Lithium battery health state estimation method and device, electronic equipment and storage medium

Abstract

The invention belongs to the technical field of electrochemical energy storage and intelligent diagnosis, and provides a lithium battery health state estimation method and device, electronic equipment and a storage medium, and the method comprises the steps: collecting and preprocessing first multi-source time sequence operation data; constructing a GLGRU neural network into which a variant of a Grunwald-Liouville fractional order difference operator is introduced, and constructing a GLGRU neural network into which the variant of the Grunwald-Liouville fractional order difference operator is introduced; a GLGRU neural network is trained by optimizing model parameters and fractional order parameters and by means of a projection gradient descent constraint mechanism; performing joint optimization processing on hyper-parameters of the GLGRU neural network through an Optuna automatic hyper-parameter optimization framework and a pruning method to obtain a lithium battery health state estimation model; according to the technical scheme, real-time estimation is carried out through the lithium battery health state estimation model, the health state estimation result is obtained, and the accuracy of battery health state estimation is improved according to the technical scheme.

Description

Technical Field

[0001] This invention relates to the field of electrochemical energy storage and intelligent diagnostics, and in particular to a method, apparatus, electronic device, and storage medium for estimating the health status of lithium batteries. Background Technology

[0002] With the rapid development of new energy vehicles and renewable energy systems, lithium-ion batteries, as their core energy storage unit, directly impact the overall performance and economic benefits of the system due to their safety, reliability, and lifespan. Batteries gradually degrade during long-term charge-discharge cycles, exhibiting capacity decay and increased internal resistance. This process leads to reduced battery range and potential safety hazards. Therefore, accurate estimation of the state of health of lithium batteries has become a key technical aspect of battery management systems (BMS).

[0003] Existing SOH estimation methods mainly include model-based methods, experiment-based methods, and data-driven methods. Methods based on electrochemical models or equivalent circuit models can reflect the battery aging process at the mechanistic level, but the modeling is complex and highly dependent on parameters, making it difficult to adapt to the diverse environments and operating conditions in real-world applications. Experiment-based capacity or impedance methods, while directly reflecting degradation characteristics, often rely on complete charge-discharge cycles or complex test conditions, making them unsuitable for real-time online monitoring. In contrast, data-driven deep learning methods, with their powerful feature extraction and nonlinear mapping capabilities, have emerged as a leading direction in recent research for achieving high-precision SOH estimation. For example, in their paper "Joint Estimation of SOC-SOH of Lithium-ion Batteries Based on Multi-channel Feature Fusion of LSTM-Transformer [J / OL]. Journal of Railway Science and Engineering, 1-13 [2025-10-13]", Yu Tianjian et al. proposed a joint network architecture (MLTA-Net) based on the fusion of adaptive weighted multi-channel long short-term memory network (LSTM) and Transformer, which achieved high-precision joint estimation of the state of charge (SOC) and state of health (SOH) of lithium-ion batteries. In their paper "Battery SOH Estimation Method Based on Fusion Physical Model and Data-Driven Model. Power Technology, 2025, 49(06):1183-1191," Yan Haodi et al. proposed a fusion structure of a single-sequence nonlinear dynamic (Malthus) physical lifetime model and a data-driven model. By embedding Malthus physical model constraints in the neural network, they achieved the estimation of SOH of lithium-ion batteries.

[0004] Among numerous neural network models, recurrent neural networks (RNNs) and their variants excel in handling time series forecasting problems. Gated recurrent units (GRUs), in particular, are widely used for battery state-of-the-art (SOH) estimation due to their simple structure and high computational efficiency. However, traditional GRUs typically employ integer-order state update mechanisms, limiting their memory characteristics to a fixed time step. This makes it difficult to fully capture the multi-scale dependencies and non-local memory effects exhibited by long-term battery degradation. When faced with long-cycle aging, temperature disturbances, and complex load conditions, traditional GRU models are prone to prediction delays and dynamic distortion. Summary of the Invention

[0005] Aimed at at least in solving one of the technical problems existing in the prior art, the present invention provides a method, apparatus, electronic device and storage medium for estimating the health status of lithium batteries, thereby improving the efficiency of battery health estimation.

[0006] One aspect of the present invention provides a method for estimating the state of health of a lithium battery, comprising:

[0007] Collect the first multi-source time-series operation data of the first lithium-ion battery, perform preprocessing and normalization on the first multi-source time-series operation data, and obtain the second multi-source time-series operation data;

[0008] Constructing a GLGRU neural network involves introducing a variant of the Grünwald-Liouville fractional difference operator into the hidden state update of the standard GRU, so that the hidden state update includes a fractional weighted cumulative term for historical time steps.

[0009] The second multi-source time-series running data is trained using the GLGRU neural network, wherein the training includes optimizing the model parameters and fractional-order parameters and using a projection gradient descent constraint mechanism to project the updated fractional-order parameters to the feasible region.

