Lithium battery thermal runaway intelligent control method and system

Abstract

The invention provides a lithium battery thermal runaway intelligent control method and system. The system comprises a multi-source sensing unit, an intelligent processing and decision-making unit and a precise execution and feedback unit. Acquiring battery state data by adopting a multi-mode sensor; the data passes through a deep decoupling multi-attention optimization network, so that fault monomer accurate positioning, thermal runaway remaining time early warning and energy equivalent evaluation are realized; dynamic optimization of fire extinguishing decisions is realized by taking model prediction control as a framework; and accurate, partitioned and quantitative spraying of a fire extinguishing agent is realized through the flexible partitioned fire extinguishing execution mechanism. According to the method, early warning can be completed at the initial stage of battery thermal runaway, a fault cell is accurately positioned, and then the fault cell is effectively inhibited by adopting a fire extinguishing technology, so that the safe and reliable performance of a battery system is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of lithium ion battery safety management, and particularly relates to a lithium battery thermal runaway intelligent control method and system. BACKGROUND

[0002] Lithium ion batteries are widely used in electric vehicles, high-power energy storage power stations and other important energy storage devices due to their high energy density, long cycle life and other advantages. However, under extreme conditions such as overcharging, internal short circuit and mechanical abuse, the battery may trigger thermal runaway, leading to a chain of exothermic reactions, causing the battery temperature to rise, toxic and flammable gases to be emitted, or the battery to explode and catch fire, resulting in personal injury and economic loss. In order to solve the current battery problems, we urgently need to propose a comprehensive protection system that combines multi-modal perception, efficient intelligent algorithms, and optimization control technology, which can timely, accurately, and automatically identify the safety state of the battery and take proactive safety protection measures to improve the intrinsic safety of the battery. SUMMARY

[0003] The technical problem to be solved by the present application is to provide a lithium battery thermal runaway intelligent control method and system to overcome the shortcomings of the prior art. The method proposes a set of proactive safety protection scheme that integrates multi-source perception, intelligent prediction, optimized decision-making, and precise execution to solve the problems of existing lithium battery thermal runaway early warning lag, fuzzy positioning, rough fire extinguishing, complex model difficult to deploy in real time, and disconnection between perception, decision-making, and execution.

[0004] The present application first provides a lithium battery thermal runaway intelligent control method, comprising the following steps:

[0005] Step 1: Synchronously collect the original multi-modal state data of the battery module through the sensor array arranged in the battery module, and perform data cleaning and standardization preprocessing; the original multi-modal state data includes voltage, current, temperature, characteristic gas concentration, and vibration signal; Step 2: Perform adaptive sliding window spectral filtering on the original multi-modal state data to remove noise, and extract key features through a multi-scale feature fusion method to generate a fusion feature dataset; Step 3: Input the fusion feature dataset into a deep decoupling multi-attention optimization network model; the deep decoupling multi-attention optimization network model outputs abnormal single body positioning probability distribution, thermal runaway remaining warning time, and potential energy equivalent evaluation results through parallel processing of the positioning branch, the time sequence warning branch, and the energy regression branch, respectively, and obtains comprehensive prediction results after uncertainty weighted fusion; Step 4, judge whether the early warning time is lower than the preset threshold value; if yes, trigger the fire extinguishing decision, calculate the total demand of fire extinguishing agent based on the comprehensive prediction result through the nonlinear dose mapping model, and solve the optimization problem in the rolling time domain by using the model predictive control framework to generate partition fire extinguishing control instructions; Step 5, according to the partition fire extinguishing control instructions, drive the corresponding flexible fire extinguishing actuator to carry out quantitative and directional fire extinguishing; collect environmental feedback data after fire extinguishing and evaluate the effect to form a closed-loop control.

[0006] Step 2 includes:

[0007] Step 2.1, data preprocessing;

[0008] The preprocessing includes adaptive sliding window spectral filtering AWSF and multi-scale feature fusion MSFF, and the calculation formula is:

[0009] (1),

[0010] (2),

[0011] Wherein, is the dynamic window length at time t; is the dynamic cutoff frequency at time t: is the local signal-to-noise ratio estimate, is the reference window length, is a preset parameter; sigmoid is a nonlinear activation function; exp is a natural exponential function;

[0012] Step 2.2, multi-scale feature fusion and deep association extraction;

[0013] Multi-scale feature fusion is performed on the filtered multi-modal data; the feature association graph is constructed by calculating the mutual information I(m;n) between different modal features m and n, and the edge weight is The calculation formula is:

[0014] (3),

[0015] Wherein is the linear correlation coefficient between the modes; is the relationship bandwidth parameter;

[0016] And the original feature is enhanced by using the graph convolution network GCN to extract the deep association feature : (4),

[0017] Wherein, is the adjacent matrix with self-connection, is the corresponding degree matrix, is the activation function, and the final fused feature dataset is obtained.

[0018] In step 3, the positioning branch is used for accurate positioning of abnormal monomers, including:

[0019] The gated cross-modal attention mechanism is adopted; first, the attention weights between different modal features are calculated : (5),

[0020] where Softmax is a normalized exponential function; T represents transposition; respectively represent the query matrix and the key matrix respectively are the query weight matrix and the key weight matrix;

[0021] and a gating vector g is introduced for dynamic information selection and fusion: (6),

[0022] where, is a trainable weight matrix; respectively are the temperature feature tensor and the gas feature tensor, is a trainable bias vector; the fused feature is represented as ; is the Hadamard product; value matrix is the voltage feature tensor;

[0023] Finally, the graph attention network is used to aggregate the battery module topology information, and the abnormal probability distribution of each monomer in the battery module is output ;

[0024] The time series early warning branch is used for thermal runaway remaining time prediction, including:

[0025] The stacked causal dilated convolution structure is adopted to capture long-term temporal dependence; the output of the lth layer (“lth layer” is a general mathematical index, which specifically refers to the lth causal dilated convolution processing unit stacked in order in the time series early warning branch sub-neural network. Its core function is to analyze the input data with an exponentially expanding time window as the layer number l increases, thereby realizing long-range, multi-scale temporal modeling of the thermal runaway evolution process) at time t is: (7),

[0026] where is the i th weight parameter of the l th convolution kernel; ReLU is the activation function, and the dilation rate ; the final output is the predicted thermal runaway remaining time and the dynamic risk index R(t).​ the bias term of the first layer;

[0027] The energy regression branch is used for potential energy equivalent assessment, including:

[0028] Introducing physical consistency constraints in the regression network; the energy regression branch regresses the base energy equivalent through a multi-layer perception and reaction parameters, and calculates the physical consistency loss Constraints are imposed on network training: (8),

[0029] wherein, is the physical loss weight coefficient; is the derivative of the predicted temperature with respect to time; is the heat capacity; is the pre-exponential factor; is the activation energy (prior), is the product of the gas constant and the predicted temperature; is the exponential term related to the predicted energy; hA is the heat dissipation coefficient; is the ambient temperature; after physical layer correction, the final thermal runaway potential energy equivalent estimate is output .

