A battery cluster-level fire linkage control method and system based on BMS

Abstract

The present invention discloses a battery cluster-level fire linkage control method and system based on a BMS. The method includes: predicting thermal runaway based on the real-time voltage, temperature, and gas concentration data of the battery cluster through a multi-source feature fusion network, and outputting a thermal runaway prediction level; based on the thermal runaway prediction level, utilizing a dynamic priority mapping algorithm, outputting a control instruction set containing execution priority and spatial positioning according to the battery cluster topology; utilizing a dual-channel redundant signal verification mechanism to execute firefighting actions based on the control instruction set, and outputting a real-time action feedback signal; and updating the parameter weights of the multi-source feature fusion network through a reinforcement learning model based on the real-time action feedback signal and the residual state data of the battery cluster for use in the next cycle of fire linkage control, thereby forming a closed-loop control link. Using the embodiments of the present invention, battery cluster-level fire linkage control can be implemented to enhance the safety protection capability of the battery system.

Description

Technical Field

[0001] The present invention belongs to the field of fire protection technology, and in particular to a battery cluster-level fire protection linkage control method and system based on a BMS. Background Art

[0002] With the widespread adoption of smart battery technology, particularly in electric vehicles, power plant energy storage, and large-scale energy storage systems, the safety of battery clusters, as core energy storage units, has become a growing concern. During operation, battery clusters can experience thermal runaway due to overcharging, short circuits, and elevated temperatures, leading to serious fires and explosions, posing a significant threat to equipment safety and personnel life.

[0003] Traditional battery management systems (BMS) are primarily responsible for monitoring battery status, controlling charging and discharging, and providing basic protection. However, in the face of sudden thermal runaway events, these systems lack the necessary response speed and coordinated control capabilities, making it difficult to implement rapid and effective fire warnings and coordinated firefighting measures. Furthermore, existing technologies often rely on a single information source for monitoring and lack the ability to integrate multi-source information, making it difficult to accurately determine the level of thermal runaway and its spread. Summary of the Invention

[0004] The purpose of the present invention is to provide a battery cluster-level fire linkage control method and system based on BMS to address the deficiencies in the prior art and to achieve battery cluster-level fire linkage control to enhance the safety protection capability of the battery system.

[0005] One embodiment of the present application provides a battery cluster-level fire linkage control method based on a BMS, the method comprising:

[0006] Based on the real-time voltage, temperature and gas concentration data of the battery cluster, thermal runaway prediction is performed through a multi-source feature fusion network. The multi-source feature fusion network integrates the SOC gradient change characteristics and the SOH decay curve characteristics to output the thermal runaway prediction level;

[0007] Based on the thermal runaway prediction level, a dynamic priority mapping algorithm is used to collaboratively plan the aerosol injection range and the thermal runaway isolation area according to the battery cluster topology, and a control instruction set including execution priority and spatial positioning is output;

[0008] According to the control instruction set, the aerosol spray fire extinguishing device and the thermal runaway isolation device are driven by a pulse trigger circuit powered by a residual power supply, a dual-channel redundant signal verification mechanism is used to execute firefighting actions, and a real-time action feedback signal is output;

[0009] Based on the real-time action feedback signal and the battery cluster residual state data, the parameter weights of the multi-source feature fusion network are updated through a reinforcement learning model for use in the next cycle of fire linkage control to form a closed-loop control link.

[0010] Optionally, the thermal runaway prediction is performed based on the real-time voltage, temperature and gas concentration data of the battery cluster through a multi-source feature fusion network, wherein the multi-source feature fusion network fuses the SOC gradient change characteristics and the SOH decay curve characteristics to output the thermal runaway prediction level, including:

[0011] Collect the time series fluctuation data of each cell voltage in the battery cluster, the spatial distribution heat map of the temperature sensor, and the gas concentration change rate, synchronize and align the timestamps, and generate a cross-dimensional original monitoring matrix;

[0012] The original monitoring matrix is ​​corrected for outliers, and the SOC gradient change characteristics are extracted using the sliding window difference method. The SOH decay curve is fitted based on the cyclic aging experimental data to generate the SOC-SOH joint degradation feature vector.

[0013] ‌Build a multi-source feature fusion network, input the SOC-SOH joint degradation feature vector and real-time temperature and gas concentration data into a bidirectional gated recurrent unit, dynamically assign weight coefficients for the voltage, temperature, and gas channels through a feature-level attention mechanism, and output the fused thermal runaway feature code;

[0014] The thermal runaway feature encoding is input into the pre-trained multi-level classifier, and the warning level is divided according to the threshold boundaries of historical thermal runaway cases. The thermal runaway prediction level with confidence is generated and the spatiotemporal location labels are marked.

[0015] Optionally, based on the thermal runaway prediction level, a dynamic priority mapping algorithm is used to collaboratively plan the aerosol injection range and the thermal runaway isolation area according to the battery cluster topology, and output a control instruction set containing execution priority and spatial positioning, including:

[0016] Analyze the spatiotemporal labels of thermal runaway prediction levels, construct an adjacency matrix based on the physical topology of the battery cluster, simulate the maximum energy diffusion range of the thermal runaway propagation path, and generate a potential risk propagation map;

[0017] In the potential risk propagation map, with the battery cell predicted to be at an emergency level as the center, the aerosol spray coverage radius and the safe isolation distance between adjacent clusters are calculated to generate the initial control area boundary;

[0018] ‌A dynamic priority mapping algorithm is used to perform weighted optimization of the initial control area, combining the parallel string relationship of the battery cluster and the distribution of heat dissipation channels;

[0019] ‌ Based on the weighted optimization results, the concentric circle scope of the aerosol injection is divided, the deployment coordinates of the thermal runaway isolation device are marked, and a draft of the three-dimensional space control instructions is generated;

[0020] ‌ Conflict detection is performed on the draft three-dimensional space control instructions, overlapping areas of the equipment motion range are eliminated, and the final control instruction set containing execution priority and spatial positioning is output.

[0021] Optionally, according to the control instruction set, the aerosol spray fire extinguishing device and the thermal runaway isolation device are driven by a pulse trigger circuit powered by a residual power supply, a dual-channel redundant signal verification mechanism is used to execute fire fighting actions, and a real-time action feedback signal is output, including:

[0022] Activate the emergency power supply module of the BMS, use a capacitor energy storage pulse generator to generate a high-voltage trigger signal, encode the control instruction set into two modulated waveforms with a phase difference of 90 degrees, and transmit them to the execution terminal through an independent channel;

[0023] A dual-channel redundancy check module is deployed on the aerosol spray device to demodulate and cross-check the two modulated waveforms. When the instruction consistency exceeds the threshold, an execution permission signal is generated; otherwise, a self-check loop is triggered.

[0024] The high-speed solenoid valve is driven according to the execution permission signal, the injection angle and dosage are adjusted according to the spatial positioning parameters of the control instruction set, and the telescopic deflector of the thermal runaway isolation device is synchronously activated to form a physical barrier;

[0025] The infrared camera captures the diffusion pattern of the fire extinguishing agent, and combined with the pressure sensor to feedback the airflow intensity, an action execution effect evaluation matrix is ​​generated as a real-time action feedback signal;

[0026] The residual of the real-time action feedback signal and the expected control target is calculated. If the deviation exceeds the safety threshold, secondary injection is triggered and the fault code is updated to the BMS log system.

[0027] Optionally, based on the real-time action feedback signal and the battery cluster residual state data, the parameter weights of the multi-source feature fusion network are updated through a reinforcement learning model for use in the fire linkage control of the next cycle to form a closed-loop control link, including:

[0028] Integrate real-time action feedback signals with battery cluster residual state data to construct a multi-dimensional reinforcement learning state vector;

[0029] Design a reward function based on policy gradient, where positive rewards include thermal runaway suppression efficiency and resource consumption ratio, and negative penalties include the number of misoperations and equipment loss value, to generate a dynamic reward score;

[0030] An asynchronous advantage actor-critic algorithm is used to update the parameters of the multi-source feature fusion network. The actor network optimizes the feature weight distribution strategy, and the critic network corrects the thermal runaway level prediction deviation.

[0031] The updated network parameters are encrypted and synchronized to all BMS nodes, and the fire linkage control cycle counter is reset to form a closed-loop control link from state perception to strategy optimization.

[0032] Another embodiment of the present application provides a battery cluster-level fire linkage control system based on a BMS, the system comprising:

[0033] A prediction module is used to predict thermal runaway based on the real-time voltage, temperature, and gas concentration data of the battery cluster through a multi-source feature fusion network. The multi-source feature fusion network integrates the SOC gradient change characteristics and the SOH decay curve characteristics to output a thermal runaway prediction level;

[0034] a planning module for collaboratively planning the aerosol spray range and the thermal runaway isolation area based on the thermal runaway prediction level and a dynamic priority mapping algorithm according to the battery cluster topology, and outputting a control instruction set including execution priority and spatial positioning;

[0035] an execution module, configured to drive the aerosol spray fire extinguishing device and the thermal runaway isolation device according to the control instruction set through a pulse trigger circuit powered by a residual power supply, execute firefighting actions using a dual-channel redundant signal verification mechanism, and output a real-time action feedback signal;

[0036] An updating module is used to update the parameter weights of the multi-source feature fusion network through a reinforcement learning model based on the real-time action feedback signal and the battery cluster residual state data, so as to be used for the fire linkage control in the next cycle and form a closed-loop control link.

[0037] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute any of the above methods when run.

[0038] Yet another embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the above methods.

[0039] Compared with the prior art, the present invention provides a battery cluster-level fire linkage control method based on BMS. According to the real-time voltage, temperature and gas concentration data of the battery cluster, thermal runaway prediction is performed through a multi-source feature fusion network, and a thermal runaway prediction level is output; based on the thermal runaway prediction level, a dynamic priority mapping algorithm is used to output a control instruction set including execution priority and spatial positioning according to the battery cluster topology; according to the control instruction set, a dual-channel redundant signal verification mechanism is adopted to execute fire actions and output a real-time action feedback signal; based on the real-time action feedback signal and the residual state data of the battery cluster, the parameter weights of the multi-source feature fusion network are updated through a reinforcement learning model for use in the next cycle of fire linkage control, forming a closed-loop control link, thereby realizing battery cluster-level fire linkage control to enhance the safety protection capability of the battery system. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A hardware structure block diagram of a computer terminal for a BMS-based battery cluster-level fire linkage control method provided in an embodiment of the present invention;

[0041] Figure 2 A flow chart of a BMS-based battery cluster-level fire linkage control method provided in an embodiment of the present invention;

[0042] Figure 3 A schematic structural diagram of a BMS-based battery cluster-level fire linkage control system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0044] The embodiment of the present invention first provides a battery cluster-level fire linkage control method based on BMS. The method can be applied to electronic devices such as computer terminals, specifically ordinary computers.

