Battery thermal management adaptive method and system based on deep learning

Abstract

The invention relates to the technical field of battery management, and discloses a battery thermal management adaptive method and system based on deep learning, and the system comprises a data collection module which is used for obtaining the real-time operation data of a battery; the failure mode diagnosis module is used for identifying specific failure modes such as internal short circuit risk and reversible concentration polarization of the single cell on line by using models such as a convolutional neural network; the joint prediction module is used for synchronously predicting future temperature and health risk indexes by using a bidirectional long-short-term memory network; the adaptive control module selects an optimal control strategy according to the failure mode and reconstructs a model prediction control problem; the instruction generation module is used for solving a model prediction control problem and generating a multi-domain cooperative control instruction of power management and heat management; and the closed-loop feedback module is used for correcting the control deviation. Through deep fusion of a deep learning diagnosis prediction model and a self-adaptive model prediction control framework, conversion from passive response to active pre-judgment is realized.

Description

Technical Field

[0001] This invention relates to the field of battery management technology, specifically to an adaptive method and system for battery thermal management based on deep learning. Background Technology

[0002] Lithium-ion batteries are widely used in electric vehicles, consumer electronics, and large-scale energy storage systems due to their high energy density and long cycle life. The battery management system (BMS) is a core component ensuring the safe, reliable, and efficient operation of the battery system, with thermal management being one of its key functions. The performance, lifespan, and safety of lithium-ion batteries are extremely sensitive to operating temperature; excessively high or low temperatures can lead to performance degradation, shortened lifespan, and even serious safety incidents such as internal short circuits and thermal runaway.

[0003] Existing battery thermal management systems are typically based on threshold-triggered reactive control logic. Specifically, the system monitors the battery system temperature using temperature sensors. When the monitored temperature exceeds a preset safety threshold, it triggers preset cooling or heating actions, such as activating a fan, water pump, or heating element, to maintain the battery temperature within a preset ideal operating range.

[0004] However, this reactive control mechanism inherently suffers from hysteresis. The control system's response occurs after the thermal event has already occurred and been detected by sensors. For rapidly developing early failure modes such as internal short circuits, this delay leads to untimely intervention, failing to effectively suppress further deterioration of the thermal anomaly, or even prevent thermal runaway. The system's intervention always lags behind the physical process of failure, lacking foresight and proactivity.

[0005] Furthermore, existing technologies lack specificity in their control strategies. Their control logic typically uses temperature exceeding limits as the sole criterion for judgment, failing to identify the underlying physicochemical causes of temperature anomalies online, such as lithium plating, internal micro-short circuits, or concentration polarization. Consequently, the system cannot execute optimized intervention measures for specific failure modes; all thermal anomalies trigger similar or identical general cooling responses, leading to suboptimal control behavior and even producing unintended effects under certain operating conditions (such as concentration polarization requiring resting recovery).

[0006] Existing technologies also suffer from insufficient synergy between electrical and thermal management. Typically, electrical management systems (such as power limiting) and thermal management systems (such as cooling system control) operate as two relatively independent subsystems, lacking unified and forward-looking collaborative optimization of their control objectives and action sequences. For example, the system may only address temperature increases through coarse power reduction, failing to find the optimal operating point that maintains battery performance while meeting safety boundaries through joint electrothermal regulation. This functional separation limits the overall control performance and safety margin of the system.

[0007] Therefore, existing technologies exhibit limitations in dealing with the complex and dynamically changing internal states of battery systems, such as delayed response, single strategy, and lack of multi-domain coordination. There is an urgent need for an active management method that can diagnose early failure modes online and implement proactive, adaptive, and multi-domain coordinated control. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a deep learning-based adaptive method and system for battery thermal management. This solves the problems of existing technologies, such as difficulty in proactively intervening in early thermal failures due to lag, single and nonspecific control strategies due to the inability to identify the root causes of failure modes online, and lack of coordinated optimization control between electrical management and thermal management, which are independent of each other.

[0009] The first aspect of this invention provides a deep learning-based adaptive method for battery thermal management. By combining online failure mode diagnosis, dual-track joint prediction, and adaptive model predictive control, it achieves precise, feedforward-oriented, multi-domain coordinated regulation of individual battery cells. The method includes the following steps:

[0010] Step 1: Real-time data acquisition

[0011] Real-time operating data of each individual cell in the battery system is collected, and the real-time operating data includes at least voltage timing data and current timing data.

[0012] Step 2: Online identification of individual cell failure modes

[0013] Based on the voltage and current time-series data collected in step 1, the failure mode of each individual cell is identified online. Specifically, this step involves: first, inputting the voltage and current time-series data into a pre-trained convolutional neural network to extract an impedance fingerprint feature vector that characterizes the internal electrochemical properties of the individual cell. The impedance fingerprint feature vector is constructed as follows:

[0014] Z fp =[R o ,R ct Z sei Z diff ];

[0015] Among them, Z fp R is the impedance fingerprint feature vector; o R is an estimated ohmic internal resistance, characterizing the total resistance of the electrolyte, diaphragm, and current collector; ct Z is an estimated value for charge transfer resistance, characterizing the kinetics of the electrochemical reaction at the electrode-electrolyte interface; sei Z represents the impedance characteristic parameter in the mid-frequency range, characterizing the properties of the solid electrolyte interface film; diffThe impedance characteristic parameter in the low-frequency band characterizes the diffusion process of lithium ions inside the electrode material.

[0016] Subsequently, the extracted impedance fingerprint feature vector Z fp The rate of change of the impedance fingerprint feature vector over time is input to a classifier, which outputs the current failure mode of the single cell.

[0017] Step 3: Simultaneous Prediction of Temperature and Health Risk Index

[0018] Based on the failure modes of the individual battery cells identified in step 2 and the real-time operating data collected in step 1, an enhanced input is constructed, and a joint prediction model is used to synchronously output the predicted temperature trajectory and health risk index trajectory of the individual battery cells within a future time window.

[0019] The enhanced input is configured as follows:

[0020] X enh =[M onehot D realtime Z fp ];

[0021] Among them, X enh For the enhanced input; M onehot This represents the unique thermal encoding of the failure mode of the individual battery cell; D realtime The real-time operating data; Z fp The impedance fingerprint feature vector is denoted as .

