Battery temperature and internal resistance collaborative estimation method driven by pure data extraction and physical information fusion

Abstract

A method for co-estimating battery temperature and internal resistance driven by pure data extraction and physical information fusion includes: 1) acquiring operating data of lithium-ion batteries, offline identification of parameters of the second-order equivalent circuit model, extracting the initial dynamic resistance and capacitance parameters of the battery, and calculating the preliminary heat generation characteristics of the battery; 2) establishing an adaptive jump mode decomposition model, decoupling the acquired signals, and extracting pure signals that accurately reflect the evolution trend of the battery's true state; 3) constructing a physical information network (PINN) integrating electrothermal coupling laws, which includes a dynamic internal resistance prediction sub-network and a temperature prediction main network incorporating a channel attention mechanism; 4) constructing a PINN network loss function by combining the battery thermodynamic equation into physical loss terms, data loss, and smoothness loss, and using a multi-stage dynamic weighting strategy to balance the contribution of the loss; and achieving real-time joint estimation of the battery's transient core temperature and dynamic internal resistance through the trained network.

Description

Technical Field

[0001] This invention relates to a method and system for co-estimating battery temperature and internal resistance based on pure data extraction and physical information fusion. Background Technology

[0002] With the widespread adoption and large-scale application of new energy vehicles and energy storage technologies, traditional power battery thermal management is transforming into a vehicle-cloud collaborative intelligent battery management model. The state of thermal (SOT) of a battery directly reflects the internal electrochemical reaction state, and accurately acquiring changes in SOT is a fundamental prerequisite for achieving efficient thermal management and thermal runaway early warning. Currently, existing technologies for acquiring the internal temperature of lithium-ion batteries mainly include direct measurement methods, traditional mechanistic model observation methods, and pure data-driven prediction methods. Direct measurement methods suffer from response lag and installation difficulties; mechanistic models have high computational dimensionality and are difficult to identify parameters; data-driven prediction methods rely on massive amounts of data and lack physical mechanisms. Specifically, in practical applications under complex operating conditions, the above-mentioned existing technologies have the following three significant drawbacks: For raw data processing, traditional methods cannot handle non-stationary signal interference under complex operating conditions. Existing technologies mostly employ direct input or basic filtering algorithms, which tend to erase true physical characteristics when processing signals containing noise and transient jumps. In addition, traditional filtering algorithms usually require single-step parameter settings for individual signals, making them difficult to implement in practical applications.

[0003] For battery temperature estimation, existing technologies often treat the thermal model as an independent entity for thermodynamic analysis and the electrical parameters as known constants. Traditional prediction methods often use the electrical characteristic parameters obtained offline as static independent inputs, ignoring the dynamic characteristics of internal resistance as a function of state of charge (SOC) and temperature, leading to distortion in heat generation calculations.

[0004] For data-driven prediction models, traditional pure data-driven prediction methods suffer from black-box limitations, and forcing predictions based on data patterns can easily violate physical laws. Traditional models rely solely on data for forced fitting, resulting in insufficient generalization ability under extreme conditions and a tendency to violate the law of conservation of energy. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a collaborative estimation method for battery temperature and internal resistance driven by pure data extraction and physical information fusion. Based on Physical Information Network (PINN) and Jumping Mode Decomposition (JMD), this invention effectively solves the problems of high signal-noise interference in lithium-ion batteries under complex operating conditions, lack of physical interpretability of pure data-driven models, and high computational complexity of traditional physical models by deeply fusing particle swarm optimization, JMD, and physical information neural networks. This invention enables high-precision and robust real-time estimation of transient core temperature and dynamic internal resistance within the battery, even under computationally limited conditions.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: A method for co-estimating battery temperature and internal resistance based on pure data extraction and physical information fusion includes the following steps: S1: Obtain operating data such as current, voltage and ambient temperature of lithium-ion batteries under actual working conditions; apply the Whale Optimization Algorithm (WOA) to quickly identify the parameters of the second-order equivalent circuit model offline in order to extract the initial dynamic resistance and capacitance parameters of the battery, and then calculate the preliminary heat generation characteristics of the battery. S2: Establish an adaptive jump mode decomposition (JMD) model based on particle swarm optimization (PSO) algorithm to decouple the acquired signal and decompose the original signal into trend component, fluctuation component, jump term and residual noise, thereby extracting a pure signal that accurately reflects the evolution trend of the battery's true state. S3: Construct a physical information network PINN that integrates electrothermal coupling laws. The PINN network includes a dynamic internal resistance prediction sub-network for real-time fine-tuning of internal resistance parameters, and a temperature prediction main network that introduces a channel attention mechanism. The main network performs real-time temperature prediction based on JMD output data. The dynamic internal resistance prediction sub-network uses ECM parameters as initial values ​​and performs real-time dynamic internal resistance prediction based on the output of the temperature prediction main network. S4: Establish a loss function constraint model; by combining the battery thermodynamic equation into physical loss terms, data loss, and smoothness loss, construct the PINN network loss function, and adopt a multi-stage dynamic weight strategy to balance the contributions of different losses at different stages; through the trained network, realize the real-time joint estimation of the transient core temperature and dynamic internal resistance of lithium-ion batteries.

