pressure through the linkage pressure relief valve, activate the emergency transfer mechanism, move the energy storage compartment or isolate the battery pack, and continue monitoring until the hazard is eliminated.

[0011] In a preferred embodiment of the present invention, the silica nanostructured alcohol-resistant foam fire extinguishing agent meets the following requirements: 28-day biodegradation rate ≥86%, acute inhalation toxicity LC50 >6.206mg / L, skin irritation score 0, safe fire extinguishing under 36kV power frequency voltage, and leakage current ≤0.06mA.

[0012] In a preferred embodiment of the present invention, the data acquisition frequency in step S1 is dynamically adjusted to 1Hz under normal operating conditions; When any modal data reaches 80% of the warning threshold of GB38031-2025, it will automatically be increased to 10Hz; When the warning threshold is reached, the frequency will be increased to 20Hz.

[0013] In a preferred embodiment of the present invention, the preset period in step S6 can be manually adjusted by the management terminal. During incremental training, the historical best model parameters are retained. If the accuracy of the new training model decreases by ≥0.5%, it automatically rolls back to the historical best parameters.

[0014] In a preferred embodiment of the present invention, it further includes S7, fault self-diagnosis: real-time monitoring of the sensor array and communication link status, with one main sensor and one redundant sensor configured at each key monitoring point. When the main sensor data exceeds the normal range or there is no data output for three consecutive acquisition cycles, the redundant sensor is automatically activated to replace it, and a device fault alarm is issued at the same time to ensure continuous sensing.

[0015] This invention addresses the shortcomings of the prior art and has the following beneficial effects: (1) This invention synchronously collects six-dimensional data, including temperature, pressure, electrolyte characteristic gas, smoke, electrical parameters, and vibration, through a distributed sensor array. This fully covers the entire chain of potential hazards, from the precursors of chemical decomposition (electrolyte characteristic gas) → physical morphological changes (pressure, vibration) → thermal runaway development (temperature, electrical parameters) → combustion spread (smoke), breaking through the limitations of existing technologies that rely on only 3-4 types of modal data. Combined with the high-precision parameter design of the sensor array, such as the detection limit of hydrogen fluoride ≤0.5ppm and the vibration measurement range of 10~1000Hz, it can accurately capture weak signals in the early stage of thermal runaway. With the fault self-diagnosis mechanism, through one main and one backup redundant sensor and clear switching logic, it avoids the perception interruption caused by sensor failure, significantly reduces the false alarm rate of hazard identification, and controls the false alarm rate at an extremely low level, laying the foundation for subsequent accurate handling.

[0016] (2) This invention adopts a ResNet-Attention-CNN-LSTM hybrid architecture, extracting spatial features through a convolutional neural network and temporal features through a long short-term memory network. Combined with the staged weight allocation logic of the cross-modal attention module, it achieves dynamic adaptation and deep fusion of features of each modality. In the early stage of potential hazards, it focuses on gas and temperature features; in the development stage, it strengthens temperature, pressure, and vibration features; and in the combustion stage, it focuses on smoke, temperature, and pressure features. The total weight of each stage is 1, avoiding feature redundancy or omission of key information. At the same time, the transfer and incremental hybrid training strategy can quickly adapt to multiple scenarios such as electric vehicles and electrochemical energy storage chambers and different battery types. The initial recognition accuracy is ≥99%, and the accuracy is improved by ≥0.1% after each incremental training. Compared with the existing single algorithm model, the recognition accuracy, anti-interference ability and scenario adaptability have all achieved a leapfrog improvement, and the response speed is more in line with the actual application requirements.

[0017] (3) This invention constructs a three-level closed-loop disposal strategy, triggering differentiated disposal actions for general, major, and critical hazards respectively. The logic is clear and feasible: general hazards are enhanced with monitoring and early warning; major hazards are activated with cooling and personnel alerts; critical hazards are linked with high-voltage power outage, environmental extinguishing agent spraying, pressure relief, and emergency transfer. The technical path for emergency transfer is clearly defined, including electric roller transfer of the energy storage compartment and mechanical isolation of the battery pack, which solves the problem of fragmented disposal and lack of clear implementation logic in existing technologies. During the disposal process, the data acquisition frequency is dynamically adjusted. Under normal operating conditions, 1Hz is used to take into account energy consumption. When the warning threshold is 80%, it is increased to 10Hz and when the threshold is reached, it reaches 20Hz to ensure high-frequency data support during the development stage of the hazard. At the same time, the data collected after disposal is fed back to the model for optimization, forming a closed-loop of perception-identification-disposal-optimization. It strictly conforms to the "Safety Requirements for Power Batteries for Electric Vehicles" standard, which greatly improves the accuracy and timeliness of hazard control.

[0018] (4) The silica nanostructured alcohol-resistant foam fire extinguishing agent selected in this invention meets the environmental protection requirements of ≥86% biodegradation rate in 28 days, low acute inhalation toxicity, and 0% skin irritation. It can also safely extinguish fires while energized under 36kV power frequency voltage with a leakage current ≤0.06mA, overcoming the shortcomings of traditional fire extinguishing agents that are toxic, corrosive, and unsuitable for occupied areas and enclosed spaces. This design makes this method applicable not only to new energy vehicles and electrochemical energy storage chambers, but also to scenarios with high environmental and safety requirements such as subways, ships, and enclosed machine rooms. Its application scope is significantly broader than existing technologies, giving it greater industrial value.