[0022] The joint prediction model is a bidirectional long short-term memory network model, which contains two independent output heads, used to output the predicted temperature trajectory and health risk index trajectory, respectively.

[0023] Step 4: Adaptive Control Strategy Selection

[0024] Based on the failure modes of the individual battery cells identified online in step 2, a currently applicable control strategy is selected from a predefined adaptive control strategy library. The adaptive control strategy library L... strategy The composition is as follows:

[0025] L strategy ={S normal ,S emergency ,S repair};

[0026] Among them, L strategy For the aforementioned adaptive control strategy library; S normal For normal operating conditions, the standard operating strategy; S emergencyFor emergency intervention strategies used in operating conditions with internal short-circuit risk; S repair This is an active repair strategy for reversible concentration polarization states.

[0027] Step 5: Dynamic Reconstruction of the Model Predictive Control Problem

[0028] Based on the control strategy selected in step 4, the objective function J and constraints of the model predictive control are dynamically reconstructed, resulting in the dynamically reconstructed model predictive control problem. The dynamic reconstruction of the objective function J specifically involves:

[0029] When selecting S normal When the objective function J is reconstructed into J normal :

[0030] Formula: J normal =min∑[w T ·||T i (t(-T ref || 2 +w E ·P total (t)];

[0031] Among them, J normal The objective function is T under the normal operating strategy; i (t) represents the predicted temperature of cell i at a future time t; T ref The target temperature; P total (t) represents the total power consumption of the system; w T The weighting coefficient for the temperature tracking term; w E This represents the weighting coefficient for the system's economic efficiency.

[0032] When selecting S emergency When the objective function J is reconstructed into J emergency :

[0033] Formula: J emergency =min∑[w γ ·(γ i ) 2 +w T,adj ·||T adj (t)-T safe || 2 ];

[0034] Among them, J emergency The objective function under the emergency intervention strategy; γ i T is the active current suppression coefficient applied to high-risk single cell i; adj (t) represents the predicted temperature of adjacent cells to the high-risk single cell; T safe For safe temperature threshold; w γ The weighting coefficient for the current suppression term; wT,adj This represents the weighting coefficient for the temperature control terms of adjacent battery cells.

[0035] When selecting S repair When the objective function J is reconstructed into J repair :

[0036] Formula: J repair =min∑[w HRI ·||HRI i (t)|| 2 ];

[0037] Among them, J repair The objective function under the active repair strategy; HRI i (t) represents the health risk index of a single battery cell i at future time t; w HRI This represents the weighting coefficient for the health risk index item.

[0038] Step 6: Generation and Issuance of Multi-Domain Cooperative Control Commands

[0039] Solve the dynamically reconstructed model predictive control problem obtained in step 5, and generate and issue multi-domain collaborative control commands covering thermal management and electrical management. The multi-domain collaborative control commands include: a refrigerant flow vector command Q, accurate to the individual battery cell level, for thermal management. vec Fan speed command ω fan and compressor operating condition command C state ; and the active current suppression coefficient γ applied to a specific individual cell for electrical management. i and restorative pulse current parameter I pulseo .

[0040] Step 7: Dynamic adjustment of control loop gain parameters

[0041] The actual temperature of the individual battery cell after the execution of the multi-domain collaborative control command in step 6 is collected, and the gain parameter of the control loop is dynamically adjusted using an adaptive law based on the deviation between the actual temperature and the predicted temperature trajectory. The adjustment process of the gain parameter is defined by the following formula:

[0042] K p,i (t)=K p0 +k i ·∫e i (τ)dτ;

[0043] Among them, K p,i (t) represents the control gain parameter of a single cell i at the current time t; K p0 k is the initial gain parameter. i e is the integral adjustment coefficient for cell i; i(τ) represents the deviation between the actual temperature and the predicted temperature of a single cell i over a past time period τ; ∫e i (τ)dτ is the integral of the deviation signal over time.

[0044] A second aspect of the present invention provides a deep learning-based adaptive battery thermal management system, the system comprising:

[0045] The data acquisition module is used to collect real-time operating data of a single battery cell, including voltage and current timing data.

[0046] The failure mode diagnosis module is used to identify the failure mode of the individual battery cell online based on the voltage and current timing data.

[0047] The joint prediction module is used to construct an enhanced input based on the failure mode and the real-time operating data, and simultaneously output the temperature trajectory and health risk index trajectory of the individual cell within a future time window;

[0048] The adaptive control module is used to select a currently applicable control strategy from a predefined adaptive control strategy library based on the identified failure modes, and dynamically reconstruct the objective function and constraints of the model predictive control based on the selected control strategy.

[0049] The instruction generation module is used to solve the dynamically reconstructed model predictive control problem and generate and issue multi-domain collaborative control instructions covering thermal management and electrical management.

[0050] The closed-loop feedback module is used to collect the actual temperature of a single cell after the control command is executed, and dynamically adjust the gain parameters of the control loop based on the deviation between the actual temperature and the predicted temperature trajectory.

[0051] This invention provides a deep learning-based adaptive method and system for battery thermal management. It offers the following advantages:

[0052] 1. This invention achieves accurate identification of specific failure modes by extracting impedance fingerprint feature vectors that characterize the internal electrochemical properties of a single cell online. This enables the thermal management system to anticipate early faults such as internal short circuits or lithium plating, transforming control from a passive temperature response to proactive intervention at the source of risk, thereby improving the operational safety of the battery system.

[0053] 2. This invention achieves adaptive switching of control strategies by selecting a strategy from a predefined adaptive control strategy library based on the identified failure modes and dynamically reconstructing the objective function of model predictive control. This approach can perform optimized multi-domain collaborative control for different operating conditions such as normal, emergency, or repair, avoiding the limitations of traditional fixed strategies in dealing with changing operating conditions and achieving a dynamic balance between safety, performance, and economy.