[0030] In step 3, the outputs of the positioning branch, the timing warning branch and the energy regression branch are fused through adaptive weighting based on uncertainty, and the final comprehensive prediction result is generated through three mechanisms of calibration factor, sharpening gate and adaptive temperature coefficient, the formula is:

[0031] (9),

[0032] wherein, is the adaptive fusion weight of the i-th branch output; is the comprehensive prediction output vector; is the normalized uncertainty estimate of the i-th branch; is the uncertainty calibration factor of the i-th branch; is the adaptive temperature coefficient of the i-th branch; is the proportional coefficient of the saturation gate function; is the original output of the i-th branch; is the saturation gate function; is the decision sharpening index.

[0033] In step 4, the total demand of extinguishing agent is calculated through the following nonlinear dose mapping model :

[0034] (10),

[0035] wherein, is the fire extinguishing agent base effectiveness coefficient; is the time urgency decay coefficient; are the energy equivalent influence coefficient and the risk change rate influence coefficient, respectively; is the thermal runaway residual warning time; is the energy equivalent-based exponential influence term; is the time derivative of the dynamic risk index (risk change rate).

[0036] Step 4 further comprises:

[0037] Step 4.1, state modeling and objective function definition;

[0038] A discrete state space model is established with the target area temperature , characteristic gas concentration , and residual fire extinguishing agent amount as state variables, and the state vector is ; the following optimization objective function is defined:

[0039] (11),

[0040] wherein, is the safety state reference value, is the state deviation penalty weight matrix, is the control quantity penalty weight matrix, is the terminal state penalty weight matrix, is the prediction time domain; is the maximum control step index in the prediction time domain is the system state at the future k+i time predicted at the time k; is the control input calculated at the time k to be applied at the future k+i time; is the system state at the prediction time domain terminal (k+i) predicted at the time k;

[0041] Step 4.2, constraint condition integration;

[0042] The total demand amount of fire extinguishing agent is taken as a key resource constraint, and combined with the system physical limit to form a complete constraint set of the optimization problem:

[0043] (12),

[0044] wherein, represents the system state vector predicted at the future i+1 step at the time k; , ​respectively denote the predicted state transition matrix and control input matrix of the i-th step at the current control period k; denotes the predicted i-th step control amount; is the control amount upper limit; Δt is the control period; is the amount of extinguishing agent consumed before the current control period; , respectively denote the lower and upper limits of the state amount; is the control input at the future k+i time;

[0045] Step 4.3, online rolling optimization solution;

[0046] At each control period , based on the latest system state and the prediction model , the numerical optimization solver is used to solve the optimization problem online to obtain the optimal control sequence in the prediction time domain;

[0047] Step 4.4, instruction generation and output;

[0048] The first element of the optimal control sequence is taken as the control instruction of the current period, i.e. the partition extinguishing agent flow instruction.

[0049] In step 4.1, the model predictive control framework adopts an adaptive weight adjustment strategy; the weight matrix and R are dynamically adjusted according to the warning time and the risk index :

[0050] (13),

[0051] (14),

[0052] wherein, is the reference weight, is the adjustment coefficient; the reference weight matrix of the control amount penalty; is the time derivative of the dynamic risk index R(t).

[0053] In step 4, a closed-loop correction mechanism is also included:

[0054] After each control period is executed, the actually measured system state is compared with the state predicted in the last period to calculate the prediction error ; the error is usedThe state space model parameters in the subsequent prediction time domain are obtained by Kalman filter or moving average method online fine-tuning is performed.

[0055] In step 5, the flexible fire extinguishing actuator comprises:

[0056] Partitioned storage tank unit: independently designed in upper and lower partitions, each partition is filled with hot aerosol fire extinguishing agent; the partitioned storage tank unit is equipped with pressure sensors and temperature sensors for real-time monitoring of internal state;

[0057] Flexible connecting pipeline: high-temperature-resistant and corrosion-resistant flexible metal bellows or silica gel pipe is used to connect the storage tank and the nozzle; this flexible design allows the pipeline to bend and stretch within a certain range to adapt to the complex spatial layout and possible deformation of the battery pack;

[0058] Multi-degree-of-freedom nozzle array: each partition is equipped with an adjustable nozzle connected by a spherical hinge with two rotational degrees of freedom (pitch and yaw), which can be manually or by a micro motor preset in the initialization stage to ensure that the spray covers all battery cells in the partition without dead angles; the nozzle is provided with an atomizing core to fully atomize the fire extinguishing agent;

[0059] High-speed electromagnetic valve: a high-speed response electromagnetic valve is installed at the outlet of each storage tank, which is directly controlled by the intelligent processing and decision unit to realize rapid and accurate release of the fire extinguishing agent;

[0060] Self-adaptive sealing device: a self-adaptive sealing ring is provided between the nozzle and the battery box mounting hole, which is made of memory alloy and high-temperature-resistant rubber composite material, and can automatically adjust the sealing pressure when the temperature changes to ensure long-term sealing reliability;

[0061] In step 5, the working process of the flexible partition fire extinguishing mechanism comprises:

[0062] Instruction receiving: the control circuit of the corresponding partition receives the spraying instruction from the MPC controller, including the spraying time and flow;

[0063] Valve starting: the high-speed electromagnetic valve is opened within milliseconds, and the fire extinguishing agent is driven by the built-in pressurized gas in the storage tank to flow to the nozzle array through the flexible pipeline;

[0064] Directional atomizing spray: the fire extinguishing agent is crushed into micron-sized particles by the atomizing core inside the nozzle, forming aerosol smoke, and accurately covering the faulty battery and its surrounding area according to the preset angle; the flexible pipeline can buffer pressure impact and reduce vibration transmission;

[0065] Flow monitoring and adjustment: through PWM (pulse width modulation) control of the electromagnetic valve, combined with storage tank pressure feedback, accurate adjustment of flow is realized to match the spraying curve optimized by MPC;

[0066] Feedback and shutdown: after reaching the predetermined injection amount, the electromagnetic valve is closed; the extinguishing effect is monitored in real time by temperature and gas sensors, and the data is fed back to the MPC controller to determine whether additional injection is needed.

