Lithium-ion battery soh and rul prediction method based on hierarchical parameter search

Abstract

This invention discloses a method for predicting SOH and RUL of lithium-ion batteries based on hierarchical parameter search: S1, collect feature data of lithium-ion batteries during charge-discharge cycles; S2, preprocess the feature data and construct tensors to generate a three-dimensional input tensor containing time steps within the cycle and historical dimensions across cycles; S3, divide the data into training set, validation set, and test set; S4, construct a backbone neural network model based on CNN-GRU; S5, construct independent and parallel SOH prediction heads and RUL prediction heads, and decouple the multi-task features for output; S6, construct a three-level hierarchical parameter search space containing a structural layer, a dynamic layer, and a multi-task calibration layer, and search for the optimal hyperparameters and multi-task loss function weights of the network; S7, configure the network using the optimal hyperparameter combination, train the network using the training set, stop training when the performance indicators on the validation set tend to stabilize, and obtain a joint prediction model; perform prediction on the test set to obtain the prediction results of SOH and RUL of lithium-ion batteries.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of battery health management and artificial intelligence, specifically involving a dual-task prediction method for lithium-ion battery state of health (SOH) and remaining service life (RUL) based on hierarchical parameter search and deep neural network. Background Technology

[0002] Lithium-ion batteries are widely used in electric vehicles and energy storage systems. Their State of Health (SOH) characterizes the current degree of battery degradation, while Remaining Useful Life (RUL) describes the number of usable cycles remaining before reaching the failure threshold. Achieving high-precision joint prediction of SOH and RUL is of great significance for developing maintenance strategies and ensuring system safety.

[0003] In recent years, data-driven deep learning technology has made significant progress in battery health prediction. However, existing technologies still face the following challenges: First, joint prediction of SOH and RUL is a typical multi-task learning problem. The two have different evolution scales, which can easily lead to gradient interference and task conflicts during training, making it difficult to manually set the optimal loss weights. Second, the performance of deep neural networks is highly dependent on hyperparameters (such as the number of network layer nodes, learning rate, batch size, etc.). Existing methods often rely on experience to perform grid search or random search, which is inefficient and difficult to reach the global optimum. Finally, a single network structure often cannot simultaneously take into account the local feature extraction of physical waveforms such as voltage and current within a single cycle, as well as the global modeling of degradation trends over long sequences across cycles. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, this invention proposes a method for predicting the State of Harm (SOH) and Rullow Usage (RUL) of lithium-ion batteries based on hierarchical parameter search. This invention designs a CNN-GRU deep network as the core model and introduces a multi-task dual-predictor mechanism. Simultaneously, a hierarchical parameter search framework is constructed to automatically optimize the network structure, training dynamic parameters, and multi-task loss weights, thereby improving the accuracy and robustness of joint prediction.

[0005] The present invention adopts the following technical solution:

[0006] A method for predicting SOH and RUL of lithium-ion batteries based on hierarchical parameter search includes the following steps:

[0007] S1. Battery Characteristic Data Acquisition. This involves acquiring characteristic data of the lithium-ion battery during charge-discharge cycles. The raw characteristic data dimensions are voltage, current, and capacity.

[0008] S2. Preprocess and construct tensors for battery feature data. Through interpolation alignment and sliding window mechanism, generate a three-dimensional input tensor containing time steps within the cycle and historical dimensions across cycles.

[0009] S3. Divide the constructed dataset into training set, validation set and test set according to the preset ratio;

[0010] S4. Construct a backbone neural network model based on CNN-GRU (Convolutional Neural Network-Gated Recurrent Unit). The CNN module consists of 7 convolutional layers, which are specifically used to handle the local degradation features of the single-cycle waveform of the battery. The GRU module consists of 2 recurrent network layers, which model the long-term degradation evolution law across cycles.

[0011] S5. After the network recursive module (i.e., GRU module, gated recurrent unit), construct independent and parallel dual prediction heads, namely the SOH prediction head and the RUL prediction head, to decouple the multi-task features for output.

[0012] S6. Construct a three-level hierarchical parameter search space containing a structural layer, a dynamics layer, and a multi-task calibration layer, and use an optimization algorithm to automatically search for the optimal hyperparameters and weights of the multi-task loss function of the network.