[0019] (5) This invention features a flexible parameter control mechanism. The preset period for dynamic model optimization can be manually adjusted via the management terminal, while retaining historically optimal model parameters. Automatic rollback occurs when accuracy decreases by ≥0.5%, avoiding performance degradation caused by incremental training overfitting or parameter drift. Combined with fault self-diagnosis and redundant sensor design, it ensures that the perception continuity is not affected when the sensor array and communication link fail, significantly improving the overall reliability of the system. It can adapt to the fire safety monitoring needs under long-term complex working conditions and reduce subsequent operation and maintenance costs. Attached Figure Description

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments; Figure 1 This is a flowchart of a preferred embodiment of the intelligent identification method for fire safety hazards based on multimodal perception.

[0021] Figure 2 This is a flowchart illustrating the logic of a preferred embodiment of the intelligent identification method for fire safety hazards based on multimodal perception. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0024] like Figure 1-2 As shown, the intelligent identification method for fire safety hazards based on multimodal perception is as follows.

[0025] S1. Six-dimensional multimodal sensing data acquisition: A distributed sensor array is employed to achieve full coverage of the physical, chemical, and electrical characteristics of potential hazards. It accurately collects six types of data: temperature, pressure, electrolyte characteristic gas concentration, smoke concentration, electrical parameters, and vibration. The sensor parameters and deployment logic are strictly matched to technical requirements. Specifically: Temperature monitoring: Distributed fiber optic temperature sensors are deployed with one sensing point for every two battery cells, covering key areas of the battery pack / energy storage compartment. The temperature measurement range is -40℃ to 1200℃ with an accuracy of ±0.5℃. It can accurately capture local high temperature points and temperature spread trends, meeting the full-range monitoring needs from the precursor to thermal runaway (50℃+) to the combustion stage (1000℃+). Pressure monitoring: High-precision pressure sensors are deployed on the top of the battery pack housing and in the energy storage compartment, with a range of 0~5MPa and a response time of ≤10ms. They monitor pressure changes caused by thermal runaway gas eruption in real time to prevent explosions caused by pressure accumulation. Electrolyte characteristic gas monitoring: The electrochemical gas sensor simultaneously monitors carbon monoxide, hydrogen fluoride, and hydrocarbon compounds. The detection limit for hydrogen fluoride is ≤0.5ppm and the response time is ≤2s, which can accurately capture early chemical precursors of thermal runaway and significantly improve early identification capabilities. Smoke monitoring: The laser smoke sensor is deployed at the top of the monitoring area, with a detection concentration range of 0.001~20%obs / m. It has strong anti-dust interference ability and can quickly identify smoke particles produced by combustion, avoiding the false triggering problem of traditional photoelectric sensors. Electrical parameter monitoring: By integrating sensors through the battery management system, the system can collect real-time data on individual battery cell voltage, total voltage, and charging / discharging current, accurately identifying short-circuit precursors such as sudden voltage drops and abnormal current fluctuations. Vibration monitoring: Piezoelectric vibration sensors are deployed at the bottom of the battery pack and the energy storage chamber shell. The measurement range is 10~1000Hz and the sensitivity is ≥10mV / g. They can capture vibration signals generated by cell expansion and rupture to help determine the stage of thermal runaway physical changes.

[0026] The data acquisition frequency is dynamically adjusted to balance real-time performance and energy consumption control: the acquisition frequency is 1Hz under normal operating conditions; when any modal data reaches 80% of the warning threshold of GB38031-2025, it is automatically increased to 10Hz; when the warning threshold is reached, it is further increased to 20Hz to ensure that high-frequency data support can be obtained during the development stage of potential hazards.

[0027] S2, Timing Synchronization Preprocessing: To address the issues of inconsistent multimodal data formats and poor time synchronization, a three-step preprocessing strategy is adopted to ensure data reliability and time-series consistency. Specifically: Outlier removal: A combination of the 3σ criterion and sliding window is used to remove instantaneous interference and extreme outliers from the sensor, thus avoiding the impact of noisy data on subsequent feature fusion and recognition accuracy; Data normalization: By using the min-max normalization method, the data of each modality are uniformly mapped to the [0,1] interval, which effectively eliminates the interference of different units such as temperature (°C), pressure (MPa), and concentration (ppm), laying the foundation for multimodal feature fusion; Time alignment: Based on high-precision timestamp synchronization technology, the six-dimensional sensor data is precisely aligned according to the time axis to ensure that the time deviation is ≤10ms, thus ensuring the collaborative effectiveness of multimodal data.

[0028] S3, Hybrid Architecture Feature Fusion: A hybrid architecture integrating CNN, LSTM, and Attention is adopted to address the technical limitation of a single model in simultaneously extracting spatial and temporal features, achieving efficient feature fusion. Specifically: Spatial feature extraction: Through the convolutional and pooling layers of the convolutional neural network, local spatial features of each modality of data are accurately extracted, such as the location of temperature peaks, gas concentration gradients, and smoke distribution range, providing support for hazard location. Temporal feature extraction: By using the forget gate, input gate, and output gate of the Long Short-Term Memory Network, the dynamic change trend of each modality data is captured, such as the temperature rise rate in 10 seconds, voltage fluctuation period, and pressure rise slope, so as to predict the development trend of potential hazards. Cross-modal attention weighting: An adaptive attention module is constructed to dynamically allocate feature weights for each modality according to defined logic, ensuring that the sum of the weights for each stage is 1: In the early warning stage (excessive concentration of characteristic gases, slow temperature rise), gas features are weighted at 0.35, temperature at 0.25, and electrical parameters at 0.2, with the remaining weights allocated to pressure and smoke modes; in the development stage (sudden temperature rise, pressure rise), temperature features are weighted at 0.4, pressure at 0.3, and vibration at 0.15, with the remaining weights allocated to other modes; in the combustion stage (smoke production, open flame appearance), smoke features are weighted at 0.4, temperature at 0.3, and pressure at 0.2, with the remaining weights allocated to other modes. This avoids feature redundancy or omission of key information caused by single weights, ultimately outputting a unified dimensional fused feature vector, ensuring accurate identification in the future.

