Lithium battery residual life prediction method based on multi-frequency decomposition and hybrid neural network

Abstract

The application provides a lithium battery residual life prediction method based on multi-frequency decomposition and a hybrid neural network, and relates to the technical field of lithium battery residual life prediction. The lithium battery residual life prediction method effectively captures local fluctuations and long-term trends in the battery degradation process by introducing a component decomposition mechanism and designing appropriate prediction models for different frequency components, significantly improving the accuracy and robustness of residual life prediction. Different prediction models are designed for different components according to their complexity differences, avoiding redundant calculations caused by using complex models on all components, and achieving a balance between prediction accuracy and computational efficiency. The super parameter optimization strategy based on hierarchical reinforcement learning can automatically and efficiently find the optimal parameter combination according to different battery data characteristics, overcoming the shortcomings of high computational cost and poor adaptability of traditional optimization methods.

Description

Technical Field

[0001] This invention relates to the field of lithium battery remaining life prediction technology, specifically to a lithium battery remaining life prediction method based on multi-frequency decomposition and hybrid neural networks. Background Technology

[0002] As a core component of modern energy storage systems, the health status and remaining lifespan of lithium-ion batteries directly impact the operational safety, maintenance costs, and economics of electronic devices, electric vehicles, and renewable energy storage systems. Since battery performance inevitably degrades during cycling, accurately predicting the remaining lifespan of batteries is crucial for achieving condition-based maintenance, optimizing energy management strategies, and improving system reliability.

[0003] Capacity regeneration is a significant challenge during battery degradation. This phenomenon manifests as a brief, localized increase in capacity within an overall degradation trend, primarily due to dynamic changes in the solid electrolyte interfacial film and the redistribution of lithium-ion concentration. This regeneration can mask the true degradation trend of the battery, thus affecting the judgment and accuracy of predictive models.

[0004] Currently, most lithium-ion battery life prediction methods considering capacity regeneration employ frequency domain signal processing techniques for data decomposition and make predictions based on the decomposition results. However, these methods typically focus only on high-frequency and low-frequency components, neglecting the characteristics of mid-frequency components. Existing prediction models lack sufficient design optimization when predicting different frequency components, leading to insufficient accuracy. Furthermore, existing models often conflict between prediction accuracy and computational efficiency; optimizing one often comes at the expense of the other. Regarding hyperparameter optimization, traditional Bayesian optimization methods are computationally time-consuming and resource-intensive. Therefore, developing a comprehensive data decomposition method, combined with a targeted prediction model and supporting staged hyperparameter adjustment, will significantly improve the accuracy of lithium-ion battery life prediction, reduce prediction time and cost, and achieve a balance between prediction accuracy and computational efficiency. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a lithium battery remaining life prediction method based on multi-frequency decomposition and hybrid neural networks, which solves the problems of low accuracy and computational efficiency in lithium battery life prediction.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for predicting the remaining life of a lithium battery based on multi-frequency decomposition and hybrid neural networks, the method comprising:

[0008] S1. Data Collection

[0009] Historical capacity data of lithium-ion batteries during multiple charge-discharge cycles are collected. Each data sample includes the number of cycles and its corresponding discharge capacity value. The data is organized and managed in time series format to form a complete battery life dataset.

[0010] S2. Data Preprocessing

[0011] Original capacity sequence The standardized capacity sequence obtained after standardization ;

[0012] S3. Data Decomposition and Component Classification

[0013] For normalized capacity sequences Decompose the sample and calculate the sample entropy and Clustering is used to classify components into three categories: high-frequency, mid-frequency, and low-frequency. The high-frequency components are then further classified. Secondary decomposition;

[0014] S4. Multi-model collaborative prediction

[0015] The corresponding prediction models are used to predict the three components respectively;

[0016] S5. Hyperparameter Optimization in Hierarchical Reinforcement Learning

[0017] All decomposition and prediction parameters used in the prediction process are optimized by adopting a hierarchical reinforcement learning strategy to optimize parameters in stages and at different levels.

[0018] S6. Final Prediction

[0019] Using the optimal combination of hyperparameters obtained through hierarchical reinforcement learning, the entire process from S3 to S4 is re-executed to generate the final, accurate prediction of remaining battery life.

[0020] S7. Adaptive Update

[0021] Whether the model needs to be retrained depends on the current state of the battery.

[0022] Preferably, in S1, the complete battery life dataset is represented as follows:

[0023]

[0024] in, For the first One cycle;

[0025] For the first The battery discharge capacity measured over one cycle;

[0026] This represents the total number of cycles.

[0027] Preferably, in S2, the normalized capacity sequence for:

[0028]

[0029] Among them, the Standardized capacity value per cycle for:

[0030]

[0031]

[0032]

[0033] in, For the first The original capacity value for each cycle;

[0034] The mean of the original capacity sequence;

[0035] denoted as the standard deviation of the original capacity sequence.

