Lithium ion battery internal temperature estimation method based on machine learning and data generation

Abstract

The invention relates to a lithium ion battery internal temperature estimation method based on machine learning and data generation, and belongs to the technical field of batteries. The method solves the problems that traditional internal temperature estimation data acquisition is high in cost, the error between a simulation model and an actual system is large, and the generalization ability of the model is insufficient. According to the technical scheme, the method comprises the following steps: establishing an electrothermal coupling model for a target battery type such as a cylindrical or square battery to generate simulation data covering various parameters and working conditions; pre-training a gated circulation unit neural network or a long-short-term memory neural network based on the simulation data set, and capturing a mapping relation of input features; the model is finely tuned through an unsupervised domain adaptive transfer learning algorithm by using actual battery unmarked data, including pseudo tag generation and feature alignment to compensate simulation errors; and finally verifying the effectiveness of the method by combining a real battery experiment. Accurate, robust and adaptive internal temperature estimation is realized, and the generalization ability of the model is ensured.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of batteries, and relates to a lithium ion battery internal temperature estimation method based on machine learning and generated data. BACKGROUND

[0002] Lithium ion batteries are widely used in electric transportation and grid energy storage systems due to their high energy density, high power density and long cycle life. Large-scale deployment of lithium ion batteries requires fine management to ensure their safe and reliable performance. Among them, accurate and real-time monitoring of temperature status is crucial, especially internal temperature is a key indicator affecting battery safety. Under high-rate operating conditions, the internal temperature of lithium ion batteries can be 10 degrees Celsius or more higher than the surface temperature, which can easily cause local overheating and accelerate aging. However, temperature sensors installed on the surface of the battery cannot track the dynamic changes of the internal temperature in time. Although embedded sensing technology attempts to directly measure the internal temperature, technical challenges and manufacturing costs limit its popularization in commercial applications. Therefore, online estimation methods based on current, voltage and other measurement data have become a more implementable alternative in current battery management systems.

[0003] Real-time internal temperature estimation methods are mainly divided into three categories: impedance-based methods, thermal model-based methods and data-driven methods. Impedance-based methods use the mapping relationship between volume average temperature and battery impedance parameters in a specific frequency range to achieve estimation, but rely on high-frequency signal acquisition equipment. Thermal model-based methods have become the mainstream technology by capturing the thermal dynamic characteristics of lithium ion batteries. Due to the high computational complexity of full-order thermal models described by partial differential equations, researchers have developed various control-oriented thermal models, such as thermal equivalent circuit models, polynomial approximation models and spectral-Galerkin models. These models can achieve high accuracy when the parameters are set properly, and can improve robustness combined with closed-loop observers. However, model parameter calibration relies on a large amount of experimental data and is sensitive to changes in actual operating conditions. Data-driven methods have become increasingly popular in recent years, such as long short-term memory recurrent neural networks that use their powerful nonlinear mapping capabilities to learn internal temperature relationships directly from measurement signals. However, the effectiveness of this type of method is highly dependent on the quantity and quality of data, and obtaining labeled data covering all operating conditions through experiments is costly and inefficient.

[0004] To improve data efficiency, some studies attempt to incorporate physical knowledge into machine learning algorithms. A typical case is the physics-informed neural network, which adds a residual term of the physical equation to the loss function during training, enabling the model to both fit the data and adhere to the physical law. For example, a physics-informed neural network is developed to simulate the spatiotemporal temperature distribution of a cylindrical battery. However, physics-informed neural networks are usually trained based on well-defined physical laws and accurate parameters, and their generalization ability under complex practical conditions has not been fully verified. Another method of integrating physical knowledge is to generate synthetic datasets through physics-based simulations to cover a variety of operating scenarios at low cost. Although this method solves the problem of data scarcity, there is an inevitable mismatch between the physical model and the actual system, resulting in simulation-to-real errors. This error will cause the trained machine learning model to perform poorly when applied to real batteries.

[0005] In summary, the existing technology faces three challenges: high cost of data acquisition, errors introduced by simplified physical models, and domain differences between simulation and reality. This limits the accuracy and robustness of lithium-ion battery internal temperature estimation methods in practical applications. Therefore, there is an urgent need to develop a method that combines physical modeling and machine learning to minimize data requirements while compensating for domain differences through an adaptive mechanism to achieve high-precision and strong-generalization temperature estimation. The present invention is proposed to provide reliable technical support for intelligent battery management. SUMMARY

[0006] In view of the above, the present invention aims to provide a lithium-ion battery internal temperature estimation method based on machine learning and generated data.