[0045] The following describes it in detail by taking running on a computer terminal as an example. Figure 1 The hardware structure block diagram of a computer terminal for a BMS-based battery cluster-level fire linkage control method provided in an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.

[0046] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, which, when executed, can cause the processor to execute any one of the battery cluster-level fire linkage control methods based on the BMS.

[0047] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0048] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any one of the battery cluster-level fire linkage control methods based on the BMS.

[0049] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 1 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0050] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0051] See also Figure 2 , an embodiment of the present invention provides a battery cluster-level fire linkage control method based on BMS, which may include the following steps:

[0052] S201, based on the real-time voltage, temperature, and gas concentration data of the battery cluster, thermal runaway prediction is performed using a multi-source feature fusion network. The multi-source feature fusion network integrates the SOC gradient change characteristics and the SOH decay curve characteristics to output a thermal runaway prediction level. Specifically, this may include:

[0053] S2011: Collect the time series fluctuation data of each cell voltage in the battery cluster, the spatial distribution heat map of the temperature sensor, and the gas concentration change rate, synchronize and align the timestamps, and generate a cross-dimensional original monitoring matrix;

[0054] This step uses high-precision data acquisition and multi-source heterogeneous data alignment technology to construct a basic data set for full-dimensional condition monitoring of battery clusters, providing standardized input for subsequent feature fusion.

[0055] Multi-source data acquisition system deployment: Voltage time-series fluctuation data acquisition: The BMS's battery management unit (BMU) is equipped with a 16-bit ADC (Analog-to-Digital Converter) to collect real-time voltage data for each cell within the battery cluster at a 1kHz sampling rate. For example, a battery cluster in a certain energy storage power station consists of 96 cells connected in series (nominal voltage 3.2V). Each BMS node (such as the TI BQ76952 chip) monitors 12 cells, and voltage data is uploaded to the main control unit via the CAN bus (baud rate 500kbps).

[0056] Temperature Spatial Distribution Heat Map Generation: Eight DS18B20 digital temperature sensors (accuracy ±0.5°C) are evenly deployed within each module in the battery cluster, arranged in a 4×2 grid. The main control unit constructs a two-dimensional temperature field based on the sensor coordinates and generates a real-time heat map with a 0.1°C resolution using a bilinear interpolation algorithm. For example, the interpolated temperature at the center of a module (X=2.5, Y=1.5) is 45.3°C.

[0057] Gas Concentration Rate Monitoring: An NDIR (non-dispersive infrared) gas sensor (such as the Figaro TGS823) is installed on top of the battery cluster to collect CO, H2, and VOC concentrations at a 5-second interval. The gas concentration rate of change (ΔC / Δt) is calculated using the backward difference method. For example, if the CO concentration is 50 ppm at t = 10:00 and 65 ppm at t = 10:05, ΔC / Δt = (65 - 50) / 300 seconds = 0.05 ppm / s.

[0058] Timestamp synchronization and alignment technology:

[0059] Global clock synchronization: The master control unit uses a GPS module (such as the U-blox NEO-M8N) to obtain the PPS (pulse per second) signal as a reference clock source (with an error of <1μs). Each BMS and sensor node synchronizes its local clock using the NTPv4 (Network Time Protocol) protocol, ensuring system-wide timestamp alignment accuracy of <10ms.

[0060] Cross-dimensional data alignment: The raw data streams of voltage (1kHz), temperature (1Hz), and gas (0.2Hz) are interpolated onto a unified time axis (e.g., 1-second granularity) by timestamp. For example, the voltage data sampled at t=10:00:00.500 seconds is linearly interpolated to match the temperature data's timestamp of t=10:00:00.

[0061] Original monitoring matrix construction: The matrix dimensions are [time step × feature dimension]. For example, a 10-minute monitoring window (600 seconds) corresponds to 600 rows, each containing 96 cell voltages, 64 temperature points (8 sensors × 8 interpolation points), and three gas concentrations and rates of change, for a total of 164 columns. Data is stored in Float32 format, occupying a memory size of 600 × 164 × 4, approximately 393 KB.

[0062] Application Example: In a cascaded energy storage project, battery cluster B-07 experienced local overheating (sensor T23 reported 48°C), while the voltage of cell V33 dropped by 0.15V (3.05V to 2.90V). At t = 14:23:05 seconds, the main control unit aligned the voltage, temperature, and gas data, generating the following eigenvalues ​​for the corresponding row in the original monitoring matrix: voltage column V33 = 2.90V, temperature column T23 = 48.0°C, and CO concentration = 120 ppm (ΔC / Δt = 0.2 ppm / s), flagging this as a potential thermal runaway event.

[0063] In S2012, the original monitoring matrix is ​​corrected for outliers, and the SOC gradient change characteristics are extracted using the sliding window difference method. The SOH decay curve is fitted based on the cyclic aging experimental data to generate the SOC-SOH joint degradation feature vector.

[0064] This step quantifies the dynamic coupling relationship between battery health status and state of charge through joint modeling of data cleaning and degradation characteristics, providing key degradation indicators for thermal runaway prediction.

[0065] ‌Outlier Correction Strategy‌:

[0066] Voltage anomaly detection: This function eliminates voltage outliers based on the 3σ principle (three times the standard deviation). For example, if the normal voltage fluctuation range of a cell is 2.8V-3.6V (μ = 3.2V, σ = 0.1V), and a sampled value of 2.5V (exceeding μ-3σ = 2.9V) is replaced with the median value of the previous 10-second sliding window (e.g., 3.15V).

[0067] Temperature Spatial Smoothing: Performs median filtering on sudden changes in the heat map (temperature differences > 5°C). For example, if the temperature at a point is 50°C and the temperatures of its eight neighboring areas are [45, 46, 48, 47, 52, 49, 46, 44], with a median of 47°C, the temperature at that point will be corrected to 47°C.

[0068] Gas Data Compensation: To account for gas sensor response delays (e.g., a 20-second lag for CO sensors), an ARIMA (Autoregressive Integrated Moving Average) model is used to predict real-time values. For example, if the actual reading at time t is 80 ppm and the model predicts 95 ppm, the predicted value prevails.

[0069] ‌SOC gradient feature extraction‌:

[0070] Sliding Window Difference Method: Using a 30-second window (30 voltage sampling points), calculate the SOC change rate ΔSOC / Δt. SOC estimation uses the ampere-hour integration method: SOC(t)=SOC(t0)+∑(I×Δt) / Capacity, where I is the current (from the Hall effect sensor, with an accuracy of ±1A) and Capacity is the rated capacity (e.g., 280Ah). For example, if the average current within the window is -50A (discharge), then ΔSOC=(-50A×30s) / 280Ah≈-0.0446 (i.e., a 4.46% decrease in SOC).

[0071] Gradient change feature encoding: A ΔSOC sequence (e.g., 20 ΔSOC values ​​within a 600-second window) is fed into a one-dimensional convolutional layer (kernel size = 5, stride = 5) to extract local gradient patterns. For example, a convolutional output feature of [0.12, -0.08, 0.05] indicates a rapid decrease in SOC followed by a rebound.

[0072] ‌SOH decay curve fitting‌:

[0073] Cycle aging data: Capacity decay curves are recorded through laboratory accelerated aging tests (e.g., 1000 charge-discharge cycles at 0.5C). A cubic polynomial fit is used: SOH(n)=a×n³ + b×n² + c×n + d, where n is the number of cycles. For example, if a battery has a SOH of 92% at n=500 cycles, the fitting parameters are a=-1.2e-7, b=0.0003, c=-0.021, and d=100.

[0074] Real-time SOH estimation: Calculates the theoretical SOH based on the current cycle count (from the BMS log) and the fitted curve, then compares it with the actual capacity (calibrated through full charge and discharge cycles) to generate a residual compensation term. For example, if the theoretical SOH is 88% and the measured capacity is 85% of the rated capacity, the compensation term is -3%.

[0075] Joint feature vector generation: The SOC gradient feature (3D) and the SOH decay feature (4D: theoretical SOH, compensation value, cycle count, and capacity residual) are concatenated into a 7-dimensional vector, which is then normalized and used as input. For example, the vector value is [0.12, -0.08, 0.05, 0.85, -0.03, 620, -0.02].

[0076] Application Example: In a power battery recycling scenario, the SOC of battery cluster C-12 dropped dramatically from 65% to 58% (ΔSOC = -7%) within 30 seconds, exceeding the normal decay rate (expected ΔSOC ≈ -2%). The sliding window differencing method detected the abnormal gradient pattern (convolution feature = [-0.25, 0.1, 0.05]). Combined with the SOH fitting results (theoretical SOH = 76%, measured = 70%), a joint feature vector was generated: [-0.25, 0.1, 0.05, 0.76, -0.06, 800, -0.07], triggering a thermal runaway warning.

[0077] In S2013, a multi-source feature fusion network was constructed. The SOC-SOH joint degradation feature vector and real-time temperature and gas concentration data were input into a bidirectional gated recurrent unit. The weight coefficients of the voltage, temperature, and gas channels were dynamically assigned through a feature-level attention mechanism, and the fused thermal runaway feature code was output.

[0078] This step uses a deep learning model to achieve adaptive fusion of multi-source heterogeneous features, capture the spatiotemporal propagation pattern of thermal runaway, and improve the accuracy of early warning.

[0079] Multi-source feature fusion network architecture:

[0080] ‌Input layer design‌: The network accepts three types of input:

[0081] Degraded features: 7-dimensional SOC-SOH joint vector (see previous step), mapped to 16 dimensions by the fully connected layer; Temperature data: 64-dimensional heat map flattened into a vector, compressed to 16 dimensions by CNN (convolution kernel 3×3, stride 2); Gas data: 3 gas concentrations and change rates (6 dimensions), mapped to 8 dimensions by the fully connected layer.