[0036] Preferably, S3 specifically includes:

[0037] S3.1. Primary decomposition

[0038] The complete set empirical mode decomposition method is used to preprocess the standardized capacity sequence. Decomposed into a series of intrinsic mode functions and residual terms:

[0039]

[0040] in, For the first One intrinsic mode function;

[0041] This is the final residual term;

[0042] S3.2. Component Classification

[0043] Calculate the sample entropy of all intrinsic mode function components obtained in the above steps, and then use... Clustering algorithm, preset number of clusters All intrinsic mode functions are classified into three categories: high frequency, mid frequency, and low frequency.

[0044] S3.3. Secondary decomposition

[0045] For the high-frequency components obtained from the above clustering, further processing is performed... The method performs secondary noise reduction.

[0046] Preferably, in S4, the high-frequency component prediction model is input after... The decomposed sub-components are used to capture the dependencies between time series data through a dual feature extraction encoder. After feature fusion, the data is input into the decoder and finally mapped through flattening layers, fully connected layers, etc., to output the predicted value.

[0047] The intermediate frequency component prediction model includes: a simplified input mapping layer, position encoding, and a 1-layer... The encoder, the linear transformation layer for feature mapping, and the output layer;

[0048] The low-frequency component prediction model employs an attention mechanism. Model;

[0049] The prediction results of the three models for high, medium, and low frequency components are superimposed to obtain the final battery capacity prediction sequence:

[0050]

[0051] Based on this fused capacity prediction sequence, the remaining lifespan of the battery is calculated according to the preset failure threshold.

[0052] Preferably, in the high-frequency component prediction model, the feature extraction encoder includes: a sensor encoder and a time-step encoder;

[0053] The sensor encoder is composed of a multi-head parallel computation mechanism and a feedforward neural network, combined with layer normalization and residual connections.

[0054] The formula for calculating the single-head attention score in the self-attention mechanism is as follows:

[0055]

[0056] in, For query matrix;

[0057] The key matrix;

[0058] It is a value matrix;

[0059] The dimension of the key is used for scaling to avoid gradient anomalies;

[0060] maskTo block invalid locations as needed;

[0061] T This indicates matrix transpose;

[0062] The results of multiple single-head attention scores are concatenated and then transformed linearly to generate the final attention features;

[0063] The specific formula for a feedforward neural network is:

[0064]

[0065] in, A This is the output of the attention mechanism;

[0066] This is the weight matrix for the first layer of linear transformation;

[0067] This is the bias for the first-level linear transformation;

[0068] GELU is the activation function for Gaussian error linear units, and its formula is:

[0069]

[0070] in, It is the cumulative distribution function of the standard normal distribution;

[0071] This is the weight matrix for the second-level linear transformation;

[0072] This is the bias for the second-level linear transformation;

[0073] Layer normalization is applied to the attention output and the feedforward network output, using the following formula:

[0074]

[0075] in, The mean;

[0076] Standard deviation;

[0077] It is the minimum value;

[0078] Learnable scaling factor

[0079] As a bias factor, it can stabilize training and accelerate convergence;

[0080] The logic of time-step encoders and sensor encoders is similar. The main difference is that time-step encoders add position encoding before multi-head attention. The specific formula is as follows:

[0081]

[0082]

[0083] in, For location index;

[0084] For dimension indexing;

[0085] For model dimensions;

[0086] The data obtained by fusing the features from the two encoders is input into the decoder, which generates a hidden representation; the final output of the decoder at the last time step is:

[0087]

[0088] in, This is the final output of the entire high-frequency prediction model decoder module;

[0089] L The length of the input sequence;

[0090] It is the domain;

[0091] The function of the flattening layer is to flatten this two-dimensional tensor It is compressed into a one-dimensional vector to connect standard fully connected layers, and its formula is:

[0092]

[0093] in, Flatten To be The tensor is reshaped into a length of ;

[0094] The fully connected layer maps high-dimensional features to specific predicted values, and the formula is as follows:

[0095]

[0096] in, This represents the final predicted value for the high-frequency components;

[0097] This is the weight matrix of the fully connected layer;

[0098] This is the bias vector of the fully connected layer;

[0099] Represents matrix multiplication;

[0100] The calculation formula for the intermediate frequency component prediction model is as follows:

[0101]

[0102]

[0103]

[0104]

[0105] in, The input sequence is...