[0007] Preferably, the construction of the second-order equivalent model and parameter identification in step S1 specifically includes: S1-1: Real-time acquisition of operating data such as current, voltage and ambient temperature of lithium-ion batteries under actual working conditions; S1-2: Transforming pure data prediction into the construction of a physical model, and then using the physical model for subsequent prediction and calculation, can further improve the prediction accuracy; a second-order equivalent RC model is adopted as the state model describing the battery's electrical characteristics. This model includes an ideal voltage source representing the open-circuit voltage, an ohmic internal resistance characterizing the transient voltage drop, and two series-connected resistor-capacitor networks, which respectively represent the electrochemical polarization and concentration polarization processes. S1-3: Assuming the input current remains constant within a small sampling period, first calculate the network's time constant and discrete coefficients; simultaneously, the identified internal resistance will be used as the initial value for the subsequent network model. S1-4: Based on the dynamic polarization voltage and internal resistance parameters obtained from the second-order RC model, the internal heat generation process of the battery is decomposed, and the dynamic electrical characteristics of the battery are coupled with thermodynamic heat generation.

[0008] Preferably, step S2 establishes a JMD model to decouple and purify the original signal, specifically including the following steps: S2-1: JMD treats the original signal as a superposition of the following three components with different physical properties: a mode component representing the smooth changes in the signal; a component responsible for capturing discontinuous and sudden jumps in the signal; and a residual term representing random high-frequency noise that cannot be explained by the above model. Based on the signal decomposition in the above formula, the solution process of JMD is transformed into a joint minimization problem; S2-2: By optimizing the JMD model using the PSO algorithm, the measurement random noise and jump interference in the original signal are effectively decoupled from the trend term reflecting the true state of the battery, thereby obtaining a pure signal that can truly characterize the electrothermal dynamic process.

[0009] Preferably, the structure of the temperature prediction main network of the PINN network in step S3 is as follows: starting from the input layer, a first hidden layer, a first channel attention mechanism module, a second hidden layer, a second channel attention mechanism module, a third hidden layer, and an output layer are sequentially connected in series; wherein, the input layer receives a 4-dimensional feature vector. t , SOC , T amb , I trend The first, second, and third hidden layers are all fully connected layers, using the hyperbolic tangent function as the activation function to extract feature information from the input signal. The first and second channel attention mechanism modules both employ a squeeze and excitation network architecture, consisting of a global average pooling layer and a two-layer fully connected bottleneck structure with a scaling factor of 16. Within the bottleneck structure, a modified linear unit and a sigmoid function are sequentially used for nonlinear mapping to recalibrate and weight the channel features, achieving adaptive enhancement of key features. The output layer is composed of fully connected layers and outputs the battery transient predicted temperature. Since different inputs contribute differently to temperature changes under different environments, a channel attention mechanism, SE-Block, is introduced into the temperature prediction main network. The input feature channels are recalibrated and then input into a deep residual network for nonlinear mapping to predict the battery temperature.

[0010] Furthermore, the structure of the dynamic internal resistance prediction sub-network of the PINN network described in step S3 is as follows: two hidden layers and one output layer are sequentially connected in series from the input layer; wherein, the input layer receives a 3-dimensional feature vector, including SOC,T amb and transient predicted temperature fed back in real time by the temperature prediction main network. T pre The hidden layers are all composed of fully connected layers and all use the hyperbolic tangent function as the nonlinear activation function to perform nonlinear decoupling and state mapping of the input features; the output layer consists of a one-dimensional fully connected layer and a sigmoid activation function, which outputs the real-time dynamic internal resistance. The dynamic internal resistance prediction subnetwork takes the real-time extracted SOC, ambient temperature features, and the predicted temperature output by the temperature prediction main network as feature inputs, and adaptively outputs the dynamic internal resistance through forward propagation.

[0011] Preferably, the PINN network described in step S3 constructs a collaborative estimation mechanism with feedback capability through bidirectional information transmission between the temperature prediction master network and the dynamic internal resistance prediction subnetwork. During a single forward propagation, the temperature prediction master network estimates the temperature based on the current input. The feature vector is then fed back to the dynamic internal resistance prediction subnetwork to estimate the dynamic internal resistance under physical conditions. .