[0007] To achieve the above-mentioned purpose, the present invention provides the following technical solutions:

[0008] A lithium-ion battery internal temperature estimation method based on machine learning and generated data, comprising the following steps:

[0009] S1: Establish a matched electro-thermal coupling model for the target battery type, generate simulation data covering multiple operating scenarios by scanning model parameters and test conditions;

[0010] S2: Based on the simulation dataset generated in S1, pre-train a machine learning model to capture the mapping relationship between input features and output;

[0011] S3: Based on unlabeled data of the actual battery, use an unsupervised domain adaptive transfer learning algorithm to fine-tune the pre-trained model to compensate for the error between simulation and reality;

[0012] S4: Use experimental data of real batteries under different operating conditions and convective cooling conditions to verify the estimation results.

[0013] Further, in the S1, a polynomial approximation (PA) thermal model combined with an equivalent circuit model (ECM) is established for cylindrical batteries; for lithium iron phosphate positive electrode material batteries, an equivalent circuit model with a first-order resistor-capacitor pair with hysteresis (first-order RC with hysteresis ECM) is used for the electrical model; for ternary positive electrode material batteries, a first-order resistor-capacitor pair equivalent circuit model (first-order RC ECM) is used for the electrical model; model parameters are randomly generated within a preset range through a Latin hypercube sampling strategy, and test conditions include standardized test cycles of urban and high-speed working conditions.

[0014] Further, in the S2, the machine learning model uses a gated recurrent unit recurrent neural network (GRU-RNN) or a long short-term memory recurrent neural network (LSTM-RNN); input features include current, voltage, ambient temperature, surface temperature, and heat generation rate, and the output is the core temperature or the temperature of a key position of the battery.

[0015] Further, in the S3, domain adaptive transfer learning includes two stages of pseudo-label generation and feature alignment:

[0016] The pseudo-label is generated by inputting the target domain current and cooling temperature through the electro-thermal coupling model, and the selection criteria satisfy and where δ is 25% of the maximum absolute error of the surface temperature, i.e.

[0017] In the feature alignment stage, the maximum mean discrepancy (MMD) and correlation alignment (CORAL) are used to evaluate the feature distribution difference between the source domain and the target domain, and the MMD calculation formula is:

[0018]

[0019] where n is the number of source domain samples, m is the number of target domain samples, x i is the source domain data point, y j is the target domain data point, and the kernel function γ is the bandwidth parameter, and the CORAL calculation formula is d is the data feature dimension, and C X is the source domain covariance matrix, and C YThe target domain covariance matrix, ||·||F F The Frobenius norm.

[0020] Further, the state space representation of the PA thermal model is:

[0021] y = Cx + Du

[0022] Where the state variable The input variable u = [Q T f ] T The output variable y = [T c T s ] T , The average temperature, The temperature gradient, Q is the heat generation rate, T f The ambient temperature, T c The core temperature, T s The surface temperature, A, B, C, D are system coefficient matrices, and the heat generation rate Q is calculated according to the Bernardi formula: Where I is the current, V t The terminal voltage, U oc The open circuit voltage, SOC is the state of charge, T is the battery temperature, The partial derivative of the open circuit voltage with respect to temperature.

[0023] Further, the forward propagation of the gated recurrent unit neural network satisfies:

[0024] r t = σ (W r · [h t-1 , x t ] + b r )

[0025] z t = σ (W z · [h t-1 , x t ] + b z )

[0026]

[0027] Where r t Is the reset gate output, z t The update gate output, The candidate hidden state, h t The current hidden state, h t-1 The previous hidden state, x t The current input feature, W r , W z , Wh is a weight matrix, b r , b z , b h is a bias matrix, σ is a Sigmoid activation function, and tanh is a hyperbolic tangent activation function.