[0082] Bidirectional GRU (Gated Recurrent Unit): Each channel (degradation, temperature, and gas) is independently fed into a bidirectional GRU layer (hidden units = 32) to capture temporal dependencies. For example, the GRU output dimension for the degradation channel is 64 (32 for each direction), the temperature channel outputs 64, and the gas channel outputs 32.

[0083] ‌Attention Mechanism‌:

[0084] Feature-level attention: Generates weight coefficients for each channel. Calculation: Weight = Softmax(MLP(GRU output)), where the MLP (Multi-layer Perceptron) has 16 hidden units. For example, the weight of the degraded channel is 0.6, the temperature is 0.3, and the gas is 0.1.

[0085] ‌Spatiotemporal Attention‌: Assign weights to the spatial locations of the heatmap within the temperature channel. For example, the center region has a weight of 0.8, and the edges have a weight of 0.2.

[0086] Thermal runaway signature code generation:

[0087] Weighted fusion: The GRU outputs of each channel are multiplied by their weights and then concatenated. For example, the degenerate dimension is 64×0.6 = 38.4, the temperature is 64×0.3 = 19.2, and the gas is 32×0.1 = 3.2. The total dimension is 60.8, which is rounded to 61.

[0088] Feature compression: Generate a high-dimensional encoding through a fully connected layer (61→256 dimensions) using the ReLU activation function. For example, the encoded value at a certain moment is [0.8, -0.2, 1.5, ..., 0.3] (256 dimensions).

[0089] ‌Normalization‌: Use Layer Normalization to normalize the encoding to prevent gradient explosion.

[0090] Application Example: During a thermal runaway event, the temperature channel detected a rise in the module's center temperature to 60°C (weight = 0.8), the CO concentration in the gas channel reached 200 ppm (weight = 0.15), and the SOC gradient showed a sudden drop (weight = 0.7). After feature fusion, the spatiotemporal attention weight for the high-temperature region was increased to 0.9, and the resulting thermal runaway feature encoding highlighted the coordinated anomaly between temperature and SOC.

[0091] In S2014, the thermal runaway feature encoding is input into a pre-trained multi-level classifier, and the warning level is divided according to the threshold boundaries of historical thermal runaway cases. The thermal runaway prediction level with confidence is generated and the spatiotemporal location labels are marked.

[0092] This step maps abstract features into actionable early warning signals and locates the source of risk through classification decisions and confidence assessment.

[0093] ‌Multi-class classifier design‌:

[0094] ‌Level division‌: The warning level is divided into 3 levels:

[0095] Level 1 (low risk): The feature code overlaps with normal operating conditions by >70%, with a confidence level of <60%. Level 2 (medium risk): The code deviates from normal but does not reach the thermal runaway threshold, with a confidence level of 60%-85%. Level 3 (high risk): The code exceeds the historical case threshold, with a confidence level of >85%.

[0096] Classifier structure: Random forest (200 trees) cascaded with support vector machine (SVM):

[0097] First level: Random forest screening of suspected cases (recall rate > 95%); Second level: SVM (kernel function = RBF, C = 1.0) accurate classification, output Level 1-3.

[0098] Confidence calculation: Based on the classification probability (Random Forest vote share) and the distance between the feature encoding and the decision boundary (SVM Margin). For example, if a sample has a Random Forest vote share of 92% (Level 3) and an SVM Margin of 1.5 (> the threshold of 1.2), the final confidence is 90%.

[0099] ‌Space-time location marker‌:

[0100] Risk source location: Based on the area with the highest weight in the temperature heat map (e.g., X=3, Y=2) and the location of the abnormal SOC cell (e.g., V45), the risk source coordinates are marked in three-dimensional space (cluster number + module row and column number + cell index).

[0101] Spread Prediction: Based on the diffusion speed of historical cases (e.g., 0.5 m / s), estimate the range of modules that may be affected within the next 5 minutes. For example, if the current risk source is located in module M5, it is predicted that it will spread to module M4, M6, and M7.

[0102] Application Example: In a certain energy storage power station, the center temperature of module M3 in battery cluster D-09 reached 58°C. The SOC gradient feature encoding triggered a Random Forest Level 3 classification (88% confidence), resulting in an SVM Level 3 classification (Margin = 1.8) with a confidence level of 91%. The system identified module M3 (coordinates X = 2, Y = 1) as the source of risk and issued a warning that adjacent modules M2 and M4 might be affected within three minutes.

[0103] This step collects real-time data on the battery cluster's voltage, temperature, and gas concentration. This data is then integrated using a multi-source feature fusion network, combining the gradient characteristics of the SOC (State of Charge) (reflecting battery charge and discharge anomalies) with the decay curve characteristics of the SOH (State of Health) (indicating battery aging). The network employs a bidirectional gated recurrent unit (BiGRU) to extract temporal correlations and dynamically assigns weights to different data channels through an attention mechanism. Ultimately, a multi-level classifier outputs a thermal runaway prediction level (e.g., low risk, medium risk, high risk) and labels the specific spatiotemporal location. This addresses the misjudgment or missed detection issues inherent in traditional thermal runaway warning systems, which rely on a single parameter (e.g., temperature). Multi-dimensional data fusion improves prediction accuracy, enabling early identification of local or overall thermal runaway risks in the battery cluster, providing a basis for subsequent precise firefighting coordination.

[0104] S202, based on the thermal runaway prediction level, using a dynamic priority mapping algorithm, and in accordance with the battery cluster topology, collaboratively planning the aerosol spray range and the thermal runaway isolation area, and outputting a control instruction set including execution priority and spatial positioning; specifically, this may include:

[0105] S2021: Analyze the spatiotemporal labels of thermal runaway prediction levels, construct an adjacency matrix based on the physical topology of the battery cluster, simulate the maximum energy diffusion range of the thermal runaway propagation path, and generate a potential risk propagation map;

[0106] This step analyzes the spatiotemporal properties of the thermal runaway prediction results, combines them with the physical connection relationship of the battery cluster, and constructs a mathematical model for the propagation of thermal runaway. This accurately predicts the scope of risk spread and provides a spatial decision-making basis for subsequent fire extinguishing and isolation.

[0107] ‌Spatiotemporal label parsing technology‌:

[0108] Spatiotemporal label structure: Spatiotemporal labels are extracted from the thermal runaway prediction level output in the above steps. The labels contain four parts of information:

[0109] Timestamp: The moment when the thermal runaway prediction is triggered (e.g., 2024-07-15 14:23:05.500);

[0110] ‌Spatial coordinates‌: The three-dimensional location code of the risk source, such as battery cluster number C-07, module number M3 (row number 2, column number 3), and cell index V45;

[0111] Risk level: Level 1 / 2 / 3 (e.g. Level 3);

[0112] Confidence: The confidence percentage of the prediction result (e.g. 91%).

[0113] Label parsing process: The main control unit matches label fields using regular expressions. For example, the string "C-07_M3(2,3)_V45_L3_91%" is parsed to extract the cluster number C-07, module M3 coordinates (X=2, Y=3), cell V45, risk level 3, and confidence level 91%.

[0114] Battery Cluster Topology Modeling:

[0115] Physical topology definition: A battery cluster consists of multiple modules connected in parallel or series, with each module containing several single cells. For example, a battery cluster consists of six modules connected in parallel (each group has 16 cells connected in series), with a total voltage of 16 × 3.2V = 51.2V and a capacity of 280Ah × 6 = 1680Ah.

[0116] Adjacency Matrix Construction: Based on the electrical connection relationship between modules (e.g., parallel string connection) and physical spacing (e.g., module spacing of 0.5 meters), an undirected graph adjacency matrix is ​​constructed. Matrix element A[i][j] represents the connection strength between module i and module j. If they are directly connected in parallel, A[i][j] = 1; if the spacing is greater than 1 meter, A[i][j] = 0.2. For example, if module M3 is adjacent to M2 and M4 and connected in parallel, A[3][2] = A[3][4] = 1; if it is 1.2 meters away from M1, A[3][1] = 0.2.

[0117] Thermal runaway propagation simulation algorithm:

[0118] Energy Diffusion Model: Assuming the total energy released by thermal runaway, Q, is determined by the cell capacity (e.g., 3.2V / 280Ah) and SOC, the calculation formula is Q = 3.2V × 280Ah × SOC × 3600s (unit: joules). For example, if a cell's SOC is 80%, then Q = 3.2 × 280 × 0.8 × 3600 ≈ 2.3 × 10^6 J.

[0119] Propagation Path Simulation: Using an improved Dijkstra algorithm, starting from the risk source, the energy propagation path is calculated along the adjacency matrix weights (electrical connection strength + heat dissipation conditions). For example, if module M3 experiences thermal runaway, energy diffuses to M2 and M4 through the parallel copper busbars (weight 0.9) and simultaneously radiates through the air (weight 0.1) to the adjacent module M5.

[0120] ‌Risk Propagation Map Generation‌: The propagation paths are superimposed on the 3D battery cluster model, and the risk intensity is indicated using a color gradient. For example, the red area (energy > 1×10^6 J) covers modules M3, M2, and M4, while the orange area (energy > 5×10^5 J) spreads to M5.

[0121] Application Example: In a certain energy storage power station, cell V45 in module M3 of battery cluster C-07 triggered a Level 3 alert (91% confidence level). After parsing the spatiotemporal tags, the main control unit loaded C-07's topology data: six modules connected in parallel, with 0.5 meters between modules. A Dijkstra algorithm simulation revealed that the thermal runaway energy spread to modules M2 and M4 within 5 seconds and to M5 within 10 seconds. The generated risk map showed that the energy in the core area of ​​M3 reached 2.3×10^6 J, while that of M2 / M4 reached 1.8×10^6 J, and that of M5 reached 0.9×10^6 J.

[0122] S2022: In the potential risk propagation map, with the battery cell predicted to be at an emergency level as the center, calculate the aerosol spray coverage radius and the safe isolation distance between adjacent clusters to generate the initial control area boundary;

[0123] This step is based on the quantitative data of the risk propagation map and combined with the performance parameters of the fire extinguishing equipment to dynamically define the initial scope of aerosol injection and isolation to ensure rapid suppression of thermal runaway and prevent cross-cluster spread.