[0010] The hyperparameters of the GLGRU neural network are jointly optimized using the Optuna automated hyperparameter optimization framework and pruning method to obtain a lithium battery health status estimation model.

[0011] The second lithium-ion battery is estimated using the lithium battery health status estimation model to obtain the health status estimation result, which includes the SOH value or available remaining capacity.

[0012] According to the lithium battery health status estimation method, first multi-source time-series operational data of the first lithium-ion battery is collected, and preprocessing and normalization are performed on the first multi-source time-series operational data to obtain second multi-source time-series operational data, including:

[0013] Collect first multi-source timing operation data of the first lithium-ion battery, wherein the first multi-source timing operation data includes time interval, control voltage, control current, charging current and discharging current;

[0014] The first multi-source time-series running data is subjected to denoising, normalization, and sliding window segmentation to obtain the second multi-source time-series running data.

[0015] According to the lithium battery health state estimation method, the construction of the GLGRU neural network includes introducing a variant of the Grünwald-Liouville fractional difference operator into the hidden state update of the standard GRU, such that the hidden state update includes a fractional weighted cumulative term from historical time points, including:

[0016] The hidden states of the update gate and hidden gate of the standard GRU are superimposed with weighted information from multiple historical moments through a variant of the Grünwald-Liouville fractional difference operator, resulting in a GLGRU layer composed of multiple fractional gated units stacked together.

[0017] According to the lithium battery health status estimation method, the method further includes:

[0018] The update equation for the hidden state of the GLGRU neural network is as follows:

[0019]

[0020] in, It is the truncation order, indicating that the hidden state of history is considered up to the nth truncation order. item; For the fractional-order parameters of the GLGRU network; , and They are , and The One portion, To update the door, For a moment For the hidden state, For a moment The candidate hidden state; For the generalized binomial coefficients:

[0021]

[0022] Among them, fractional-order parameters and truncation order The convergence condition is .

[0023] According to the lithium battery health state estimation method, training the second multi-source time-series runtime data using the GLGRU neural network includes:

[0024] Initialize the GLGRU neural network to obtain initial values ​​for the fractional-order parameters and the truncation order. The truncation order is used to characterize the time-series running data of historical moments that the model parameters can use. The model parameters include weights and biases.

[0025] The initial values ​​of the fractional-order parameters and the initial values ​​of the cutoff order are adjusted using a convergence criterion function and based on the theoretical region of convergence.

[0026] When the GLGRU neural network is trained on the second multi-source time-series running data, the optimizer updates the model parameters and fractional-order parameters according to the gradient direction after each backpropagation.

[0027] After the optimizer is updated, it uses a convergence test function to determine whether to perform a projection operation on the fractional-order parameters. If the condition is met, the least squares approximation method is used for projection; otherwise, it automatically reverts to the nearest stable value.

[0028] Record the trajectory of the fractional parameter after each projection. If the trajectory shows an out-of-bounds projection a preset number of times, roll back the fractional parameter to the previous convergence state.

[0029] According to the lithium battery health state estimation method, the hyperparameters of the GLGRU neural network are jointly optimized using the Optuna automated hyperparameter optimization framework and pruning method to obtain the lithium battery health state estimation model, including:

[0030] Based on the objective function, a set of parameter combinations is randomly selected from the search space, where the search parameters of the search space include the learning rate, hidden layer dimension, sliding window length, truncation order, and fractional order parameters, and the objective function is to minimize the RMSE of the validation set.

[0031] Based on the parameter combination, at each experiment during training, the projective gradient descent constraint is executed and the validation set loss is recorded. The Optuna automated hyperparameter optimization framework is used to perform pruning based on the validation set loss, and the pruning results are recorded.

[0032] The pruning results are sampled, and Bayesian optimization driven by the TPE algorithm is used to process the sampled results to obtain candidate hyperparameters;

[0033] After repeating the experiment a preset number of times or after performance convergence, a lithium battery health state estimation model is obtained.

[0034] According to the lithium battery health state estimation method, Bayesian optimization driven by the TPE algorithm is used to process the sampling results to obtain candidate hyperparameters, including:

[0035] Experiments for obtaining sampling results The performance threshold is determined using the TPE algorithm. Divided into and :

[0036]

[0037] in, In the hyperparameter vector The losses obtained below Indicates the first of all trials A small loss value;

[0038] by As a performance threshold, the kernel density is used to determine the two conditional probability density functions. and :

[0039]

[0040] By using two conditional probability density functions and The ratio determines the candidate hyperparameters :

[0041]

[0042] in, It is the preset hyperparameter search space.

[0043] Another aspect of the present invention provides a lithium battery health status estimation device, comprising:

[0044] The first module is used to collect the first multi-source time-series operation data of the first lithium-ion battery, and to perform preprocessing and normalization on the first multi-source time-series operation data to obtain the second multi-source time-series operation data.