[0028] Further, the loss function of the fine-tuning process is a weighted quadruple:

[0029]

[0030] wherein, α , β, γ are non-negative weight coefficients, is a source domain regression loss; is a target domain regression loss; L MMD is a maximum mean difference loss, L CORAL is a correlation alignment loss, is a loss term amplitude scaling coefficient;

[0031]

[0032] is a source domain surface temperature, is a core temperature true value, is a source domain surface temperature prediction value, is a core temperature true value prediction value, and λ1 is a weight factor;

[0033]

[0034] is a target domain surface temperature true value, is a target domain surface temperature true value prediction value, is a pseudo-label temperature, and λ2 is a weight factor.

[0035] The present application has the following beneficial effects:

[0036] (1) By combining a physical-based battery model and a machine learning algorithm, a two-stage framework of "simulation data pre-training + real data transfer learning" is constructed. The architecture generates synthetic data covering all operating conditions using an electro-thermal coupling model, completely solving the problem of high dependence of data-driven methods on experimental samples and reducing data acquisition costs. At the same time, unsupervised domain adaptive transfer learning technology effectively compensates for the domain difference between simulation and reality, improving the temperature estimation accuracy under dynamic operating conditions.

[0037] (2) Develop a polynomial approximation thermal model and equivalent circuit model coupling scheme for cylindrical batteries, and a Chebyshev-Galerkin thermal model and equivalent circuit model coupling scheme for square batteries. This differentiated modeling strategy breaks through the limitations of traditional single models, enabling accurate temperature estimation for lithium iron phosphate batteries (with hysteresis effect) and ternary lithium batteries (without hysteresis effect), covering mainstream application scenarios such as electric vehicles and energy storage stations.

[0038] (3) Use gated recurrent unit neural networks or long short-term memory recurrent neural networks to capture strong non-linear relationships, and combine maximum mean difference and correlation alignment double-index feature alignment mechanisms to control the absolute error of core temperature estimation within a certain range. Even under 5C ultra-high rate charging and discharging conditions, it can still track internal temperature changes in real time.

[0039] Other advantages, objects, and features of the present application will be apparent to those skilled in the art from the following specification, and it is intended to be covered by the following claims. The objects and other advantages of the present application can be achieved and obtained by the following description. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to make the purposes, technical solutions and advantages of the present application clearer, the preferred detailed description of the present application will be combined with the drawings as follows, wherein:

[0041] Figure 1 The overall flowchart for the internal temperature estimation of the lithium ion battery of the present application;

[0042] Figure 2 The schematic diagram of the thermal and electronic model of the present application;

[0043] Figure 3 The schematic diagram of the soft package battery shape and thermocouple arrangement in the embodiment of the present application;

[0044] Figure 4 The coupling principle diagram of the electric and thermal sub-models of the lithium ion battery of the present application;

[0045] Figure 5 The principle diagram of the GRU-RNN in the embodiment of the present application;

[0046] Figure 6 The domain self-adaptive principle diagram of the present application. DETAILED DESCRIPTION

[0047] The present application is illustrated by way of example and not limitation in the figures of the accompanying drawings, in which like references indicate similar elements, and in which: BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The accompanying drawings, which are included to provide a further understanding only, illustrate embodiments of the application and together with the description explain the principles of the application. In the drawings:

[0049] Corresponding same or similar reference numerals in the drawings of the present application indicate same or similar components; in the description of the present application, it should be understood that the orientation or position relationship indicated by terms such as "upper", "lower", "left", "right", "front", "back" and the like are based on the orientation or position relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and therefore the terms describing the position relationship in the drawings should not be understood as limiting the present application, and the above terms should be understood according to the specific meaning by the person skilled in the art according to the specific situation.

[0050] Referring to Figure 1 , the lithium-ion battery internal temperature estimation method based on machine learning and generated data includes the following steps:

[0051] S1: Establish a matched electro-thermal coupling model for the selected battery type, and use the physical model to scan the model parameters and test working conditions to generate simulation data covering multiple operating scenarios;

[0052] S2: Pre-train the machine learning model based on the simulation data set generated in S1, to capture the underlying mapping relationship between the feature input and the output;

[0053] S3: Based on the unlabeled data of the actual battery, use an unsupervised domain adaptive transfer learning algorithm to fine-tune the pre-trained machine learning model, and compensate for the simulation error between simulation and reality;

[0054] S4: Verify the method by using experimental data of real batteries under different operating conditions and convective cooling conditions.