[0029] S4. Inference of the model trained using a hybrid transfer and incremental training method: The hazard identification model is built on a ResNet-Attention-CNN-LSTM hybrid architecture. It addresses the challenges of adapting to multiple scenarios and battery types, as well as continuously optimizing accuracy, through a hybrid training strategy that combines transfer and incremental training. Specifically: Multi-scenario sample library construction: collect data on the entire stage of thermal runaway of lithium iron phosphate / ternary lithium batteries, fire data of energy storage compartments / electric vehicle battery packs, and false triggering data in different environments (high temperature / humidity / electromagnetic interference), with a cumulative sample size of ≥100,000 groups. The sample labels are formulated with reference to the "Safety Requirements for Power Batteries for Electric Vehicles" standard and fire hazard classification specifications. Transfer learning adaptation: First, the model is pre-trained based on public fire datasets (such as the NFPA FireDynamicsSimulator dataset), and then fine-tuned through a multi-scenario sample library to make the initial recognition accuracy of the model ≥99% in different scenarios such as electric vehicles and energy storage cabins, solving the pain point of traditional models requiring a large number of scenario-specific samples. Incremental training optimization: Employ mini-batch gradient descent and weight freezing strategies, only updating the parameters of the fully connected layers of the model to effectively avoid overfitting. Incremental training is performed every 1000 sets of feedback samples or every 1-3 months. After each incremental training, the model accuracy improves by ≥0.1%, ensuring that the model adapts to changes in the scene and the performance degradation caused by sensor aging.

[0030] Actual testing showed that the model's inference response time was ≤2.5s, the recognition accuracy was ≥99.8%, and the false alarm rate was ≤0.2%, which is significantly better than existing technologies. It can accurately output the type of hidden danger (precursor to battery thermal runaway, electrolyte combustion, metal short circuit fire) and the level of hidden danger (general hidden danger, major hidden danger, major hidden danger).

[0031] S5. Tiered closed-loop processing: Based on the GB38031-2025 standard and the fire hazard classification specification, a three-level closed-loop response strategy is constructed to achieve precise, safe, and environmentally friendly handling of potential hazards. General Hazard Handling: When a general hazard is identified (such as local temperature of 50-60℃ or slightly excessive concentration of characteristic gas), an audible and visual alarm signal is issued, and the warning information is pushed to the management terminal through the communication module. At the same time, the data acquisition frequency is increased to 10Hz for continuous monitoring. If there is no relief within 10 minutes, the level is automatically upgraded. Handling of major hazards: When a major hazard is identified (such as a temperature that continues to rise to 60-80℃ or a doubling of the concentration of characteristic gas), start the cooling system (using heat pump cooling or coolant circulation), shut down non-essential electrical equipment, issue a personnel evacuation warning, output a hazard status assessment every 5 seconds, and upgrade to a major hazard if the temperature / gas concentration continues to rise. Major Hazard Handling: When a major hazard is identified (such as a sudden temperature rise ≥80℃, sudden pressure change, or detection of smoke / open flame), the high-voltage power is immediately cut off, the solenoid valve is opened to directionally spray the silicone nanostructured alcohol-resistant foam extinguishing agent, and the spray pressure is controlled at 0.3~0.5MPa. At the same time, the pressure relief valve is activated to release internal pressure, and the emergency transfer mechanism (energy storage compartment relocation or battery pack isolation) is initiated. Continuous monitoring continues until the hazard is eliminated. The silicone nanostructured alcohol-resistant foam extinguishing agent meets the following requirements: 28-day biodegradability ≥86%, acute inhalation toxicity LC50 >6.206mg / L, skin irritation score 0, safe live-line extinguishing under 36kV power frequency voltage, leakage current ≤0.06mA. It solves the problems of existing extinguishing agents being toxic, corrosive, and non-degradable, and is suitable for populated areas and confined spaces.

[0032] Multimodal data is collected simultaneously during the disposal process to form feedback samples, providing data support for incremental model training and constructing a closed-loop link of identification-disposal-feedback.

[0033] S6. Dynamic Model Optimization: Based on the sample data fed back after processing, the model is dynamically optimized: incremental training is carried out regularly to optimize the model weight parameters and threshold settings, ensuring that the recognition accuracy continues to improve with the accumulation of application scenarios; the preset period (1-3 months) can be manually adjusted through the management terminal. The historical best model parameters are retained during the incremental training process. If the accuracy of the new training model decreases by ≥0.5%, it will automatically roll back to the historical best parameters to avoid performance degradation caused by parameter drift.

[0034] S7. Fault Self-Diagnosis: To ensure the continuity of the sensing link, a fault self-diagnosis step is added: real-time monitoring of the sensor array and communication link status. Each key monitoring point is configured with one main sensor and one redundant sensor. When the main sensor data exceeds the normal range for three consecutive acquisition cycles or there is no data output, the redundant sensor is automatically activated to replace it, and a device fault alarm is issued at the same time. This effectively avoids missed reports due to sensor failure and improves the overall reliability of the system.