[0012] Preferably, in step S4, the loss function construction and network training specifically include: S4-1: Construct a thermodynamic physical residual model based on energy conservation. First, perform automatic time-division differentiation on the transient predicted temperature output by the temperature prediction master network to obtain the transient temperature rise rate of the battery. By combining the current trend component extracted from JMD decomposition with the dynamic internal resistance output by the internal resistance prediction subnetwork, the dynamic Joule heat generation rate inside the battery is calculated. Then, based on the law of conservation of energy, the product of the battery core heat capacity and the temperature rise rate is constrained to be equal to the difference between the dynamic Joule heat generation rate and the convective heat dissipation rate, thereby constructing the thermodynamic physical residual boundary. S4-2: Construct a comprehensive objective function that includes multiple physical constraints; first, calculate the mean square error loss between the predicted temperature and the actual temperature. This is used to ensure the accuracy of the basic fit: Secondly, a physical residual loss term is introduced. Through the Sum of squares is performed to establish partial differential constraints between the neural network output and the underlying thermodynamic equations, ensuring that the model always conforms to the battery thermodynamic equilibrium. To address the continuous, non-abrupt evolution of internal resistance with SOC, a smoothness loss term is introduced. By taking the partial derivative of the internal resistance generated by the subnetwork with respect to the state of charge and applying gradient penalty, non-physical internal resistance fluctuations caused by data noise can be effectively suppressed, ensuring that the identification results conform to physical laws. Finally, a comprehensive objective function incorporating multi-stage dynamic weights is established; S4-3: Perform multi-stage dynamic weight optimization and joint estimation of the output, and use the gradient descent optimization algorithm to optimize the comprehensive objective function. Minimize optimization and simultaneously update network weights and hot parameters to be optimized. and During the optimization process, a multi-stage dynamic weight control mechanism is implemented to balance gradient competition. In the early stages of training, weights are assigned to the data fitting error. Maximum value, and physical constraint weights and Setting it to zero is a stage designed to avoid gradient explosion or dead zone problems caused by the high non-convexity of the early physical loss function, enabling the network to quickly establish preliminary feature maps and converge to the vicinity of the real data domain. In the later stages of training, a physical confinement strategy is implemented to decay. And improve and The solution space is constrained by a second time using thermodynamic and smoothness physical penalty gradients, forcing the model to converge to the global optimal solution that conforms to the underlying energy conservation and electrochemical evolution law; After optimization, the online operating data collected in real time under varying operating conditions and processed by modal decomposition is input into the optimal network. The model can then output the transient core temperature of the battery and the physically reliable dynamic internal resistance in real time with high fidelity, thus completing the joint estimation.

[0013] The beneficial effects of this invention are: 1. An adaptive JMD method based on PSO is proposed to decouple non-stationary operating signals. It effectively removes high-frequency noise and step pulses from sensor-acquired signals under complex operating conditions, extracts pure trend components that reflect the true physical evolution of the battery, and improves the anti-interference capability of system state estimation while reducing the unnecessary computational burden on the neural network.

[0014] 2. PINN, based on the law of conservation of energy, is used to solve the core temperature estimation problem under complex operating conditions. A multi-stage weight optimization is performed using a comprehensive objective function that includes thermodynamic residuals, forcing the model's solution space to follow the underlying electrochemical and thermodynamic evolution of the battery. This effectively avoids the problems of lack of physical interpretability and large generalization errors in purely data-driven models.

[0015] 3. A joint prediction architecture was designed, consisting of a main network with an integrated channel attention mechanism and a dynamic internal resistance prediction subnetwork, which are bidirectionally coupled. By dynamically fine-tuning the internal resistance parameters in real time and feeding them back to the heat generation calculation in a closed loop, the nonlinear evolution of internal resistance with environmental and operating conditions was accurately captured, achieving high-precision joint estimation of transient core temperature and dynamic internal resistance. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall system of the present invention.

[0017] Figure 2 This is a flowchart of the data preprocessing process of the present invention.

[0018] Figure 3 This is a diagram of the PINN network training process of the present invention.

[0019] Figure 4 This is the network structure diagram of the present invention.

[0020] Figure 5 This is a diagram showing the results of the temperature data mode decomposition of this invention.

[0021] Figure 6 This is the temperature prediction curve of the present invention.

[0022] Figure 7 This is the dynamic internal resistance prediction diagram of the present invention. Specific implementation methods

[0023] The invention will now be further described with reference to the accompanying drawings.

[0024] Reference Figure 1 A method for estimating the temperature of a lithium-ion battery based on physical information neural networks and mode decomposition includes the following steps: S1: As Figure 1 As shown in the top left, operating data such as current, voltage, and ambient temperature of the lithium-ion battery under actual working conditions are obtained. The state of charge (SOC) of the battery is calculated using the ampere-hour integration method, and the Whale Optimization Algorithm (WOA) is applied to perform offline parameter identification on the second-order equivalent circuit model to extract the initial dynamic resistance and capacitance parameters of the battery, and then calculate the preliminary heat generation characteristics of the battery.

[0025] S2: As Figure 1 As shown in the upper right corner, an adaptive JMD model based on PSO is established to decouple the acquired non-stationary current, voltage, and temperature signals. The original signal is decomposed into trend components, fluctuation components, jump terms, and residual noise, thereby extracting a pure signal that accurately reflects the actual evolution trend of the battery.

[0026] S3: As Figure 4 As shown, a physical information network PINN integrating electrothermal coupling laws is constructed. The PINN network includes a dynamic internal resistance prediction sub-network for real-time fine-tuning of internal resistance parameters, and a main prediction network that introduces a channel attention mechanism, which is used to output the inferred result of the core temperature of the lithium-ion battery.

[0027] S4: As Figure 1As shown in the lower left corner, a loss function constraint model is established. By transforming the thermodynamic equation into a physical residual loss term and combining data fitting loss and smoothness loss, a comprehensive objective function is constructed. A multi-stage dynamic weighting strategy is used to train and update the neural network. The trained network enables real-time joint estimation of the transient core temperature and dynamic internal resistance of a lithium-ion battery.