[0055] Please refer to Figure 2 , the S1, for cylindrical battery PA thermal model combined with the ECM of the establishment of electro-thermal coupling model; for square battery CG thermal model combined with the ECM of the establishment of electro-thermal coupling model. For lithium iron phosphate cathode material battery, the electrical model selects a first-order RC with hysteresis ECM; for ternary cathode material battery, the electrical model selects a first-order RC ECM.

[0056] Specifically, the control equation of the PA thermal model is expressed as follows:

[0057]

[0058] Where, ρ, c p and k t respectively represent the battery volume average density, specific heat capacity and radial thermal conductivity. V b is the battery volume, Q represents the battery heat generation rate. Subject to the following boundary conditions:

[0059]

[0060] Where, T f is the external environment temperature, h is the convection coefficient, R is the battery outer diameter. The battery heat generation rate is calculated according to the classical Bernardi heat generation formula:

[0061]

[0062] The polynomial approximation model is to use polynomial expression to approximate the battery radial temperature distribution, which is a reduced order form of the above control equation, and can be expressed in state space form:

[0063]

[0064] y=Cx+Du

[0065] Where, u=[Q T f ] T , y=[T c T s ] T Respectively represent the state, input and output of the battery thermal system. And Respectively represent the battery volume average temperature and temperature gradient. T c And T s Respectively represent the core temperature and surface temperature of the battery. The coefficient matrix of the PA model is specifically expressed as follows:

[0066]

[0067]

[0068] where a is the thermal diffusivity of the battery, expressed as

[0069] For large-size soft-pack batteries, the control equations of the 2-D CG thermal model are expressed as follows:

[0070] pC h D t T-k x D xx T-k y D yy T=q

[0071] Subject to the following boundary conditions:

[0072]

[0073] where T(x, y, t) and q(x, y, t) represent the battery's spatial and temporal distribution temperature and volumetric heat generation rate. p, C h and k are the battery's volume-averaged density, specific heat capacity, and anisotropic thermal conductivity. Considering the lamination manufacturing process of soft-pack batteries, the thermal conductivity can be approximately considered to be consistent in the x and y directions. and and w and l represent the width and length of the battery. Subscripts r, l, t, and b refer to the right, left, top, and bottom boundaries of the battery, h x ∈ [h r -h l ] and h y ∈ [h t -h b ] represent the equivalent heat dissipation coefficients of the battery in the x and y directions (taking into account the effects of convective and radiative heat dissipation). In this invention, it is specified that the values of the heat dissipation coefficients are all positive, and the heat dissipation directions in the Cartesian coordinate system are distinguished by “±”. T ∞,x = [T r,∞ , T l,∞ ] and T ∞,y = [T t,∞ , T b,∞ ] represent the external cooling environment temperatures in both directions of the battery, assuming that the cooling environment around the battery being tested in the oven is uniform.

[0074] The solution to the original two-dimensional unsteady heat conduction problem can be decomposed into an auxiliary temperature function that satisfies the homogeneous boundary conditions and an arbitrary temperature function that satisfies the original non-homogeneous boundary conditions, i.e.

[0075]

[0076] where, Solving by the CG approximation method, T e Solving by constructing an arbitrary function using Chebyshev polynomials.

[0077] Using the CG approximation method and spatiotemporal separation technique, the original two-dimensional unsteady heat conduction equation can be transformed into a control-oriented state-space expression:

[0078]

[0079] Where E, A, B, and C are system matrices, and T e Let x be an arbitrary time-invariant temperature function related to spatial location, and let x = (χ) 00 ,χ 10 ,…,χ N0 ,χ 01 ,χ 11 ,…,χ N1 ,χ 0N ,χ 1N ,…,χ NN ) T The system input is u = [q 1]. T q represents the heat production rate per unit volume. Specifically, the matrix representations of each system are as follows:

[0080] For ease of expression, variables are defined.

[0081]

[0082] Then there is

[0083] E(i,j)=ρC p (ψ j ,ψ i )

[0084]

[0085] Where i,j=0,1,···,N.

[0086] The system output y can be defined according to the actual control system requirements. Generally, thermal management systems require the thermal model to accurately monitor the battery's highest temperature, maximum temperature gradient, and volume average temperature. For example... Figure 3 As shown, this invention extracts T3 to T9 and the volume average temperature T. b The output temperature of this thermal model, i.e.