[0106] The tensor is mapped and then its dimensions are adjusted to fit. Model;

[0107] Weights of the linear layer

[0108] For the bias of the linear layer;

[0109] Used for each position in time series data The positional coding is generated using a simplified periodic positional coding form, which assigns different periodic coding features to even-numbered positions using a sine function;

[0110] It is through including Layer encoder The intermediate feature representation obtained after processing; The encoder adds position-encoded input sequences Encode the sequence, extract its feature information, and output the result. It includes features processed by mechanisms such as self-attention, which are used for subsequent decoding and prediction;

[0111] This refers to the predicted value of the mid-frequency prediction model, passed through the weights of the fully connected layer. and Features of encoder output Calculated;

[0112] The calculation formula for the low-frequency component prediction model is as follows:

[0113]

[0114]

[0115]

[0116]

[0117] in, It means The layer in the first The hidden state of each cycle contains information up to the current moment. The layer stores information about the input sequence for subsequent computation and context transmission.

[0118] The low-frequency component time series is in the first... Input samples for each cycle;

[0119] yes Layer at time The hidden state is the memory information from the previous moment;

[0120] It is the attention weight, used to measure the attention weight. Hidden state of each cycle The importance of the overall characteristics, through After normalization, the values ​​range from 0 to 1, and the sum of the attention weights at all times is 1;

[0121] This refers to the context vector, which is processed through attention weights. Hidden state at each time point The weighted summation yields an integrated set of important information about the hidden states at different times, highlighting the characteristics of key time points;

[0122] It is the prediction output of the low-frequency prediction model;

[0123] Used for context vectors Perform a linear transformation;

[0124] It is the rectified linear unit activation function, for Perform a nonlinear transformation, setting values ​​less than 0 to 0 and keeping values ​​greater than 0 unchanged, thereby increasing the model's nonlinear expressive power.

[0125] It is a regularization operation;

[0126] Used for passing Activation function calculation and The data after regularization is then linearly transformed again to obtain the final predicted value.

[0127] Preferably, S5 specifically includes:

[0128] S5.1. State Awareness

[0129] The system obtains the current status;

[0130] S5.2. High-level decision-making

[0131] The PPO agent selects a high-level action to execute based on its policy network according to its current state.

[0132] S5.3. Low-level execution

[0133] Based on the selected high-level action, execute the corresponding low-level action and adjust the relevant hyperparameters accordingly.

[0134] S5.4. Performance Evaluation and Reward Calculation

[0135] Run the prediction framework using the new hyperparameters to calculate the reward. The reward function is calculated using the following formula:

[0136]

[0137] in, For the current round compared to the baseline Improvement amount;

[0138] for The historical standard deviation;

[0139] This represents the average accuracy of the predicted direction.

[0140] This is a stability penalty term, which measures the imbalance in the prediction performance of each component;

[0141] S5.5. State Update and Strategy Learning

[0142] New status and rewards R Feedback is given to the PPO agent to update its policy network and value network;

[0143] This process is repeated a preset number of times, and the final output is the hyperparameter combination that performs best in the optimization process.

[0144] Preferably, S7 specifically includes:

[0145] S7.1. Triggering Mechanism and Data Collection

[0146] The system presets a trigger interval, which can be based on time or a cycle. Once the trigger condition is met, the module automatically starts the update process.

[0147] The system collects the latest battery operation data from the battery management system since the last update, mainly including the latest cycle capacity measurement value and its corresponding cycle number;

[0148] S7.2. Data Buffering and Sequence Reconstruction

[0149] The system maintains a fixed-length first-in-first-out (FIFO) data buffer to store historical data for a recent period of time.

[0150] Newly collected data is added to the buffer while the oldest data points are removed to keep the total length of the buffer constant; thus, an updated time series reflecting the latest state of the battery is formed.

[0151] S7.3. Update Strategy Decisions

[0152] Based on the significance of data changes, the system intelligently selects between full-process updates or rapid fine-tuning to balance computational costs and prediction accuracy.

[0153] S7.3.1. Change Detection

[0154] Calculate the statistical difference between the new data and recent historical data; calculate a mean dissimilarity based on a sliding window, using the following formula:

[0155]

[0156] in, The average capacity of the newly added data points;

[0157] To update the average capacity of windows of the same length in the front buffer;

[0158] To update the standard deviation of the contents of windows of the same length in the front buffer;

[0159] S7.3.2. Strategy Selection

[0160] Two thresholds are set, and the threshold for determining no significant change is: The threshold for determining a significant change is: ;

[0161] When the battery degradation status is determined to have changed significantly, the system triggers a full-process update mode.

[0162] When this is determined to be a normal degradation fluctuation, the system triggers the rapid fine-tuning mode;

[0163] If the status does not change significantly, no update will be made this time, and the system will wait for the next trigger cycle.

[0164] The full-process update mode starts from S1 and executes all steps completely.