[0054] 3. This invention collects the actual temperature of a single cell after the control command is executed and compares it with the predicted temperature trajectory. Then, it dynamically adjusts the gain parameters of the control loop to form a complete adaptive closed loop. This closed-loop feedback mechanism can continuously correct the deviation between the model prediction and the actual operating conditions, effectively suppress the decline in control performance caused by battery aging or model mismatch, and ensure the control accuracy and robustness of the thermal management method throughout its entire life cycle. Attached Figure Description

[0055] Figure 1 This is an overall functional block diagram of the adaptive thermal management system for the battery system of the present invention;

[0056] Figure 2 This is an overall flowchart of the adaptive thermal management method of the present invention;

[0057] Figure 3 This is an internal functional block diagram of the failure mode diagnosis module of the present invention;

[0058] Figure 4 This is an internal functional block diagram of the joint prediction module of the present invention;

[0059] Figure 5 This is a flowchart illustrating the strategy selection and problem reconstruction process of the adaptive control module of the present invention.

[0060] Figure 6 This is a schematic diagram of the emergency intervention strategy workflow of the present invention;

[0061] Figure 7 This is a schematic diagram of the active repair strategy workflow of the present invention.

[0062] The system comprises: 1. Battery system; 2. Actuator; 10. Data acquisition module; 20. Failure mode diagnosis module; 30. Joint prediction module; 40. Adaptive control module; 50. Command generation module; and 60. Closed-loop feedback module. Detailed Implementation

[0063] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] See attached document Figure 1 , Figure 1 This is an overall functional block diagram of a battery system adaptive thermal management system according to an embodiment of the present invention. The present invention provides a deep learning-based adaptive battery thermal management system, which interacts with and physically controls a battery system 1 and a set of actuators 2. The system may include: a data acquisition module 10, a failure mode diagnosis module 20, a joint prediction module 30, an adaptive control module 40, an instruction generation module 50, and a closed-loop feedback module 60.

[0065] The data acquisition module 10 is used to acquire real-time operating data of each individual cell in the battery system 1. The real-time operating data includes voltage timing data, current timing data, and temperature data. The data acquisition module 10 is connected to the failure mode diagnosis module 20 and the joint prediction module 30, respectively, and is used to transmit the acquired real-time operating data to these two modules.

[0066] The failure mode diagnosis module 20 has its input connected to the data acquisition module 10. The failure mode diagnosis module 20 receives voltage and current timing data and identifies the failure modes of individual cells online based on the data. Specifically, this process involves extracting an impedance fingerprint feature vector from the voltage and current timing data using a pre-trained convolutional neural network, the structure of which is as follows:

[0067] Z fp =[R o ,R ct Z sei Z diff ];

[0068] Among them, Z fp R is the impedance fingerprint feature vector; o R is an estimated ohmic internal resistance, characterizing the total resistance of the electrolyte, diaphragm, and current collector; ct Z is an estimated value for charge transfer resistance, characterizing the kinetics of the electrochemical reaction at the electrode-electrolyte interface; sei Z represents the impedance characteristic parameter in the mid-frequency range, characterizing the properties of the solid electrolyte interface film; diff The impedance characteristic parameter in the low-frequency band characterizes the diffusion process of lithium ions inside the electrode material.

[0069] Subsequently, a classifier analyzes the obtained impedance fingerprint feature vector and its rate of change over time, outputting an identifier representing the current failure mode. The output of the failure mode diagnosis module 20 is connected to the joint prediction module 30 and the adaptive control module 40, respectively.

[0070] The joint prediction module 30 has its input terminals connected to the data acquisition module 10 and the failure mode diagnosis module 20, respectively. The joint prediction module 30 receives real-time operating data and failure mode identifiers to construct an enhanced input, which is configured as follows:

[0071] X enh =[M onehot D realtime Z fp ];

[0072] Among them, X enh For enhanced input; M onehot The unique thermal encoding representation of the failure modes of a single cell; D realtime For real-time running data; Z fp This is the impedance fingerprint feature vector.

[0073] The joint prediction module 30 processes the enhanced input through a bidirectional long short-term memory network model, and simultaneously outputs the predicted temperature trajectory and health risk index trajectory of a single battery cell within a future time window. The output of the joint prediction module 30 is connected to the adaptive control module 40 and the closed-loop feedback module 60, respectively.

[0074] The adaptive control module 40 has its inputs connected to the failure mode diagnosis module 20 and the joint prediction module 30, respectively. Based on the received failure mode identifier, the adaptive control module 40 selects a currently applicable control strategy from a predefined adaptive control strategy library containing normal operation strategies, emergency intervention strategies, and active repair strategies. Subsequently, based on the selected control strategy, it dynamically reconstructs the objective function J and constraints of the model predictive control. The output of the adaptive control module 40 is connected to the instruction generation module 50 for transmitting the reconstructed control problem.

[0075] The instruction generation module 50 has its input connected to the adaptive control module 40 and its output connected to the actuator 2. The instruction generation module 50 receives the reconstructed control problem and solves it using a numerical optimization solver to generate multi-domain coordinated control instructions covering thermal and electrical management. When the control strategy is in normal operation, the objective function J has the following form:

[0076] Formula: J normal =min∑[w T ·||T i (t(-T ref ||2 +w E ·P total (t)];

[0077] Among them, J normal The objective function is T under the normal operating strategy; i (t) represents the predicted temperature of cell i at a future time t; T ref The target temperature; P total (t) represents the total power consumption of the system; w T The weighting coefficient for the temperature tracking term; w E This represents the weighting coefficient for the system's economic efficiency.

[0078] When the control strategy is emergency intervention, the objective function J takes the form:

[0079] Formula: J emergency =min∑[w γ ·(γ i ) 2 +w T,adj ·||T adj (t)-T safe || 2 ];

[0080] Among them, J emergency The objective function under the emergency intervention strategy; γ i T is the active current suppression coefficient applied to high-risk single cell i; adj (t) represents the predicted temperature of adjacent cells to the high-risk single cell; T safe For safe temperature threshold; w γ The weighting coefficient for the current suppression term; w T,adj This represents the weighting coefficient for the temperature control terms of adjacent battery cells.

[0081] When the control strategy is active repair, the objective function J takes the form:

[0082] Formula: J repair =min∑[w HRI ·||HRI i (t)|| 2 ];

[0083] Among them, J repair The objective function under the active repair strategy; HRI i (t) represents the health risk index of a single battery cell i at future time t; w HRI This represents the weighting coefficient for the health risk index item.