[0067] The application also provides a lithium battery thermal runaway intelligent control system for executing the method, comprising:

[0068] A multi-source perception unit is configured to perform step 1 of the method, and synchronously collects raw multi-modal state data of the battery module through a sensor array arranged in the battery module;

[0069] An intelligent processing and decision unit is configured to perform steps 2 to 4 of the method, pre-processes the raw multi-modal state data, fuses features, and uses a deep decoupling multi-attention optimization network model for comprehensive prediction, and generates partitioned fire extinguishing control instructions based on a model predictive control framework when the trigger condition is met;

[0070] A precise execution and feedback unit is configured to perform step 5 of the method, and drives the flexible fire extinguishing execution mechanism to extinguish the fire according to the partitioned fire extinguishing control instructions, and collects feedback data to form a closed-loop control.

[0071] The application has the following advantages: through the fusion of multi-modal sensing information of lithium batteries and the decoupling analysis based on a deep decoupling attention network, the early warning and suppression of lithium battery thermal runaway are completed, and the active safety protection capability of the battery system is effectively improved, mainly reflected in the following points:

[0072] (1) The timeliness and accuracy of early warning are greatly improved: through the fusion of signals such as voltage, current, temperature, characteristic gas, vibration, and the use of adaptive sliding window filtering and multi-scale feature fusion, weak and small abnormal signals in the early stage of thermal runaway can be captured, breaking through the time lag problem of traditional single threshold alarm; based on the deep decoupling multi-attention optimization network (DMAO-Net), fault location, remaining time and energy evaluation are performed in parallel, and an uncertainty weighted fusion method is used to achieve better early warning purpose, improving the early warning reliability and time margin.

[0073] (2) Higher fault location accuracy and precise intervention: the DMAO-Net method using the gating cross-modal attention mechanism and the graph attention network can accurately locate the abnormal battery monomer and output a probabilistic positioning result, and the fault location accuracy has jumped from the former "module level" to the latter "monomer level", which will effectively support subsequent partitioned directional fire extinguishing.

[0074] (3) Using the model predictive control (MPC) framework to decompose the fire extinguishing decision-making process into multiple simplified sub-time domain problems, in the rolling optimization process, the fire extinguishing strategy is dynamically optimized through the early warning output, and the allocation scheme of the fire extinguishing agent and the injection time of each injection are solved in the sub-time domain, a series of fire extinguishing agent and fire extinguishing agent injection strategies are solved on the basis of the current optimal and gradually advancing, and combined with the nonlinear dose mapping model and the adaptive weight adjustment strategy, on the one hand, the fire extinguishing agent is minimized, and on the other hand, the normal battery is avoided during the fire extinguishing process.

[0075] (4) The flexible actuator has fast and reliable response: the sub-area independent storage tank, flexible connecting pipeline and multi-degree-of-freedom adjustable nozzle are selected, the fire extinguishing system has strong space adaptability and good physical robustness, realizes millisecond-level response and millisecond-level turning speed, and completes the "perception-decision-execution-feedback" closed-loop safety control. The selection of "perception-decision-execution-feedback" depends on the intelligent judgment of the nearby sufficient liquid supply and the scientific flame covering opportunity.

[0076] (5) Using the anti-knowledge distillation model compression technology, etc., the parameter quantity of the original complex network is compressed to less than 5M, and the second-level reasoning is realized on the edge computing terminal, which can fully meet the application scene of the vehicle-mounted, energy storage and other industries with high time efficiency requirement, so as to realize the leap from "cloud intelligence" to "edge intelligence". BRIEF DESCRIPTION OF DRAWINGS

[0077] The above and / or other aspects of the present application will become more apparent by describing in detail the present application with reference to the accompanying drawings, as follows.

[0078] Figure 1 is the system overall architecture diagram of the present application, which shows the three-layer hardware architecture (perception layer, calculation and decision-making layer, execution and feedback layer) of the system of the present application and the data flow and control flow therebetween.

[0079] Figure 2 is the DMAO-Net network model structure diagram of the present application. The internal structure and data fusion path of the three parallel branches (positioning, timing, energy) in the deep decoupling multi-attention optimization network are clearly depicted.

[0080] Figure 3 is the flexible partition fire extinguishing mechanism arrangement schematic diagram (side view) of the present application. The physical layout and corresponding relationship of the battery module and the upper and lower partition fire extinguishing units (storage tank, nozzle) are shown.

[0081] Figure 4 is the model predictive control closed loop fire extinguishing logic flow chart. The complete intelligent decision-making process from data acquisition to closed loop execution is illustrated in the form of a flow chart. DETAILED DESCRIPTION

[0082] The embodiment of the present application proposes a lithium battery thermal runaway intelligent control method, comprising the following steps:

[0083] Step 1: Through the sensor array arranged in the battery module, the original multi-modal state data of the battery module is synchronously collected, and data cleaning and standardization preprocessing are performed; the original multi-modal state data includes voltage, current, temperature, characteristic gas concentration and vibration signal;

[0084] Step 2: Multi-source data acquisition and adaptive preprocessing

[0085] In order to improve the data quality and retain the early fault characteristics, an adaptive sliding window spectrum filtering technology is used to carry out noise reduction processing on the original data. The filter window length and the cut-off frequency are dynamically adjusted according to the real-time estimated local signal-to-noise ratio , and the specific calculation formula is as follows:

[0086] (1),

[0087] (2),

[0088] wherein is the local signal-to-noise ratio estimation, is the reference window length, is a preset parameter.

[0089] Subsequently, multi-scale feature fusion is carried out on the filtered multi-modal data; the feature correlation graph is constructed by calculating the mutual information I(m;n) between different modal features m and n, and the edge weight is calculated according to the following formula:

[0090] (3),

[0091] Then, the original features are enhanced by using a graph convolution network to extract deep correlation features: (4),

[0092] wherein, is the adjacency matrix with self-connection, is the corresponding degree matrix, is a trainable weight matrix, and σ is an activation function; finally, the fusion feature dataset is obtained.

[0093] The implementation process of step 3 intelligent prediction based on deep decoupling multi-attention optimization network is as follows:

[0094] The fusion feature dataset obtained in step 2 is input into a deep decoupling multi-attention optimization network. The network includes three functionally decoupled and parallel processed neural network branches:

[0095] (3.1) Positioning branch: using a gated cross-modal attention mechanism, first calculate the attention weights between different modal features:

[0096] (5),

[0097] where are projections from different modalities respectively. A gating vector is introduced to dynamically select key modal information:

[0098] (6),

[0099] where, is a trainable weight matrix, is a trainable bias vector. The fused feature representation is:

[0100] (7),

[0101] Finally, the graph attention network (GAT) is used to aggregate the battery module topology information, and the abnormal probability distribution of each battery node is output .