[0165] The rapid fine-tuning mode, in order to save computational resources, only performs online fine-tuning on the already trained prediction model; specifically, it includes:

[0166] Fine-tuning data

[0167] Use all or part of the latest data in the data buffer as the fine-tuning dataset;

[0168] Fine-tuning method

[0169] A small learning rate is used to update the weights of the three prediction models with a small number of iterations of gradient descent; the loss function is typically the mean squared error.

[0170]

[0171] in, To fine-tune the loss value;

[0172] For each sample in the fine-tuning dataset;

[0173] This represents the total number of data samples for fine-tuning.

[0174] This is the actual capacity;

[0175] For predicted capacity;

[0176] The gradient descent algorithm is used during fine-tuning to continuously and incrementally adjust the internal parameters of the three prediction models in order to minimize the value of the loss function.

[0177] Dynamic learning rate

[0178] The learning rate decays dynamically with each fine-tuning epoch, and the calculation formula is as follows:

[0179]

[0180] in, For fine-tuning the rounds;

[0181] To initially fine-tune the learning rate;

[0182] This represents the attenuation rate, with a value greater than 0 and less than 1.

[0183] S7.4. Prediction Result Output and Model Version Management

[0184] Output results: Regardless of the update mode used, the system will eventually output an updated prediction of remaining useful life based on the latest data and model.

[0185] Version management: The system archives the model parameters, hyperparameters, and key performance indicators after this update and adds a timestamp; this helps track the evolution of model performance and allows for a quick rollback to a previous stable version if problems occur during the update.

[0186] A lithium battery remaining life prediction system based on multi-frequency decomposition and hybrid neural network, the lithium battery remaining life prediction system includes: a data collection module, a data preprocessing module, a data decomposition and component classification module, a multi-model collaborative prediction module, a hyperparameter optimization module, a final prediction module and an adaptive update module;

[0187] The data collection module is used to collect historical operating data of lithium-ion batteries, including capacity data and the corresponding number of cycles, and store them in a certain format.

[0188] The data preprocessing module is used to standardize the collected lithium battery data in preparation for subsequent input into the prediction model.

[0189] The data decomposition and component classification module is used to decompose the capacity data of lithium batteries and calculate the sample entropy and component classification. Clustering is used to classify components, and high-frequency components are then... Secondary decomposition;

[0190] The multi-model collaborative prediction module is used to predict the three components obtained by decomposition using the corresponding prediction models.

[0191] The hyperparameter optimization module is used to optimize all decomposition parameters and prediction parameters used in the prediction process. It adopts a hierarchical reinforcement learning strategy to optimize parameters in stages and at different levels.

[0192] The final prediction module is used to predict the final result based on the optimal parameters obtained by hyperparameter optimization.

[0193] The adaptive update module is used to determine whether the model needs to be retrained based on the current state of the battery.

[0194] This invention provides a method for predicting the remaining life of lithium batteries based on multi-frequency decomposition and hybrid neural networks. Compared with existing technologies, it has the following advantages:

[0195] In this invention, the lithium battery remaining life prediction method introduces a component decomposition mechanism and tailors prediction models for different frequency components, effectively capturing local fluctuations and long-term trends in the battery degradation process, significantly improving the accuracy and robustness of remaining life prediction. Appropriate prediction models are designed for the complexity differences of different components, avoiding redundant calculations caused by using complex models uniformly on all components, thus achieving a balance between prediction accuracy and computational efficiency. The hyperparameter optimization strategy based on hierarchical reinforcement learning can automatically and efficiently find the optimal parameter combination according to different battery data characteristics, overcoming the shortcomings of traditional optimization methods such as high computational cost and poor adaptability. Attached Figure Description

[0196] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.

[0197] Figure 1 This is a flowchart of the lithium battery remaining life prediction method in an embodiment of the present invention.

[0198] Figure 2 This is a flowchart of hierarchical reinforcement learning in an embodiment of the present invention.

[0199] Figure 3 This is a schematic diagram of the lithium battery remaining life prediction system in an embodiment of the present invention. Detailed Implementation

[0200] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0201] This application provides a lithium battery remaining life prediction method based on multi-frequency decomposition and hybrid neural networks, which solves the problems of low accuracy and computational efficiency in lithium battery life prediction.

[0202] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0203] Example:

[0204] like Figure 1 ,Figure 2 As shown, this invention provides a method for predicting the remaining life of a lithium battery based on multi-frequency decomposition and hybrid neural networks. The method for predicting the remaining life of a lithium battery includes:

[0205] S1. Data Collection

[0206] Collect historical capacity data of lithium-ion batteries during multiple charge-discharge cycles. Each data sample includes the number of cycles and its corresponding discharge capacity value.