[0045] The second module is used to construct the GLGRU neural network, which includes introducing a variant of the Grünwald-Liouville fractional difference operator into the hidden state update of the standard GRU, so that the hidden state update includes a fractional weighted cumulative term of the historical time step.

[0046] The third module is used to train the second multi-source time-series running data through the GLGRU neural network, wherein the training includes optimizing the model parameters and fractional-order parameters and the projection gradient descent constraint mechanism, and projecting the updated fractional-order parameters to the feasible region.

[0047] The fourth module is used to jointly optimize the hyperparameters of the GLGRU neural network using the Optuna automated hyperparameter optimization framework and pruning method to obtain a lithium battery health status estimation model.

[0048] The fifth module is used to estimate the health status of the second lithium-ion battery using the lithium battery health status estimation model, and to obtain the health status estimation result, wherein the health status estimation result includes the SOH value or the available remaining capacity.

[0049] Another aspect of the present invention provides an electronic device, including a processor and a memory;

[0050] The memory is used to store programs;

[0051] The processor executes the program to implement the method as described above.

[0052] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the methods described above.

[0053] The beneficial effects of this invention are as follows: By introducing the Grünwald–Liouville fractional difference operator, the network possesses non-integer order memory decay characteristics, effectively characterizing the long-term dependence and multi-scale dynamic relationships existing in the battery degradation process; combined with the PGD projection constraint strategy, parameter divergence and numerical instability are avoided in the fractional-order parameter optimization process, ensuring the controllability and convergence speed of network training; the Optuna framework is used for automated hyperparameter optimization, reducing manual intervention, improving the model's adaptability to different battery types and working environments, achieving robust SOH estimation across datasets, and improving the accuracy of battery health state estimation. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the lithium battery health status estimation process according to an embodiment of the present invention.

[0055] Figure 2 This is a schematic diagram of the multi-source time-series running data preprocessing process according to an embodiment of the present invention.

[0056] Figure 3 This is a standard GRU model diagram according to an embodiment of the present invention.

[0057] Figure 4 This is an improved GLGRU model according to an embodiment of the present invention.

[0058] Figure 5 This is a schematic diagram of the GLGRU neural network training process according to an embodiment of the present invention.

[0059] Figure 6 This is a schematic diagram of the fractional-order parameters and PGD convergence control of the GLGRU neural network training according to an embodiment of the present invention.

[0060] Figure 7 This is a schematic diagram of the training and optimization process of the GLGRU neural network according to an embodiment of the present invention.

[0061] Figure 8 This is a schematic diagram of the automated hyperparameter optimization process of the GLGRU neural network according to an embodiment of the present invention.

[0062] Figure 9 This is a schematic diagram of a lithium battery health status estimation device according to an embodiment of the present invention. Detailed Implementation

[0063] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings. Throughout the description, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions. In the following description, suffixes such as "module," "part," or "unit" used to denote elements are used only for the purpose of illustrative purposes and have no specific meaning in themselves. Therefore, "module," "part," or "unit" can be used interchangeably. Terms such as "first," "second," etc., are used only to distinguish technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the sequential relationship of the indicated technical features. In the following description, the consecutive reference numerals for method steps are for ease of review and understanding. Adjusting the implementation order of steps, in conjunction with the overall technical solution of the present invention and the logical relationship between the various steps, will not affect the technical effect achieved by the technical solution of the present invention. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0064] refer to Figure 1 , Figure 1 This is a schematic diagram of the lithium battery health status estimation process according to an embodiment of the present invention, which includes, but is not limited to, steps S100~S500:

[0065] S100: Collect the first multi-source timing data of the first lithium-ion battery, perform preprocessing and normalization on the first multi-source timing data, and obtain the second multi-source timing data.

[0066] In some embodiments, the first multi-source time-series running data is the collected raw data, and the first lithium battery is the object of training data collection.

[0067] In some embodiments, reference Figure 2 The schematic diagram of the multi-source time-series runtime data preprocessing process shown includes, but is not limited to, steps S110~S120:

[0068] S110, collect the first multi-source timing operation data of the first lithium-ion battery, wherein the first multi-source timing operation data includes time interval, control voltage, control current, charging current and discharging current.

[0069] S120, perform denoising, normalization and sliding window segmentation on the first multi-source time series running data to obtain the second multi-source time series running data.

[0070] In some embodiments, after data acquisition is completed, the following preprocessing operations are performed on the raw data:

[0071] (1) Outlier removal and noise reduction: The moving average or triple σ method is used to remove abrupt changes;

[0072] (2) Normalization: Use min-max normalization to scale the input features to the [0,1] interval;

[0073] (3) Sliding window construction: Construct a fixed-length input window (length is WindowSize) according to the time step to form corresponding samples of the input sequence and the prediction target;

[0074] (4) Training / validation set partitioning: The dataset is partitioned in chronological order (60% training, 40% validation) to ensure that future information is not used during the testing phase.