[0013] S7. Configure the network using the optimal hyperparameter combination obtained from the search, and perform closed-loop training on the network using the training set. Stop training when the performance indicators on the validation set tend to stabilize, and obtain the final joint prediction model. Make predictions on the test set to obtain the prediction results of SOH and RUL of lithium-ion batteries. Further performance evaluation can be performed.

[0014] As a preferred approach, in step S1, the original data dimensions are the voltage, current, and capacity data of the discharge phase in a complete charge-discharge cycle. The true SOH label and global RUL label are calculated by reading and recording the capacity of the first cycle. The standardization process uses the Z-score standardization method to standardize each feature dimension separately to eliminate the influence of dimensions.

[0015] As a preferred option, in step S2, the preprocessing and tensor construction are specifically as follows: one-dimensional linear interpolation is used to align the waveform of each cycle to a fixed length of sampling points; the format of the three-dimensional input tensor is (number of feature channels x time steps within a cycle x cross-cycle historical dimension), where the sliding window size of the cross-cycle historical dimension is fixed at 10, that is, the waveform features of 10 consecutive cycles are stacked.

[0016] As a preferred option, in step S2, one-dimensional linear interpolation is used to align the waveform of each cycle to 100 sampling points of a fixed length; the sliding window size is set to 10, and the waveform features of 10 consecutive cycles are stacked to generate a data tensor input with dimensions (Batch, 10, 100, number of channels).

[0017] As a preferred approach, in step S3, the dataset is the MIT-Stanford dataset (public dataset), with a split ratio of 70% for training, 15% for validation, and 15% for testing, maintaining consistency in the data distribution. As a preferred approach, in step S4, the specific structure of the CNN-GRU backbone network is as follows: the CNN module contains 7 convolutional layers; the first 3 layers are used for fast dimensionality reduction and feature smoothing, and the last 4 layers, in conjunction with pooling layers, extract higher-order nonlinear degenerate features; the GRU module contains 2 gated recurrent units, used to receive the temporal feature vectors compressed by the convolutional modules.

[0018] As a preferred embodiment, in step S5, the SOH prediction head specifically maps the time step features point by point and outputs an SOH sequence of length 10; the RUL prediction head specifically flattens the time dimension into a one-dimensional vector and then outputs a single lifetime scalar through a linear layer.

[0019] As a preferred option, in step S5, the two independently parallel prediction heads adopt different aggregation strategies: the SOH prediction head adopts a many-to-many mapping strategy to directly output the real-time SOH estimation sequence of each cycle within the historical window; the RUL prediction head adopts a global average pooling mechanism to aggregate and compress the time dimension, and outputs a single scalar representing the predicted remaining cycle lifetime.

[0020] As a preferred option, in step S6, the search parameters for the structure layer are specifically: the number of GRU hidden layer nodes is searched between 32, 64, and 128.

[0021] As a preferred option, in step S6, the parameters of the three-level hierarchical parameter search space include: the structural layer covers the number of hidden nodes in the GRU layer and the Dropout ratio (representing the cyclic dropout method, which is introduced between GRU layers to prevent overfitting); the dynamics layer covers the basic learning rate (1e-4 to 1e-2) and the batch size (16 to 128); and the multi-task calibration layer covers the dynamic weighting coefficient between the SOH prediction loss and the RUL prediction loss.

[0022] As a preferred option, step S6, which involves automated hierarchical parameter search, mainly includes the following steps:

[0023] S6.1 Construct a three-level hierarchical parameter search space and clarify the parameter boundaries of the structural layer, dynamics layer and multi-task calibration layer;

[0024] S6.2 Utilize the tree-structured Parzen estimator (TPE) algorithm for automatic parameter optimization iteration;

[0025] S6.3 Calculate the total loss function on the validation set to evaluate the performance of each parameter combination, record and output the best-performing hyperparameter combination for configuring the final network.

[0026] Preferably, in step S6, the search parameters for the dynamics layer are: a base learning rate ranging from 1e-4 to 1e-2, and a batch size ranging from 16 to 128.

[0027] Preferably, in step S6, the weighting coefficient α of the search parameters of the multi-task calibration layer ranges from 0.1 to 0.9, and the total loss function is defined as L = α*L_SOH + (1-α)*L_RUL, which is used to balance the dimensions of the two and the gradient.