[0028] In step S1, refer to Figure 1 The S1 part preprocesses the data and provides the input basis for the subsequent construction of the JMD model and PINN network, specifically including the following steps: S1-1: Real-time acquisition of time-series operating data of lithium-ion batteries, including current, voltage, and ambient temperature, under actual operating conditions.

[0029] S1-2: A second-order RC model is used to describe the internal electrical state of the battery. This model includes an ideal voltage source representing the open-circuit voltage, an ohmic internal resistance characterizing the transient voltage drop, and two series-connected resistive-capacitive networks. These two networks are used to characterize the electrochemical polarization and concentration polarization processes, respectively. The model's electrical state equations and terminal voltage prediction equations are described as follows:

[0030] In the formula, Terminal voltage, Open circuit voltage, For ohmic internal resistance, This represents the system's charging and discharging current. The two parallel RC circuits represent the short-time and long-time constants of the step response, respectively. This indicates the concentration polarization effect of the battery. This indicates the electrochemical polarization effect inside a lithium battery. and These represent the electrochemical polarization voltage and concentration polarization voltage over continuous time, respectively.

[0031] S1-3: Since the battery data is set to charge and discharge within a certain sampling period, in order to further transform the battery data into physical features in the constructed model, this step assumes that the input current remains constant within a small sampling period and first calculates the network's time constant and discrete coefficients. The time constant is equal to the product of the corresponding resistance and capacitance, and the discrete update coefficients are determined by both the sampling time interval and the time constant. Simultaneously, the internal resistance identified in this step will be used as the initial value for subsequent optimization. The discretization parameter formula is expressed as follows:

[0032] In the formula, and Represents the time constant of the polarization network. Indicates the sampling frequency. and as well as and This represents the attenuation coefficient of the polarization voltage eccentricity step.

[0033] Iterative updates of the state equation can be achieved using discrete coefficients. For any time in the discrete time series, the polarization voltage at the current time is equal to the sum of the natural decay of the polarization voltage at the previous time and the polarization increment caused by the current at the previous time. The iterative equation is as follows:

[0034] After obtaining the dynamic polarization voltage at the current moment, the parameters of the model can be identified according to the above formula.

[0035] To ensure the accuracy of equivalent circuit parameter identification, an objective function is constructed using the root mean square error (RMSE) index. The optimal set of discrete coefficient parameters is obtained through iterative optimization based on the WOA algorithm. The objective function is described as follows:

[0036] In the formula, for k The terminal voltage predicted by the time-matter model. Let represent the root mean square error objective function to be optimized. N This represents the total number of sampling points. By minimizing this objective function, the software can accurately extract the dynamic resistance and capacitance parameters of the battery as the initial input to the PINN network.

[0037] S1-4: Calculating heat generation is a prerequisite for realizing the PINN network. Only by accurately calculating the dynamic heat generation rate can the physical boundary residuals be derived, thereby ensuring that the output of the neural network conforms to the underlying thermodynamic mechanism. The dynamic polarization voltage and internal resistance parameters obtained from the second-order RC model parameter decomposition can break down the internal heat generation process of the battery. The internal heat generation mechanism of the battery includes three main heat sources, and the overall Joule heat generation rate characteristic equation is described as follows:

[0038] In the formula, This indicates the total heat production rate. This represents ohmic heat, the heat generated when current flows directly through the ohmic internal resistance; first polarization heat. The heat generated by the corresponding current on the solid electrolyte interface film and charge transfer impedance; the heat of second polarization. This corresponds to the concentration polarization heat generated during ion diffusion. This equation couples the dynamic electrical characteristics of the battery with thermodynamic heat generation.

[0039] Step S2 establishes a JMD model to decouple the original signal, such as Figure 2As shown, the specific steps include: S2-1: JMD treats a complex original signal as a superposition of the following three components with completely different physical properties:

[0040] In the formula, This represents the mode components, i.e., the smoothed variations in the signal. The software will... Considered a trend item. It is a fluctuation term with a specific center frequency. This represents the jump term, which is specifically responsible for capturing discontinuous, sudden steps in the signal. (t) represents the residual term, which is random high-frequency noise that cannot be explained by the above model.

[0041] Based on the signal decomposition in the above equation, the solution process for JMD is transformed into a joint minimization problem, the core objective function of which is as follows:

[0042] In the formula, α This represents the bandwidth limitation parameter, which controls the smoothness of the signal; β The sparsity penalty coefficient determines the degree to which the software measures transitions and noise. This indicates the expected minimum jump height; only jumps exceeding this minimum height will be classified as jumps by the software. k This indicates the number of mode decomposition layers; choosing an appropriate value achieves complete decoupling in terms of frequency. It indicates the center frequency, which points to the dominant frequency of the battery's dynamic response.