[0102] (3.2) Time series warning branch: a stacked causal dilated convolution structure is used to capture long-term temporal dependencies; the output of the lth layer at time t is : (8),

[0103] where the dilation rate ; finally output the predicted thermal runaway remaining time and dynamic risk index R(t);

[0104] (3.3) Energy regression branch: introduce physical consistency constraints in the regression network; this branch uses a multi-layer perceptron to regress the equivalent energy and reaction parameters, and calculates the physical consistency loss Lphysics to constrain network training: (9),

[0105] where is the activation energy (prior), is the heat capacity, and hA is the heat dissipation coefficient. After the physical layer correction, the final estimated value of the equivalent energy of the thermal runaway .

[0106] (3.4) The outputs of the three branches are fused into the final comprehensive prediction result through an adaptive weighting fusion based on uncertainty, a calibration factor, a sharpening gate, and a triple mechanism of adaptive temperature coefficient, as follows:

[0107] (10),

[0108] where, is the normalized uncertainty, is the uncertainty calibration factor is the saturation gate function, is the decision sharpening index, is the adaptive temperature coefficient, is the output of each branch.

[0109] The implementation process of the intelligent fire extinguishing decision based on model predictive control in step 4 is as follows:

[0110] When the warning time is lower than the preset threshold, the fire extinguishing decision is triggered. First, the total demand of fire extinguishing agent is calculated through a nonlinear dose mapping model:

[0111] (11),

[0112] where, is the basic coefficient, which is the basic effectiveness coefficient of the fire extinguishing agent, calibrated by the characteristics of the fire extinguishing agent and the standard fire experiment; is the time urgency decay coefficient, which controls the influence rate of the warning time on the demand,

[0113] are the energy equivalent influence coefficient and the risk change rate influence coefficient, respectively, determined by historical data regression analysis.

[0114] The process of optimization decision and instruction generation through the model predictive control framework in step 4 is as follows:

[0115] (4.1) State modeling and objective function definition: a discrete state space model is established with the target area temperature , the characteristic gas concentration , and the remaining fire extinguishing agent amount as state variables, and the state vector is ; the optimization objective function is defined to minimize the system state deviation and control amount consumption in the prediction time domain:

[0116] (12),

[0117] where, is the safety state reference value, is the weight matrix, is the prediction horizon;

[0118] (4.2) Constraint integration: the total extinguishing agent demand is taken as a key resource constraint, combined with system physical limits, to form a complete set of constraints for the optimization problem:

[0119] (13),

[0120] where, is the amount of extinguishing agent consumed before the current control period;

[0121] The implementation process of the command generation and output in step 4 is as follows:

[0122] The first element of the optimal control sequence is taken as the control command for the current period, i.e., the partition extinguishing agent flow command, which is output to the precise execution and feedback unit.

[0123] The numerical optimization solver based on the interior point method or the efficient set method is used to ensure that the solution is completed within the preset control period (≤1 second) under the limited computing power of the edge computing device, meeting the real-time requirements.

[0124] The model predictive control framework adopts an adaptive weight adjustment strategy; the weight matrix and R are dynamically adjusted according to the warning time and the risk index :

[0125] (14),

[0126] (15),

[0127] where, is the baseline weight, is the adjustment coefficient; this strategy makes the optimization objective more inclined to quickly reduce the system state deviation when the warning time is urgent or the risk is rapidly increasing.

[0128] The implementation process of the online rolling optimization solution in step 4 is as follows:

[0129] At each control period , based on the latest system state and the prediction model , a numerical optimization solver is used to solve the above constrained optimization problem online to obtain the optimal control sequence in the prediction horizon;

[0130] The closed-loop correction mechanism in step 5 is as follows:

[0131] After each control cycle is executed, the actual measured system state is compared with the state predicted in the previous cycle, and the prediction error is calculated ; using the error, the state space model parameters in the subsequent prediction time domain are fine-tuned online through a Kalman filter or a moving average method to reduce the impact of model mismatch and improve the accuracy of rolling optimization.

[0132] Further, the specific details of step 5 are as follows:

[0133] As Figure 3 shown, the fire extinguishing actuator of the present application adopts a flexible partition design, which is one of the key innovations of the present application. This mechanism not only realizes precise spatial positioning of fire extinguishing, but also has high adaptability and reliability.

[0134] 1. Mechanism structure composition

[0135] The flexible partition fire extinguishing mechanism mainly includes the following core components:

[0136] Partitioned tank unit: The fire extinguishing agent tank adopts an independent design of upper and lower partitions, and each partitioned tank has a capacity of 300-500ml, which is filled with hot aerosol fire extinguishing agent. The tank is equipped with a pressure sensor and a temperature sensor for real-time monitoring of the internal state.

[0137] Flexible connecting pipeline: A flexible metal bellows or silicone tube that is resistant to high temperature and corrosion is used to connect the tank and the spray head. This flexible design allows the pipeline to bend and stretch within a certain range to adapt to the complex spatial layout and possible deformation of the battery pack.

[0138] Multi-degree-of-freedom spray head array: Each partition is equipped with 2-4 adjustable spray heads. The spray head is connected with a spherical hinge and has two rotational degrees of freedom (pitch and yaw), which can be manually or through a micro motor to preset the angle in the initialization stage, ensuring that the spray covers all battery cells in the partition without dead angles. The spray head is provided with an atomizing core to fully atomize the fire extinguishing agent.

[0139] High-speed electromagnetic valve: A high-speed response electromagnetic valve (response time <10ms) is installed at the outlet of each tank, which is directly controlled by the intelligent processing and decision unit to realize the rapid and accurate release of the fire extinguishing agent.

[0140] Self-adaptive sealing device: A self-adaptive sealing ring is provided between the spray head and the battery box mounting hole. The sealing ring is made of a composite material of memory alloy and high-temperature resistant rubber, which can automatically adjust the sealing pressure when the temperature changes to ensure long-term sealing reliability.

[0141] 2. Working principle of flexible mechanism​

[0142] When the intelligent processing and decision unit outputs the positioning of the DMAO-Net After determining the partition (upper or lower) where the faulty battery is located, the following process is performed:

[0143] 1. Instruction reception: The control circuit corresponding to the partition receives the injection instruction from the MPC controller, including the injection duration and flow.