[0028] As a preferred option, in step S7, the criterion for model convergence is that the total loss curve on the validation set tends to stabilize.

[0029] As a preferred option, in step S7, the performance evaluation metrics include the root mean square error (RMSE) of SOH on the test set and the symmetric mean absolute percentage error (sMAPE) of RUL, in order to perform a quantitative analysis of the joint prediction performance and output the best-performing hyperparameter combination for actual deployment.

[0030] Compared with the prior art, the beneficial effects of the present invention are:

[0031] (1) This invention uses a 7-layer CNN as a convolutional module to deeply mine the local degradation details of a single-cycle waveform, and combines it with a 2-layer GRU as a recursive module to capture long-term capacity decay memory. This combination effectively takes into account both microscopic physical characteristics and macroscopic degradation trends.

[0032] (2) The present invention decouples multi-task output with dual prediction heads. In view of the different temporal characteristics of SOH and RUL, a many-to-many SOH tracking head and a global pooling-based RUL trend head are designed respectively, which avoids information loss caused by forcibly unifying the output dimension and significantly improves the collaborative prediction performance of dual tasks.

[0033] (3) This invention solves the parameter tuning problem through hierarchical parameter search. It incorporates network nodes, learning dynamics, and multi-task loss weights into a unified three-level hierarchical parameter search framework. Through automated optimization, it effectively solves the problem of uneven gradient distribution in multi-task learning, completely eliminates inefficient manual trial and error, and improves the generalization ability and development efficiency of the model. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart of a preferred embodiment of the present invention for predicting the SOH and RUL of a lithium-ion battery based on hierarchical parameter search;

[0036] Figure 2 This is a schematic diagram of a CNN-GRU network and a dual prediction head structure according to a preferred embodiment of the present invention (wherein, Convolutional module refers to the convolutional module, Conv+ReLU refers to a convolutional layer plus an activation layer, MaxPool refers to the max pooling layer; Recurrent module refers to the recurrent module, GRU is a gated unit; the third box is the dual prediction head, where SOH hesd refers to the SOH prediction head, per-timestep means output step by step, RUL head refers to the RUL prediction head, and AvgPool over T means output step by step).

[0037] Figure 3 This is a diagram illustrating the operational logic and hierarchical architecture of the hierarchical parameter search space in a preferred embodiment of the present invention (the left side represents the search algorithm TPE + Median Pruner tree structure search + median pruning; the hierarchical classification in the middle is divided into 3 layers: the first layer is structural plasticity constraint, used to detect global parameter configuration; the second layer is differential evolution dynamics, used to determine the learning rate ratio of different network layers; the third layer is multi-task objective calibration, used to coordinate the gradient contribution between different prediction tasks; TPEtrial configurations refer to the parameter configuration of the TPE experiment, which transmits the parameters of this round to the main network, evaluates the adaptation loss, and determines whether to continue using the search algorithm to obtain a new batch of parameters based on the loss evaluation of this round). Detailed Implementation

[0038] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0039] This embodiment provides a method for predicting the SOH and RUL of lithium-ion batteries based on hierarchical parameter search, referring to... Figure 1 The diagram shown is a flowchart of the method. (Refer to...) Figure 2 This is a schematic diagram of a CNN-GRU network and a dual prediction head structure. The image is processed according to the flowchart of the method of the present invention to illustrate in detail the effect of the method of the present invention on improving detection accuracy.

[0040] This embodiment presents a method for predicting the State of Health (SOH) and Rullow U (RUL) of lithium-ion batteries based on hierarchical parameter search, comprising the following steps:

[0041] S1. Battery Characteristic Data Acquisition. This involves acquiring characteristic data of the lithium-ion battery during charge-discharge cycles. The raw characteristic data dimensions are voltage, current, and capacity.

[0042] S2. Preprocess and construct tensors for battery feature data. Through interpolation alignment and sliding window mechanism, generate a three-dimensional input tensor containing time steps within the cycle and historical dimensions across cycles.