[0087] y = [T3 T4 T5 T6 T7 T8 T9 T] b ] T

[0088] T e =[T e,3 T e,4 T e,5 T e,6 Te,7 T e,8 T e,9 T e,b ] T

[0089] Specifically, the output matrix C and the arbitrary temperature function T e is expressed as follows:

[0090]

[0091]

[0092] where i,j = 0,1,···,N.

[0093] The control equation of the first-order RC ECM is expressed as follows:

[0094]

[0095] V t,k = U oc,k -U p,k -R s,k I k

[0096] where I and U t represent the loading current and terminal voltage of the battery, respectively. For the current direction, this study defines charging as negative and discharging as positive. R s , R p , and C p represent the ohmic internal resistance, polarization resistance, and polarization capacitance of the battery, respectively. U p and U oc represent the polarization voltage and open-circuit voltage of the battery, respectively. η and C b are the Coulombic efficiency and the current available capacity of the battery, respectively (usually, η = 0.98 is taken). In actual use, C b needs to be calibrated in real time according to the average temperature output by the SOT estimator. SOC is the state of charge of the battery, which is calculated according to the ampere-hour integration method.

[0097] The control equation of the first-order RC hysteresis ECM is expressed as follows:

[0098]

[0099] V T,k = U oc,SOC,k -R s,k I k -U p,k +h k

[0100] where H is the maximum hysteresis voltage at a certain SOC, denotes the decay factor, and γ is a positive number that adjusts the decay rate of the hysteresis voltage.

[0101] Referring to Figure 4 , the mutual coupling between the electro-thermal model is realized through the temperature dependence of electrical parameters. Specifically, the dynamic heat generation of the battery during operation affects the evolution of the battery temperature, which in turn affects the electrical parameters. The volumetric heat generation rate of the battery is approximated using the overpotential-based method, and the thermal-dependent electrical parameters, such as the open-circuit voltage and the open-circuit voltage In addition to temperature dependence, these parameters also exhibit strong dependence on SOC.

[0102] Referring to Figure 5 , in S2, since GRU and LSTM have structural similarities and are both RNNs, only GRU-RNN is used as an example for illustration. The feature input includes measurable quantities such as current, voltage, ambient temperature, and surface temperature, as well as non-measurable quantities such as heat generation rate, and the model output includes the core temperature of the battery (cylindrical battery) or the temperature at key positions required for optimization control (square battery).

[0103] Specifically, the GRU memory unit includes a reset gate r t and an update gate z t , which are determined by the hidden state h t-1 at the previous time and the input x t at the current time. Using these two control gates, GRU-RNN can determine which needs to be updated and which needs to be reset, thereby capturing the long-term correlation of sequence data. The forward propagation mechanism of the GRU layer can be represented by the following equation set:

[0104]

[0105] where σ and tanh represent the gating activation functions. σ denotes the sigmoid function, represented as σ = 1 / (1 + exp(-x)), and tanh denotes the hyperbolic tangent function, represented as tanh = (exp(x) - exp(-x)) / (exp(x) + exp(-x)). denotes the linear activation function. ⊙ and [x, x] represent element multiplication and matrix concatenation, respectively. W r , W z , W h , and W0 are the weight matrices of the reset gate, the update gate, the short-time output, and the output layer, respectively. Similarly, b r , b z , b h , and b0 are the corresponding bias matrices. r t and z trespectively denote the outputs of the reset gate and the update gate. and h t respectively denote the candidate hidden state and the hidden state. y t denotes the output of the GRU. In the training phase, the mean squared error (MSE) of all outputs is chosen as the loss function of the GRU-RNN, as follows:

[0106]

[0107] where M and N represent the length of the predicted output of the GRU-RNN at a certain time step and the length of the sequence data, respectively. i,j and denote the estimated value and the reference value of the temperature T i at time step j. The GRU-RNN constantly updates each weight and bias in the network by backpropagation using the loss function gradient. The present application selects the Adam optimization algorithm to minimize the loss function, because Adam, as the most typical gradient-based optimization method, can adaptively adjust the learning rate and does not require excessive storage space.