[0207] The data is organized and managed in a time-series format. A complete battery life dataset is represented as follows:

[0208]

[0209] in, For the first One cycle;

[0210] For the first The battery discharge capacity measured over one cycle;

[0211] This represents the total number of cycles.

[0212] S2. Data Preprocessing

[0213] Original capacity sequence through The standardized capacity sequence obtained after standardization for:

[0214]

[0215] Among them, the Standardized capacity value per cycle for:

[0216]

[0217]

[0218]

[0219] in, For the first The original capacity value for each cycle;

[0220] The mean of the original capacity sequence;

[0221] denoted as the standard deviation of the original capacity sequence.

[0222] S3. Data Decomposition and Component Classification

[0223] S3.1. Primary decomposition

[0224] The complete set empirical mode decomposition method is used to preprocess the standardized capacity sequence. Decompose into a series of intrinsic mode functions and residual terms; the specific steps are: to Adaptive white Gaussian noise is injected multiple times and then subjected to empirical mode decomposition followed by ensemble averaging to suppress mode aliasing, thereby obtaining multiple intrinsic mode function components distributed from high frequency to low frequency:

[0225]

[0226] in, For the first One intrinsic mode function;

[0227] This is the final residual term, representing the long-term trend in the signal;

[0228] The first intrinsic mode function (IMF1) and the first residual The calculation formula is:

[0229]

[0230]

[0231] for , No. p individual residuals and the The calculation formula is:

[0232]

[0233]

[0234] in, This indicates the extraction of the first intrinsic mode function from the input signal;

[0235] Indicates the first Noise extraction One intrinsic mode function;

[0236] Indicates the first The Gaussian white noise that conforms to a standard normal distribution is added next;

[0237] Indicates the first The weighting coefficients of the white noise added in the next iteration;

[0238] This represents the overall average frequency.

[0239] S3.2. Component Classification

[0240] Calculate the sample entropy of all intrinsic mode function components obtained in the above steps, and then use... Clustering algorithm, preset number of clusters All intrinsic mode functions are classified into three categories: high frequency, medium frequency, and low frequency.

[0241] S3.2.1. Sample Entropy Calculation

[0242] For time series , build m dimensional vector, n The length of the time series is calculated using the following formula:

[0243]

[0244] Distance between two vectors:

[0245]

[0246] To calculate the number of similar vectors, define... For all Under the condition of satisfying distance The proportion of the number of vectors to the total number of possibilities; The formula for calculating the set similarity threshold is:

[0247]

[0248] Based on each Calculate the overall average value. The formula is:

[0249]

[0250] Increase the dimension to Repeat the above steps to obtain .

[0251] Finally, the sample entropy is:

[0252]

[0253] The larger the sample entropy value, the more complex and irregular the sequence is, corresponding to higher frequency components.

[0254] S3.2.2. Clustering

[0255] Based on the calculated entropy value, The method clusters the component data to obtain high-frequency, mid-frequency, and low-frequency data.

[0256] The preset number of clusters is the preset number of clusters. , For the first A set of clusters, The feature vector corresponding to each component, For the first The cluster centers of each cluster are determined, and based on this, the objective function is:

[0257]

[0258] Cluster centers are the mean of all data points, calculated using the following formula:

[0259]

[0260] in, It is a cluster The number of data points in the data.

[0261] Each data point is assigned to the cluster containing the nearest cluster center, using the following formula:

[0262]

[0263] Ultimately, this method is used to identify high-frequency components, mid-frequency components, and low-frequency components.

[0264] S3.3. Secondary decomposition

[0265] For the high-frequency components obtained from the above clustering, further processing is performed... The method performs secondary denoising to separate the real electrochemical fluctuations from the residual noise;

[0266] The original signal to be decomposed is The decomposed modal components are , k Here are the modal numbers, and the center frequencies of each mode are... ;

[0267] Construct a constrained variational problem; the objective function is the sum of the squares of the frequency domain norms of each mode after processing, and its mathematical expression is:

[0268]

[0269] in, It is the impulse function;

[0270] The imaginary unit;

[0271] Indicates time Find the partial derivative;

[0272] It is a convolution operation;

[0273] for Norm.

[0274] The constraint condition is that the sum of the modal components equals the original signal.

[0275]

[0276] To solve the constrained variational problem, Lagrange multipliers are introduced. and penalty factor α Constructing the augmented Lagrange function l :

[0277]

[0278] in, Indicates inner product operation;

[0279] Iterative updates are performed using the alternating direction multiplier method. Continue until the convergence condition is met;

[0280] renew , for Fourier transform, yes Fourier transform, yes Fourier transform; updating in the frequency domain Fourier coefficients :

[0281]

[0282] Based on the updated Calculate the center frequency Get the updated value :

[0283]

[0284] Update Lagrange multipliers Fourier coefficients The formula is:

[0285]

[0286] in, It is the iteration step size;

[0287] Repeat the above update steps until the changes in each variable in two consecutive iterations are less than the threshold, which satisfies the convergence criterion. At this point, the result is... The decomposed modal components are the complete modal components. Signal decomposition process.