[0084] The resulting multi-domain collaborative control command is sent to actuator 2, which includes the electric water pump and compressor of the refrigerant circulation system, as well as the relevant control units in the electrical management system.

[0085] The closed-loop feedback module 60 has its input terminals connected to the joint prediction module 30 and the temperature sensor of the battery system 1, respectively. The closed-loop feedback module 60 receives the predicted temperature trajectory and the actual temperature of the individual battery cells after the execution of the multi-domain collaborative control command. Based on the deviation between the two, the closed-loop feedback module 60 dynamically adjusts the gain parameter of the control loop through an adaptive law and feeds the adjusted gain parameter back to the control loop in the command generation module 50. The adjustment process of the gain parameter is defined by the following formula:

[0086] K p,i (t)=K p0 +k i ·∫e i (τ)dτ;

[0087] Among them, K p,i (t) represents the control gain parameter of a single cell i at the current time t; K p0 k is the initial gain parameter. i e is the integral adjustment coefficient for cell i; i (τ) represents the deviation between the actual temperature and the predicted temperature of a single cell i over a past time period τ.

[0088] Through the connection and information transmission between the above modules, this system constitutes a complete technical solution from state perception, failure diagnosis, trajectory prediction to adaptive control decision-making and closed-loop feedback.

[0089] See attached document Figure 2 , Figure 2 This is an overall flowchart of an adaptive thermal management method according to an embodiment of the present invention. In one embodiment of the present invention, the data acquisition module 10 and the corresponding data acquisition steps are described. The data acquisition module 10 establishes communication with the battery management unit of the battery system 1 through one or more hardware interfaces. The communication follows the Controller Area Network (CAN) bus protocol or other automotive Ethernet protocols.

[0090] The data acquisition module 10 periodically acquires real-time operating data for each individual battery cell and organizes this data into structured data frames. The acquired real-time operating data includes: individual cell voltage, individual cell current, and the values ​​from temperature sensors located on the surface of the individual cell. To meet the requirements of subsequent failure mode diagnosis for dynamic characteristic analysis, the sampling frequency for individual cell voltage and current is set to 10 Hz or higher. The sampling frequency for individual cell temperature is set to 1 Hz.

[0091] Before sending the collected raw data to subsequent modules, the data acquisition module 10 performs a data preprocessing procedure. The data preprocessing procedure includes two steps: data cleaning and data normalization. The data cleaning step uses a median filter to remove outliers introduced by instantaneous sensor fluctuations or communication interference.

[0092] The data normalization step linearly scales the cleaned data to a specified interval to meet the input requirements of the subsequent neural network model. For any input feature X, its normalized value Xnormal is... norm Calculated using the following formula:

[0093] formula:

[0094] Among them, X norm X represents the normalized eigenvalues; X represents the original eigenvalues; X min X is the minimum value of this feature within the preset operating range; max This is the maximum value of the feature within the preset operating range.

[0095] After preprocessing, the data acquisition module 10 sends the real-time running data packet containing normalized voltage time-series data, current time-series data, and temperature data to the failure mode diagnosis module 20 and the joint prediction module 30 respectively via the internal data bus.

[0096] See attached document Figure 3 , Figure 3 This is an internal functional block diagram of a failure mode diagnosis module according to an embodiment of the present invention. In one embodiment of the present invention, the failure mode diagnosis module 20 and the corresponding online failure mode identification steps are described. The input terminal of the failure mode diagnosis module 20 is connected to the data acquisition module 10 and is used to receive preprocessed voltage and current timing data.

[0097] The core function of the failure mode diagnosis module 20 is to estimate the impedance characteristics that reflect the internal electrochemical state of the cell based on the high-frequency voltage and current dynamic response, and then determine the failure mode of a single cell. This process includes two sequentially executed steps: impedance fingerprint feature vector extraction and failure mode classification.

[0098] In the impedance fingerprint feature vector extraction step, the failure mode diagnosis module 20 inputs voltage and current time-series data within a fixed time window as a two-dimensional data matrix into a pre-trained convolutional neural network. The convolutional neural network is designed to capture the frequency domain information contained in the dynamic response of voltage to current excitation. Through multiple convolution and pooling operations, the convolutional neural network automatically extracts key features from the time-series data and finally outputs a four-dimensional impedance fingerprint feature vector through a fully connected layer. The impedance fingerprint feature vector is constructed as follows:

[0099] Z fp =[R o ,R ct Z sei Z diff ];

[0100] Among them, Z fp R is the impedance fingerprint feature vector; o R is an estimated ohmic internal resistance, characterizing the total resistance of the electrolyte, diaphragm, and current collector; ct Z is an estimated value for charge transfer resistance, characterizing the kinetics of the electrochemical reaction at the electrode-electrolyte interface; sei Z represents the impedance characteristic parameter in the mid-frequency range, characterizing the properties of the solid electrolyte interface film; diff The impedance characteristic parameter in the low-frequency band characterizes the diffusion process of lithium ions inside the electrode material.

[0101] In the failure mode classification step, the failure mode diagnosis module 20 uses a classifier to analyze the extracted impedance fingerprint feature vector. To capture the gradual change of failure modes, the data input to the classifier includes not only the impedance fingerprint feature vector Z at the current time. fp It also includes the rate of change ΔZ of the impedance fingerprint feature vector over time. fp (t). The rate of change is calculated as follows:

[0102] Formula: ΔZ f p(t) = Z fp (t)-Z fp (t-Δt);

[0103] Where, ΔZ fp (t) represents the rate of change of the impedance fingerprint feature vector at the current time t; Z fp (t) is the impedance fingerprint feature vector at the current time t; Z fp (t-Δt) is the impedance fingerprint feature vector at the time of the previous calculation cycle; Δt is the time interval of the calculation cycle.

[0104] The classifier is a pre-trained support vector machine model. The support vector machine model is based on the input [Z...]. fp (t),ΔZ fp The combined features (t) output a discrete failure mode identifier. In one embodiment, the failure mode identifier is defined as follows: 0 represents normal operating condition, 1 represents internal short-circuit risk condition, and 2 represents reversible concentration polarization state.