[0144] 2. Valve start: The high-speed electromagnetic valve is opened within milliseconds, and the extinguishing agent is driven by the pressurized gas built-in the storage tank, flowing to the nozzle array through flexible pipelines.

[0145] 3. Directional atomization injection: The extinguishing agent is crushed into micron-sized particles by the atomization core inside the nozzle, forming aerosol smoke, and accurately covering the faulty battery and its surrounding area according to the preset angle. The flexible pipeline can buffer pressure impact and reduce vibration transmission.

[0146] 4. Flow monitoring and adjustment: Through PWM (Pulse Width Modulation) control of the electromagnetic valve, combined with storage tank pressure feedback, accurate adjustment of the flow is realized to match the injection curve optimized by the MPC.

[0147] 5. Feedback and shutdown: After reaching the predetermined injection amount, the electromagnetic valve is closed. The system monitors the extinguishing effect in real time through temperature and gas sensors, and feeds back the data to the MPC controller to determine whether additional injection is needed.

[0148] The specific operation steps and examples of the lithium battery thermal runaway intelligent control method described in the present application are as follows:

[0149] Step 1: Multi-source data acquisition and adaptive preprocessing

[0150] First, real-time data is obtained from the sensor array arranged in the battery module. Temperature monitoring uses distributed fiber Bragg grating sensors with a sampling frequency of 1 Hz at each cell surface and key points of the module; voltage and current data are obtained synchronously through the CAN bus of the battery management system at a frequency of 10 Hz; CO and H2 sensitive characteristic gas sensors are arranged near the pressure relief valve at the top of the module, with a sampling frequency of 0.5 Hz; vibration signals are collected by MEMS accelerometers attached to the side panels of the module, with a frequency of 2 kHz. After all the raw data is collected into the edge computing unit (using Huawei Atlas 200 AI acceleration module), adaptive sliding window spectral filtering is performed. The filtering window length and the cutoff frequency are dynamically adjusted according to the real-time estimated local signal-to-noise ratio , and the specific calculation formula is:

[0151] (16),

[0152] (17),

[0153] wherein, is the reference window length, for example, set to 100 sampling points; is the adjustment coefficient, which is empirically set to 0.5, 0.1 and 1.0 respectively; is the historical mean and standard deviation of the signal-to-noise ratio. This process can effectively suppress high-frequency noise and periodic interference while completely preserving the low-frequency trend characteristics related to early thermal runaway.

[0154] Step 2: Multimodal feature fusion and key feature extraction and between different modal features and the edge weights of the feature correlation graph are constructed according to the following formula:

[0155] (18),

[0156] wherein, is the bandwidth parameter, usually set to 0.5. Then, the original features are enhanced using a graph convolution network:

[0157] (19),

[0158] wherein, is the adjacency matrix with self-connection, is the corresponding degree matrix, is the trainable weight matrix, is the ReLU activation function. At the same time, the system automatically calculates a number of expert experience features such as "maximum temperature rise rate (°C / min)", "CO to H2 concentration ratio", "200-500Hz frequency band vibration energy", etc. These automatically extracted deep features and artificially defined experience features together form a high-dimensional fusion feature set, which is used as the input of the deep prediction model.

[0159] Step 3: DMAO-Net model construction, training and lightweight deployment

[0160] The features obtained in step 2 are predicted using a deep decoupling multi-attention optimization network (DMAO-Net). This network contains three parallel and functionally decoupled branches:

[0161] The positioning branch adopts a 2-layer gated cross-modal attention (GCMA) mechanism with a hidden layer dimension of 128, which is used to fuse the temperature, gas, and voltage modalities. The cross-modal attention weight is calculated as follows:

[0162] (20),

[0163] and the gating vector dynamically selects the key information, and finally outputs the anomaly probability of each battery cell .

[0164] The timing warning branch contains 4 layers of causal dilated convolution layers with dilation rates of 1, 2, 4, and 8, respectively, and a convolution kernel size of 3. The output of the first layer at time t is:

[0165] (21),

[0166] This structure can exponentially expand the receptive field and effectively capture long-term temporal dependencies, outputting the warning time and risk index .

[0167] The energy regression branch contains 3 layers of fully connected layers, and introduces a physical information constraint in the output layer. The physical consistency loss function is:

[0168] (22),

[0169] This loss forces the predicted temperature trajectory to conform to the simplified reaction kinetics described by the Arrhenius equation, outputting the corrected energy equivalent . The outputs of the three branches are fused through an adaptive weighting based on uncertainty, as well as a calibration factor, sharpening gating, and adaptive temperature coefficient triple mechanism to generate the final comprehensive prediction result, with weights . The network is trained using historical thermal runaway experiment data, with the total loss . To meet the real-time requirements of edge deployment, the knowledge of the trained complex model (teacher network) is transferred to a simplified student network using adversarial knowledge distillation technology. The final deployed model has a parameter size of less than 5M, and the inference period on the Atlas 200 platform does not exceed 2 seconds.

[0170] Step 4: Fire extinguishing decision optimization and execution based on model predictive control (MPC)

[0171] When the warning time output by the DMAO-Net model in step 3 is less than the preset action threshold (e.g., 300 seconds), the intelligent fire extinguishing decision is triggered. First, calculate the estimated total demand for fire extinguishing agent as follows: ​​

[0172] (23),

[0173] wherein, are model parameters calibrated by a large number of experiments, for example, can be set to 100, 0.01, 0.8 and 10 respectively. Subsequently, the model predictive control (MPC) module is started to establish a discrete state space model with the target area temperature, gas concentration and residual extinguishing agent amount as state variables with a control period of 1 second. In each control period , a rolling optimization problem of step prediction horizon is solved:

[0174] (24)

[0175] The future optimal control sequence is obtained by online optimization using the efficient solver OSQP, and the flexible partition fire extinguishing mechanism is driven by the MPC decision instruction. The high-speed electromagnetic valve corresponding to one of the upper and lower zones (response time <10 ms) is opened, and the aerosol extinguishing agent pre-pressurized to 2.5 MPa in the storage tank is sent to the multi-degree-of-freedom nozzle (Dv90 <50 μm) located at a preset angle through the flexible metal bellows pipe, and the sprayed aerosol extinguishing agent is accurately sprayed to the corresponding partition on fire through the flexible bellows pipe; the electromagnetic valve opening is controlled by PWM signal, and then the injection flow is closed-loop regulated. For details, please see the original text.