[0035] Example 1: Application of Fire Safety Hazard Identification System for Large-Scale Electrochemical Energy Storage Power Stations This embodiment is applied to a 100MWh large-scale electrochemical energy storage power station. The power station uses lithium iron phosphate battery packs and is divided into 20 energy storage compartments, each containing 12 battery packs. It needs to achieve intelligent identification and handling of potential hazards such as thermal runaway, short circuit, and electrolyte leakage throughout the entire process. The specific implementation process is as follows: S1. Six-dimensional multimodal sensing data acquisition: Distributed sensor arrays are deployed according to the power station layout. Distributed fiber optic temperature sensors are configured in each battery pack with one sensing point for every two cells, covering key areas such as the positive and negative electrodes of the battery pack and cell clusters, with a temperature measurement range of -40℃ to 1200℃ and an accuracy of ±0.5℃. Two high-precision pressure sensors are deployed on the top and in each partition of each energy storage compartment, with a range of 0~5MPa and a response time of ≤10ms. One electrochemical gas sensor is configured for every three battery packs to simultaneously monitor carbon monoxide, hydrogen fluoride, and hydrocarbon compounds, with a detection limit of ≤0.5ppm for hydrogen fluoride and a response time of ≤2s. Laser smoke sensors are deployed at the four corners of the top of each energy storage compartment, with a detection concentration range of 0.001~20%obs / m. The battery management system integrates sensors to collect the voltage of each individual battery cell, the total voltage, and the charging and discharging current in real time. Piezoelectric vibration sensors are deployed at the bottom of each battery pack and in the middle of the energy storage compartment shell, with a measurement range of 10~1000Hz and a sensitivity of ≥10mV / g. Data acquisition frequency is dynamically adjusted: 1Hz under normal operating conditions, increased to 10Hz when any sensor data reaches 80% of the warning threshold of GB38031-2025, and increased to 20Hz when the warning threshold is reached.

[0036] S2. Time-series synchronization preprocessing: Outliers are removed by combining the 3σ criterion with the sliding window method, eliminating instantaneous noise data caused by wind and sand and electromagnetic interference; temperature, pressure, gas concentration and other different dimensional data are uniformly mapped to the [0,1] interval through min-max standardization; based on high-precision timestamp synchronization technology, the six-dimensional data are aligned according to the time axis, and the time deviation is controlled to ≤10ms to ensure effective collaboration of multimodal data.

[0037] S3. Hybrid Architecture Feature Fusion: A hybrid architecture of CNN-LSTM-Attention is adopted. CNN extracts spatial features such as temperature peak location, gas concentration gradient, and smoke distribution range; LSTM captures temporal features such as temperature 10s heating rate, voltage fluctuation period, and pressure rise slope; the adaptive attention module dynamically allocates weights: in the early stage of potential hazards, gas features have a weight of 0.35, temperature features have a weight of 0.25, electrical parameters have a weight of 0.2, and the remaining 0.2 is allocated to the pressure (0.1), smoke (0.05), and vibration (0.05) modes; in the development stage of potential hazards, temperature features have a weight of 0.4, pressure features have a weight of 0.3, vibration features have a weight of 0.15, and the remaining 0.15 is allocated to the gas (0.08), electrical parameters (0.05), and smoke (0.02) modes; in the combustion stage, smoke features have a weight of 0.4, temperature features have a weight of 0.3, pressure features have a weight of 0.2, and the remaining 0.1 is allocated to the other modes, outputting a unified dimension fused feature vector.

[0038] S4. Hybrid Training of Transfer and Incremental Model Inference: A multi-scenario sample library was constructed, collecting data on the entire stage of thermal runaway of lithium iron phosphate batteries, fire data of energy storage compartments, and false triggering data under high temperature / humidity / electromagnetic interference environments, accumulating 120,000 sets of samples. The labels refer to the "Safety Requirements for Power Batteries for Electric Vehicles" and the fire hazard classification standard. Based on the NFPA FireDynamicsSimulator public dataset, the model was pre-trained and then fine-tuned through the sample library, achieving an initial recognition accuracy of 99.2%. A mini-batch gradient descent and weight freezing strategy was adopted, with incremental training performed every 1,000 sets of feedback samples. After each training, the accuracy improved by ≥0.1%. The measured model inference response time was 2.2s, the recognition accuracy was 99.9%, and the false alarm rate was 0.15%, which can accurately identify the type and level of hazards.

[0039] S5. Tiered Closed-Loop Response: For general hazards (such as localized temperature of 52℃ or slightly excessive hydrogen fluoride concentration), an audible and visual alarm is issued and a warning is pushed to the operation and maintenance terminal. The data collection frequency is increased to 10Hz for continuous monitoring. If there is no relief within 10 minutes, the response is escalated. For larger hazards (such as temperature rising to 68℃ or hydrogen fluoride concentration doubling), the energy storage tank coolant circulation cooling system is activated, non-essential electrical equipment is shut down, and personnel evacuation warnings are issued. The status is assessed every 5 seconds, and the response is escalated if indicators continue to rise. For major hazards (such as a sudden temperature rise to 85℃, sudden pressure change, or smoke detection), high-voltage power is immediately triggered, the directional spraying of the silicone nanostructured alcohol-resistant foam extinguishing agent is initiated (spraying pressure 0.4MPa), the pressure relief valve is activated to release pressure, the energy storage tank isolation mechanism is activated, and monitoring continues until the hazard is resolved. This extinguishing agent has a 28-day biodegradability rate of 88%, meeting the power plant's environmental protection requirements and allowing for safe extinguishing of fires while the power is on. The response data is simultaneously used as feedback samples to support incremental model training.