[0075] S200, constructing a GLGRU neural network, including a variant of the Grünwald-Liouville fractional difference operator introduced into the hidden state update of the standard GRU, so that the hidden state update includes a fractional weighted cumulative term of the historical time step.

[0076] In some embodiments, the hidden states of the update gate and the hidden gate of the standard GRU are superimposed on the weighted information of multiple historical moments through a variant of the Grünwald-Liouville fractional difference operator to obtain a GLGRU layer composed of multiple fractional gated units stacked together.

[0077] In some embodiments, reference Figure 3 The standard GRU model shown and Figure 4 The GLGRU model shown in this embodiment is an improvement on the standard GRU model.

[0078] The GLGRU model proposed in this invention differs from the traditional GRU. GLGRU introduces Grünwald-Liouville fractional-order differencing during the hidden state update process, achieving fractional-order memory modeling of the battery degradation process by superimposing weighted information from multiple historical moments. Through weighted accumulation of historical data, GLGRU can more accurately simulate the nonlinear degradation characteristics of the battery throughout its entire lifespan.

[0079] In some embodiments, the model as a whole consists of an input layer, a GLGRU layer, and an output layer, specifically:

[0080] The input layer receives preprocessed feature sequences (time interval, control voltage, control current, and charge / discharge voltage and current, etc.).

[0081] GLGRU layer: Composed of multiple fractional-order gated units stacked together, it is the core for realizing long-term dependency modeling;

[0082] Output layer: Linearly maps the hidden state to the SOH estimate, realizing the prediction of battery health state.

[0083] In some embodiments, a standard GRU network consists of update gates and reset gates:

[0084] (1)

[0085] (2)

[0086] in, and These represent updating the door and resetting the door, respectively. The value range is [0,1], used to balance past and current information. Control the degree of influence of the previous hidden state; and , and , That is the corresponding weight matrix. and It is the corresponding bias. This represents the Sigmoid activation function.

[0087] According to the reset door And the previous hidden state Can calculate candidate hidden states :

[0088] (3)

[0089] Where tanh is the hyperbolic tangent activation function. and That is the corresponding weight matrix. That is the corresponding bias.

[0090] By updating the door The hidden state in the previous moment and candidate hidden state The current hidden state output can be calculated. :

[0091] (4)

[0092] In some embodiments, the GLGRU neural network is implemented using Grünwald-Liouville fractional difference:

[0093] hypothesis function yes For a differentiable function on , its GL derivative is calculated as follows:

[0094] (5)

[0095] in, It is a fractional-order parameter, taking any positive real number value, and h is the step size. These are the coefficients of the generalized binomial, calculated using the following formula:

[0096] (6)

[0097] Based on the continuous form of the derivative of GL, discretizing it yields the following difference form:

[0098] (7)

[0099] Where K is the truncation order.

[0100] In the GRU discrete-time series model, after normalizing the time, we can set the time step h=1. Therefore, at this point, we only need to consider a fixed fractional order. The above formula can be simplified to:

[0101] (8)

[0102] When the fractional order When k=1, the coefficients of the generalized binomial are only 0 when k=0, and not 0 when k=1; all other terms are 0. Therefore, the GL difference degenerates into an integer-order difference (first order), i.e.:

[0103] (9)

[0104] But under normal circumstances At this point, multiple weighting steps will occur, causing GL difference to exhibit fractional-order characteristics.

[0105] In some embodiments, the construction of the GLGRU model and its convergence conditions involve a simple transformation of the traditional GRU hidden layer formula, revealing that the GRU model uses a simple first-order difference:

[0106] (10)

[0107] in, For the fractional-order parameters of the GLGRU network; , and They are , and The One portion, To update the door, For a moment For the hidden state, For a moment The candidate hidden state.

[0108] By using GL fractional differencing, the original first-order differencing is extended to an adjustable fractional-order differencing form, and the hidden state update equation of the GLGRU model is constructed as follows:

[0109] (11)

[0110] Where K is the truncation order, indicating that the historical hidden state is considered up to the Kth term. This represents the order of the GLGRU network.

[0111] When fractional parameter When =1, it degenerates into the standard GRU update form, thus balancing the computational efficiency of the traditional model with the long-term memory characteristics of the fractional model.

[0112] To make GLGRU converge, the fractional order The cutoff length K must satisfy the convergence condition:

[0113] (12)

[0114] As can be seen from formula (12), region R is the convergent feasible region of the GLGRU model. Within this region, the weighted accumulation of historical information by the fractional difference term of the model is convergent, thereby ensuring that the dynamic evolution of the hidden state is within a controllable range.