[0176] The present application will take an electric vehicle battery pack composed of 96 21700 type lithium ion cells as an example to explain the working principle, process and results of the battery pack. When continuously fast charging, the 25th cell located in the upper partition abnormally heats up due to internal micro-short circuit.

[0177] At the early stage t=120s, through adaptive sliding window spectral filtering and multi-scale feature fusion on the multi-modal data collected by the multi-source sensor array, the following features are extracted: the temperature rise rate of the 25th cell reaches 0.5℃ / min (the average value of the adjacent cells is 0.1℃ / min), the concentration change rate of the characteristic gas CO reaches 0.03ppm / s, and the vibration energy in the frequency band of 200-500Hz reaches 0.08g². Although these features are relatively weak, they are always in an enhanced state.

[0178] The fused features are taken as input into a deep decoupling multi-attention optimization network, and three branches are processed in parallel under the network, which are a positioning branch, a timing early warning branch and an energy regression branch. The output of the positioning branch is the probability of the abnormality of the No. 25 battery cell, which is 0.82 (the threshold is 0.7); the output of the timing early warning branch is the remaining time of thermal runaway, which is 85 seconds; and the output of the energy regression branch is the potential released energy equivalent, which is about 0.8 MJ. Finally, the final prediction result is obtained by weighting and fusing the uncertainty of the network prediction output.

[0179] According to the prediction result, intelligent fire extinguishing decision is started, and after calculation by a nonlinear dose mapping model, the total amount of extinguishing agent is about 150 ml. Based on the model predictive control framework, an optimal control sequence is obtained through rolling horizon optimization in a 1 s control cycle, which indicates that the accurate spraying time of the upper partition is about 5 s, and the spraying amount is about 105 ml. Then, the command is issued to drive the upper partition fire extinguishing mechanism to act. Subsequently, the high-speed electromagnetic valve is opened, the pre-pressurized 2.0 Mpa thermal aerosol extinguishing agent in the tank is sent into the nozzle with a sighting angle through a flexible pipeline, and then atomized into Dv90 < 60 μm particles, and then sprayed out. The return data after spraying (t = 125 s) is: the maximum temperature of the target battery cell decreases from 42.5℃ to 40.8℃, and the temperature rise rate is 0.2℃ / min; the CO concentration is stable at 15 ppm and the growth rate is significantly reduced. The system judges that the fire tends to be stable and enters the monitoring state. The whole upper partition uses about 21% of the extinguishing agent, and the remaining resources of the lower partition are on standby.

[0180] Finally, the early inhibition of thermal runaway is realized, and compared with the traditional single threshold scheme, the warning time is increased from 20-40 seconds to 60-90 seconds, the ability from module level warning to single cell level warning is realized, the extinguishing agent consumption is reduced by about 40-60% compared with the previous traditional full package spraying method, and the average reasoning time of the edge device of the system is 1.8 seconds. The example verifies that the "multi-modal early perception-intelligent fusion prediction-optimization decision execution" closed loop is beneficial to realize the early discovery, accurate positioning, drug saving, short time, low cost and other characteristics of lithium battery thermal runaway protection.

[0181] In one embodiment of the present application, a lithium battery thermal runaway intelligent control method and system are proposed, which fuses multi-modal sensing information such as voltage, current, temperature, gas, vibration, etc. into a deep decoupling multi-attention optimization network (DMAO-Net) for early and accurate early warning of thermal runaway; and adopts a model predictive control (MPC) framework and a flexible partition fire extinguishing mechanism to realize self-perception, self-decision and self-execution, thereby completing the entire intelligent safety protection closed loop; the DMAO-Net completes accurate positioning of abnormal single cells, thermal runaway remaining time prediction and energy release quantification through decoupled positioning, early warning and energy evaluation branches; the MPC controller is used to combine the predicted value to rollingly optimize the amount of fire extinguishing agent; and the flexible partition execution mechanism is used to accurately and quantitatively release the fire extinguishing agent. The method can solve the problems of early warning lag, unclear positioning and extensive fire extinguishing mode of traditional methods, and can greatly improve the active safety protection capability of the lithium battery system.

[0182] Figure 1 A three-layer hardware architecture from perception to execution of the system and data flow and control flow therebetween are described in detail:

[0183] The perception layer (data acquisition) uses a multi-source sensor array (including voltage, current, temperature, gas, vibration, etc.) to acquire the state raw data of the battery.

[0184] The calculation and decision layer (edge intelligent processing) is the entire core processing part, after the received data is accepted from the perception layer, it enters the adaptive filtering and feature extraction module - DMAO-Net - deep decoupling multi-attention optimization network prediction model - MPC decision maker - model predictive control in turn, and finally forms the process from data to processing, from risk to prediction, to decision making, and fire extinguishing strategy making.

[0185] The implementation and feedback layer (accurate physical intervention) is manifested as the "upper zone fire extinguishing unit" and the "lower zone fire extinguishing unit" separated by a partition; each unit includes a storage tank, a valve, a spray head, etc.; this layer makes the corresponding partition mechanism execute the fire extinguishing task according to the MPC instruction, and feeds back the battery state data after the fire extinguishing is completed to the decision layer, constituting a closed loop control.

[0186] Figure 2 The model structure diagram of the deep decoupling multi-attention optimization network (DMAO-Net) is shown

[0187] Shared feature encoding: first, the basic features of the input multi-modal feature tensor are obtained through a shared bottom layer encoding network.

[0188] Decoupled parallel processing: three functionally decoupled branches, namely gated cross-modal attention (GCMA), which is responsible for fault location output Pnode; time causal dilated convolution (TCausal-DC) branch, which is responsible for time warning output Tw, R(t); and physical information energy regression (PIER) branch, which is responsible for energy equivalent evaluation output Qe.

[0189] Uncertainty weighted fusion: the output derived from the branch is finally sent to the uncertainty weighted fusion layer, where the uncertainty change inherent in each branch prediction is used to construct the weighting coefficient, and the relevant position, warning, energy and risk index in the comprehensive prediction information are weighted and fused.

[0190] Figure 3 The physical layout and corresponding relationship of the flexible partition fire extinguishing actuator and the battery module are shown:

[0191] Spatial partition layout: the battery module array can be expressed as: the battery module array is divided into "upper zone" and "lower zone" two separate protection areas.

[0192] Mechanism component correspondence: each zone is equipped with a set of relatively independent fire extinguishing units, including fire extinguishing agent storage tanks (such as S1, S2) and a set of spray arrays arranged in each zone through flexible pipelines (such as N1 / N2 / M1 / M2), the figure shows that the angle of the spray head can be adjusted to cover all the battery monomers in the corresponding zone.