[0105] Finally, the failure mode diagnosis module 20 sends the calculated failure mode identifiers to the joint prediction module 30 and the adaptive control module 40 via the internal data bus, respectively, as direct inputs for subsequent prediction and control decisions.

[0106] See attached document Figure 4 , Figure 4 This is an internal functional block diagram of a joint prediction module according to an embodiment of the present invention. In one embodiment of the present invention, the joint prediction module 30 and the corresponding synchronous prediction steps for temperature and health risk indices are described. The input terminals of the joint prediction module 30 are connected to the data acquisition module 10 and the failure mode diagnosis module 20, respectively.

[0107] The function of the joint prediction module 30 is to generate the temperature trajectory and health risk index trajectory of a single cell within a preset time window in the future, based on current and historical real-time operating data and combined with the identified individual cell failure modes.

[0108] Upon receiving the data, the joint prediction module 30 first constructs an enhanced input. This enhanced input integrates heterogeneous data from different modules to form a feature set containing multi-dimensional information. The enhanced input is structured as follows:

[0109] X enh =[M onehot D realtime Z fp ];

[0110] Among them, X enh For enhanced input; M onehot This represents the unique thermal encoding of the individual cell failure mode received from the failure mode diagnosis module 20; D realtime This refers to real-time operational data received from the data acquisition module 10, including information such as voltage, current, and temperature; Z fp The impedance fingerprint feature vector received from the failure mode diagnosis module 20.

[0111] Subsequently, enhanced input X enh The data is fed into a pre-trained bidirectional long short-term memory (LSTM) network model. The bidirectional LSM network model contains one or more hidden layers, and its core feature is that, for the input time-series data, the model transmits and processes information in both the forward and reverse directions of time within each hidden layer. This bidirectional structure allows the model to utilize information before and after that moment when calculating the output at any given time, thereby capturing the bidirectional dependencies of the data in the time dimension.

[0112] The bidirectional long short-term memory (LSTM) network model's structure is designed with two independent output heads in the final output layer. The first output head is a fully connected layer used to map the high-dimensional features extracted by the bidirectional LSM network layer to a temperature trajectory sequence within the future prediction time window. The second output head is another independent fully connected layer used to map the same high-dimensional features to a health risk index trajectory sequence within the future prediction time window. The health risk index is a quantitative indicator whose value is positively correlated with the probability of thermal runaway or irreversible performance degradation in a single battery cell.

[0113] Through this dual-output design, the joint prediction module 30 can simultaneously generate two related but physically different prediction results based on the same set of inputs and core network calculations. Finally, the module sends the generated predicted temperature trajectory and predicted health risk index trajectory to the adaptive control module 40 and the closed-loop feedback module 60 respectively via the internal data bus.

[0114] See attached document Figure 5 , Figure 5 This is a flowchart illustrating the strategy selection and problem reconstruction process of an adaptive control module according to an embodiment of the present invention. In one embodiment, the adaptive control module 40 and the corresponding adaptive control strategy selection and model predictive control problem reconstruction steps are described. The input of the adaptive control module 40 receives a failure mode identifier output from the failure mode diagnosis module 20.

[0115] The adaptive control module 40 internally stores a predefined adaptive control strategy library. The adaptive control strategy library is structured as follows:

[0116] L strategy ={S normal ,S emergency ,S repair};

[0117] Among them, L strategy For adaptive control strategy library; S normal For normal operating conditions, the standard operating strategy; S emergency For emergency intervention strategies used in operating conditions with internal short-circuit risk; S repair This is an active repair strategy for reversible concentration polarization states.

[0118] The adaptive control module 40 executes a deterministic selection logic based on the received failure mode identifier. When the failure mode identifier is 0 (normal operating condition), S is selected. normal When the failure mode identifier is 1 (internal short circuit risk condition), select S. emergency When the failure mode identifier is 2 (reversible concentration polarization state), select S. repair .

[0119] After selecting an appropriate control strategy, the core function of the adaptive control module 40 is to dynamically reconstruct the objective function and constraints of the model predictive control. This reconstruction process involves adjusting the mathematical expression of the optimization problem according to the specific control objective of the selected strategy.

[0120] When the normal operating strategy S is selected normal At that time, the adaptive control module 40 sets the objective function of the reconstructed model predictive control to J. normal The optimization goal of this approach is to track the target temperature while also considering the economic efficiency of system operation. Simultaneously, the temperature and current constraints for all individual battery cells are set to standard upper and lower operating limits.

[0121] When choosing emergency intervention strategy S emergency At that time, the adaptive control module 40 switches the objective function of the reconstructed model predictive control to J. emergency Form. This objective function introduces an active current suppression coefficient γ for high-risk individual cells. i As a strong penalty, suppressing heat generation in the cell is given the highest priority. At the same time, the adaptive control module 40 modifies the constraints, adding forced cooling requirements for adjacent cells of the high-risk cell to the constraint set. Specifically, the predicted temperature upper limit of these adjacent cells is set to a safe threshold lower than the normal operating upper limit.

[0122] When selecting the active repair strategy S repair At that time, the adaptive control module 40 switches the objective function of the reconstructed model predictive control to J. repair The primary optimization objective is to minimize the predicted health risk index trajectory. Simultaneously, the adaptive control module 40 modifies the constraints, tightening the temperature constraints of the individual battery cells requiring repair to a preset safe repair temperature range, such as [20℃, 35℃], to ensure the repair process is carried out in a suitable temperature environment.

[0123] Finally, the adaptive control module 40 sends the model predictive control problem definition, which includes the specific form of the selected objective function, the weight coefficients of each optimization term, and the updated set of constraints, as a structured data packet to the instruction generation module 50.

[0124] In one embodiment of the present invention, the steps for generating and issuing the instruction generation module 50 and the corresponding multi-domain collaborative control instructions are described. The input terminal of the instruction generation module 50 is connected to the adaptive control module 40, and its output terminal is connected to the system actuator 2 via a control bus.

[0125] The function of the instruction generation module 50 is to receive the model predictive control problem dynamically reconstructed by the adaptive control module 40, solve the optimization problem, and generate a set of executable multi-domain cooperative control instructions. This process is repeated in each control cycle.