[0115] In some embodiments, fractional order parameters The range of values ​​is It can be obtained through automatic learning or manual setting during the training process, and the preferred value of the fractional order parameter is between 0.2 and 0.8.

[0116] S300 trains the second multi-source time-series data using a GLGRU neural network. The training includes optimizing the model parameters and fractional-order parameters and using a projection gradient descent constraint mechanism to project the updated fractional-order parameters into the feasible region.

[0117] In some embodiments, reference Figure 5 The diagram showing the training process of the GLGRU neural network includes, but is not limited to, steps S310 to S350:

[0118] S310, initialize the GLGRU neural network to obtain the initial values ​​of the fractional-order parameters and the initial value of the truncation order. The truncation order is used to characterize the time-series running data of the historical moments that the model parameters can use. The model parameters include weights and biases.

[0119] S320 adjusts the initial values ​​of the fractional-order parameters and the initial value of the cutoff order based on the convergence judgment function and the theoretical convergence region.

[0120] When the S330 GLGRU neural network is trained on the second multi-source time-series running data, the optimizer updates the model parameters and fractional-order parameters according to the gradient direction after each backpropagation.

[0121] S340 After the optimizer is updated, it uses a convergence judgment function to determine whether to perform a projection operation on the fractional-order parameters. If the condition is met, the least squares approximation method is used for projection; otherwise, it automatically reverts to the nearest stable value.

[0122] S350 records the trajectory of the fractional-order parameter after each projection. If the trajectory of the parameter exceeds the limit a preset number of times, the fractional-order parameter is rolled back to the previous convergence state.

[0123] In some embodiments, reference Figure 6 The diagram shows the fractional-order parameters of the GLGRU neural network training and the convergence control of PGD.

[0124] It should be noted that in the traditional gradient descent process, if the fractional-order parameter... If the model is updated outside the convergence region, it may exhibit non-physical divergence. To address this issue, this embodiment of the invention introduces the PGD (Projected Gradient Descent) mechanism, which uses a combination of "update + projection" strategy to ensure parameter stability.

[0125] Specifically, fractional-order parameters are gradually guaranteed during training and model design. With truncation order Satisfy the theoretical region of convergence Requirements to avoid numerical instability, while supporting A learnable or hyperparameter search strategy is employed. Based on the convergence boundary derived theoretically, a convergence criterion function is created in the model, used during model initialization and at each iteration. After the update, all calls are made, and the steps are as follows:

[0126] (1) Parameter update: After each backpropagation, the optimizer updates the model parameters according to the gradient direction, including weights, biases and fractional parameters. ;

[0127] (2) Parameter projection: After updating, the convergence judgment function is used to determine whether the projection operation needs to be performed immediately. ,Will Projecting back to the theoretical region of convergence Within this framework, ensure that it satisfies the fractional stability condition;

[0128] (3) Projection strategy, implemented using the projection least squares approximation method, if If the value exceeds the boundary, it will automatically revert to a nearby stable value;

[0129] (4) Rollback mechanism: Record after each projection The trajectory of change; if out-of-bounds projections occur multiple times in a row, roll back to the previous convergence state.

[0130] In some embodiments, the initialization strategy is achieved by giving... Set an initial value and verify it using a convergence criterion function during initialization. If the condition is not met, adjust the initial value or... The process continues until the convergence condition is met before proceeding to the next step. When the value is large, the fractional accumulation will consume a lot of memory; by using sparse truncation ( By dividing the model into blocks, the amount of computation can be reduced, thereby improving the efficiency of model training.

[0131] In some embodiments, dynamic validation during training is also included, performed each time during a training loop. or During updates, a convergence check function is invoked. If an out-of-bounds or near-boundary condition is detected, subsequent constraint steps are triggered, and the data is recorded. , The time series of the convergence criterion function facilitates backtracking analysis and visual diagnosis.

[0132] Embodiments of the present invention transform the theoretical projection convergence region R into engineering steps that can be executed step-by-step within a deep learning framework, ensuring... The convergence condition is satisfied after each update, along with other constrained parameters. The convergence criterion function is called immediately after each parameter update step in the training iteration to check if the convergence is satisfied. When the boundary is exceeded or approaching, the PGD projection function is called. Projection is performed. Under this constraint, the historical weighted sum of the model is strictly limited to the convergence interval, ensuring the boundedness of the hidden state during its temporal evolution.

[0133] The S400 uses the Optuna automated hyperparameter optimization framework and pruning method to jointly optimize the hyperparameters of the GLGRU neural network, thus obtaining a lithium battery health state estimation model.

[0134] In some embodiments, reference Figure 7 The diagram showing the training and optimization process of the GLGRU neural network includes, but is not limited to, steps S410 to S430:

[0135] S410, according to the objective function, randomly select a set of parameter combinations from the search space, where the search parameters of the search space include the learning rate, hidden layer dimension, sliding window length, truncation order, and fractional order parameters, and the objective function is to minimize the RMSE of the validation set.