[0124] ‌Aerosol spray parameter calculation‌:

[0125] Coverage Radius Modeling: The coverage radius R (meters) of an aerosol spray device (such as the HATS-200) is related to the spray pressure P (kPa) and the spray angle θ (degrees) as R = 0.12 × P × sin θ. For example, when P = 500 kPa and θ = 60°, R = 0.12 × 500 × sin 60° ≈ 0.12 × 500 × 0.866 ≈ 52 meters (the actual coverage radius needs to be scaled based on the battery cluster size, for example, the actual coverage radius is 5.2 meters).

[0126] Dose Requirement Estimation: The amount of extinguishing agent used (D, grams) is positively correlated with the risk level. The formula is D = Level × BaseDose. Level 3 corresponds to BaseDose = 200g. For example, for a Level 3 alert, D = 3 × 200 = 600g.

[0127] Dynamic Radius Adjustment: If multiple risk sources are adjacent (distance < 2R), their coverage areas are merged. For example, if the risk sources of modules M3 and M2 are 4 meters apart (R = 5 meters), the combined coverage radius increases to 7 meters.

[0128] ‌Safety Isolation Distance Calculation‌:

[0129] Thermal runaway propagation speed: According to historical data, the propagation speed of thermal runaway in air is V_spread = 0.5m / s, and in conductors (such as copper busbars) V_spread = 2m / s.

[0130] Safety distance formula: Safety isolation distance L = Max(V_spread × t_response, L_min), where t_response is the system response time (e.g., 5 seconds) and L_min is the physical isolation margin for the device (e.g., 0.5 meters). For example, L in the airborne direction = 0.5 × 5 + 0.5 = 3.0 meters, and L in the copper busbar direction = 2 × 5 + 0.5 = 10.5 meters.

[0131] Cross-cluster isolation strategy: If the distance between adjacent battery clusters is less than L, isolation devices must be deployed between the clusters. For example, if the distance between clusters C-07 and C-08 is 8 meters (less than 10.5 meters in the copper busbar direction), isolation devices must be deployed between the clusters using a deflector.

[0132] Initial Control Zone Generation:

[0133] Geometric boundary definition: With the coordinates of the risk source (e.g., M3 (X=2, Y=3)) as the center, draw an aerosol coverage circle (radius 5.2 meters) and a safety isolation rectangle (length = 2L, width = 2L).

[0134] ‌3D Spatial Mapping‌: Map the 2D geometric boundaries to the battery cluster's 3D coordinate system (X, Y, Z), where the Z axis represents the module layer height (e.g., 0.2 meters per layer). For example, module M3 has a Z axis of 1.2 meters (layer 6), and its control area is a cylinder (radius 5.2 meters, height 0.2 meters).

[0135] ‌Boundary data encapsulation‌: The initial control area is stored in JSON format, including fields such as center coordinates, radius, isolation distance, timestamp, etc. For example:

[0136] {

[0137] "cluster": "C-07",

[0138] "center": {"X": 2, "Y": 3, "Z": 1.2},

[0139] "radius": 5.2,

[0140] "isolation_distance": {"air": 3.0, "conductor": 10.5},

[0141] "timestamp": "2024-07-15 14:23:05.500"

[0142] }.

[0143] Application Example: For a thermal runaway event in module M3 of cluster C-07, the aerosol spray radius is set to 5.2 meters (covering modules M3, M2, and M4), and the safety isolation distance is set to 3 meters in the air direction and 10.5 meters in the copper busbar direction. The initial control area consists of a cylinder (5.2-meter radius, Z=1.2 meters) and a rectangular isolation zone (21 meters long and 6 meters wide, covering the corridor between clusters C-07 and C-08).

[0144] S2023 uses a dynamic priority mapping algorithm to perform weighted optimization of the initial control area based on the parallel string relationship of the battery cluster and the distribution of heat dissipation channels;

[0145] This step uses a dynamic priority algorithm to comprehensively consider electrical topology and thermodynamic characteristics to optimize the allocation strategy of fire extinguishing and isolation resources, improve firefighting efficiency and reduce the risk of false operation.

[0146] ‌Parallel string relationship modeling‌:

[0147] String weight distribution:

[0148] The electrical impact weight of parallel modules is W_electric = 1 / (1 + Distance × R_parallel), where Distance is the module spacing (meters) and R_parallel is the parallel resistance (e.g., 0.01Ω). For example, if the distance between modules M3 and M2 is 0.5 meters, then W_electric = 1 / (1 + 0.5 × 0.01) = 0.995.

[0149] Fault Propagation Probability: According to Ohm's law, the fault current I_fault is inversely proportional to the parallel resistance, and the propagation probability P_propagate = I_fault / I_max. For example, if I_max = 100A and I_fault = 80A, then P_propagate = 0.8.

[0150] ‌Heat Path Assessment‌:

[0151] Heat Dissipation Coefficient Calculation: The relationship between heat dissipation capacity (H) (W / °C), fan speed (V) (m / s), and heat sink area (A) (m²) is: H = 15 × V × A^0.5. For example, if the wind speed is 2 m / s and the area is 0.5 m², then H = 15 × 2 × 0.707 ≈ 21.2 W / °C.

[0152] Thermal resistance factor: Define the thermal resistance factor K_thermal = 1 / (1+H×Δt), where Δt is the duration of thermal runaway (in seconds). For example, if H = 21.2W / °C and Δt = 10 seconds, then K_thermal = 1 / (1+21.2×10) = 0.0045.

[0153] Dynamic priority mapping algorithm:

[0154] Weighted optimization objective: Priority Score = α × W_electric + β × P_propagate + γ × (1 - K_thermal), where α, β, and γ are weight coefficients (default α = 0.6, β = 0.3, and γ = 0.1). For example, if a module has W_electric = 0.995, P_propagate = 0.8, and K_thermal = 0.0045, then Score = 0.6 × 0.995 + 0.3 × 0.8 + 0.1 × (1 - 0.0045) = 0.597 + 0.24 + 0.0995 ≈ 0.936.

[0155] Real-time adjustment mechanism: If a module's temperature rise rate exceeds 5°C / s (as detected by the thermal map), its γ weight is automatically increased to 0.3 and its α weight is reduced to 0.4. For example, Score = 0.4×0.995 + 0.3×0.8 + 0.3×0.995 ≈ 0.398+0.24+0.299 ≈ 0.937.

[0156] Optimization Result Output: Sort by score from high to low, prioritizing firefighting actions in high-scoring areas. For example, module M3 score = 0.936 (highest), M2 score = 0.892, M4 score = 0.885, and M5 score = 0.752.

[0157] Application Example: During the optimization process for cluster C-07, module M3, with a score of 0.936, was assigned the highest priority due to its direct connection to the fault source and poor heat dissipation (H = 15W / °C). Module M5, with a score of 0.752, was reassigned to secondary processing due to its greater distance and excellent heat dissipation (H = 30W / °C). The algorithm changed the aerosol spray dose from an even split to 400g for M3, 300g for M2, 300g for M4, and 100g for M5.

[0158] S2024: Divide the concentric circle scope of aerosol injection according to the weighted optimization results, mark the deployment coordinates of the thermal runaway isolation device, and generate a draft of the three-dimensional space control instructions;

[0159] This step maps the optimized priority into executable space control instructions, and realizes the coordinated operation of fire extinguishing and isolation through layered spraying and precise positioning.

[0160] ‌Concentric Circle Scope Division‌:

[0161] ‌Level Definition‌: The control area is divided into three levels according to the priority score:

[0162] Core layer (Score ≥ 0.9): Radius 3 meters, injection dose accounts for 60% (e.g. M3: 400g);

[0163] Middle layer (0.7≤Score<0.9): 5-meter radius, 30% of the dose (e.g., M2: 300g);

[0164] ‌Outer layer (Score < 0.7): Radius 7 meters, dose accounting for 10% (e.g. M5: 100g).

[0165] Dose Gradient Allocation: Dose is allocated within the same layer based on the module score ratio. For example, in the middle layer, if M2 score is 0.892 and M4 score is 0.885, the total dose is 600g × 30% = 180g. The M2 dose is 0.892 / (0.892 + 0.885) × 180, which is ≈ 90g, and the M4 dose is ≈ 90g.

[0166] ‌Isolation Device Deployment Strategy‌:

[0167] Coordinate Calculation: The deployment location of an isolation device (such as a telescopic deflector) must meet two conditions:

[0168] Located at the boundary of the safe isolation distance between clusters (for example, 3 meters in the air direction); away from cooling channels and maintenance entrances.

[0169] ‌Three-dimensional coordinate annotation‌: For example, between clusters C-07 and C-08, the coordinates of the four anchor points for deploying the deflector are (10, 3, 1.2), (10, 5, 1.2), (12, 5, 1.2), and (12, 3, 1.2), forming a rectangular isolation zone.

[0170] Draft control instructions generation:

[0171] ‌Command Structure‌: Each command contains the device type (aerosol spray / isolation device), spatial coordinates, action parameters, and execution priority (1-100). For example:

[0172] {

[0173] "device": "Aerosol Injector #23",

[0174] "type": "fire extinguishing",

[0175] "coordinates": {"X": 2, "Y": 3, "Z": 1.2},

[0176] "radius": 3.0,

[0177] "dose": 400,

[0178] "priority": 95

[0179] },

[0180] {

[0181] "device": "Deflector #07",

[0182] "type": "Isolation",

[0183] "coordinates": [

[0184] {"X": 10, "Y": 3, "Z": 1.2},

[0185] {"X": 10, "Y": 5, "Z": 1.2},

[0186] {"X": 12, "Y": 5, "Z": 1.2},

[0187] {"X": 12, "Y": 3, "Z": 1.2}

[0188] ],

[0189] "action": "extend",

[0190] "priority": 90

[0191] }.

[0192] ‌Time synchronization mark‌: All instructions are appended with a unified timestamp (such as 14:23:06.200) to ensure coordinated execution.

[0193] Application example: In the control instruction draft generated by cluster C-07, aerosol sprayer #23 sprays 400g of fire extinguishing agent at coordinates (2,3,1.2) (priority 95), and deflector #07 deploys at coordinates (10,3,1.2)-(12,5,1.2) (priority 90), covering the safety isolation area in the direction of the copper busbar.

[0194] ‌S2025, perform conflict detection on the draft three-dimensional space control instructions, eliminate overlapping areas of the device action range, and output the final control instruction set including execution priority and spatial positioning.

[0195] This step ensures that the actions of fire extinguishing and isolation equipment do not interfere with each other through spatial conflict detection and priority arbitration, and optimizes the execution sequence to maximize fire fighting efficiency.