[0043] S2-2: Because the decoupling effect of the JMD model depends on α , β , , k To avoid repeatedly adjusting these four parameters, this step uses the PSO optimization algorithm to combine potential parameters. α , β , , k Treating each particle as a single entity, a multi-particle optimization process is implemented, using a fitness function to control the optimization direction of individual particles within the population, iterating until the population optimum is reached. The fitness function employs a series of evaluation metrics, including sample entropy. Compared to variance, permutation entropy, and approximate entropy, sample entropy, while computationally more complex, exhibits stronger robustness and no self-matching bias, giving it a relative advantage in this non-stationary signal environment dominated by pulse signals. The fitness function is primarily composed of the following weighted components:

[0044] The specific components of each part are as follows:

[0045] In the formula, Represents sample entropy, used to find the optimal... k To ensure the simplicity of the physical quantities, among which B It is a length of m The template is at a distance r The number of inner matches, A It is a length of m +1 match count. Represents the correlation coefficient, used for constraints. This measures the degree of decoupling between the trend term and the jump / fluctuation term, where r This is the residual signal after removing the trend. Indicates second-order curvature penalty, preventing α When the value is small, it gets stuck at the boundary value. This represents the modal overlap penalty, used to measure the spectral independence between different modes and prevent over-decomposition.

[0046] By constructing the JMD model optimized based on the PSO algorithm, the measurement random noise and jump interference in the original signal are effectively decoupled from the trend term reflecting the true state of the battery, thereby obtaining a pure signal that can truly characterize the electrothermal dynamic process.

[0047] Furthermore, in step S3, the constructed PINN architecture with integrated electrothermal coupling law, such as Figure 1 As shown in section S3, the specific steps include: S3-1: Because the battery's internal resistance is affected by SOC, ambient temperature, and temperature changes, such as Figure 1 In the lower half of the S3 network, an internal resistance prediction subnetwork based on a multilayer perceptron is constructed. This subnetwork takes the real-time extracted SOC, ambient temperature features, and the predicted temperature output by the main network as feature inputs, and adaptively outputs the dynamic internal resistance through forward propagation. Its characteristic equation is:

[0048] In the formula, for The state of charge of the battery at any given time; for The ambient temperature at that moment; The transient predicted temperature is fed back in real time by the main network; The real-time dynamic internal resistance predicted for the subnetwork; For predicting the internal resistance of a multilayer sensor subnetwork function; This refers to all learnable weights and bias parameters within the internal resistance prediction subnetwork.

[0049] S3-2: Because different inputs contribute differently to temperature changes under different environments, a channel attention mechanism (SE-Block) is introduced into the main temperature prediction network. The input feature channels are recalibrated and then fed into a deep residual network for nonlinear mapping to predict the battery temperature. Its characteristic equation is:

[0050] In the formula, The battery temperature predicted by the main network; The current trend component is extracted via JMD; This is the main network function for temperature prediction; To infer the weight parameters of the main network neurons based on temperature; These are the weight parameters for the channel attention mechanism module.

[0051] S3-3: This invention constructs a collaborative estimation mechanism with feedback capability through bidirectional information transmission between the main network and sub-networks. During a single forward propagation, the main network estimates the temperature based on the current input. The feature vector is then fed back to the sub-network to estimate the dynamic internal resistance under physical conditions. .

[0052] During the training and validation phases, the dynamic internal resistance output by the subnetwork is substituted into the thermodynamic equation in subsequent step S4 to calculate heat generation and physical residuals. The model requires that the predicted temperature and impedance heat generation remain consistent with the law of conservation of energy. Through joint optimization of electrothermal coupling, the model can overcome the black-box effect of the pure data-driven mode, achieve synchronous physical identification of the dynamic internal resistance of the battery, and significantly improve the model's generalization performance and effectiveness under varying operating conditions and noise environments.

[0053] In step S4, a loss function constraint model based on the law of energy conservation and a multi-stage joint estimation method are established, specifically including the following process: S4-1: Construct a thermodynamic physical residual model based on energy conservation. First, perform automatic time-dimension differentiation on the transient predicted temperature output by the main network to obtain the transient temperature rise rate of the battery.

[0054] By combining the current trend component extracted from the jump mode decomposition with the dynamic internal resistance output by the internal resistance prediction subnetwork, the dynamic Joule heat generation rate inside the battery is calculated. The calculation formula is as follows:

[0055] in This represents the transient temperature rise rate, i.e., the partial derivative of the network-predicted temperature with respect to time. for Dynamic Joule heat production rate at any given time; For current trend components; The output internal resistance of the subnetwork is estimated using the internal resistance estimation method. After obtaining the heat generated by the battery, the convective heat dissipation rate between the battery and the external environment is deduced based on the temperature difference between the predicted temperature and the ambient temperature.

[0056] in For convective heat dissipation from the environment; The global convective thermal resistance is optimized synchronously during network training. Then, based on the law of conservation of energy, the product of the battery core heat capacity and the rate of temperature rise is constrained to be equal to the difference between the dynamic Joule heat generation rate and the convective heat dissipation rate, thereby constructing the thermodynamic physical residual boundary:

[0057] In the formula For the core heat capacity of the battery; This represents the physical boundary residual.

[0058] S4-2: Construct a comprehensive objective function incorporating multiple physical constraints. To ensure that the prediction results closely match the measured observation data and follow electrochemical laws, a detailed analysis of the loss term is required. First, calculate the mean square error loss between the predicted and actual temperatures. This is used to ensure the accuracy of the basic fit:

[0059] Secondly, a physical residual loss term is introduced. Through the By performing a summation of squares, a partial differential constraint is established between the neural network output and the underlying thermodynamic equations. This ensures that the model always conforms to the battery thermodynamic equilibrium.