[0136] S420 performs projective gradient descent constraints and records the validation set loss at each experiment during training, based on the parameter combination. The Optuna automated hyperparameter optimization framework is then used to perform pruning based on the validation set loss, and the pruning results are recorded.

[0137] S430: The pruning results are sampled, and Bayesian optimization driven by the TPE algorithm is used to process the sampled results to obtain candidate hyperparameters.

[0138] S440, after repeating the experiment a preset number of times or after performance convergence, obtains the lithium battery health state estimation model.

[0139] In some embodiments, RMSE is the root mean square error, and the objective function is to minimize the root mean square error index of the validation set.

[0140] In some embodiments, reference Figure 8 The diagram illustrates the automated hyperparameter optimization process of the GLGRU neural network. The Optuna framework is used to automatically search for and optimize key hyperparameters. The process is as follows:

[0141] (1) Establish the objective function, defining the objective as minimizing the RMSE of the validation set. The objective function includes the complete model building, training and validation process.

[0142] (2) Set the search space. The search parameters include the learning rate, hidden layer dimension, sliding window length, fractional truncation order K, and fractional order. Each experiment randomly samples a set of parameter combinations.

[0143] (3) Bayesian optimization driven by TPE algorithm:

[0144] For a given past trial TPE based on performance threshold It is divided into:

[0145] (13)

[0146] in, In the hyperparameter vector The losses obtained below Indicates the first of all trials The minimum loss value is used as the performance threshold.

[0147] It should be noted that the division results (Good experimental results) and (Poor experimental results) are automatically selected based on the experiments already conducted, therefore, each time a new experimental round begins... and The value may be different.

[0148] The two conditional probability density functions are then determined using kernel density estimation (KDE):

[0149] (14)

[0150] The next set of candidate hyperparameters with high potential performance is generated by maximizing the probability ratio between the two. ,Right now:

[0151] (15)

[0152] in This is the preset hyperparameter search space.

[0153] (4) Mid-term pruning mechanism: periodically evaluate and verify the loss during training. If the performance is lower than the median, terminate the experiment early to reduce computational resource consumption.

[0154] (5) Optimal parameter extraction and application: When the preset number of trials or performance convergence is reached, the system outputs the optimal parameter combination and automatically uses it for the retraining of the final GLGRU model.

[0155] It should be noted that the objective function needs to encapsulate the complete model building, training, and validation process, with each call representing an independent trial. Inside the objective function, the trial objects provided by Optuna are used to dynamically sample various hyperparameters, and performance metrics are returned after training is complete.

[0156] In some embodiments, at the start of each trial, the GLGRU model is re-instantiated based on the parameters obtained from the trial sampling, and the parameters are passed into the model initialization function so that the network structure and fractional order mechanism are automatically adjusted with sampling.

[0157] Each trial executes the complete training and validation process:

[0158] (1) Set the optimizer according to the sampling learning rate;

[0159] (2) Use PGD to perform several epochs (referring to the process of the model performing one forward and one backward propagation on the entire training dataset);

[0160] (3) Record the validation set loss after each epoch;

[0161] (4) Report the interim results of the current trial to Optuna. This is used for subsequent pruning decisions.

[0162] (5) Mid-term pruning mechanism: Optuna automatically calculates the median performance of trials in the same phase. If the validation performance of the current trial is consistently lower than the median, pruning is triggered to terminate the trial early. Pruning events are recorded and used for posterior distribution updates in the next sampling.

[0163] In embodiments of this invention, the Optuna automated hyperparameter optimization framework is introduced during the model training phase to achieve adaptive search and dynamic adjustment of key network parameters and training parameters. This phase combines Tree-based Bayesian Optimization (TPE) with Median Pruning to achieve rapid convergence of model performance and optimal combination of hyperparameter configurations while ensuring training efficiency.

[0164] S500 estimates the health status of the second lithium-ion battery using a lithium battery health status estimation model, and obtains the health status estimation result, which includes the SOH value or available remaining capacity.

[0165] In some embodiments, the second lithium-ion battery is the lithium-ion battery to be estimated or detected. It should be noted that the second lithium-ion battery needs to be preprocessed in step S100 and input into the final GLGRU model to predict and output the SOH value or remaining usable capacity prediction result in real time for battery management system or lifetime decision support.

[0166] Figure 9 This is a schematic diagram of a lithium battery health status estimation device according to an embodiment of the present invention. The device includes a first module 910, a second module 920, a third module 930, a fourth module 940, and a fifth module 950.