[0196] ‌Conflict Detection Algorithm‌:

[0197] Spatial hashing grid division: The three-dimensional space of the battery cluster is divided into 0.5m × 0.5m × 0.2m cube grids, and each grid records the device action status (idle / occupied).

[0198] ‌Action Area Projection‌: Projects the aerosol ejection area (cylinder) and the isolation device area (cuboid) onto the grid. For example, the cylinder of aerosol ejector #23 covers the grid G(2-5,3-8,1.2), and the cuboid of deflector #07 covers the grid G(10-12,3-5,1.2).

[0199] Overlapping Area Detection: If the action grids of two devices intersect, a conflict is detected. For example, if the coverage grid G(10-13,3-5,1.2) of injector #24 overlaps with the grid G(10-12,3-5,1.2) of shroud #07, a conflict is triggered.

[0200] Conflict resolution strategies:

[0201] Priority Arbitration: High-priority devices are prioritized, while low-priority devices have their parameters adjusted (e.g., reducing radius, delaying execution). For example, if Aerosol Injector #24 (priority 85) conflicts with Deflector #07 (priority 90), the Deflector will take precedence, reducing the radius of Injector #24 from 5 meters to 4 meters.

[0202] Spatial Parameter Optimization: Use gradient descent to iteratively adjust device coordinates or ranges until overlap is eliminated. For example, adjust the X coordinate of shroud #07 from 10-12 to 10-11.5 to avoid the coverage area of ​​injector #24.

[0203] ‌Final instruction set generation‌:

[0204] ‌Execution sequence arrangement‌: Sort by priority from high to low and add a time delay (e.g., high-priority devices execute immediately, low-priority devices delay 0.5 seconds). For example: [

[0206] {"device": "Aerosol Injector #23", "delay_ms": 0},

[0207] {"device": "Deflector#07", "delay_ms": 500},

[0208] {"device": "Aerosol Injector #24", "delay_ms": 1000}

[0209] ].

[0210] Logging and rollback mechanism: If a conflict cannot be completely resolved (e.g., due to physical space limitations), a fault code (e.g., ERR_Conflict_Unresolved) is recorded and manual intervention is triggered.

[0211] Based on the predicted high-risk battery locations and the physical topology of the battery cluster (such as parallel string relationships and heat dissipation channel distribution), a dynamic priority mapping algorithm simulates the thermal runaway propagation path, calculates the aerosol spray range (coverage radius) and isolation area (safety distance), and performs spatial optimization based on the battery cluster layout. Through weighted optimization (such as prioritizing core strings and avoiding heat dissipation blind spots), the algorithm generates three-dimensional control instructions, including spray angle, dosage, and isolation device deployment coordinates. This algorithm also resolves equipment action conflicts, avoiding the resource waste and response delays caused by traditional "one-size-fits-all" firefighting strategies. Through dynamic priority adjustment and coordinated spatial planning, it precisely suppresses the spread of thermal runaway, maximizes firefighting efficiency, and minimizes the impact on healthy batteries.

[0212] S203: According to the control instruction set, the aerosol spray fire extinguishing device and the thermal runaway isolation device are driven by a pulse trigger circuit powered by a residual power supply, a dual-channel redundant signal verification mechanism is used to execute firefighting actions, and a real-time action feedback signal is output. Specifically, this may include:

[0213] S2031 activates the emergency power supply module of the BMS, uses a capacitor energy storage pulse generator to generate a high-voltage trigger signal, encodes the control instruction set into two modulated waveforms with a phase difference of 90 degrees, and transmits them to the execution terminal through an independent channel;

[0214] This step realizes emergency power supply through capacitor energy storage technology, uses phase difference modulation to ensure the reliability and anti-interference of signal transmission, and provides high-precision trigger instructions for the fire execution terminal.

[0215] Emergency power module activation logic:

[0216] The BMS has a built-in supercapacitor bank (parameters: 6 2.7V / 100F capacitors connected in series, total capacity 16.2V / 16.7F). When the main power fails, the capacitor bank discharges through a MOSFET switch (model IRF3205), with an output voltage of 12V-16.2V and a continuous power supply time of ≥30 seconds.

[0217] Pulse Generator Design: Capacitor energy is converted into high-voltage pulses via a boost converter circuit (topology: boost converter). Key parameters include a pulse width of 1ms, a peak voltage of 300V, and a repetition rate of 10Hz. For example, triggering an aerosol spray device requires three consecutive pulses (1ms pulse width, 50ms interval).

[0218] Control command encoding and modulation:

[0219] Two-channel phase difference modulation: The control instruction set (JSON format) is converted into a binary data stream and modulated using QPSK (Quadrature Phase Shift Keying). The two carrier waves have a phase difference of 90° (I channel 0°, Q channel 90°). For example, the instruction "injection dose 400g" corresponds to the binary code "010001101101". After modulation, the I channel waveform (frequency 10kHz, amplitude 5V) and the Q channel waveform (frequency 10kHz, amplitude 5V, phase delay 1 / 4 cycle) are generated.

[0220] Independent channel transmission: The I-channel signal is transmitted via the CAN bus (physical layer twisted pair, 120Ω impedance), and the Q-channel signal is transmitted via the RS-485 bus (differential signal, shielded twisted pair), avoiding signal loss caused by electromagnetic interference.

[0221] ‌Execute terminal signal reception‌:

[0222] Signal Demodulation Circuit: The execution terminal deploys the AD8302 phase detection chip, which coherently demodulates the two signals and restores the original binary instruction. For example, when the I signal is 5V at 0° and the Q signal is 5V at 90°, the AD8302 outputs a logic level of "11," corresponding to the highest priority injection instruction.

[0223] Application Example: When a thermal runaway occurs at an energy storage power station, the BMS emergency power module activates, discharging the capacitor bank to 16.2V and boosting the voltage to 300V to drive the pulse generator. The control command "shroud coordinates (10, 3, 1.2)" is modulated using QPSK. The I signal is transmitted via the CAN bus and the Q signal via RS-485. The execution terminal then demodulates and recovers the command with an error rate of <0.1%.

[0224] S2032: A dual-channel redundancy check module is deployed at the aerosol spray device to demodulate and cross-check the two modulated waveforms. When the instruction consistency exceeds the threshold, an execution permission signal is generated; otherwise, a self-check loop is triggered.

[0225] This step ensures command integrity through dual-channel redundant verification, combines a self-check mechanism to prevent false operations, and improves the fault tolerance of the fire protection system.

[0226] ‌Redundancy check module design‌:

[0227] Dual-channel data analysis: The I and Q signals are received via the STM32F407's CAN controller and RS-485 interface, respectively. They are then parsed into binary instructions and stored in buffers (Buffer 1 and Buffer 2, each 1KB).

[0228] Cross-matching algorithm: Compare Buffer1 and Buffer2 byte by byte, calculating the consistency ratio: Consistency = number of matching bytes / total number of bytes × 100%. For example, if the content of Buffer1 is "010001101101" and the content of Buffer2 is "010001101101", the consistency is 100%; if the content of Buffer2 is "010001101100", the consistency ratio is 11 / 12 × 100% ≈ 91.7%.

[0229] ‌Execute permission logic‌:

[0230] Threshold determination: Set the consistency threshold to 95%. When Consistency ≥ 95%, the execution permission signal (high level 5V) is triggered; otherwise, the self-test circuit is triggered. For example, if Consistency = 91.7%, a self-test request (error code E01) is sent to the BMS.

[0231] Self-test loop process: The self-test includes signal path testing (e.g., testing the CAN bus terminal resistance at 120Ω), power supply voltage testing (required to be ≥11V), and solenoid valve impedance testing (normal range: 20-30Ω). If the self-test passes, the command is resent; if it fails, a fault log is recorded (e.g., "ERR_CAN_BUS_OFF").

[0232] ‌Anti-interference enhancement technology‌:

[0233] Hamming Code Error Correction: Hamming Code is added during the instruction encoding phase, adding 3 check bits for every 4 bits of data to correct 1-bit errors. For example, the original data "0100" is encoded as "0100011". The receiving end uses the check bits to detect and correct single-bit flips.

[0234] Application Example: During a fire alarm trigger, two bytes of command I were lost due to electromagnetic interference (Consistency = 83.3%). The redundancy check module triggered a self-check and discovered abnormal CAN bus impedance (measured 60Ω). After switching to the backup channel, the command was resent, completing the error correction in 120ms.

[0235] S2033: Drive the high-speed solenoid valve according to the execution permission signal, adjust the injection angle and dosage according to the spatial positioning parameters of the control instruction set, and simultaneously activate the telescopic deflector of the thermal runaway isolation device to form a physical barrier;

[0236] This step uses high-precision actuators to achieve the coordinated action of directional injection of fire extinguishing agent and physical isolation, ensuring the timeliness and spatial coverage of thermal runaway suppression.

[0237] High-speed solenoid valve control technology:

[0238] Drive Circuit Design: The solenoid valve (model SMC VQD212) is driven by an H-bridge circuit and receives a PWM signal (1kHz frequency, 10%-90% duty cycle) to adjust flow. For example, a 50% duty cycle corresponds to a 50% valve opening and a flow rate of 20 L / min. The injection dose is calculated as: flow rate × time. For a 2-second injection time, the dose is 20 L / min × 2 / 60, which equals 0.67 L.

[0239] Spatial Positioning Parameter Analysis: Extracts the three-dimensional coordinates (X, Y, Z) and the spray angle (Azimuthθ, Elevationφ) from the control command. For example, the coordinates (2, 3, 1.2) correspond to module M3, with θ = 45° (horizontal) and φ = 30° (vertical elevation). The nozzle direction is adjusted using a stepper motor (28BYJ-48, 5.625° step angle).

[0240] Thermal runaway isolation device action:

[0241] Retractable shroud control: The shroud is powered by a pneumatic cylinder (at 0.6 MPa pressure) at a deployment speed of 0.5 m / s, forming a U-shaped barrier (1.5 m × 1 m × 0.5 m). For example, a shroud at the deployment coordinates (10, 3, 1.2) fully deploys within 500 ms, isolating the inter-cluster heat conduction path.

[0242] Synchronous Timing Control: The activation signals for the solenoid valve and the deflector are synchronized by an FPGA (Xilinx Spartan-6) with an error of less than 1ms. For example, the solenoid valve opens at t=0ms and the deflector activates at t=10ms, ensuring that the timing of fire extinguishing agent injection and physical barrier formation is aligned.