[0108] Referring to Figure 6 , in the S3, the artificially synthesized data provides the core temperature or the key temperature data as the label of the machine learning model for supervised learning of the model, updates the pre-trained model by means of the domain adaptive transfer learning algorithm, compensates for the simulation error between the simulation and the actual, and improves the generalization ability of the model. The domain adaptive process includes two stages: pseudo-label selection and feature alignment. The pseudo-label is used to generate the key temperature label in the target domain to guide the retraining; in the feature alignment stage, the maximum mean difference (MMD) and the correlation alignment (CORAL) are used to evaluate the distribution difference between the high-dimensional features of the source domain and the target domain, so as to achieve more effective feature alignment.

[0109] Specifically, by inputting the same current and cooling temperature in the target domain to the thermal sub-model while keeping the thermal parameters in the source domain the same, the pseudo-label of the key temperature is generated. The present application takes the PA thermal model as an example for illustration.

[0110]

[0111] wherein, is the generated pseudo-label of the core temperature at time k. is the length of the unlabeled data, and the superscript T represents the target domain. The selection criterion of the pseudo-label is determined according to the simulation error of the surface temperature, and is as follows:

[0112]

[0113] wherein, is the pseudo-label of the core temperature in the target domain, Early data length to meet the requirement of surface temperature simulation accuracy (since it is an open-loop model, the estimation error of PA model for surface temperature and core temperature will gradually increase over time) Surface temperature calculated by target domain PA model, Surface temperature measured by target domain, δ is the selection threshold of pseudo label, which is assumed to be 25% of the maximum absolute error of surface temperature:

[0114]

[0115] After the pseudo label selection is completed, the data in the source domain and the target domain used for domain adaptation can be represented as:

[0116]

[0117] Where, N s and respectively refer to the number of samples in the source domain and the number of pseudo label samples. It is worth mentioning that the number of samples in the source domain may be much larger than the number of unlabeled samples in the target domain , which causes the imbalance problem in the domain adaptation process. Therefore, generally samples in the source domain are randomly selected, and

[0118] In feature alignment, in order to evaluate the distribution difference between the extracted high-dimensional features in the source domain and the target domain, a multi-order statistical index is adopted, including maximum mean difference (MMD) and correlation alignment (CORAL) index, to achieve more effective feature alignment. Given two data sets and subject to probability distribution p and q, the MMD between the two data sets can be expressed as:

[0119]

[0120] Where H represents a reproducing kernel Hilbert space RKHS, sup(·) is the upper bound of the input set, and Φ(·) is a nonlinear fitting function from the original data space to RKHS. Because the feature space of Φ(·) has infinite dimension, the mean square value of MMD can be calculated using the following characteristic kernel:

[0121]

[0122] Where is the two norm operation of RKHS, k(·,·) is the kernel function of RKHS, which is represented by Gaussian radial basis function as follows:

[0123]

[0124] Where γ represents the bandwidth parameter.

[0125] Compared with MMD, CORAL calculates the second-order statistical indicators of the source domain and the target domain, as follows:

[0126]

[0127] wherein, is the Frobenius norm, d is the feature dimension of data, C X and C Y are the covariances of two data sets, calculated as follows:

[0128]

[0129] wherein, 1 is a column vector with all elements being 1.

[0130] In order to retrain the GRU-RNN based on the proposed domain adaptation strategy, four loss terms are considered: (1) the regression loss in the source domain (2) the regression loss in the target domain and the domain difference loss of high-dimensional features between the source domain and the target domain: (3) L MMD and (4) L CORAL The regression loss in the two domains can be calculated as follows, respectively:

[0131]

[0132] wherein, λ1 and λ2 determine the weight of different regression losses. Since the GRU-RNN model estimates the surface temperature and the core temperature simultaneously, the regression loss of the surface temperature as an index in the target domain can be used as the proxy supervision in the data augmentation process, so as to promote the retraining. The final loss function can be summarized as:

[0133]

[0134] wherein, α, β and γ are all non-negative loss function weight coefficients. and are scaling factors of different loss terms, used to eliminate the amplitude difference. In order to avoid manual adjustment of these weight coefficients and improve the flexibility of the model in retraining under different scenarios, the present application adopts the homoscedastic uncertainty method to realize the automatic allocation of the weight of each task by maximizing the Gaussian likelihood function.

[0135] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the purpose and scope of the technical solutions, which should be covered in the scope of the claims of the present application.