[0060]

[0061] To address the continuous, non-abrupt evolution of internal resistance with SOC, a smoothness loss term is introduced. By taking the partial derivative of the internal resistance generated by the subnetwork with respect to the state of charge and applying gradient penalty, non-physical internal resistance fluctuations caused by data noise can be effectively suppressed, ensuring that the identification results conform to physical laws.

[0062] Finally, a comprehensive objective function incorporating multi-stage dynamic weights is established. :

[0063] In the formula, This represents the total number of sampling points; The mean square error of the data fitting; for The actual measured temperature at that moment; For smoothness loss; This is the physical residual loss term; The overall objective function; , , These are the dynamic weighting factors assigned to the corresponding loss terms.

[0064] S4-3: As Figure 3 As shown, multi-stage dynamic weight optimization and joint estimation of the output are performed, and the gradient descent optimization algorithm is used to optimize the comprehensive objective function. Perform minimization optimization, and simultaneously update network weights and global hot parameters to be optimized. and A multi-stage dynamic weight control mechanism is implemented during the optimization process to balance gradient competition.

[0065] In the early stages of training, weights are assigned to the data fitting error. Maximum value, and physical constraint weights and Setting it to zero is a stage designed to avoid gradient explosion or dead zone problems caused by the high non-convexity of the early physical loss function, enabling the network to quickly establish preliminary feature maps and converge to the vicinity of the real data domain.

[0066] In the later stages of training, a physical confinement strategy is implemented to decay. And improve and By using thermodynamic and smoothness physical penalty gradients to impose secondary constraints on the solution space, the model is forced to converge to the global optimal solution that conforms to the underlying energy conservation and electrochemical evolution laws.

[0067] After optimization, the online operating data collected in real time under varying operating conditions and processed by modal decomposition is input into the optimal network. The model can then output the transient core temperature of the battery and the physically reliable dynamic internal resistance in real time with high fidelity, thus completing the joint estimation of the multi-dimensional state.

[0068] To enable those skilled in the art to better understand this invention, this paper uses experimental data from lithium battery HPPC operating conditions under different ambient temperatures as examples for simulation, including the following components: 1. Data Preprocessing To improve subsequent prediction accuracy, data preprocessing is performed, aiming to transform the original discrete operating data into feature inputs with clear physical meaning and high signal-to-noise ratio through the fusion of multiple algorithms. First, the State of Charge (SOC) is calculated using the ampere-hour integration method, and a second-order RC equivalent circuit model is constructed. The WOA algorithm is then used for offline identification to obtain electrical characteristic parameters. Based on this, to address the non-stationary and high-noise characteristics of the signal under complex operating conditions, JMD is used for decoupling processing. PSO is used to adaptively optimize bandwidth limiting parameters, sparse penalty coefficients, transition heights, and the number of modal layers. By constructing a composite fitness function that includes sample entropy, correlation coefficients, and second-order curvature penalties, the original signal is decoupled and reconstructed. Finally, a pure current trend component that accurately reflects the true state evolution of the battery is extracted, effectively eliminating the interference of sensor noise and step pulses on subsequent temperature estimation and dynamic internal resistance identification.

[0069] Part 2: Training Data and Network Hyperparameter Settings

[0070] Before the physical information neural network designed in this invention is applied to the vehicle terminal for joint state estimation, its main sub-network needs to be trained and optimized using experimental data to obtain network parameters suitable for complex and variable operating conditions. The training data consists of battery HPPC experimental data covering five different ambient temperatures (0℃, 10℃, 20℃, 30℃, and 40℃), and some reconstructed sample data after JMD processing is shown below. Figure 5 As shown.

[0071] The operating data was divided into training and validation sets in an 8:2 ratio. The network received the normalized time, reconstructed trend current, SOC, and ambient temperature four-dimensional features as input state information, and then... Figure 3 The PINN training process, as shown, involves multi-stage iterative iterations, synchronously updating the parameters of the master neural network and global thermodynamic variables until training is complete. The temperature prediction curve and dynamic internal resistance prediction results output after training optimization are shown below. Figure 6 and Figure 7 As shown.

[0072] After multiple trials and comparative verifications, this invention sets the small-batch sampling size in network model training to 2048, and the total number of iterations for optimization to 1000. The initial learning rate of the optimizer is set to 0.002, and multi-step decay is performed in the 400th and 800th iterations, with a decay coefficient of 0.5 each time. To prevent violations of thermodynamic principles, the boundary constraints of physical parameters are set between 20.0 and 150.0 for the battery core heat capacity, between 5.0 and 80.0 for the global convective thermal resistance, and a cutoff of 0.1 for the upper limit of dynamic internal resistance prediction. These settings are implemented in iterations 1 to 199. Setting from round 200 onwards .

[0073] IV. Results Analysis and Comparison

[0074] Figure 5 The original data and PSO-JMD reconstructed data under different ambient temperatures were compared. The PSO-JMD mechanism separates the intrinsic thermodynamic evolution trajectory from non-stationary random noise, explicitly suppressing noise caused by hardware such as sensors. Unlike traditional recursive filtering that inherently introduces time delay, the jump threshold constraint ensures a strictly zero phase shift under current excitation. This decoupling process extracts high-fidelity thermal inertia boundaries, providing physically consistent input for subsequent PINN optimization.