[0040] S6. Dynamic Model Optimization: Incremental training is performed monthly to optimize model weights and thresholds, retaining historically optimal parameters. If the accuracy of the new model decreases by ≥0.5%, it will automatically roll back to ensure continuous improvement in recognition accuracy.

[0041] S7. Fault Self-Diagnosis: Each key monitoring point is equipped with primary and redundant dual sensors. When the primary sensor shows abnormal data or no output for three consecutive acquisition cycles, it automatically switches to the redundant sensor and issues a device fault alarm, avoiding missed detections due to sensor failure and improving system reliability. After the system was put into operation, the early identification time of potential hazards in the energy storage power station was advanced to 30 minutes before thermal runaway, and no incidents of hazard escalation occurred.

[0042] Example 2: Application of Fire Safety Hazard Identification Device for Electric Vehicle Power Battery Packs This embodiment is applied to the power battery pack of new energy electric vehicles (ternary lithium battery, capacity 80kWh). It needs to be adapted to multiple operating conditions such as vehicle driving, charging, and stationary conditions, and realize the real-time identification and rapid handling of potential hazards such as short circuits, thermal runaway, and electrolyte leakage in the battery pack. The specific implementation process is as follows: S1. Six-dimensional multimodal sensing data acquisition: Sensor deployment is optimized for the compact structure of the battery pack. Distributed fiber optic temperature sensors are deployed inside the battery pack with one sensing point for every two cells, covering key areas such as cell clusters and busbars, with a temperature measurement range of -40℃ to 1200℃ and an accuracy of ±0.5℃. A high-precision pressure sensor is deployed at the center of the top of the battery pack shell, with a range of 0~5MPa and a response time of ≤10ms, to monitor pressure changes caused by thermal runaway gas eruptions. A small electrochemical gas sensor is deployed on the side of the battery pack to simultaneously monitor carbon monoxide and hydrogen fluoride. Hydrocarbon compounds, with a detection limit of ≤0.5ppm for hydrogen fluoride and a response time of ≤2s; a laser smoke sensor is deployed on the inner top of the battery pack, resistant to dust and moisture interference, with a detection concentration range of 0.001~20%obs / m; sensors are integrated through the on-board battery management system (BMS) to collect real-time data on individual cell voltage, total voltage, and charging / discharging current, identifying voltage drops and abnormal current fluctuations; two piezoelectric vibration sensors are deployed at the bottom of the battery pack, with a measurement range of 10~1000Hz and a sensitivity of ≥10mV / g, to capture vibration signals of cell expansion and rupture. The data acquisition frequency is dynamically adjusted: 1Hz under driving and stationary conditions, and 5Hz by default under charging conditions; when any modal data reaches 80% of the warning threshold, it increases to 10Hz, and when it reaches the warning threshold, it increases to 20Hz, balancing real-time performance and on-board energy consumption control.

[0043] S2. Timing synchronization preprocessing: The 3σ criterion and sliding window method are used to remove abnormal data caused by vehicle bumps and electromagnetic interference; the dimensions of each modal data are unified to the [0,1] interval through min-max standardization; based on the vehicle-mounted high-precision time synchronization module, the six-dimensional data are aligned according to the time axis, and the time deviation is controlled to ≤10ms to ensure the consistency of data timing.

[0044] S3. Hybrid Architecture Feature Fusion: A hybrid architecture of CNN-LSTM-Attention is adopted. CNN extracts spatial features such as temperature peak location, gas concentration distribution, and vibration signal amplitude; LSTM captures temporal features such as temperature rise rate, voltage fluctuation period, and pressure change trend; the adaptive attention module dynamically allocates weights according to the working condition: in the early stage of potential hazards (such as slight excess of characteristic gas during charging), the gas feature weight is 0.35, the electrical parameter weight is 0.25, the temperature feature weight is 0.2, and the remaining weights are allocated to the pressure, smoke, and vibration modes; in the stage of potential hazard development (such as a sudden temperature rise during driving), the temperature feature weight is 0.4, the vibration feature weight is 0.2, the pressure feature weight is 0.25, and the remaining weights are allocated to the other modes; in the combustion stage, the smoke feature weight is 0.4, the temperature feature weight is 0.3, the pressure feature weight is 0.2, and the remaining weights are allocated to the other modes, outputting a fused feature vector.

[0045] S4. Transfer and Incremental Hybrid Training Model Inference: A multi-scenario sample library was constructed, collecting data on the entire stage of thermal runaway of ternary lithium batteries, electric vehicle battery pack fire data, and false triggering data under different driving conditions (high speed / low speed / climbing) and charging conditions, with a total sample size of 110,000 sets. Based on the pre-trained model using a publicly available fire dataset, after fine-tuning with the sample library, the initial recognition accuracy reached 99.1%. Using a mini-batch gradient descent and weight freezing strategy, incremental training was performed every 2 months, and the model was optimized by combining vehicle driving feedback handling samples, with an accuracy improvement of ≥0.1% after each training. The measured model inference response time was 2.0s, the recognition accuracy was 99.85%, and the false alarm rate was 0.18%, which can accurately distinguish between three types of hidden dangers and their corresponding levels: battery thermal runaway precursors, electrolyte combustion, and metal short circuit fires.