[0288] S4. Multi-model collaborative prediction

[0289] High-frequency component prediction model

[0290] The high-frequency component prediction model is an improvement. The model, whose input is processed The decomposed sub-components are captured by a dual feature extraction encoder to capture the dependencies between time series data. After feature fusion, the data is input into the decoder to focus on the prediction target. Finally, the data is mapped through flattening layers, fully connected layers, etc., to output the predicted value.

[0291] The feature extraction encoder consists of a sensor encoder and a time-step encoder. The sensor encoder captures the dependencies between feature dimensions through multi-head parallel computation. It mainly consists of a multi-head attention mechanism and a feedforward neural network, combined with layer normalization and residual connections to improve the stability of the model.

[0292] The self-attention mechanism, by calculating the weights at each time step in the sequence, enables the model to focus on the historical information most crucial for prediction. The formula for calculating the single-head attention score is:

[0293]

[0294] in, For query matrix;

[0295] The key matrix;

[0296] It is a value matrix;

[0297] The dimension of the key is used for scaling to avoid gradient anomalies;

[0298] mask To mask invalid positions as needed, such as time series padding bits;

[0299] T This indicates that a matrix is ​​being transposed.

[0300] The results of multiple single-head attention scores are concatenated and then transformed linearly to generate the final attention features.

[0301] Feedforward neural networks are used to enhance nonlinear feature representation and uncover deeper patterns. The specific formula is as follows:

[0302]

[0303] in, This is the output of the attention mechanism;

[0304] This is the weight matrix for the first layer of linear transformation;

[0305] This is the bias for the first-level linear transformation;

[0306] The activation function for the Gaussian error linear unit is given by the following formula:

[0307]

[0308] in, It is the cumulative distribution function of the standard normal distribution;

[0309] This is the weight matrix for the second-level linear transformation;

[0310] This is the bias for the second-level linear transformation.

[0311] Layer normalization is applied to the attention output and the feedforward network output, using the following formula:

[0312]

[0313] in, The mean;

[0314] Standard deviation;

[0315] It is the minimum value;

[0316] Learnable scaling factor

[0317] As an offset factor, it can stabilize training and accelerate convergence.

[0318] Time-step encoders and sensor encoders are similar in logic, the main difference being that time-step encoders add position encoding before multi-head attention; because It does not inherently possess time-series awareness capabilities. Instead, it generates position codes using sine and cosine functions, incorporating time sequence information into the input sequence. The specific formula is as follows:

[0319]

[0320]

[0321] in, For location index;

[0322] For dimension indexing;

[0323] For model dimensions;

[0324] The data resulting from the feature fusion of the outputs of the two encoders is input into the decoder, which (also composed of multi-head attention, feedforward network, and other layers) generates a hidden representation; the final output of the decoder at the last time step is:

[0325]

[0326] in, This is the final output of the entire high-frequency prediction model decoder module;

[0327] The length of the input sequence;

[0328] It is the domain;

[0329] The function of the flattening layer is to flatten this two-dimensional tensor It is compressed into a one-dimensional vector to connect standard fully connected layers, and its formula is:

[0330]

[0331] in, To be The tensor is reshaped into a length of ;

[0332] The fully connected layer maps high-dimensional features to specific predicted values, and the formula is as follows:

[0333]

[0334] in, This represents the final predicted value for the high-frequency components;

[0335] This is the weight matrix of the fully connected layer;

[0336] This is the bias vector of the fully connected layer;

[0337] This represents matrix multiplication.

[0338] Intermediate frequency component prediction model

[0339] For intermediate frequency components, since their complexity is lower compared to high frequency components, a simplified approach is possible. The model is sufficient to capture its main operating mode, and it mainly consists of a simplified input mapping layer, positional encoding, and a 1-layer... The encoder, the linear transformation layer for feature mapping, and the output layer; its calculation formula is:

[0340]

[0341]

[0342]

[0343]

[0344] in, The input sequence is...

[0345] The tensor is mapped and then its dimensions are adjusted to fit. Model;

[0346] Weights of the linear layer

[0347] This is the bias for the linear layer.

[0348] Used for each position in time series data Positional codes are generated to introduce temporal sequence information, enabling the model to perceive the relative positional relationships of elements in the sequence. Here, a simplified periodic positional coding form is adopted, using a sine function to assign different periodic coding features to even-numbered positions.