[0243] Application Example: In response to a thermal runaway event in module M3, the solenoid valve adjusts the nozzle direction according to the command (θ=45°, φ=30°), spraying 400g of KL-5 aerosol fire extinguishing agent (flow rate 25L / min x 1.6 seconds). Simultaneously, the deflector deploys at coordinates (10, 3, 1.2), isolating the thermal radiation path between clusters C-07 and C-08.

[0244] S2034 uses an infrared camera to capture the diffusion pattern of the fire extinguishing agent, combined with the airflow intensity feedback from the pressure sensor, to generate an action execution effect evaluation matrix as a real-time action feedback signal;

[0245] This step quantifies the effectiveness of firefighting actions through multi-sensor fusion technology, providing data support for closed-loop control.

[0246] Fire Extinguishing Agent Dispersion Monitoring:

[0247] Infrared Imaging Technology: A FLIR A315 infrared camera (320×240 resolution, temperature range -20°C to 150°C) captures images at 30fps and identifies areas covered by the fire extinguishing agent based on temperature gradients. For example, effective coverage is determined when the temperature difference ΔT in the aerosol diffusion area is ≥5°C (25°C ambient to 20°C in the extinguishing area).

[0248] Image Processing Algorithm: Use the OpenCV library to binarize the infrared image (threshold ΔT ≥ 5°C). Calculate coverage: Coverage = Number of valid pixels / Total number of pixels × 100%. For example, if 5,000 pixels meet the standard and the total number of pixels is 76,800, then Coverage ≈ 6.5%.

[0249] Airflow intensity feedback:

[0250] Pressure Sensor Deployment: An MPX5700AP differential pressure sensor (range 0-700 kPa) is installed at the nozzle outlet to measure real-time pressure P (in kPa). Airflow intensity Q (L / min) is calculated using the formula: Q = K × √(P / ρ), where K = 0.07 (calibration factor) and ρ = 1.2 kg / m³ (air density). For example, when P = 500 kPa, Q = 0.07 × √(500 / 1.2) ≈ 0.07 × 20.4 ≈ 1.43 L / s, which is 85.8 L / min.

[0251] ‌Evaluation Matrix Generation‌:

[0252] Matrix data structure: The evaluation matrix is ​​a 5×5 table containing timestamp, coverage, airflow intensity, temperature suppression rate, and device status code. For example:

[0253] Timestamp: 2024-07-15 14:23:07.500;

[0254] Coverage: 65%;

[0255] Flow: 85.8L / min;

[0256] Temp_Reduction: 12℃;

[0257] Status_Code: 0x0000 (Normal).

[0258] Application Examples:

[0259] After a certain injection, the infrared image showed a coverage rate of 65% (target 70%), the pressure sensor reported an airflow intensity of 85.8 L / min (target 90 L / min), and the evaluation matrix marked the status code 0x0001 (insufficient flow), triggering the subsequent optimization strategy.

[0260] ‌S2035, calculate the residual between the real-time action feedback signal and the expected control target. If the deviation exceeds the safety threshold, secondary injection is triggered and the fault code is updated to the BMS log system.

[0261] This step dynamically adjusts the firefighting strategy through residual analysis and realizes system self-optimization in combination with fault logs.

[0262] ‌Residual calculation logic‌:

[0263] Deviation Quantification Method: Coverage deviation ΔC = |actual coverage - target coverage|, and flow deviation ΔQ = |actual flow - target flow|. For example, if the target coverage is 70% and the actual is 65%, ΔC = 5%; if the target flow is 90 L / min and the actual is 85.8 L / min, ΔQ = 4.2 L / min.

[0264] Safety threshold settings: ΔC_max = 10%, ΔQ_max = 15L / min. If ΔC > 10% or ΔQ > 15L / min, the system is considered to have failed and secondary injection is triggered.

[0265] Secondary injection strategy:

[0266] Incremental Adjustment Rule: Secondary Dose = Original Dose × (1 + ΔC / Target Coverage). For example, if the initial spray is 400g (ΔC = 5%), the secondary dose = 400 × (1 + 5 / 70) = 400 × 1.071 ≈ 428g.

[0267] ‌Execution Priority Improvement‌: The secondary injection instruction priority has been increased from 80 to 95, preempting system resources to ensure timely response.

[0268] ‌Fault Log Management‌:

[0269] Log format specification: Log entries contain timestamp, device ID, fault type, and action. For example: 2024-07-15 14:23:08.200 | Aerosol sprayer #23 | Low coverage (65% vs. 70%) | Secondary spray 428g.

[0270] ‌Encrypted synchronization mechanism‌: Logs are encrypted using AES-256 and synchronized to all BMS nodes via the LoRa wireless module (frequency band 433MHz, rate 1kbps).

[0271] Application Example: After a fire extinguishing operation, an evaluation shows ΔC = 12% (exceeding the 10% threshold). The system automatically triggers a secondary injection of 428g, increasing coverage and recording the fault log "ERR_COVERAGE_UNDER."

[0272] The BMS's emergency power supply (capacitor energy storage) generates a high-voltage pulse signal, encoding the control command into two modulated waveforms with a 90° phase difference, which are then transmitted to the execution terminal via independent channels. A dual-channel redundancy check module demodulates and cross-verifies the signal to ensure command consistency before triggering the aerosol spray (adjusting dosage and angle) and the isolation device (physical isolation with a retractable deflector). Simultaneously, an infrared camera and pressure sensor provide real-time feedback on the extinguishing agent's diffusion and airflow intensity, generating an action execution evaluation matrix. This redundant signal verification and real-time feedback mechanism addresses the problem of false actions caused by signal interference or equipment failure in traditional firefighting operations, improving the reliability of firefighting actions and the traceability of execution results.

[0273] S204: Based on the real-time action feedback signal and the battery cluster residual state data, the parameter weights of the multi-source feature fusion network are updated through a reinforcement learning model to be used for the fire linkage control in the next cycle, forming a closed-loop control link. Specifically, this may include:

[0274] S2041, integrating real-time action feedback signals with battery cluster residual state data to construct a multi-dimensional reinforcement learning state vector;

[0275] This step constructs the state space of reinforcement learning by fusing multi-source data, providing comprehensive environmental perception input for strategy optimization.

[0276] ‌Data integration logic‌:

[0277] Real-time action feedback signals: This includes the percentage of aerosol spray coverage (e.g., 65%), airflow intensity reported by the pressure sensor (e.g., 85.8 L / min), temperature suppression rate identified by the infrared camera (e.g., ΔT = 12°C), and device status code (e.g., 0x0000 indicates normal).

[0278] Battery cluster residual status data: Collects post-fire cell voltage (e.g., 3.2V / cell), remaining capacity (SOC=45%), state of health (SOH=78%), and inter-cluster insulation resistance (e.g., ≥100MΩ).

[0279] Normalization: All data is mapped uniformly to the range [0, 1]. For example, for a voltage range of 2.5V-4.2V, 3.2V corresponds to a normalized value of (3.2-2.5) / (4.2-2.5)=0.41; a coverage of 65% corresponds to 0.65.

[0280] ‌State vector construction method‌:

[0281] Dimension Definition: The state vector is 14-dimensional, including: action feedback data (4 dimensions): coverage, airflow intensity, temperature suppression rate, and device status code; battery residual data (6 dimensions): average cell voltage, SOC, SOH, insulation resistance, maximum temperature difference within the cluster, and residual gas concentration; spatiotemporal label (4 dimensions): X / Y / Z coordinates of the thermal runaway cluster and timestamp difference (the difference in seconds between the current time and the event trigger time).

[0282] Example: The state vector is [0.65, 0.86, 0.48, 0.0, 0.41, 0.45, 0.78, 0.92, 0.15, 0.03, 2.1, 3.5, 1.2, 120], corresponding to the 14 dimensions mentioned above.

[0283] Application Example: After a fire in an energy storage system is extinguished, the "temperature suppression rate" in the state vector is 0.48 (corresponding to ΔT = 12°C / 25°C ambient temperature), the "maximum temperature difference within the cluster" is 0.15 (corresponding to 4.5°C / 30°C full scale), and the device status code is 0x0000 (normalized to 0.0). This is used for subsequent reinforcement learning training.

[0284] S2042: Design a reward function based on policy gradients, where positive rewards include thermal runaway suppression efficiency and resource consumption ratio, and negative penalties include the number of misoperations and equipment loss value, to generate a dynamic reward score.

[0285] This step uses a quantitative reward mechanism to guide the reinforcement learning model to optimize decision-making and balance fire extinguishing effect and cost loss.

[0286] Positive Reward Calculation Rules:

[0287] Thermal runaway suppression efficiency (R1): Defined as (actual suppression time / theoretical maximum suppression time) × 100%. For example, if the theoretical maximum suppression time is 5 seconds and the actual suppression time is 6 seconds, then R1 = 5 / 6 × 100 ≈ 83.3%, corresponding to a reward of +8.3.

[0288] Resource Consumption Ratio (R2): Calculates the ratio of fire extinguishing agent usage to the theoretical minimum usage. For example, if the theoretical minimum usage is 300g and the actual usage is 400g, then R2 = 300 / 400 = 0.75, resulting in a +7.5 bonus.

[0289] Negative penalty calculation rules:

[0290] ‌Number of false triggers (P1): Records the number of times the fire protection device was triggered unnecessarily. For example, if the device was triggered unnecessarily once, the penalty value is -5.0.

[0291] Equipment loss (P2): Calculated based on the number of solenoid valve actuations (lifespan: 100,000) and the mechanical wear of the air deflector (lifespan: 5,000). For example, if a single actuation consumes 0.01% (1 / 10,000) of the solenoid valve's lifespan and 0.02% (1 / 5,000) of the air deflector's lifespan, then P2 = -(0.01 + 0.02) × 100 = -3.0.

[0292] Dynamic reward scoring formula: Total reward R_total = α × (R1 + R2) + β × (P1 + P2), where α = 0.6 (effect weight) and β = 0.4 (loss weight).

[0293] Example: If R1 = 8.3, R2 = 7.5, P1 = -5.0, and P2 = -3.0, then R_total = 0.6 × (8.3 + 7.5) + 0.4 × (-5.0 - 3.0) = 0.6 × 15.8 + 0.4 × (-8.0) = 9.48 - 3.2 = 6.28.