[0046] S5. Tiered Closed-Loop Response: For general hazards (such as local temperature of 55℃ when stationary, slight voltage fluctuations), an onboard audible and visual alarm is triggered, a warning message is pushed to the owner's mobile app, and the data collection frequency is increased to 10Hz for continuous monitoring. If there is no improvement within 10 minutes, the response is escalated. For larger hazards (such as temperature rising to 72℃ while driving, hydrogen fluoride concentration doubling), the onboard system automatically reduces charging power (charging condition) or limits vehicle speed (driving condition), activates the battery pack cooling system (heat pump cooling), and reminds the owner to park at the nearest station. For major hazards (such as temperature suddenly rising to 82℃, smoke detection, sudden pressure change), a high-voltage power cut-off is immediately triggered, disconnecting the battery pack from the vehicle's electrical system. The battery pack's built-in silicone nanostructured alcohol-resistant foam fire extinguishing agent spray device (spray pressure 0.35MPa) is activated, and the battery pack pressure relief valve is activated to release pressure. At the same time, an emergency rescue message is pushed to the nearest service station. This fire extinguishing agent is suitable for enclosed vehicle spaces, is non-toxic and non-corrosive, and can safely extinguish fires under 36kV power frequency voltage. The processed data is simultaneously uploaded to the cloud and used as incremental training samples for the model.

[0047] S6. Dynamic Model Optimization: Every two months, incremental training data is synchronized through the cloud to optimize model parameters and thresholds, retaining historically optimal parameters. If the accuracy of the new model decreases by ≥0.5%, it will automatically roll back, adapting to scenarios such as sensor aging and battery performance degradation after long-term vehicle use.

[0048] S7. Fault Self-Diagnosis: Key sensing points within the battery pack (temperature, pressure, gas sensors) are equipped with primary and redundant dual sensors. When the primary sensor generates abnormal data or no output for three consecutive acquisition cycles, it automatically switches to the redundant sensor and simultaneously alerts the owner via the app to perform maintenance, preventing missed detections due to sensor malfunctions and ensuring driving safety. After this device is installed, the accuracy of identifying potential hazards in the vehicle's power battery pack is significantly improved, and the incidence of thermal runaway accidents is reduced by more than 95%.

[0049] Example 3: Application of Fire Safety Hazard Identification System for Sealed Energy Storage Cabin (Data Center Support) This embodiment is applied to a closed energy storage compartment (50MWh capacity, lithium iron phosphate battery pack) for a data center. This energy storage compartment has a closed structure and limited ventilation, requiring careful prevention of potential hazards such as thermal runaway, gas accumulation, and short circuits, while also meeting the uninterrupted power supply requirements of the data center. The specific implementation process is as follows: S1. Six-dimensional multimodal sensing data acquisition: A sensor array is deployed based on the structural characteristics of the sealed energy storage compartment. Distributed fiber optic temperature sensors are configured in each battery pack with one sensor point for every two cells, covering key areas around the cells, busbars, and cooling pipes. The temperature measurement range is -40℃ to 1200℃, with an accuracy of ±0.5℃. Two high-precision pressure sensors are deployed in each zone (every four battery packs constitute one zone), located at the top and side of each zone respectively, with a range of 0~5MPa and a response time ≤10ms. Two electrochemical gas sensors are deployed in each zone to simultaneously monitor carbon monoxide, hydrogen fluoride, and... Hydrocarbon compounds, with a detection limit of ≤0.5ppm for hydrogen fluoride and a response time of ≤2s, are monitored to capture early gas leak signals in real time. One laser smoke sensor is deployed every 5㎡ on the top of the energy storage compartment, resistant to dust interference, with a detection concentration range of 0.001~20%obs / m. Sensors are integrated into the battery management system to collect real-time data on individual cell voltage, total voltage, charging / discharging current, and loop resistance, accurately identifying short-circuit precursors. Piezoelectric vibration sensors are deployed at the bottom of each battery pack and at the four corners of the energy storage compartment shell, with a measurement range of 10~1000Hz and a sensitivity ≥10mV / g, capturing cell expansion and shell vibration signals. The data acquisition frequency is dynamically adjusted: 1Hz under normal operating conditions, increasing to 10Hz when data reaches 80% of the warning threshold, and further increasing to 20Hz when the warning threshold is reached, simultaneously triggering the energy storage compartment ventilation system to assist in controlling gas accumulation.

[0050] S2. Time-series synchronization preprocessing: The 3σ criterion and sliding window method are combined to remove abnormal data caused by dust and electromagnetic interference in the closed space; the dimensions of each modal data are unified to the [0,1] interval through min-max standardization to eliminate the impact of dimensional differences on fusion accuracy; based on high-precision timestamp synchronization technology, the six-dimensional data are precisely aligned according to the time axis with a time deviation of ≤10ms to ensure the effectiveness of multimodal data collaborative analysis.

[0051] S3. Hybrid Architecture Feature Fusion: A hybrid architecture of CNN-LSTM-Attention is adopted. CNN extracts spatial features such as temperature distribution gradient, peak gas concentration location, and smoke diffusion range to support the localization of hazards in confined spaces. LSTM captures temporal features such as temperature change rate, pressure rise slope, and voltage fluctuation pattern to predict the development trend of hazards. The adaptive attention module dynamically allocates weights: In the early stage of hazards (excessive gas concentration, slow temperature rise), gas features are weighted at 0.35, temperature features at 0.25, and pressure features at 0.15, with the remaining weights allocated to electrical parameters, smoke, and vibration modes; In the development stage of hazards (sudden temperature rise, pressure accumulation), temperature features are weighted at 0.4, pressure features at 0.3, and electrical parameters at 0.15, with the remaining weights allocated to other modes; In the combustion stage (smoke generation, open flame appearance), smoke features are weighted at 0.4, temperature features at 0.3, and pressure features at 0.2, with the remaining weights allocated to other modes, outputting a unified dimension fused feature vector to ensure the accuracy of hazard identification in confined spaces.