[0043] S3. Divide the constructed dataset into training set, validation set and test set according to the preset ratio;

[0044] S4. Construct a backbone neural network model based on CNN-GRU, wherein the CNN module consists of 7 convolutional layers, which is specifically used to handle the local degradation features of the single-cycle waveform of the battery; the GRU module consists of 2 recurrent layers, which models the long-term degradation evolution law across cycles.

[0045] S5. After the network recursive module, construct independent and parallel dual prediction heads, namely the SOH prediction head and the RUL prediction head, to decouple the multi-task features for output.

[0046] S6. Construct a three-level hierarchical parameter search space containing a structural layer, a dynamics layer, and a multi-task calibration layer, and use an optimization algorithm to automatically search for the optimal hyperparameters and weights of the multi-task loss function of the network.

[0047] S7. Configure the network using the optimal hyperparameter combination obtained from the search, and perform closed-loop training on the CNN-GRU network using the training set. Stop training when the performance indicators on the validation set tend to stabilize, and obtain the final joint prediction model. Make predictions on the test set to obtain the prediction results of SOH and RUL of lithium-ion batteries. Further performance evaluation can be performed.

[0048] The following is a more detailed description of each step.

[0049] In step S1, the collection of battery data specifically includes the following key points:

[0050] S1.1 For the voltage, current, and capacity data during the discharge phase of a complete charge-discharge cycle, the Z-score normalization method is used to standardize each feature dimension separately:

[0051]

[0052] in, For the first The sampling point at the th sampling point Raw readings on each characteristic dimension (voltage, current, or capacity), and These are the mean and standard deviation of the feature dimension on the global training set, respectively.

[0053] S1.2, in the The battery's true health status label during each charge-discharge cycle. With Remaining Useful Life Label Calculate strictly according to the following formula:

[0054]

[0055]

[0056] in, For the first The actual discharge capacity measured during each cycle of discharge phase; This represents the actual capacity of the battery in its first cycle under its initial nominal condition. This refers to the total number of cycles in the battery's complete lifecycle when it reaches a specified failure threshold (such as 80% of its initial capacity).

[0057] In step S2, data preprocessing and tensor construction specifically include the following key points:

[0058] S2.1 Use one-dimensional linear interpolation to align the waveform of each cycle to 100 sampling points of a fixed length.

[0059] S2.2. Set the sliding window size to 10, stack the waveform features of 10 consecutive cycles, and generate a data tensor input with dimensions (Batch, 10, 100, number of channels). That is, generate a three-dimensional input tensor. In the middle, the number of sliding window cycles Sampling point length Number of feature channels (Covering voltage, current, and capacity).

[0060] Furthermore, in step S3, the dataset is divided using a stratified sampling strategy, with the specific proportions being 70% for the training set, 15% for the validation set, and 15% for the test set.

[0061] In step S4, the specific implementation of the CNN-GRU hybrid model is as follows:

[0062] S4.1, CNN section: Contains a 7-layer convolutional network. The first 3 layers are used for fast dimensionality reduction, and the last 4 layers, combined with max pooling layers, extract higher-order features. The output is compressed into a temporal feature vector that retains the historical cyclic dimension. Let the single-cycle input matrix be... After the first three layers of fast dimensionality reduction convolutions and the last four layers of high-order nonlinear convolutions (in conjunction with max pooling layers), the single-cycle waveform is compressed into a feature vector. :

[0063]

[0064] S4.2. Feed the output of the convolutional module into a 2-layer GRU network to extract the temporal dependencies and decay dynamics between the 10 historical loops. Then, process the feature vector sequence of the 10 consecutive loops within the window. Input a 2-level gated recurrent unit (GRU). At time step... GRU gating mechanism and hidden state Updates follow the following operators:

[0065]

[0066]

[0067]

[0068]

[0069] The above formulas are the basic formulas for the GRU gated cyclic unit model, where:

[0070] t represents the current time step index, corresponding to the t-th historical loop within the window; This represents the input vector at the current time step, i.e., the feature vector extracted by the convolutional module (CNN) for the t-th cycle; This represents the hidden state of the previous time step, carrying the timing information of the previous t-1 cycles; Indicates the hidden state after the current time step update; This indicates a candidate hidden state, which is equivalent to new memory content to be written at the current moment.