[0075] The results of temperature prediction training and validation are as follows Figure 6 As shown in the figure, the red dashed line represents the temperature predicted by the PINN main network, the black solid line represents the actual temperature, and the blue dashed line represents the boundary between the training set and the validation set. Within the validation set interval, the predicted curve and the actual temperature curve show a high degree of consistency, indicating that the improved PINN model has good temperature prediction capabilities.

[0076] In addition, the network can predict changes in battery internal resistance in real time, and the results are as follows: Figure 7 As shown, the internal resistance increases significantly at the moment of current pulse application, indicating that the model can effectively respond to changes in current excitation and reflects the transient dynamic characteristics of the internal resistance. This demonstrates that the subnetwork has successfully learned the physical laws governing the change of internal resistance with operating conditions. Furthermore, from the overall trend, the dynamic internal resistance exhibits nonlinear characteristics across different SOC ranges: it remains relatively stable in the medium SOC range, while showing an increasing trend in the low or high SOC regions. This phenomenon is consistent with the typical characteristics of lithium-ion batteries.

[0077] In this specification, the illustrative descriptions of the invention are not necessarily directed at the same embodiments or examples. Those skilled in the art can combine and integrate the different embodiments or examples described in this specification. Furthermore, the content described in the embodiments of this specification is merely an enumeration of implementation forms of the inventive concept, and the scope of protection of the invention should not be regarded as limited to the specific forms stated in the embodiments. The scope of protection of the invention also includes equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for co-estimating battery temperature and internal resistance based on pure data extraction and physical information fusion, characterized in that, Includes the following steps: S1: Obtain operating data such as current, voltage and ambient temperature of lithium-ion batteries under actual working conditions; apply the Whale Optimization Algorithm (WOA) to quickly identify the parameters of the second-order equivalent circuit model offline in order to extract the initial dynamic resistance and capacitance parameters of the battery, and then calculate the preliminary heat generation characteristics of the battery. S2: Establish an adaptive jump mode decomposition (JMD) model based on particle swarm optimization (PSO) algorithm to decouple the acquired signal and decompose the original signal into trend component, fluctuation component, jump term and residual noise, thereby extracting a pure signal that accurately reflects the evolution trend of the battery's true state. S3: Construct a physical information network PINN that integrates electrothermal coupling laws. The PINN network includes a dynamic internal resistance prediction sub-network for real-time fine-tuning of internal resistance parameters, and a temperature prediction main network that introduces a channel attention mechanism. The main network performs real-time temperature prediction based on JMD output data. The dynamic internal resistance prediction sub-network uses ECM parameters as initial values ​​and performs real-time dynamic internal resistance prediction based on the output of the temperature prediction main network. S4: Establish a loss function constraint model; by combining the battery thermodynamic equation into physical loss terms, data loss, and smoothness loss, construct the PINN network loss function, and adopt a multi-stage dynamic weight strategy to balance the contributions of different losses at different stages; through the trained network, realize the real-time joint estimation of the transient core temperature and dynamic internal resistance of lithium-ion batteries.

2. The battery temperature and internal resistance co-estimation method based on pure data extraction and physical information fusion as described in claim 1, characterized in that, Step S1, which involves constructing a second-order equivalent model and identifying parameters, specifically includes: S1-1: Real-time acquisition of operating data such as current, voltage and ambient temperature of lithium-ion batteries under actual working conditions; S1-2: A second-order equivalent RC model is used as the state model to describe the battery's electrical characteristics. This model includes an ideal voltage source representing the open-circuit voltage, an ohmic internal resistance characterizing the transient voltage drop, and two series-connected resistive-capacitive networks. These two resistive-capacitive networks represent the electrochemical polarization and concentration polarization processes, respectively. S1-3: Assuming the input current remains constant within a small sampling period, first calculate the network's time constant and discrete coefficients; simultaneously, the identified internal resistance will be used as the initial value for the subsequent network model. S1-4: Based on the dynamic polarization voltage and internal resistance parameters obtained from the second-order RC model, the internal heat generation process of the battery is decomposed, and the dynamic electrical characteristics of the battery are coupled with thermodynamic heat generation to obtain the heat generation rate.

3. The battery temperature and internal resistance co-estimation method based on pure data extraction and physical information fusion as described in claim 1 or 2, characterized in that, Step S2 establishes a JMD model to decouple and purify the original signal, specifically including the following steps: S2-1: JMD treats the original signal as a superposition of the following three components with different physical properties: a mode component representing the smooth changes in the signal; a component responsible for capturing discontinuous and sudden jumps in the signal; and a residual term representing random high-frequency noise that cannot be explained by the above model. Based on the signal decomposition in the above formula, the solution process of JMD is transformed into a joint minimization problem; S2-2: By optimizing the JMD model using the PSO algorithm, the measurement random noise and jump interference in the original signal are effectively decoupled from the trend term reflecting the true state of the battery, thereby obtaining a pure signal that can truly characterize the electrothermal dynamic process.