[0294] Application Example: A fire extinguishing attempt consumed 0.05% of the deflector's lifespan, but the thermal runaway suppression efficiency reached 95%, resulting in a total reward of 6.28. Another fire extinguishing attempt resulted in a reward of -2.1 due to a false trigger. The model will automatically reduce the probability of misjudgment in similar scenarios.

[0295] S2043 uses an asynchronous advantage actor-critic algorithm to update the parameters of the multi-source feature fusion network. The actor network optimizes the feature weight distribution strategy, and the critic network corrects the thermal runaway level prediction bias.

[0296] This step uses a distributed reinforcement learning framework to efficiently update model parameters and improve the accuracy of thermal runaway prediction.

[0297] ‌Asynchronous Advantage Actor-Critic (A3C) Architecture‌:

[0298] Actor Network: Input is a 14-dimensional state vector, and output is the weight adjustment (Δw1, Δw2, Δw3) for the voltage, temperature, and gas channels in the multi-source feature fusion network. For example, if the original weights are [0.3, 0.5, 0.2], the Actor Network recommends adjusting them to [+0.02, -0.01, +0.03].

[0299] Critic: Given the same state vector as the input, it outputs the deviation between the predicted value of the value function V(s) and the actual reward. For example, if the predicted value V(s) is 6.5 and the actual value R_total is 6.28, the deviation δ = 6.28 - 6.5 = -0.22.

[0300] Asynchronous Update Mechanism: Deploy 8 parallel worker threads, each of which independently interacts with the environment and periodically synchronizes gradients to the global network. For example, after Worker 1 trains 10 batches (batch_size=256) locally, it uploads the gradients to update the global model.

[0301] ‌Parameter update process‌:

[0302] Advantage calculation: A(s,a)=R_total+γ×V(s')-V(s), where γ=0.99 is the discount factor and s' is the next state. For example, if V(s)=6.5, V(s')=7.0, and R_total=6.28, then A=6.28+0.99×7.0-6.5=6.28+6.93-6.5=6.71.

[0303] Policy Gradient Update: The actor network adjusts the weight distribution policy according to the advantage value, with a learning rate of 0.001 and an RMSprop optimizer (decay factor ρ = 0.9).

[0304] ‌Value Function Modification‌: The critic network uses δ=-0.22 as the supervision signal and adopts the mean squared error (MSE) loss function to update parameters.

[0305] Application Example: During a training session, the actor network increased the gas channel weight from 0.2 to 0.23, and the critic network corrected the prediction deviation from -0.22 to +0.05, increasing the confidence level of the subsequent thermal runaway prediction by 12%.

[0306] ‌S2044, encrypt and synchronize the updated network parameters to all BMS nodes, and reset the fire linkage control cycle counter to form a closed-loop control link from state perception to strategy optimization.

[0307] This step implements distributed system collaboration through secure communication protocols, ensuring policy consistency and completing closed-loop control.

[0308] ‌Parameter encryption and synchronization mechanism‌:

[0309] Encryption Algorithm: Updated network parameters are encrypted using AES-256 (Advanced Encryption Standard, 256-bit key). For example, the parameter file (512KB) is encrypted using CBC (Cipher Block Chaining) mode, with an initialization vector (IV) of 16 bytes.

[0310] Synchronization Protocol: Encrypted parameters are published using the MQTT (Message Queuing Telemetry Transport) protocol with the topic " / BMSCluster / ModelUpdate." Each BMS node subscribes to this topic and decrypts the received data using a hardware security module (HSM).

[0311] ‌Version Control‌: The version number is incremented for each update (e.g. v1.2.3 → v1.2.4), and the hash value (SHA-256) is recorded to prevent tampering.

[0312] ‌Control cycle reset logic‌:

[0313] ‌Counter initialization‌: The fire linkage control cycle counter starts from 0, and the maximum number of cycles is set to 1000 times (about 30 days). When it is reached, the entire system model is forced to update.

[0314] ‌Event-triggered reset‌: After each successful firefighting action, the counter is reset to 0; if there are no thermal runaway events for 10 consecutive cycles, lightweight model fine-tuning is triggered (only the critic network is updated).

[0315] Application Example: During a model update, an encrypted parameter file (version v2.1.9) was broadcast to 56 BMS nodes via MQTT. Decryption took 15ms, and the system completed full node synchronization within 200ms. After the counter was reset, it entered the first monitoring cycle.

[0316] Firefighting action feedback data (such as extinguishing agent coverage and isolation effectiveness) is integrated with the battery cluster's residual state (such as voltage recovery and temperature drop rate) to construct a reinforcement learning state vector. Utilizing the Asynchronous Advantage Actor-Critic (A3C) algorithm, the actor network optimizes the feature weight allocation strategy of the multi-source feature fusion network, while the critic network corrects thermal runaway prediction bias. The reward function integrates thermal runaway suppression efficiency, resource consumption ratio, and equipment loss. Network parameters are dynamically updated and encrypted and synchronized to all BMS nodes, achieving closed-loop control from data perception to policy optimization. By continuously learning from historical firefighting cases and real-time environmental changes, the system adaptively improves thermal runaway prediction accuracy and firefighting linkage efficiency, enhancing the system's long-term robustness and adaptability.

[0317] It can be seen that according to the real-time voltage, temperature and gas concentration data of the battery cluster, thermal runaway prediction is performed through the multi-source feature fusion network, and the thermal runaway prediction level is output; based on the thermal runaway prediction level, the dynamic priority mapping algorithm is used to output the control instruction set containing execution priority and spatial positioning according to the battery cluster topology; according to the control instruction set, a dual-channel redundant signal verification mechanism is adopted to execute fire-fighting actions, and a real-time action feedback signal is output; based on the real-time action feedback signal and the residual state data of the battery cluster, the parameter weights of the multi-source feature fusion network are updated through the reinforcement learning model for use in the next cycle of fire-fighting linkage control, forming a closed-loop control link, thereby realizing battery cluster-level fire-fighting linkage control to enhance the safety protection capability of the battery system.

[0318] Another embodiment of the present invention provides a battery cluster level fire linkage control system based on BMS, see Figure 3 , the system may include:

[0319] Prediction module 301, configured to predict thermal runaway based on the real-time voltage, temperature, and gas concentration data of the battery cluster using a multi-source feature fusion network. The multi-source feature fusion network integrates SOC gradient change characteristics and SOH decay curve characteristics to output a thermal runaway prediction level;

[0320] A planning module 302 is configured to collaboratively plan the aerosol spray range and the thermal runaway isolation area based on the thermal runaway prediction level and a dynamic priority mapping algorithm according to the battery cluster topology, and output a control instruction set including execution priority and spatial positioning;

[0321] An execution module 303 is configured to drive the aerosol spray fire extinguishing device and the thermal runaway isolation device according to the control instruction set through a pulse trigger circuit powered by a residual power supply, execute firefighting actions using a dual-channel redundant signal verification mechanism, and output a real-time action feedback signal;

[0322] The updating module 304 is used to update the parameter weights of the multi-source feature fusion network through a reinforcement learning model based on the real-time action feedback signal and the battery cluster residual state data, so as to be used for the fire linkage control in the next cycle and form a closed-loop control link.

[0323] It can be seen that according to the real-time voltage, temperature and gas concentration data of the battery cluster, thermal runaway prediction is performed through the multi-source feature fusion network, and the thermal runaway prediction level is output; based on the thermal runaway prediction level, the dynamic priority mapping algorithm is used to output the control instruction set containing execution priority and spatial positioning according to the battery cluster topology; according to the control instruction set, a dual-channel redundant signal verification mechanism is adopted to execute fire-fighting actions, and a real-time action feedback signal is output; based on the real-time action feedback signal and the residual state data of the battery cluster, the parameter weights of the multi-source feature fusion network are updated through the reinforcement learning model for use in the next cycle of fire-fighting linkage control, forming a closed-loop control link, thereby realizing battery cluster-level fire-fighting linkage control to enhance the safety protection capability of the battery system.

[0324] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps of any one of the above method embodiments when running.

[0325] Specifically, in this embodiment, the above-mentioned storage medium may be configured to store a computer program for performing the following steps:

[0326] S201, based on the real-time voltage, temperature and gas concentration data of the battery cluster, thermal runaway prediction is performed through a multi-source feature fusion network, wherein the multi-source feature fusion network fuses the SOC gradient change characteristics and the SOH decay curve characteristics to output a thermal runaway prediction level;

[0327] S202, based on the thermal runaway prediction level, using a dynamic priority mapping algorithm, according to the battery cluster topology, collaboratively planning the aerosol injection range and the thermal runaway isolation area, and outputting a control instruction set including execution priority and spatial positioning;

[0328] S203, according to the control instruction set, driving the aerosol spray fire extinguishing device and the thermal runaway isolation device through the pulse trigger circuit powered by the residual power supply, using a dual-channel redundant signal verification mechanism to execute the firefighting action, and outputting a real-time action feedback signal;

[0329] S204 , based on the real-time action feedback signal and the battery cluster residual state data, the parameter weights of the multi-source feature fusion network are updated through a reinforcement learning model for use in the next cycle of fire linkage control to form a closed-loop control link.

[0330] It can be seen that according to the real-time voltage, temperature and gas concentration data of the battery cluster, thermal runaway prediction is performed through the multi-source feature fusion network, and the thermal runaway prediction level is output; based on the thermal runaway prediction level, the dynamic priority mapping algorithm is used to output the control instruction set containing execution priority and spatial positioning according to the battery cluster topology; according to the control instruction set, a dual-channel redundant signal verification mechanism is adopted to execute fire-fighting actions, and a real-time action feedback signal is output; based on the real-time action feedback signal and the residual state data of the battery cluster, the parameter weights of the multi-source feature fusion network are updated through the reinforcement learning model for use in the next cycle of fire-fighting linkage control, forming a closed-loop control link, thereby realizing battery cluster-level fire-fighting linkage control to enhance the safety protection capability of the battery system.

[0331] An embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any one of the above method embodiments.