[0126] First, the instruction generation module 50 contains a prediction model for describing the thermoelectric behavior of the battery system. In one embodiment, this prediction model is a discrete-time state-space model, with the following form:

[0127] x(k+1)=Ax(k)+Bu(k);

[0128] y(k) = Cx(k);

[0129] Where k is the discrete time step; x(k) is the system's state vector at time k, containing the temperature of all individual battery cells; u(k) is the control vector applied at time k, containing refrigerant flow rate, fan speed, and current adjustment applied to a specific battery cell; y(k) is the system's output vector, related to the state vector x(k); A, B, and C are state-space matrices describing the system's dynamic characteristics. This predictive model is used to predict the evolution trajectory of the system's state (e.g., temperature) within a future time window under a given control input sequence.

[0130] Next, the instruction generation module 50 takes the received objective function and constraints, and combines them with the above prediction model to construct an optimization problem with a finite prediction time domain. The goal of this optimization problem is to find an optimal control input sequence u(k), u(k+1),...,u(k+N). p -1), so that in the prediction time domain N p Within this range, the selected objective function J is minimized.

[0131] When the received objective function is J normal In this context, the optimization problem seeks to find the control sequence that minimizes the sum of temperature tracking error and system energy consumption while satisfying standard operating constraints.

[0132] Formula: J normal =min∑[w T ·||T i (t)-T ref || 2 +w E ·P total (t)];

[0133] Among them, J normal The objective function is T under the normal operating strategy; i (t) represents the predicted temperature of cell i at a future time t; T ref The target temperature; P total(t) represents the total power consumption of the system; w T The weighting coefficient for the temperature tracking term; w E This represents the weighting coefficient for the system's economic efficiency.

[0134] When the received objective function is J emergency In this case, the optimization problem is to solve a control sequence with the objective of minimizing the current suppression coefficient and controlling the temperature of adjacent cells.

[0135] Formula: J emergency =min∑[w γ ·(γ i ) 2 +w T,adj ·||T adj (t)-T safe || 2 ];

[0136] Among them, J emergency The objective function under the emergency intervention strategy; γ i T is the active current suppression coefficient applied to high-risk single cell i; adj (t) represents the predicted temperature of adjacent cells to the high-risk single cell; T safe For safe temperature threshold; w γ The weighting coefficient for the current suppression term; w T,adj This represents the weighting coefficient for the temperature control terms of adjacent battery cells.

[0137] When the received objective function is J repair In this context, the optimization problem seeks to solve the control sequence with the objective of minimizing the health risk index.

[0138] Formula: J repair =min∑[w HRI ·||HRI i (t)|| 2 ];

[0139] Among them, J repair The objective function under the active repair strategy; HRI i (t) represents the health risk index of a single battery cell i at future time t; w HRI This represents the weighting coefficient for the health risk index item.

[0140] Since the objective function is quadratic and the prediction model and constraints are linear, this optimization problem is a quadratic programming problem. The instruction generation module 50 integrates a quadratic programming solver to efficiently compute the optimal control sequence in each control cycle.

[0141] Based on the principle of model predictive control, after obtaining the optimal control sequence for the entire prediction time domain, only the first element of the sequence is issued as the control command for the current moment. These commands are converted into specific physical control signals and sent to actuator 2. In one embodiment, the multi-domain cooperative control command specifically includes: a refrigerant flow vector command Q for thermal management. vec Fan speed command ω fan and compressor operating condition command C state ; and the active current suppression coefficient γ applied to a specific individual cell for electrical management. i Or restorative pulse current parameter I pulse The restorative pulse current parameters are generated based on the current adjustment amount obtained from the optimization solution, combined with a preset restorative current waveform library.

[0142] In one embodiment of the present invention, the dynamic adjustment steps of the closed-loop feedback module 60 and the corresponding control loop gain parameters are described. The input terminal of the closed-loop feedback module 60 is connected to the joint prediction module 30 and the temperature sensor array in the battery system 1, respectively, and its output terminal is connected to the command generation module 50.

[0143] The function of the closed-loop feedback module 60 is to calculate the deviation between the predicted temperature trajectory output by the joint prediction module 30 and the actual temperature measured by the temperature sensor of the battery system 1, and to adjust the parameters of the control loop online based on the deviation.

[0144] In each control cycle, the closed-loop feedback module 60 receives the predicted temperature trajectory of each individual cell within a future time window generated by the joint prediction module 30, and simultaneously receives the actual temperature value measured by the temperature sensor arranged on the corresponding individual cell in the battery system 1. For each individual cell i, the closed-loop feedback module 60 calculates a temperature deviation signal e. i (t), which is defined as the difference between the actual temperature and the predicted temperature.

[0145] Based on the calculated temperature deviation signal, the closed-loop feedback module 60 uses an integral adaptive law to update the gain parameters in the control loop online. The adjustment process of the gain parameters is defined by the following formula:

[0146] K p,i (t)=K p0 +k i ·∫e i (τ)dτ;

[0147] Among them, K p,i (t) represents the control gain parameter of a single cell i at the current time t; K p0 is the initial gain parameter, a constant set during system initialization; k ie is the integral adjustment coefficient for individual cell i, used to adjust the update rate of the gain parameters; i (τ) represents the deviation between the actual temperature and the predicted temperature of a single cell i over a past time period τ; ∫e i (τ)dτ is the integral of the deviation signal over time.

[0148] In each calculation cycle, the closed-loop feedback module 60 updates the control gain parameter K of all individual cells. p,i (t) is sent to instruction generation module 50. Instruction generation module 50 uses this updated gain parameter in its internal model predictive control algorithm to adjust the strength of its control output or correct the deviation of the predictive model, thereby dynamically eliminating the observed temperature deviation. This process constitutes a closed-loop feedback correction mechanism for compensating for model mismatch and time-varying characteristics of the system.

[0149] See attached document Figure 6 , Figure 6 This is a schematic diagram of an emergency intervention strategy workflow according to an embodiment of the present invention. In a working scenario, it is assumed that the battery system 1 is initially in a normal operating state. The data acquisition module 10 continuously acquires the voltage and current timing data of each individual cell in the battery system 1 at a frequency of 10 Hz, and acquires its temperature data at a frequency of 1 Hz.