[0193] Precise targeting relationship: it can be seen that the spatial "partitioned tank-flexible pipeline-adjustable spray head" and the "corresponding partitioned battery module" have spatial targeting, which is the physical basis for precise intervention.

[0194] Figure 4 The whole process of intelligent decision-making steps of model predictive control closed-loop fire extinguishing logic is shown:

[0195] System startup and data acquisition: after starting the program, it enters the process of real-time acquisition of dynamic multi-modal sensor data of the system.

[0196] Data processing and model inference: after adaptive filtering and feature extraction of the collected data, the extracted features are input into the DMAO-Net prediction model for real-time inference.

[0197] Risk judgment and early warning triggering: judge whether the risk prediction value output by the model is greater than the threshold value, if it is less than or equal to the threshold value, return to continue data acquisition, otherwise immediately trigger the early warning signal.

[0198] Optimization decision making: when the early warning signal is sent, the MPC is started, and the state space model of the controlled object is established with the predicted output, and the optimal control instruction is obtained by solving the optimization problem in the bounded finite time domain.

[0199] Precise execution and feedback: the corresponding partition fire extinguishing equipment is driven to perform the first injection, and a certain delay is performed, the effect after fire extinguishing is observed in the delay time, and the data information of the environment of the area is collected.

[0200] Effect evaluation and closed loop adjustment: whether the fire condition is stable is evaluated according to the feedback, and the working state of the system is changed according to the evaluation result: if it is stable, the system enters the enhanced monitoring mode; if it is not stable, the current condition is updated and the MPC optimization is re-executed, the action is taken to supplement the injection amount, and the cycle of 'perception-decision-making-execution-feedback' control is performed until the fire is reduced or the warning standard is reached.

[0201] The application provides a lithium battery thermal runaway intelligent control method and system, and there are many methods and ways to specifically implement the technical scheme. The above description is only the preferred embodiment of the application, and it should be pointed out that for ordinary skilled persons in the art, some improvements and refinements can be made without departing from the principle of the application, and these improvements and refinements should also be regarded as the protection scope of the application. The components not explicitly described in the embodiment can be realized by using the existing technology.

Claims

1. A method for intelligent control of thermal runaway in lithium batteries, characterized in that, Includes the following steps: Step 1: The sensor array deployed on the battery module can simultaneously collect the original multimodal state data of the battery module, clean and standardize it, and the original multimodal state data package contains voltage, current, temperature, characteristic gas concentration and vibration signal, etc. Step 2: Adaptive sliding window spectral filtering is used for noise reduction, and multi-scale feature fusion method is used to extract key features and generate a fused feature dataset. Step 3: The fused feature dataset is fed into the deep decoupled multi-attention optimization network model. After parallel processing through the localization branch, the temporal warning branch and the energy regression branch, the output results are evaluated for the probability distribution of abnormal single-unit localization, the remaining warning time of thermal runaway and the potential energy equivalent. The uncertainty of the output results is weighted and then fused to obtain the comprehensive prediction result. Step 4: Determine whether the warning time is less than the set threshold. If it is less, trigger the fire extinguishing decision. Based on the comprehensive prediction results, apply the nonlinear dose mapping model to calculate the total demand for fire extinguishing agent, and combine the model prediction control framework to solve the optimization problem and obtain the zonal fire extinguishing control command. Step 5: Drive the flexible fire suppression actuators of each zone to achieve quantitative and directional fire suppression according to the zone fire suppression control command; after the fire is extinguished, collect environmental feedback information to evaluate the fire suppression effect and achieve closed-loop control.

2. The method according to claim 1, characterized in that, Step 2 includes: Step 2.1, Data Preprocessing; Preprocessing includes Adaptive Sliding Window Spectral Filtering (AWSF) and Multi-Scale Feature Fusion (MSFF), calculated using the following formula: （1）， （2）， in, Let be the length of the dynamic window at time t; The dynamic cutoff frequency at time t: For local signal-to-noise ratio estimation, As the baseline window length, These are preset parameters; sigmoid is a non-linear activation function; exp is the natural exponential function. Step 2.2, Multi-scale feature fusion and deep correlation extraction; Multi-scale feature fusion is performed on the filtered multimodal data; a feature association graph is constructed by calculating the mutual information I(m;n) between different modal features m and n, and edge weights are applied. The calculation formula is: （3）， in The linear correlation coefficient between modes; For relational bandwidth parameters; And the graph convolutional network (GCN) is used to process the original features. Enhancement is performed to extract deep correlation features. : (4), in, For an adjacency matrix with self-connections, This is the corresponding degree matrix. σ is a trainable weight matrix, and σ is the activation function, ultimately yielding the fused feature dataset.

3. The method according to claim 2, characterized in that, In step 3, the positioning branch is used for precise location of the abnormal single entity, including: A gated cross-modal attention mechanism is adopted; firstly, the attention weights between features of different modalities are calculated. : (5), Where Softmax is the normalized exponential function; T denotes transpose; Representing the query matrix and key matrix respectively These are the query weight matrix and the key weight matrix, respectively. Furthermore, a gating vector g is introduced for dynamic information selection and fusion: (6), in, This is a trainable weight matrix; These are the temperature characteristic tensor and the gas characteristic tensor, respectively. For trainable bias vectors; fused features Represented as ; For Hadamah accumulation; Value matrix It is the voltage characteristic tensor; Finally, the topology information of the battery module is aggregated through a graph attention network, and the anomaly probability distribution of each cell within the battery module is output. ; The timing warning branch is used for predicting the remaining time of thermal runaway, including: A stacked causal dilated convolutional structure is used to capture long-term temporal dependencies; the output of the l-th layer at time t for: (7), in The weight parameter is the i-th weight parameter of the l-th convolutional kernel; ReLU is the activation function, and the dilation rate is... The final output is the predicted remaining time of thermal runaway. and the dynamic risk index R(t); This is the bias term for the l-th layer; The energy regression branch is used for potential energy equivalent assessment, including: Physical consistency constraints are introduced into the regression network; the energy regression branch regresses the basic energy equivalent using a multilayer perceptron. And reaction parameters, and calculate physical consistency loss. Constraints on network training: (8), in, This refers to the physical loss weighting coefficient; To predict the derivative of temperature with respect to time; Heat capacity; Pre-exponential factors; For activation energy, This is the product of the gas constant and the predicted temperature. The term is an exponential term related to the predicted energy; hA is the heat dissipation coefficient. The ambient temperature is used as the output; after physical layer correction, the final estimated value of the potential energy equivalent of thermal runaway is output. .