[0349] It is through including Layer encoder The intermediate feature representation obtained after processing; The encoder adds position-encoded input sequences Encode the sequence, extract its feature information, and output the result. It includes features processed by mechanisms such as self-attention, which are used for subsequent decoding and prediction;

[0350] This refers to the predicted value of the mid-frequency prediction model, passed through the weights of the fully connected layer. and Features of encoder output Calculated.

[0351] Low-frequency component prediction model

[0352] Low-frequency components represent long-term trends, while Long Short-Term Memory Networks (LSNs)... )compared to It is more effective at capturing long-term dependencies, but traditional The model is prone to missing key information in long sequences, so an attention mechanism was designed. Adding an attention mechanism to the model helps identify key time points, makes it easier to learn long-term trends, and improves predictive performance. Furthermore, it adds attention mechanisms between layers. Layers are used to prevent overfitting; the calculation formula is as follows:

[0353]

[0354]

[0355]

[0356]

[0357] in, It means The layer in the first The hidden state of each cycle contains information up to the current moment. The layer stores information about the input sequence for subsequent computation and context transmission.

[0358] The low-frequency component time series is in the first... Input samples for each cycle;

[0359] yes Layer at time The hidden state is the memory information from the previous moment;

[0360] It is the attention weight, used to measure the attention weight. Hidden state of each cycle The importance of the overall characteristics, through After normalization, the values ​​range from 0 to 1, and the sum of the attention weights at all times is 1;

[0361] This refers to the context vector, which is processed through attention weights. Hidden state at each time point The weighted summation yields an integrated set of important information about the hidden states at different times, highlighting the characteristics of key time points;

[0362] It is the prediction output of the low-frequency prediction model;

[0363] Used for context vectors Perform a linear transformation;

[0364] It is the rectified linear unit activation function, for Perform a nonlinear transformation, setting values ​​less than 0 to 0 and keeping values ​​greater than 0 unchanged, thereby increasing the model's nonlinear expressive power.

[0365] It is a regularization operation;

[0366] Used for passing Activation function calculation and The data after regularization is then linearly transformed again to obtain the final predicted value.

[0367] Results fusion

[0368] The prediction results of the three models for high, medium, and low frequency components are superimposed to obtain the final battery capacity prediction sequence:

[0369]

[0370] Based on this fused capacity prediction sequence, the remaining lifespan of the battery is calculated according to a preset failure threshold (such as 70% of the rated capacity).

[0371] S5. Hyperparameter Optimization in Hierarchical Reinforcement Learning

[0372] A hierarchical reinforcement learning hyperparameter optimization system consists of a high-level policy and multiple low-level actions.

[0373] Its state space includes performance metrics, prediction performance of each component, current hyperparameter state, historical optimization trends, and component complexity.

[0374] Basic performance indicators include , , Four-dimensional indicators, including direction accuracy.

[0375] Component performance includes high-frequency, mid-frequency, and low-frequency components. The five-dimensional indicators, including high-frequency and low-frequency stability, reflect the predictive performance and fluctuation of each component.

[0376] The parameter state includes the normalized values ​​of the decomposed parameters and the predicted parameters.

[0377] Historical trends include four dimensions: performance trend, stability, last improvement interval, and training progress.

[0378] Component complexity mainly measures the difficulty of predicting low-frequency components and is used to guide resource allocation for parameter optimization.

[0379] Its high-level operational space is The agent-driven mechanism is responsible for selecting the optimization direction from a macroscopic perspective. For example, in adjusting decomposition parameters, the intensity of stochastic exploration in the decomposition algorithm can be controlled by increasing or decreasing the number of hyperparameter search iterations. In adjusting prediction parameters, the sufficiency of the model's fit to the data can be adjusted by increasing or decreasing the number of training epochs to avoid overfitting or underfitting.

[0380] Its underlying action space executes specific hyperparameter adjustment instructions based on the selection of the higher-level strategy. For example, increasing or decreasing... Noise addition times, adjustment Adjust the number of training rounds or modify the learning rate, etc.

[0381] S5.1. State Awareness

[0382] The system obtains the current status;

[0383] S5.2. High-level decision-making

[0384] Based on its current state, the agent selects a high-level action to execute according to its policy network;

[0385] S5.3. Low-level execution

[0386] Based on the selected high-level action, execute the corresponding low-level action and adjust the relevant hyperparameters accordingly.

[0387] S5.4. Performance Evaluation and Reward Calculation

[0388] Run the prediction framework using the new hyperparameters to calculate the reward. The reward function is calculated using the following formula:

[0389]

[0390] in, For the current round compared to the baseline Improvement amount;

[0391] for The historical standard deviation;

[0392] This represents the average accuracy of the predicted direction.