Claims

1. A method for estimating internal temperature of a lithium-ion battery based on machine learning and generated data, characterized in that: The method comprises the following steps: S1: Establish a matched electro-thermal coupling model for the target battery type, generate simulation data covering multiple operating scenarios by scanning model parameters and test conditions; S2: Based on the simulation data set generated in S1, pre-train the machine learning model to capture the mapping relationship between input features and output; S3: Based on the unlabeled data of the actual battery, use an unsupervised domain adaptive transfer learning algorithm to fine-tune the pre-trained model to compensate for the error between simulation and reality; S4: Use experimental data of real batteries under different operating conditions and convection cooling conditions to verify the estimation results.

2. The lithium-ion battery internal temperature estimation method based on machine learning and generated data according to claim 1, characterized in that: In S1, an electro-thermal coupling model is established for cylindrical batteries by combining a polynomial approximation PA thermal model with an equivalent circuit model ECM; for lithium iron phosphate positive electrode material batteries, an equivalent circuit model with a first-order resistor-capacitor pair with hysteresis is used as the electrical model; for ternary positive electrode material batteries, a first-order resistor-capacitor pair equivalent circuit model is used as the electrical model; model parameters are randomly generated within a predetermined range using a Latin hypercube sampling strategy, and test conditions include standardized test cycles of urban and high-speed working conditions.

3. The lithium-ion battery internal temperature estimation method based on machine learning and generated data according to claim 1, characterized in that: In S2, the machine learning model uses a gated recurrent unit neural network GRU-RNN or a long short-term memory recurrent neural network LSTM-RNN; input features include current, voltage, ambient temperature, surface temperature, and heat generation rate, and output is battery core temperature or key position temperature.

4. The lithium-ion battery internal temperature estimation method based on machine learning and generated data according to claim 1, characterized in that: In S3, domain adaptive transfer learning includes two stages of pseudo-label generation and feature alignment: The pseudo label is generated by inputting target domain current and cooling temperature through an electro-thermal coupling model, and the selection criterion is satisfied and wherein δ is 25% of the maximum absolute error of the surface temperature, i.e. The feature alignment stage uses maximum mean difference MMD and correlation alignment CORAL to evaluate the feature distribution difference between the source domain and the target domain, and the MMD calculation formula is: where n is the number of source domain samples, m is the number of target domain samples, x i is a source domain data point, y j is a target domain data point, and K is a kernel function, γ is a bandwidth parameter, and CORAL is calculated as d is the dimension of data features, C X is a source domain covariance matrix, C Y is a target domain covariance matrix, and ||·||F F is the Frobenius norm.

5. The method of claim 2, wherein: The state space representation of the PA thermal model is: Wherein, the state quantity Input quantity u = [Q T f ] T , output quantity y = [T c T s ] T , The average temperature, The temperature gradient, Q is the heat generation rate, T f The ambient temperature, T c The core temperature, T s The surface temperature, A, B, C, D are system coefficient matrices, and the heat generation rate Q is calculated according to Bernardi formula: Wherein I is the current, V t The terminal voltage, U oc The open circuit voltage, SOC is the state of charge, T is the battery temperature, The partial derivative of open circuit voltage with respect to temperature.

6. The method of claim 3, wherein: The forward propagation of the gated recurrent unit neural network satisfies: r t = σ(W r · [h t-1 , x t ]+ b r ) z t = σ(W z · [h t-1 , x t ]+ b z ) where r t is the reset gate output, z t is the update gate output, is the candidate hidden state, h t is the current hidden state, h t-1 is the previous hidden state, x t is the current input feature, W r , W z , W h are weight matrices, and b r , b z , b h are bias matrices. σ is the Sigmoid activation function, and tanh is the hyperbolic tangent activation function.

7. The method of claim 4, wherein: The loss function of the fine-tuning process is a weighted four-tuple: wherein, α β, γ are non-negative weight coefficients, is a source domain regression loss; is a target domain regression loss; L MMD is a maximum mean discrepancy loss, L CORAL is a correlation alignment loss, is a loss term amplitude scaling coefficient; is the source domain surface temperature, is the core temperature true value, is the source domain surface temperature prediction value, is the core temperature true value prediction value, and λ1is a weight factor. is the target domain surface temperature real value, is the prediction of the target domain surface temperature real value, is the pseudo-label temperature, and λ2is a weight factor.

Images