4. The battery temperature and internal resistance co-estimation method based on pure data extraction and physical information fusion as described in claim 1, characterized in that, The structure of the temperature prediction main network of the PINN network described in step S3 is as follows: starting from the input layer, the first hidden layer, the first channel attention mechanism module, the second hidden layer, the second channel attention mechanism module, the third hidden layer, and the output layer are sequentially connected in series; wherein, the input layer receives a 4-dimensional feature vector. t , SOC , T amb , I trend The first, second, and third hidden layers are all fully connected layers, using the hyperbolic tangent function as the activation function to extract feature information from the input signal. The first and second channel attention mechanism modules both employ a squeeze and excitation network architecture, consisting of a global average pooling layer and a two-layer fully connected bottleneck structure with a scaling factor of 16. Within the bottleneck structure, a modified linear unit and a sigmoid function are sequentially used for nonlinear mapping to recalibrate and weight the channel features, achieving adaptive enhancement of key features. The output layer is composed of fully connected layers and outputs the battery transient predicted temperature. Since different inputs contribute differently to temperature changes under different environments, a channel attention mechanism, SE-Block, is introduced into the temperature prediction main network. The input feature channels are recalibrated and then input into a deep residual network for nonlinear mapping to predict the battery temperature.

5. The battery temperature and internal resistance co-estimation method based on pure data extraction and physical information fusion as described in claim 4, characterized in that, The structure of the dynamic internal resistance prediction subnetwork of the PINN network described in step S3 is as follows: two hidden layers and one output layer are sequentially connected in series from the input layer; wherein, the input layer receives a 3-dimensional feature vector, including SOC, T amb and transient predicted temperature fed back in real time by the temperature prediction main network. T pre The hidden layers are all composed of fully connected layers and all use the hyperbolic tangent function as the nonlinear activation function to perform nonlinear decoupling and state mapping of the input features; the output layer consists of a one-dimensional fully connected layer and a sigmoid activation function, which outputs the real-time dynamic internal resistance. The dynamic internal resistance prediction subnetwork takes the real-time extracted SOC, ambient temperature features, and the predicted temperature output by the temperature prediction main network as feature inputs, and adaptively outputs the dynamic internal resistance through forward propagation.

6. The battery temperature and internal resistance co-estimation method based on pure data extraction and physical information fusion as described in claim 5, characterized in that, The PINN network described in step S3 constructs a collaborative estimation mechanism with feedback capability through bidirectional information transmission between the temperature prediction master network and the dynamic internal resistance prediction subnetwork. During a single forward propagation, the temperature prediction master network estimates the temperature based on the current input. The feature vector is then fed back to the dynamic internal resistance prediction subnetwork to estimate the dynamic internal resistance under physical conditions. .

7. The battery temperature and internal resistance co-estimation method based on pure data extraction and physical information fusion as described in claim 1 or 2, characterized in that, In step S4, the loss function construction and network training specifically include: S4-1: Construct a thermodynamic physical residual model based on energy conservation. First, perform automatic time-division differentiation on the transient predicted temperature output by the temperature prediction master network to obtain the transient temperature rise rate of the battery. By combining the current trend component extracted from JMD decomposition with the dynamic internal resistance output by the internal resistance prediction subnetwork, the dynamic Joule heat generation rate inside the battery is calculated. Then, based on the law of conservation of energy, the product of the battery core heat capacity and the temperature rise rate is constrained to be equal to the difference between the dynamic Joule heat generation rate and the convective heat dissipation rate, thereby constructing the thermodynamic physical residual boundary. S4-2: Construct a comprehensive objective function that includes multiple physical constraints; first, calculate the mean square error loss between the predicted temperature and the actual temperature. This is used to ensure the accuracy of the basic fit: Secondly, a physical residual loss term is introduced. Through the Sum of squares is performed to establish partial differential constraints between the neural network output and the underlying thermodynamic equations, ensuring that the model always conforms to the battery thermodynamic equilibrium. To address the continuous, non-abrupt evolution of internal resistance with SOC, a smoothness loss term is introduced. By taking the partial derivative of the internal resistance generated by the subnetwork with respect to the state of charge and applying gradient penalty, non-physical internal resistance fluctuations caused by data noise can be effectively suppressed, ensuring that the identification results conform to physical laws. Finally, a comprehensive objective function incorporating multi-stage dynamic weights is established; S4-3: Perform multi-stage dynamic weight optimization and joint estimation of the output, and use the gradient descent optimization algorithm to optimize the comprehensive objective function. Minimize optimization and simultaneously update network weights and hot parameters to be optimized. and During the optimization process, a multi-stage dynamic weight control mechanism is implemented to balance gradient competition. In the early stages of training, weights are assigned to the data fitting error. Maximum value, and physical constraint weights and Setting it to zero is a stage designed to avoid gradient explosion or dead zone problems caused by the high non-convexity of the early physical loss function, enabling the network to quickly establish preliminary feature maps and converge to the vicinity of the real data domain. In the later stages of training, a physical confinement strategy is implemented to decay. And improve and The solution space is constrained by a second time using thermodynamic and smoothness physical penalty gradients, forcing the model to converge to the global optimal solution that conforms to the underlying energy conservation and electrochemical evolution law; After optimization, the online operating data collected in real time under varying operating conditions and processed by modal decomposition is input into the optimal network. The model can then output the transient core temperature of the battery and the physically reliable dynamic internal resistance in real time with high fidelity, thus completing the joint estimation.

Images