[0167] The system comprises five modules: First, a module collects first multi-source time-series operational data from a first lithium-ion battery, preprocesses and normalizes this data to obtain second multi-source time-series operational data. Second, a module constructs a GLGRU neural network, incorporating a variant of the Grünwald-Liouville fractional difference operator into the hidden state update of the standard GRU, enabling the hidden state update to include a fractional-weighted cumulative term from historical time points. Third, a module trains the second multi-source time-series operational data using the GLGRU neural network, including optimizing model parameters and fractional-order parameters and applying a projection gradient descent constraint mechanism to project the updated fractional-order parameters into the feasible region. Fourth, a module jointly optimizes the hyperparameters of the GLGRU neural network using the Optuna automated hyperparameter optimization framework and pruning methods to obtain a lithium battery health state estimation model. Fifth, a module estimates the health state of the second lithium-ion battery using this model, obtaining a health state estimation result, which includes the State of Health (SOH) value or available remaining capacity.

[0168] For example, with the cooperation of the first, second, third, fourth, fifth, and sixth modules in the device, the embodiment device can implement any of the aforementioned lithium battery health state estimation methods, namely, acquiring first multi-source time-series operating data of the first lithium-ion battery, performing preprocessing and normalization on the first multi-source time-series operating data to obtain second multi-source time-series operating data; constructing a GLGRU neural network, including introducing a variant of the Grünwald-Liouville fractional difference operator into the hidden state update of the standard GRU, so that the hidden state update includes historical time... A fractional-order weighted cumulative term is used. The second multi-source time-series data is trained using a GLGRU neural network. Training includes optimizing model parameters and fractional-order parameters, and employing a projection gradient descent constraint mechanism to project the updated fractional-order parameters into the feasible region. The hyperparameters of the GLGRU neural network are jointly optimized using the Optuna automated hyperparameter optimization framework and pruning methods to obtain a lithium battery health state estimation model. This model is then used to estimate the health state of the second lithium-ion battery, yielding a health state estimation result, which includes the state of health (SOH) value or available remaining capacity. The beneficial effects of this invention are as follows: By introducing the Grünwald–Liouville fractional difference operator, the network possesses non-integer order memory decay characteristics, effectively characterizing the long-term dependence and multi-scale dynamic relationships existing in the battery degradation process; combined with the PGD projection constraint strategy, parameter divergence and numerical instability are avoided in the fractional-order parameter optimization process, ensuring the controllability and convergence speed of network training; the Optuna framework is used for automated hyperparameter optimization, reducing manual intervention, improving the model's adaptability to different battery types and working environments, achieving robust SOH estimation across datasets, and improving the accuracy of battery health state estimation.

[0169] This invention also provides an electronic device, which includes a processor and a memory;

[0170] The memory stores the program;

[0171] The processor executes a program to perform the aforementioned lithium battery health state estimation method; the electronic device has the function of carrying and running the lithium battery health state estimation software system provided in the embodiments of the present invention, such as a personal computer, minicomputer, mainframe, workstation, network or distributed computing environment, standalone or integrated computer platform, or communicating with charged particle tools or other imaging devices, etc.

[0172] This invention also provides a computer-readable storage medium storing a program that is executed by a processor to implement the lithium battery health state estimation method described above.

[0173] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this invention are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented in the embodiments of this invention. Alternative embodiments are contemplated, in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0174] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the aforementioned lithium battery health state estimation method.

[0175] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, considering the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed in the embodiments of the invention, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.

[0176] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0177] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can include, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0178] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0179] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0180] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0181] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0182] The above is a detailed description of the preferred embodiments of the present invention, but the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for estimating the state of health of a lithium battery, characterized in that, include: Collect the first multi-source time-series operation data of the first lithium-ion battery, perform preprocessing and normalization on the first multi-source time-series operation data, and obtain the second multi-source time-series operation data; Constructing a GLGRU neural network involves introducing a variant of the Grünwald-Liouville fractional difference operator into the hidden state update of the standard GRU, so that the hidden state update includes a fractional weighted cumulative term for historical time steps. The second multi-source time-series running data is trained using the GLGRU neural network, wherein the training includes optimizing the model parameters and fractional-order parameters and using a projection gradient descent constraint mechanism to project the updated fractional-order parameters to the feasible region. The hyperparameters of the GLGRU neural network are jointly optimized using the Optuna automated hyperparameter optimization framework and pruning method to obtain a lithium battery health status estimation model. The second lithium-ion battery is estimated using the lithium battery health status estimation model to obtain the health status estimation result, which includes the SOH value or available remaining capacity.

2. The lithium battery health status estimation method according to claim 1, characterized in that, The process involves collecting first multi-source time-series operational data from the first lithium-ion battery, preprocessing and normalizing the first multi-source time-series operational data to obtain second multi-source time-series operational data, including: Collect first multi-source timing operation data of the first lithium-ion battery, wherein the first multi-source timing operation data includes time interval, control voltage, control current, charging current and discharging current; The first multi-source time-series running data is subjected to denoising, normalization, and sliding window segmentation to obtain the second multi-source time-series running data.