[0150] At a certain moment, a micro-short circuit begins to form inside a single cell in battery system 1, denoted as cell k. The data acquisition module 10 detects an abnormal change in the dynamic response of the voltage and current of cell k and sends this timing data to the failure mode diagnosis module 20.

[0151] After receiving the timing data of cell k, the failure mode diagnosis module 20 extracts the impedance fingerprint feature vector Z through its internal convolutional neural network. fp Due to the presence of micro-short circuits, the extracted estimated ohmic resistance R... o It displays a value below the normal range. The failure mode diagnosis module 20 further adjusts the current Z... fp (t) and Z at the previous time step fp By comparing (t-Δt), the time rate of change ΔZ of the eigenvector is calculated. fp (t). Based on Z fp outliers of (t) and ΔZ fp The abrupt change characteristic of (t) indicates that the support vector machine classifier inside the failure mode diagnosis module 20 outputs a failure mode identifier of 1, representing an internal short circuit risk condition. fp It is simultaneously sent to the joint prediction module 30 and the adaptive control module 40.

[0152] Upon receiving the failure mode identifier 1, the adaptive control module 40 selects an emergency intervention strategy S from its predefined adaptive control strategy library. emergenacy Subsequently, the adaptive control module 40 changes the objective function of model predictive control from J... normal Dynamically refactored into J emergency Its specific form is as follows:

[0153] Formula: J emergency =min∑[w γ ·(γ i ) 2 +w T,adj ·||T adj (t)-T safe || 2 ];

[0154] Among them, J emergency Here, γ represents the objective function under the emergency intervention strategy; i is the index of a single battery cell, specifically referring to the high-risk cell k and its adjacent cells; i T is the active current suppression coefficient applied to high-risk single cell i; adj (t) represents the predicted temperature of adjacent cells to the high-risk single cell; T safe For safe temperature threshold; w γ The weighting coefficient for the current suppression term; w T,adj This represents the weighting coefficient for the temperature control terms of adjacent battery cells.

[0155] Meanwhile, in the constraints of model predictive control, the adaptive control module 40 sets the upper temperature limit of the adjacent cells of the high-risk cell k to a safe temperature threshold T. safe The redefined control problem is sent to instruction generation module 50.

[0156] Instruction generation module 50 receives J emergency This is a model predictive control problem with an objective function. Since w γ The active current suppression coefficient γ applied to cell k is set to a relatively large value in order to minimize the objective function. This is part of the optimal control sequence calculated by the quadratic programming solver within the instruction generation module 50. k This will be a significantly non-zero value. Meanwhile, in order to satisfy the temperature constraints on adjacent cells and minimize J... emergency The adjacent cell temperature term in the solution will lead to thermal management control commands that tend to maximize cooling capacity. For example, the refrigerant flow rate through cell k and its surrounding area will be set to the maximum value, and the compressor will be started.

[0157] The instruction generation module 50 sends the first control action of the optimal control sequence as the current instruction to the actuator 2. The electrical management unit in the actuator 2 suppresses the current flowing through the battery cell k according to the instruction, while the electric water pump and compressor of the thermal management system execute forced cooling instructions.

[0158] During this process, the closed-loop feedback module 60 continuously compares the actual temperature of cell k and its adjacent cells with the predicted temperature output by the joint prediction module 30. If there is a deviation between the actual temperature and the predicted temperature, the closed-loop feedback module 60 will adjust the control gain parameter K according to the deviation. p,i (t), and send the updated parameters to the instruction generation module 50 to correct the strength of the control instruction in the next control cycle.

[0159] See attached document Figure 7 , Figure 7 This is a schematic diagram of the active repair strategy workflow according to one embodiment of the present invention. In another working scenario, the workflow of the active repair strategy of the system of the present invention in a reversible concentration polarization state will be described. Assume that a single cell in battery system 1, denoted as cell j, has entered a reversible concentration polarization state due to excessive lithium-ion concentration gradient inside its electrode material after experiencing continuous high-rate discharge.

[0160] During periodic data acquisition, the data acquisition module 10 observes that the voltage response of cell j experiences an early and sharp drop at the end of the discharge phase, while its current data remains within the high-rate discharge range. This high-frequency timing data is then sent to the failure mode diagnosis module 20.

[0161] Failure Mode Diagnosis Module 20 processes the input data through its convolutional neural network to extract the impedance fingerprint feature vector Z. fp Due to the formation of concentration polarization, the low-frequency impedance characteristic parameter Z, which reflects the diffusion process of lithium ions within the electrode material, is affected. diff It exhibits a significantly increasing trend. Based on this characteristic and its rapidly increasing time-varying rate of change ΔZ fp (t), the support vector machine classifier in the failure mode diagnosis module 20 outputs a failure mode identifier of 2, representing a reversible concentration polarization state. The failure mode identifier 2 and the impedance fingerprint feature vector Z fp It is simultaneously sent to the joint prediction module 30 and the adaptive control module 40.

[0162] At the same time, the joint prediction module 30 receives identifier 2 and real-time operating data D containing voltage, current, and temperature. realtime And construct an enhanced input X enh Based on this, the bidirectional long short-term memory network model within the joint prediction module 30 predicts that, if the current operating conditions continue, the health risk index (HRI) of cell j will increase.j (t) will rise rapidly in the future time window, indicating the risk of lithium plating and irreversible capacity decay.

[0163] After receiving the failure mode identifier 2, the adaptive control module 40 selects the active repair strategy S from its strategy library. repair The adaptive control module 40 then dynamically reconstructs the objective function of the model predictive control into J. repair Its specific form is as follows:

[0164] Formula: J repair =min∑[w HRI ·||HRI i (t)|| 2 ];

[0165] Among them, J repair Here, represents the objective function under the active repair strategy; i is the index of a single battery cell, specifically cell j; HRI i (t) represents the health risk index of a single battery cell i at future time t; w HRI This represents the weighting coefficient for the health risk index item.

[0166] Simultaneously, the adaptive control module 40 modifies the constraints, tightening the temperature constraint of cell j to a preset safe repair temperature range, such as [20℃, 35℃]. The reconstructed control problem definition is sent to the instruction generation module 50.