4. The method according to claim 3, characterized in that, In step 3, the outputs of the localization branch, the time-series early warning branch, and the energy regression branch are used to generate the final comprehensive prediction result through an adaptive weighted fusion based on uncertainty, and a triple mechanism of calibration factor, sharpening gating, and adaptive temperature coefficient. The formula is as follows: （9）， in, The adaptive fusion weights are the output of the i-th branch; This is the comprehensive prediction output vector; Normalized uncertainty estimate for the i-th branch; Let be the uncertainty calibration factor for the i-th branch; Let be the adaptive temperature coefficient of the i-th branch; The proportional gain of the saturation gate function; This represents the original output of the i-th branch; It is a saturation-gated function; The decision-making sharpening index.

5. The method according to claim 4, characterized in that, In step 4, the total demand for extinguishing agent is calculated using the following nonlinear dose mapping model. : （10）， in, This refers to the basic efficiency coefficient of the extinguishing agent. The attenuation coefficient is determined by the urgency of the situation. These are the energy equivalent influence coefficient and the risk change rate influence coefficient, respectively. This refers to the remaining warning time for thermal runaway; This is an exponential effect term based on energy equivalent; This is the time derivative of the dynamic risk index.

6. The method according to claim 5, characterized in that, Step 4 also includes: Step 4.1, State Modeling and Objective Function Definition; Establish based on the temperature of the target area Characteristic gas concentration Remaining extinguishing agent For a discrete state-space model of state variables, the state vector is: Define the following optimization objective function: （11）， in, This is a safety status reference value. The state deviation penalty weight matrix is... For the control quantity penalty weight matrix, Terminal state penalty weight matrix, For prediction in the time domain; Index for the maximum number of control steps in the time domain Let k be the system state predicted at time k+i in the future. The control input calculated at time k and intended to be applied at a future time k+i; For the prediction of the time-domain terminal at time k ( The system state at any given time; Step 4.2, Constraint Integration; Total demand for fire extinguishing agents As key resource constraints, combined with system physical limitations, a complete set of constraints for the optimization problem is formed: （12）， in, This represents the system state vector predicted at time k for the (i+1)th future step. , Let represent the state transition matrix and control input matrix predicted at the i-th step in the current control period k, respectively; This represents the predicted control quantity for the i-th step. Δt is the upper limit of the control quantity; Δt is the control period. This represents the amount of extinguishing agent consumed before the current control cycle. , These represent the lower and upper limits of the state variable, respectively. For control input at time k+i in the future; Step 4.3: Online rolling optimization solution; In each control cycle Based on the latest system status and prediction models A numerical optimization solver is used to solve the optimization problem online, obtaining the optimal control sequence in the prediction time domain. ; Step 4.4, Instruction generation and output; The first element of the optimal control sequence This is the control command for the current cycle, namely the zone extinguishing agent flow command.

7. The method according to claim 6, characterized in that, In step 4.1, the model predictive control framework adopts an adaptive weight adjustment strategy; the weight matrix... And R based on the warning time and risk index Dynamic adjustment: （13）， （14）， in, As the benchmark weight, For adjustment coefficients; The baseline weight matrix for control quantity penalty; Let R(t) be the time derivative of the dynamic risk index.

8. The method according to claim 7, characterized in that, Step 4 also includes a closed-loop correction mechanism: After each control cycle is completed, the actual measured system state will be recorded. Compared with the state predicted in the previous cycle Compare and calculate the prediction error ; Utilizing error The state-space model parameters in the subsequent prediction time domain are analyzed using a Kalman filter or moving average method. Perform online fine-tuning.

9. The method according to claim 8, characterized in that, In step 5, the flexible fire extinguishing actuator includes: Zoned storage tank unit: It adopts an independent upper and lower zone design, and each zone is filled with thermal aerosol fire extinguishing agent; the zoned storage tank unit is equipped with pressure sensors and temperature sensors for real-time monitoring of the internal status; Flexible connection piping: High-temperature and corrosion-resistant flexible metal corrugated pipes or silicone pipes are used to connect the storage tank and the nozzle; Multi-degree-of-freedom nozzle array: Each zone is equipped with an adjustable nozzle, which is connected by a spherical hinge and has two rotational degrees of freedom. The angle is preset during the initialization phase; the nozzle has an atomizing core inside. High-speed solenoid valve: Each tank outlet is equipped with a high-speed response solenoid valve, which is directly controlled by the intelligent processing and decision-making unit; Adaptive sealing device: An adaptive sealing ring is provided between the nozzle and the battery box mounting hole. The adaptive sealing ring is made of a composite material of shape memory alloy and high temperature resistant rubber. In step 5, the workflow of the flexible zone fire suppression mechanism includes: Command reception: The control circuit of the corresponding zone receives injection commands from the MPC controller, including injection duration and flow rate; Valve activation: The high-speed solenoid valve opens within milliseconds, and the extinguishing agent, driven by the pressurized gas inside the storage tank, flows to the nozzle array through the flexible pipeline; Directional atomization spray: The extinguishing agent is pulverized into micron-sized particles through the atomizing core inside the nozzle, forming an aerosol smoke that precisely covers the faulty battery and surrounding area at a preset angle; Flow monitoring and regulation: The flow rate is regulated by using PWM control of the solenoid valve and combining it with tank pressure feedback to match the injection curve optimized by MPC. Feedback and shutdown: After the predetermined spray volume is reached, the solenoid valve closes; the fire extinguishing effect is monitored in real time by temperature and gas sensors, and the data is fed back to the MPC controller to determine whether supplementary spraying is needed.

10. A lithium battery thermal runaway intelligent control system, used to execute the method as described in any one of claims 1 to 9, characterized in that, include: A multi-source sensing unit is configured to perform step 1 of the method, synchronously acquiring raw multimodal state data of the battery module through a sensor array arranged in the battery module; The intelligent processing and decision-making unit is configured to perform steps 2 to 4 of the method, preprocess the original multimodal state data, fuse features, perform comprehensive prediction using a deep decoupled multi-attention optimization network model, and generate zonal fire extinguishing control commands based on the model prediction control framework when the triggering conditions are met. The precision execution and feedback unit is configured to execute step 5 of the method, drive the flexible fire extinguishing actuator to extinguish the fire according to the zone fire extinguishing control command, and collect feedback data to form a closed-loop control.

Images