[0071] For the update gate, the value is between 0 and 1. The closer the value is to 1, the more likely it is to overwrite the old information with the new information. To reset the gate, the value is between 0 and 1, with the value closer to 0 indicating the current input. The more independent it is from its previous history.

[0072] , The input weight matrix corresponds to the update gate, reset gate, and candidate state calculation, respectively, and is related to the current input. The weights of the multiplication.

[0073] The cyclic weight matrix or hidden state weights correspond to the update gate, reset gate, and candidate state calculation, respectively, and are related to the hidden state at the previous time step. The weights of the multiplication.

[0074] The bias vector is used to add a translation amount to a linear transformation, helping the model fit the data.

[0075] σ represents the sigmoid activation function, which compresses the input value to the (0, 1) interval, making it suitable as the switching ratio of a gating signal.

[0076] tanh represents the hyperbolic tangent activation function, which compresses the input values ​​to the interval (-1, 1) to generate candidate states. It serves to stabilize the numerical range.

[0077] In step S5, the implementation steps of the dual prediction head are as follows:

[0078] S5.1. After the GRU module outputs the hidden state, two fully connected branches are generated.

[0079] S5.2, the SOH prediction head maps time-step features point-by-point, outputting a sequence of length 10:

[0080]

[0081] Where b represents the bias vector. This is the bias for SOH.

[0082] S5.3 The RUL prediction head uses global average pooling to flatten the time dimension into a one-dimensional vector, and then outputs a single lifetime scalar through a linear layer:

[0083]

[0084]

[0085] in, The hidden state output by the GRU at time step t is referred to as W, which represents the total length of the time window. This refers to the output vector of global average pooling. The weight matrix of a linear layer (i.e., a fully connected layer). The bias of the linear layer. This refers to the remaining useful life value ultimately predicted by the model.

[0086] Further, in step S6, the implementation details of the hierarchical parameter search include constructing a three-layer search space: a structure layer searches for the number of GRU hidden nodes (32, 64, or 128); a dynamics layer searches for the learning rate (1e-4 to 1e-2) and batch size (16 to 128); and a multi-task calibration layer searches for the weighting coefficient α (0.1 to 0.9). The total loss function is defined as L = α * L_SOH + (1-α) * L_RUL. Automatic parameter optimization iteration is performed using the Tree Structure Parzen Estimator (TPE) algorithm. The optimization objective of the joint training is to minimize the total loss. The mean squared error loss is tracked by the SOH window. RUL lifetime endpoint variance loss Weighted composition:

[0087]

[0088]

[0089]

[0090] Where y represents the true value. This represents the predicted value.

[0091] Using the TPE algorithm on the validation set To guide closed-loop optimization, the optimal hyperparameter combination is output for actual deployment.

[0092] In step S7, the criteria for model convergence and deployment are that the total loss curve on the validation set tends to stabilize, the best model is saved as the final joint prediction model, and the performance is quantified by calculating the RMSE and sMAPE metrics on the test set.

[0093] In step S7, the performance evaluation metrics include the root mean square error (RMSE) of SOH on the test set and the symmetric mean absolute percentage error (sMAPE) of RUL, as well as the mean absolute percentage error (MAPE), the quantification formula of which is:

[0094]

[0095]

[0096]

[0097] in, The total number of samples in the test set. and The first The predicted value and the true label of each sample; The expression represents a very small positive number to ensure that the denominator is not zero.

[0098] To verify the advancement and joint prediction performance of the method proposed in this invention, experiments are conducted below.

[0099] In this experiment, the dataset was divided using a stratified sampling strategy, with the specific proportions being 70% for the training set, 15% for the validation set, and 15% for the test set. In the battery health prediction task, a closed-loop comparison experiment was conducted between the core network model of this invention and classic temporal baseline networks in the field—Long Short-Term Memory (LSTM) and Temporal Convolutional Network (TCN).

[0100] The quantitative error indices of the experimental test results are shown in Table 1.

[0101] Table 1: Comparison of prediction performance of different models for health status (SOH) and remaining useful life (RUL) at MIT-Stanford

[0102]

[0103] Based on the experimental results in Table 1, the dual-task prediction method based on hierarchical parameter search and deep neural networks proposed in this invention achieves comprehensive best performance in predicting both the state of health (SOH) and remaining service life (RUL) of lithium-ion batteries. Among all the comparison models, including Long Short-Term Memory (LSTM) networks and Temporal Convolutional Networks (TCN), this invention has the lowest error rates across all four metrics, fully validating the effectiveness and robustness of this method in capturing the dynamic characteristics of battery capacity decay.