3. The lithium battery health status estimation method according to claim 1, characterized in that, The construction of the GLGRU neural network includes introducing a variant of the Grünwald-Liouville fractional difference operator into the hidden state update of the standard GRU, so that the hidden state update includes a fractional weighted cumulative term of historical time steps, including: The hidden states of the update gate and hidden gate of the standard GRU are superimposed with weighted information from multiple historical moments through a variant of the Grünwald-Liouville fractional difference operator, resulting in a GLGRU layer composed of multiple fractional gated units stacked together.

4. The lithium battery health status estimation method according to claim 1, characterized in that, The method further includes: The update equation for the hidden state of the GLGRU neural network is as follows: in, It is the truncation order, indicating that the hidden state of history is considered up to the nth truncation order. item; For the fractional-order parameters of the GLGRU network; , and They are , and The One portion, To update the door, For a moment For the hidden state, For a moment The candidate hidden state; For the generalized binomial coefficients: Among them, fractional-order parameters and truncation order The convergence condition is .

5. The lithium battery health status estimation method according to claim 4, characterized in that, The step of training the second multi-source time-series runtime data using the GLGRU neural network includes: Initialize the GLGRU neural network to obtain initial values ​​for the fractional-order parameters and the truncation order. The truncation order is used to characterize the time-series running data of historical moments that the model parameters can use. The model parameters include weights and biases. The initial values ​​of the fractional-order parameters and the initial values ​​of the cutoff order are adjusted using a convergence criterion function and based on the theoretical region of convergence. When the GLGRU neural network is trained on the second multi-source time-series running data, the optimizer updates the model parameters and fractional-order parameters according to the gradient direction after each backpropagation. After the optimizer is updated, it uses a convergence test function to determine whether to perform a projection operation on the fractional-order parameters. If the condition is met, the least squares approximation method is used for projection; otherwise, it automatically reverts to the nearest stable value. Record the trajectory of the fractional parameter after each projection. If the trajectory shows an out-of-bounds projection a preset number of times, roll back the fractional parameter to the previous convergence state.

6. The lithium battery health status estimation method according to claim 1, characterized in that, The hyperparameters of the GLGRU neural network are jointly optimized using the Optuna automated hyperparameter optimization framework and pruning method to obtain a lithium battery health state estimation model, including: Based on the objective function, a set of parameter combinations is randomly selected from the search space, where the search parameters of the search space include the learning rate, hidden layer dimension, sliding window length, truncation order, and fractional order parameters, and the objective function is to minimize the RMSE of the validation set. Based on the parameter combination, at each experiment during training, the projective gradient descent constraint is executed and the validation set loss is recorded. The Optuna automated hyperparameter optimization framework is used to perform pruning based on the validation set loss, and the pruning results are recorded. The pruning results are sampled, and Bayesian optimization driven by the TPE algorithm is used to process the sampled results to obtain candidate hyperparameters; After repeating the experiment a preset number of times or after performance convergence, a lithium battery health state estimation model is obtained.

7. The lithium battery health status estimation method according to claim 6, characterized in that, The Bayesian optimization driven by the TPE algorithm processes the sampling results to obtain candidate hyperparameters, including: Experiments for obtaining sampling results The performance threshold is determined using the TPE algorithm. Divided into and : in, In the hyperparameter vector The losses obtained below Indicates the first of all trials A small loss value; by As a performance threshold, the kernel density is used to determine the two conditional probability density functions. and : By using two conditional probability density functions and The ratio determines the candidate hyperparameters : in, It is the preset hyperparameter search space.

8. A lithium battery health status estimation device, characterized in that, include: The first module is used to collect the first multi-source time-series operation data of the first lithium-ion battery, and to perform preprocessing and normalization on the first multi-source time-series operation data to obtain the second multi-source time-series operation data. The second module is used to construct the GLGRU neural network, which includes introducing a variant of the Grünwald-Liouville fractional difference operator into the hidden state update of the standard GRU, so that the hidden state update includes a fractional weighted cumulative term of the historical time step. The third module is used to train the second multi-source time-series running data through the GLGRU neural network, wherein the training includes optimizing the model parameters and fractional-order parameters and the projection gradient descent constraint mechanism, and projecting the updated fractional-order parameters to the feasible region. The fourth module is used to jointly optimize the hyperparameters of the GLGRU neural network using the Optuna automated hyperparameter optimization framework and pruning method to obtain a lithium battery health status estimation model. The fifth module is used to estimate the health status of the second lithium-ion battery using the lithium battery health status estimation model, and to obtain the health status estimation result, wherein the health status estimation result includes the SOH value or the available remaining capacity.

9. An electronic device, characterized in that, Including the processor and memory; The memory is used to store programs; The processor executes the program to implement the lithium battery health status estimation method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a program that is executed by a processor to implement the lithium battery health state estimation method as described in any one of claims 1-7.

Images