[0052] S4. Hybrid Training of Transfer and Incremental Model Inference: A multi-scenario sample library was constructed, collecting data on the entire stage of thermal runaway of lithium iron phosphate batteries, fire data in sealed energy storage compartments, false triggering data under high temperature / high humidity / sealed environments, and data specific to data center supporting energy storage conditions, accumulating 130,000 sets of samples. Based on the NFPAFireDynamicsSimulator public dataset, the pre-trained model, after fine-tuning with the sample library, achieved an initial recognition accuracy of 99.3%. Employing a mini-batch gradient descent and weight freezing strategy, incremental training was performed every 1,000 sets of feedback samples or at one-month intervals, combined with disposal data to optimize the model, resulting in an accuracy improvement of ≥0.1% after each training. The measured model inference response time was 2.3s, the recognition accuracy was 99.92%, and the false alarm rate was 0.12%, accurately identifying the type and level of hazards while avoiding misjudgments caused by electromagnetic interference in the data center environment.

[0053] S5. Tiered Closed-Loop Response: For general hazards (such as local temperature of 53℃ or slightly excessive hydrogen fluoride concentration), an audible and visual alarm is issued, a warning is pushed to the data center operation and maintenance platform, the data collection frequency is increased to 10Hz, the corresponding zone ventilation system is activated, and the situation is escalated if there is no relief within 10 minutes; for major hazards (such as temperature rising to 70℃, pressure slowly rising, or gas concentration doubling), the energy storage compartment zone cooling system (coolant circulation) is activated, non-essential electrical equipment in the corresponding zone is shut down, and operation and maintenance personnel are reminded to conduct on-site investigations. A hazard status assessment is output every 5 seconds, and the situation is escalated if the indicators continue to rise; for critical hazards (such as temperature suddenly rising to 83℃, pressure sudden change, or smoke detection), a high-voltage power outage is immediately triggered to ensure that the data center main circuit is not affected, the silicone nanostructured alcohol-resistant foam extinguishing agent directional spray device (spray pressure 0.45MPa) is activated, the energy storage compartment pressure relief valve and emergency ventilation system are linked to quickly release pressure and suppress the fire, and the energy storage compartment isolation mechanism is activated to prevent the fire from spreading to the data center. Monitoring continues until the hazard is resolved. This fire extinguishing agent has a biodegradability rate of 87%, is non-toxic and non-corrosive, and is suitable for the environmental protection and safety requirements of confined spaces and data centers. It can safely extinguish fires while the circuit is powered. The disposal data is simultaneously used as feedback samples to support incremental model training and build a closed-loop system.

[0054] S6. Dynamic Model Optimization: Incremental training is performed monthly to optimize model weight parameters and threshold settings, retaining historically optimal parameters. If the accuracy of the newly trained model decreases by ≥0.5%, it automatically rolls back to the historically optimal parameters. This adapts to scenarios such as sensor aging and environmental changes after long-term operation in enclosed spaces, ensuring stable recognition accuracy.

[0055] S7. Fault Self-Diagnosis: Each critical sensing point in the partition is equipped with primary and redundant dual sensors. When the primary sensor shows abnormal data or no output for three consecutive acquisition cycles, it automatically switches to the redundant sensor and simultaneously pushes an equipment fault alarm to the operation and maintenance platform, arranging for personnel to repair it. This avoids missed detections due to sensor failures and ensures the safe operation of the sealed energy storage compartment and data center. After the system was implemented, the early identification efficiency of potential hazards in the data center's supporting energy storage compartment improved by 40%, and no incidents of potential hazards affecting the data center's power supply occurred.

[0056] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention. These are all equivalent modifications and improvements made to the above embodiments based on the essential technology of the present invention, and all of these fall within the protection scope of the present invention.

Claims

1. A method for intelligent identification of fire safety hazards based on multimodal perception, characterized in that, Includes the following steps: S1. Six-dimensional multimodal sensing data acquisition: Real-time acquisition of temperature data, pressure data, electrolyte characteristic gas concentration data, smoke concentration data, electrical parameter data, and vibration data of the monitoring area through a distributed sensor array; the electrolyte characteristic gas includes carbon monoxide, hydrogen fluoride, and hydrocarbon compounds; the electrical parameters include battery cell voltage, total voltage, and charging / discharging current; the sensor array includes distributed fiber optic temperature sensors, electrochemical gas sensors, laser smoke sensors, battery management system integrated sensors, piezoelectric vibration sensors, and high-precision pressure sensors; S2. Time series synchronization preprocessing: Outliers are removed by using the 3σ criterion and sliding window, and the dimensional differences are eliminated by min-max standardization. Six-dimensional data time series alignment is achieved based on timestamp synchronization technology, with a time deviation ≤10ms, resulting in a standardized time series dataset. S3. Hybrid Architecture Feature Fusion: Spatial features of data from each modality are extracted through a convolutional neural network, and temporal dynamic features are extracted through a long short-term memory network. The cross-modal attention module is combined to dynamically allocate the weights of each feature. The weight coefficients are adaptively adjusted according to the development stage of the hidden danger, and a unified dimension fused feature vector is output. S4. Transfer and incremental hybrid training model inference: The fused feature vector is input into the pre-trained hazard identification model. The model is based on the ResNet-Attention-CNN-LSTM hybrid architecture, which is adapted to multiple scenarios and battery types through transfer learning and dynamically optimized through incremental training. Types of potential hazards: precursors to battery thermal runaway, electrolyte combustion, and metal short circuit fires; Output hazard levels: general hazard, major hazard, and serious hazard; S5. Tiered closed-loop response: Based on the level of hazard, corresponding response strategies are triggered, including tiered alarm, cooling and temperature reduction, high-voltage power cut-off, directional spraying of environmentally friendly fire extinguishing agent, pressure relief and decompression, and emergency transfer. At the same time, multimodal data after the response is collected to form feedback samples for incremental training of the model. S6. Dynamic Model Optimization: Every 1-3 months at preset intervals or when the cumulative number of feedback samples reaches a threshold of ≥1000 sets, incremental training is performed based on the feedback sample set to optimize model parameters and ensure continuous improvement in recognition accuracy.