[0167] The instruction generation module 50 receives a model predictive control problem with the objective of minimizing the health risk index. To achieve this objective, the optimal control sequence calculated by the quadratic programming solver will include a set of coordinated thermal and electrical management actions. Regarding electrical management, the solver outputs a current adjustment to eliminate concentration polarization, which is converted by the instruction generation module 50 into a specific restorative pulse current parameter I. pulse For example, the instruction may require cell j to undergo a resting period or be given a small charging current pulse. In terms of thermal management, the solver outputs thermal management instructions that maintain the temperature of cell j within the [20°C, 35°C] temperature range, such as adjusting the refrigerant flow rate or fan speed through that region.

[0168] The instruction generation module 50 will include the restorative pulse current parameter I. pulse The first control action of the optimal control sequence of thermal management commands is sent to actuator 2. The electrical management unit and thermal management unit in actuator 2 coordinate to execute these commands, performing active repair operations on cell j until the failure mode diagnosis module 20 detects a Z-type fault in cell j. diff The parameters are restored to the normal range, and the failure mode identifier 0 is output, thus restoring the system to its normal operating strategy.

Claims

1. A deep learning based adaptive method for battery thermal management, characterized in that, The method comprises the following steps: collecting real-time operation data of the single battery cell, including voltage and current time series data; online identifying the failure mode of the single battery cell based on the voltage and current time series data; constructing an enhanced input based on the failure mode of the single battery cell and the real-time operation data, and synchronously outputting the temperature trajectory and health risk index trajectory of the single battery cell in a future time window by using a joint prediction model; selecting a currently applicable control strategy from a pre-defined adaptive control strategy library according to the failure mode of the single battery cell identified online; dynamically reconstructing the objective function and constraint condition of model predictive control based on the selected control strategy, to obtain a dynamically reconstructed model predictive control problem; solving the dynamically reconstructed model predictive control problem, and generating and issuing multi-domain collaborative control instructions covering thermal management and electrical management; collecting the actual temperature of the single battery cell after the execution of the multi-domain collaborative control instructions covering thermal management and electrical management, and dynamically adjusting the gain parameter of the control loop according to the deviation between the actual temperature and the predicted temperature trajectory. 2.The deep learning based battery thermal management adaptive method of claim 1, wherein, The step of online identifying the failure mode of the single battery cell specifically comprises: using a pre-trained convolutional neural network to process the voltage and current time series data, and extracting an impedance fingerprint feature vector capable of representing the electrochemical characteristics inside the battery cell; using a classifier to analyze the impedance fingerprint feature vector and the change rate of the impedance fingerprint feature vector over time, and outputting the failure mode of the single battery cell. 3.The deep learning based battery thermal management adaptive method of claim 2, wherein, The impedance fingerprint feature vector comprises: an ohmic resistance estimation value representing the total resistance of the electrolyte and the current collector, a charge transfer resistance estimation value representing the reaction kinetics of the electrode interface, a medium frequency band impedance characteristic parameter representing the characteristics of the solid electrolyte interface film, and a low frequency band impedance characteristic parameter representing the lithium ion diffusion process. 4.The deep learning based battery thermal management adaptive method of claim 1, wherein, The joint prediction model is a bidirectional long short-term memory network model, and the enhanced input is a data set formed by integrating the one-hot encoding representation of the failure mode, the real-time operation data and the impedance fingerprint feature vector. 5.The deep learning based battery thermal management adaptive method of claim 1, wherein, The adaptive control strategy library comprises: a normal operation strategy for normal operation conditions, an emergency intervention strategy for internal short circuit risk conditions, and an active repair strategy for reversible concentration polarization states. 6.The battery thermal management adaptive method based on deep learning according to claim 5, wherein, The step of dynamically reconstructing the objective function and constraint condition of model predictive control specifically comprises: under the condition that the normal operation strategy is selected, the objective function sets tracking target temperature and ensuring system operation economy as the main optimization target; under the condition that the emergency intervention strategy is selected, the objective function switches to set inhibiting heat generation of high-risk single battery cells and reducing the temperature of adjacent single battery cells as the highest priority optimization target; under the condition that the active repair strategy is selected, the objective function switches to minimize the health risk index trajectory as the main optimization target, while maintaining the temperature of the single battery cell within a pre-set safe repair temperature zone.

7. The deep learning based battery thermal management adaptive method of claim 6, wherein, Under the condition that the emergency intervention strategy is selected, the switching of the objective function specifically comprises: An active current suppression coefficient for high-risk single cells is introduced in the objective function as a strong penalty term, and a forced cooling requirement for adjacent single cells is added in the constraint condition. 8.The deep learning based battery thermal management adaptive method of claim 1, wherein, The multi-domain collaborative control instruction includes: A refrigerant flow vector instruction, a fan speed instruction, and a compressor working condition instruction for thermal management at the single cell level; An active current suppression coefficient and a restorative pulse current parameter applied to a specific single cell for electrical management. 9.The deep learning based battery thermal management adaptive method of claim 1, wherein, The step of dynamically adjusting the gain parameter of the control loop is specifically: An adaptive law containing proportional and integral terms is used to continuously correct the value of the gain parameter based on the historical integral deviation between the actual temperature and the predicted temperature trajectory.

10. A battery thermal management adaptive system based on deep learning, characterized in that, It includes: A data acquisition module for acquiring real-time operation data of single cells, including voltage and current time series data; A failure mode diagnosis module for identifying the failure mode of the single cell online based on the voltage and current time series data; A joint prediction module for constructing an enhanced input based on the failure mode and the real-time operation data, and synchronously outputting the temperature trajectory and health risk index trajectory of the single cell within a future time window; An adaptive control module for selecting a currently applicable control strategy from a predefined adaptive control strategy library according to the identified failure mode, and dynamically reconstructing the objective function and constraint condition of model predictive control based on the selected control strategy; An instruction generation module for solving the dynamically reconstructed model predictive control problem, generating and issuing multi-domain collaborative control instructions covering thermal management and electrical management; A closed-loop feedback module for collecting the actual temperature of the single cell after the execution of the control instruction, and dynamically adjusting the gain parameter of the control loop according to the deviation between the actual temperature and the predicted temperature trajectory.

Images