[0104] In terms of health status prediction, the root mean square error (RMSE) of the SOH of this invention is only 0.98%, and the mean absolute percentage error (MAPE) is 0.83%. Both of these indicators are better than TCN's 1.09% and 0.91%, respectively. Compared with the better-performing TCN model, the error is reduced by about 10% and 9%, respectively. This indicates that the method can fit the nonlinear changes of battery capacity degradation with extremely high accuracy, providing a highly reliable quantitative basis for health status assessment.

[0105] For remaining battery life prediction, the RUL root mean square error of this invention is 63.96 cycles, and the symmetric mean absolute percentage error (sMAPE) is 7.11%, representing reductions of over 71% and 67% respectively compared to LSTM's 225.37 cycles and 21.56% error. Even compared to TCN, the errors are reduced by approximately 37% and 21%. This significant improvement is attributed to the hierarchical parameter search strategy's effective optimization of network depth, learning rate, and gating mechanism, enabling the model to stably extract long-range temporal dependencies during battery aging while avoiding overfitting and local optima. In summary, this invention not only achieves high-precision prediction on dual tasks but also demonstrates excellent generalization ability and promising engineering applications on complex operating condition datasets.

[0106] In summary, this invention discloses a method for predicting the State of Health (SOH) and Rullow Usage Limit (RUL) of lithium-ion batteries based on hierarchical parameter search, belonging to the interdisciplinary field of battery health management and artificial intelligence. The method first collects characteristic data such as voltage, current, and capacity during the charge-discharge cycle of the lithium battery, generating a three-dimensional input tensor through interpolation alignment and a sliding window mechanism. Second, it constructs a CNN-GRU-based backbone neural network, utilizing the CNN module to deeply extract local degradation details of the battery's single-cycle waveform, and combining this with the GRU module to model the long-term degradation evolution law across cycles. Subsequently, it constructs independent and parallel SOH and RUL prediction heads at the network's end, employing different aggregation strategies to decouple multi-task features for different time-series characteristics. Finally, it constructs a three-level hierarchical parameter search space containing a structural layer, a dynamic layer, and a multi-task calibration layer, using optimization algorithms to automatically search for the optimal hyperparameters and multi-task loss function weights of the network, thereby configuring and training the model in a closed loop. This invention effectively balances microscopic physical characteristics and macroscopic degradation trends, avoids information loss caused by forcibly unifying the output dimension, and solves the problem of manual trial and error parameter tuning through automated optimization, significantly improving the accuracy, generalization ability and development efficiency of dual-task collaborative prediction.

[0107] The beneficial effects of this invention can be summarized as follows:

[0108] (1) This invention uses a 7-layer CNN as a convolutional module to deeply mine the local degradation details of a single-cycle waveform, and combines it with a 2-layer GRU as a recursive module to capture long-term capacity decay memory. This combination effectively takes into account both microscopic physical characteristics and macroscopic degradation trends.

[0109] (2) The present invention decouples multi-task output with dual prediction heads. In view of the different temporal characteristics of SOH and RUL, a many-to-many SOH tracking head and a global pooling-based RUL trend head are designed respectively, which avoids information loss caused by forcibly unifying the output dimension and significantly improves the collaborative prediction performance of dual tasks.

[0110] (3) This invention solves the parameter tuning problem through hierarchical parameter search. It incorporates network nodes, learning dynamics, and multi-task loss weights into a unified three-level hierarchical parameter search framework. Through automated optimization, it effectively solves the problem of uneven gradient distribution in multi-task learning, eliminates inefficient manual trial and error, and improves the generalization ability and development efficiency of the model.

[0111] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope of the present invention.