2. The intelligent identification method for fire safety hazards based on multimodal perception according to claim 1, characterized in that: The parameters of the sensor array in step S1 satisfy the following: the temperature measurement range of the distributed optical fiber temperature sensor is -40℃ to 1200℃, and the accuracy is ±0.5℃. The electrochemical gas sensor has a detection limit of ≤0.5ppm for hydrogen fluoride and a response time of ≤2s. The laser smoke sensor has a detection concentration range of 0.001~20%obs / m; The piezoelectric vibration sensor has a measurement range of 10~1000Hz and a sensitivity of ≥10mV / g; the high-precision pressure sensor has a range of 0~5MPa and a response time of ≤10ms.

3. The intelligent identification method for fire safety hazards based on multimodal perception according to claim 1, characterized in that: The weight allocation logic of the cross-modal attention module in step S3 is as follows: in the early stage of potential hazards: the concentration of characteristic gas exceeds the standard and the temperature rises slowly, the weight of gas characteristic is 0.35, the weight of temperature characteristic is 0.25, the weight of electrical parameter is 0.2, and the remaining weight is allocated to pressure and smoke modes; The development stage of the hidden danger is characterized by a sudden increase in temperature and pressure. The weights for temperature characteristics are 0.4, pressure characteristics are 0.3, and vibration characteristics are 0.

15. The remaining weights are allocated to the other modes. Combustion stage: Smoke generation and open flame appearance. Smoke feature weight 0.4, temperature feature weight 0.3, pressure feature weight 0.2, and the remaining weights are allocated to other modes. The sum of all modal weights in each stage is 1.

4. The intelligent identification method for fire safety hazards based on multimodal perception as described in claim 1, Its features are: The training process of the hazard identification model in step S4 includes the following steps: S41. Construct a multi-scenario sample library: covering normal operating condition data, full-stage data of potential hazards, and false trigger data for scenarios such as thermal runaway of electric vehicle batteries, fire of electrochemical energy storage chamber, and short circuit of live equipment. The sample labels refer to the "Safety Requirements for Power Batteries for Electric Vehicles" standard and fire hazard classification specifications. S42. Transfer learning adaptation: The ResNet-Attention-CNN-LSTM model is pre-trained based on a public fire dataset, and then fine-tuned using a multi-scene sample library to achieve an initial recognition accuracy of ≥99% in different scenarios. S43. Incremental Training Optimization: Employ mini-batch gradient descent and weight freezing strategies, updating only the fully connected layer parameters of the model to avoid overfitting. The model accuracy improves by ≥0.1% after each incremental training.

5. The intelligent identification method for fire safety hazards based on multimodal perception according to claim 1, characterized in that: The hierarchical closed-loop processing strategy in step S5 specifically includes: General hazards: Issue audible and visual alarm signals, push early warning information to the management terminal, increase the collection frequency to 10Hz for continuous monitoring, and upgrade the level if there is no relief within 10 minutes. Major hazard: Activate the cooling system, heat pump cooling or coolant circulation, shut down non-essential electrical equipment, issue personnel evacuation warnings, and output a hazard status assessment every 5 seconds; Major hazard: Immediately trigger high-voltage power outage, control the solenoid valve to open the directional spray of silicone nanostructured alcohol-resistant foam extinguishing agent at a spray pressure of 0.3~0.5MPa, release internal pressure through the linkage pressure relief valve, activate the emergency transfer mechanism, move the energy storage compartment or isolate the battery pack, and continue monitoring until the hazard is eliminated.

6. The intelligent identification method for fire safety hazards based on multimodal perception according to claim 5, characterized in that: The silica nanostructured alcohol-resistant foam fire extinguishing agent meets the following requirements: 28-day biodegradation rate ≥86%, acute inhalation toxicity LC50 >6.206mg / L, skin irritation score 0, safe for fire extinguishing under 36kV power frequency voltage, and leakage current ≤0.06mA.

7. The intelligent identification method for fire safety hazards based on multimodal perception according to claim 1, characterized in that: The data acquisition frequency in step S1 is dynamically adjusted to 1Hz under normal operating conditions. When any modal data reaches 80% of the warning threshold of GB38031-2025, it will automatically be increased to 10Hz; When the warning threshold is reached, the frequency will be increased to 20Hz.

8. The intelligent identification method for fire safety hazards based on multimodal perception according to claim 1, characterized in that: The preset period mentioned in step S6 can be manually adjusted through the management terminal. During incremental training, the historical best model parameters are retained. If the accuracy of the new training model decreases by ≥0.5%, it will automatically roll back to the historical best parameters.

9. The intelligent identification method for fire safety hazards based on multimodal perception according to claim 1, characterized in that: It also includes S7 and fault self-diagnosis: real-time monitoring of the sensor array and communication link status. Each key monitoring point is configured with one main sensor and one redundant sensor. When the main sensor data exceeds the normal range for three consecutive acquisition cycles or there is no data output, the redundant sensor is automatically activated to replace it, and at the same time, an equipment fault alarm is issued to ensure continuous sensing.

Images