[0332] Specifically, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0333] Specifically, in this embodiment, the processor may be configured to execute the following steps through a computer program:

[0334] S201, based on the real-time voltage, temperature and gas concentration data of the battery cluster, thermal runaway prediction is performed through a multi-source feature fusion network, wherein the multi-source feature fusion network fuses the SOC gradient change characteristics and the SOH decay curve characteristics to output a thermal runaway prediction level;

[0335] S202, based on the thermal runaway prediction level, using a dynamic priority mapping algorithm, according to the battery cluster topology, collaboratively planning the aerosol injection range and the thermal runaway isolation area, and outputting a control instruction set including execution priority and spatial positioning;

[0336] S203, according to the control instruction set, driving the aerosol spray fire extinguishing device and the thermal runaway isolation device through the pulse trigger circuit powered by the residual power supply, using a dual-channel redundant signal verification mechanism to execute the firefighting action, and outputting a real-time action feedback signal;

[0337] S204 , based on the real-time action feedback signal and the battery cluster residual state data, the parameter weights of the multi-source feature fusion network are updated through a reinforcement learning model for use in the next cycle of fire linkage control to form a closed-loop control link.

[0338] It can be seen that according to the real-time voltage, temperature and gas concentration data of the battery cluster, thermal runaway prediction is performed through the multi-source feature fusion network, and the thermal runaway prediction level is output; based on the thermal runaway prediction level, the dynamic priority mapping algorithm is used to output the control instruction set containing execution priority and spatial positioning according to the battery cluster topology; according to the control instruction set, a dual-channel redundant signal verification mechanism is adopted to execute fire-fighting actions, and a real-time action feedback signal is output; based on the real-time action feedback signal and the residual state data of the battery cluster, the parameter weights of the multi-source feature fusion network are updated through the reinforcement learning model for use in the next cycle of fire-fighting linkage control, forming a closed-loop control link, thereby realizing battery cluster-level fire-fighting linkage control to enhance the safety protection capability of the battery system.

[0339] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. A battery cluster-level fire linkage control method based on BMS, characterized in that: The method comprises: Based on the real-time voltage, temperature, and gas concentration data of the battery cluster, thermal runaway prediction is performed through a multi-source feature fusion network. The multi-source feature fusion network integrates the SOC gradient change characteristics with the SOH decay curve characteristics to output a thermal runaway prediction level. The network collects the time-series fluctuation data of each cell voltage in the battery cluster, the spatial distribution thermodynamic map of the temperature sensor, and the gas concentration change rate, and synchronizes the timestamps to generate a cross-dimensional original monitoring matrix. The original monitoring matrix is ​​corrected for outliers and the SOC gradient change characteristics are extracted using a sliding window difference method. The SOH decay curve is fitted with cycle aging test data to generate a joint SOC-SOH degradation feature vector. A multi-source feature fusion network is constructed, and the SOC-SOH joint degradation feature vector is input into a bidirectional gated recurrent unit along with real-time temperature and gas concentration data. The weight coefficients of the voltage, temperature, and gas channels are dynamically assigned through a feature-level attention mechanism, and the fused thermal runaway feature code is output. The thermal runaway feature code is input into a pre-trained multi-level classifier, which divides the warning level according to the threshold boundaries of historical thermal runaway cases, generates a thermal runaway prediction level with confidence, and labels the spatiotemporal location. Based on the thermal runaway prediction level, a dynamic priority mapping algorithm is used to collaboratively plan the aerosol injection range and the thermal runaway isolation area according to the battery cluster topology, and a control instruction set including execution priority and spatial positioning is output; According to the control instruction set, the aerosol spray fire extinguishing device and the thermal runaway isolation device are driven by a pulse trigger circuit powered by a residual power supply, a dual-channel redundant signal verification mechanism is used to execute firefighting actions, and a real-time action feedback signal is output; Based on the real-time action feedback signal and the battery cluster residual state data, the parameter weights of the multi-source feature fusion network are updated through a reinforcement learning model for use in the next cycle of fire linkage control to form a closed-loop control link.

2. The method according to claim 1, characterized in that Based on the thermal runaway prediction level, a dynamic priority mapping algorithm is used to collaboratively plan the aerosol spray range and the thermal runaway isolation area according to the battery cluster topology, and a control instruction set containing execution priority and spatial positioning is output, including: Analyze the spatiotemporal labels of thermal runaway prediction levels, construct an adjacency matrix based on the physical topology of the battery cluster, simulate the maximum energy diffusion range of the thermal runaway propagation path, and generate a potential risk propagation map; In the potential risk propagation map, with the battery cell predicted to be at an emergency level as the center, the aerosol spray coverage radius and the safe isolation distance between adjacent clusters are calculated to generate the initial control area boundary; ‌A dynamic priority mapping algorithm is used to perform weighted optimization of the initial control area, combining the parallel string relationship of the battery cluster and the distribution of heat dissipation channels; ‌ Based on the weighted optimization results, the concentric circle scope of the aerosol injection is divided, the deployment coordinates of the thermal runaway isolation device are marked, and a draft of the three-dimensional space control instructions is generated; ‌ Conflict detection is performed on the draft three-dimensional space control instructions, overlapping areas of the equipment motion range are eliminated, and the final control instruction set containing execution priority and spatial positioning is output.

3. The method according to claim 2, characterized in that According to the control instruction set, the pulse trigger circuit powered by the residual power supply drives the aerosol spray fire extinguishing device and the thermal runaway isolation device, adopts a dual-channel redundant signal verification mechanism to execute the fire fighting action, and outputs a real-time action feedback signal, including: Activate the emergency power supply module of the BMS, use a capacitor energy storage pulse generator to generate a high-voltage trigger signal, encode the control instruction set into two modulated waveforms with a phase difference of 90 degrees, and transmit them to the execution terminal through an independent channel; A dual-channel redundancy check module is deployed on the aerosol spray device to demodulate and cross-check the two modulated waveforms. When the instruction consistency exceeds the threshold, an execution permission signal is generated; otherwise, a self-check loop is triggered. The high-speed solenoid valve is driven according to the execution permission signal, the injection angle and dosage are adjusted according to the spatial positioning parameters of the control instruction set, and the telescopic deflector of the thermal runaway isolation device is synchronously activated to form a physical barrier; The infrared camera captures the diffusion pattern of the fire extinguishing agent, and combined with the pressure sensor to feedback the airflow intensity, an action execution effect evaluation matrix is ​​generated as a real-time action feedback signal; The residual of the real-time action feedback signal and the expected control target is calculated. If the deviation exceeds the safety threshold, secondary injection is triggered and the fault code is updated to the BMS log system.

4. The method according to claim 3, characterized in that The method updates the parameter weights of the multi-source feature fusion network based on the real-time action feedback signal and the battery cluster residual state data through a reinforcement learning model for use in the fire linkage control of the next cycle, thereby forming a closed-loop control link, including: Integrate real-time action feedback signals with battery cluster residual state data to construct a multi-dimensional reinforcement learning state vector; Design a reward function based on policy gradient, where positive rewards include thermal runaway suppression efficiency and resource consumption ratio, and negative penalties include the number of misoperations and equipment loss value, to generate a dynamic reward score; An asynchronous advantage actor-critic algorithm is used to update the parameters of the multi-source feature fusion network. The actor network optimizes the feature weight distribution strategy, and the critic network corrects the thermal runaway level prediction deviation. The updated network parameters are encrypted and synchronized to all BMS nodes, and the fire linkage control cycle counter is reset to form a closed-loop control link from state perception to strategy optimization.

5. A battery cluster-level fire linkage control system based on BMS, characterized in that: The system comprises: The prediction module is used to predict thermal runaway based on the real-time voltage, temperature, and gas concentration data of the battery cluster through a multi-source feature fusion network. The multi-source feature fusion network fuses the SOC gradient change characteristics with the SOH decay curve characteristics to output a thermal runaway prediction level. The module collects the time series fluctuation data of each cell voltage in the battery cluster, the spatial distribution thermal map of the temperature sensor, and the gas concentration change rate, and synchronizes the timestamps to generate a cross-dimensional original monitoring matrix. The original monitoring matrix is ​​corrected for outliers and the SOC gradient change characteristics are extracted using a sliding window difference method. The SOH decay curve is fitted with the cycle aging test data to generate a SOC-SOH joint degradation feature vector. A multi-source feature fusion network is constructed, and the SOC-SOH joint degradation feature vector is input into a bidirectional gated recurrent unit along with real-time temperature and gas concentration data. The weight coefficients of the voltage, temperature, and gas channels are dynamically assigned through a feature-level attention mechanism, and the fused thermal runaway feature code is output. The thermal runaway feature code is input into a pre-trained multi-level classifier, which divides the warning level according to the threshold boundaries of historical thermal runaway cases, generates a thermal runaway prediction level with confidence, and labels the spatiotemporal location. a planning module for collaboratively planning the aerosol spray range and the thermal runaway isolation area based on the thermal runaway prediction level and a dynamic priority mapping algorithm according to the battery cluster topology, and outputting a control instruction set including execution priority and spatial positioning; an execution module, configured to drive the aerosol spray fire extinguishing device and the thermal runaway isolation device according to the control instruction set through a pulse trigger circuit powered by a residual power supply, execute firefighting actions using a dual-channel redundant signal verification mechanism, and output a real-time action feedback signal; An updating module is used to update the parameter weights of the multi-source feature fusion network through a reinforcement learning model based on the real-time action feedback signal and the battery cluster residual state data, so as to be used for the fire linkage control in the next cycle and form a closed-loop control link.

6. The system according to claim 5, characterized in that The planning module is specifically used to: Analyze the spatiotemporal labels of thermal runaway prediction levels, construct an adjacency matrix based on the physical topology of the battery cluster, simulate the maximum energy diffusion range of the thermal runaway propagation path, and generate a potential risk propagation map; In the potential risk propagation map, with the battery cell predicted to be at an emergency level as the center, the aerosol spray coverage radius and the safe isolation distance between adjacent clusters are calculated to generate the initial control area boundary; ‌A dynamic priority mapping algorithm is used to perform weighted optimization of the initial control area, combining the parallel string relationship of the battery cluster and the distribution of heat dissipation channels; ‌ Based on the weighted optimization results, the concentric circle scope of the aerosol injection is divided, the deployment coordinates of the thermal runaway isolation device are marked, and a draft of the three-dimensional space control instructions is generated; ‌ Conflict detection is performed on the draft three-dimensional space control instructions, overlapping areas of the equipment motion range are eliminated, and the final control instruction set containing execution priority and spatial positioning is output.

7. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 4 when executed.

8. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 4.

Images