Claims

1. A method for predicting SOH and RUL of lithium-ion batteries based on hierarchical parameter search, characterized in that, Including the following steps: S1. Collect characteristic data of lithium-ion batteries during charge-discharge cycles; S2. The battery feature data is preprocessed and tensor constructed sequentially through interpolation alignment and sliding window mechanism to generate a three-dimensional input tensor containing the time step within the cycle and the historical dimension across cycles. S3. Divide the data according to the preset ratio to obtain the training set, validation set and test set; S4. Construct a backbone neural network model based on CNN-GRU, wherein the CNN module consists of 7 layers of convolutional structure to process the local degradation features of the single-cycle waveform of the battery; the GRU module consists of 2 layers of recurrent network to model the long-term degradation evolution law across cycles. S5. Construct independent and parallel SOH prediction heads and RUL prediction heads to decouple the output of multi-task features; S6. Construct a three-level hierarchical parameter search space containing a structural layer, a dynamics layer, and a multi-task calibration layer, and use an optimization algorithm to automatically search for the optimal hyperparameters and weights of the multi-task loss function of the network. S7. Configure the network using the optimal hyperparameter combination obtained from the search, perform closed-loop training on the network using the training set, and stop training when the performance indicators on the validation set tend to stabilize to obtain the final joint prediction model; perform prediction on the test set to obtain the prediction results of SOH and RUL of lithium-ion batteries.

2. The method for predicting SOH and RUL of lithium-ion batteries based on hierarchical parameter search according to claim 1, characterized in that, In step S1, the battery characteristic data collected specifically includes voltage, current, and capacity data during the discharge phase of the complete life cycle aging data. The actual SOH tag and global RUL tag are calculated by reading and recording the first cycle capacity.

3. The method for predicting SOH and RUL of lithium-ion batteries based on hierarchical parameter search according to claim 1, characterized in that, In step S2, the generated three-dimensional input tensor is in the format of feature channel number x time step within period x historical dimension across periods.

4. The method for predicting SOH and RUL of lithium-ion batteries based on hierarchical parameter search according to claim 1, characterized in that, In step S3, the dataset is divided into a training set of 70%, a validation set of 15%, and a test set of 15%, while maintaining the same data distribution consistency.

5. The method for predicting SOH and RUL of lithium-ion batteries based on hierarchical parameter search according to claim 1, characterized in that, In step S4, the first three layers of the CNN module are used for rapid dimensionality reduction and feature smoothing, while the last four layers work with max pooling layers to extract high-order nonlinear degenerate features.

6. The method for predicting SOH and RUL of lithium-ion batteries based on hierarchical parameter search according to claim 1, characterized in that, In step S5, the SOH prediction head specifically maps the time step features point by point and outputs an SOH sequence of a set length; the RUL prediction head specifically flattens the time dimension into a one-dimensional vector and then outputs a single lifetime scalar through a linear layer.

7. The method for predicting SOH and RUL of lithium-ion batteries based on hierarchical parameter search according to claim 1, characterized in that, In step S6, the parameters of the hierarchical parameter search space include: the structural layer covers the number of GRU hidden layer nodes and the Dropout ratio; the dynamics layer covers the base learning rate and batch size; and the multi-task calibration layer covers the dynamic weighting coefficient between the SOH prediction loss and the RUL prediction loss.

8. The method for predicting SOH and RUL of lithium-ion batteries based on hierarchical parameter search according to claim 7, characterized in that, Step S6, the specific process of optimizing the algorithm is as follows: S6.1 Construct a three-level hierarchical parameter search space and clarify the parameter boundaries of the structural layer, dynamics layer and multi-task calibration layer; S6.2 Utilize the tree-structured Parzen estimator algorithm for automatic parameter optimization iteration; S6.3 Calculate the total loss function on the validation set to evaluate the performance of each parameter combination, record and output the best-performing hyperparameter combination for configuring the final CNN-GRU-based network model.

9. A method for predicting SOH and RUL of lithium-ion batteries based on hierarchical parameter search according to claim 7 or 8, characterized in that, In step S6, the search parameters for the dynamics layer are as follows: the basic learning rate ranges from 1e-4 to 1e-2, and the batch size ranges from 16 to 128.

10. A method for predicting SOH and RUL of lithium-ion batteries based on hierarchical parameter search according to claim 8, characterized in that, In step S6, the weighting coefficient α of the search parameters for the multi-task calibration layer ranges from 0.1 to 0.9, and the total loss function is defined as: in, , The meanings represent the losses of SOH and RUL, respectively.

Images