[0393] This is a stability penalty term, which measures the imbalance in the prediction performance of each component;

[0394] S5.5. State Update and Strategy Learning

[0395] New status and rewards Feedback to Intelligent agents are used to update their policy network and value network;

[0396] This process is repeated a preset number of times, and the final output is the hyperparameter combination that performs best in the optimization process.

[0397] S6. Final Prediction

[0398] Using the optimal combination of hyperparameters obtained through hierarchical reinforcement learning, the entire process from data decomposition to result fusion is re-executed to generate the final, accurate prediction of remaining battery life.

[0399] S7. Adaptive Update

[0400] S7.1. Triggering Mechanism and Data Collection

[0401] The system presets a trigger interval, which can be based on time or a cycle. Once the trigger condition is met, the module automatically starts the update process.

[0402] The system collects the latest battery operation data from the battery management system since the last update, mainly including the latest cycle capacity measurement value and its corresponding cycle number;

[0403] S7.2. Data Buffering and Sequence Reconstruction

[0404] The system maintains a fixed-length first-in-first-out (FIFO) data buffer to store historical data for a recent period. The length of the buffer can be set according to the actual application scenario, for example, to store data from the last 200 cycles.

[0405] Newly collected data is added to the buffer while the oldest data points are removed to keep the total length of the buffer constant; thus, an updated time series reflecting the latest state of the battery is formed.

[0406] S7.3. Update Strategy Decisions

[0407] Based on the significance of data changes, the system intelligently selects between full-process updates or rapid fine-tuning to balance computational costs and prediction accuracy.

[0408] S7.3.1. Change Detection

[0409] Calculate the statistical difference between new data and recent historical data. For example, a mean dissimilarity based on a sliding window can be calculated using the following formula:

[0410]

[0411] in, The average capacity of several newly added data points (such as the last 5 cycles);

[0412] To update the average capacity of windows of the same length in the front buffer;

[0413] To update the standard deviation of the contents of windows of the same length in the front buffer;

[0414] S7.3.2. Strategy Selection

[0415] Two thresholds are set, and the threshold for determining no significant change is: The threshold for determining a significant change is: ;

[0416] When a significant change in battery degradation status is detected (such as a sudden drop in capacity or regeneration), the system triggers a full-process update mode.

[0417] When this is determined to be a normal degradation fluctuation, the system triggers the rapid fine-tuning mode;

[0418] If the status does not change significantly, no update will be made this time, and the system will wait for the next trigger cycle.

[0419] The full-process update mode starts from S1 and executes all steps completely.

[0420] The rapid fine-tuning mode, in order to save computational resources, only performs online fine-tuning on the already trained prediction model; specifically, it includes:

[0421] Fine-tuning data

[0422] Use all or part of the latest data in the data buffer as the fine-tuning dataset;

[0423] Fine-tuning method

[0424] A small learning rate is used to update the weights of the three prediction models with a small number of iterations of gradient descent. The loss function is typically the mean squared error.

[0425]

[0426] in, To fine-tune the loss value;

[0427] For each sample in the fine-tuning dataset;

[0428] This represents the total number of data samples for fine-tuning.

[0429] This is the actual capacity;

[0430] For capacity prediction.

[0431] The gradient descent algorithm is used during fine-tuning to continuously and incrementally adjust the internal parameters of the three prediction models in order to minimize the value of the loss function.

[0432] Dynamic learning rate

[0433] The learning rate decays dynamically with each fine-tuning epoch, and the calculation formula is as follows:

[0434]

[0435] in, For fine-tuning the rounds;

[0436] The initial fine-tuning learning rate (much smaller than the learning rate used during the original training of the model);

[0437] This represents the attenuation rate, with a value greater than 0 and less than 1.

[0438] S7.4. Prediction Result Output and Model Version Management

[0439] Output results: Regardless of the update mode used, the system will eventually output an updated prediction of remaining useful life based on the latest data and model.

[0440] Version management: The system archives the model parameters, hyperparameters, and key performance indicators after this update and adds a timestamp; this helps track the evolution of model performance and allows for a quick rollback to a previous stable version if problems occur during the update.

[0441] This invention provides a lithium battery remaining life prediction system based on multi-frequency decomposition and hybrid neural network. The lithium battery remaining life prediction system includes: a data collection module, a data preprocessing module, a data decomposition and component classification module, a multi-model collaborative prediction module, a hyperparameter optimization module, a final prediction module, and an adaptive update module.

[0442] The data collection module is used to collect historical operating data of lithium-ion batteries, including capacity data and the corresponding number of cycles, and store them in a certain format.

[0443] The data preprocessing module is used to standardize the collected lithium battery data in preparation for subsequent input into the prediction model.

[0444] The data decomposition and component classification module is used to decompose the capacity data of lithium batteries and calculate the sample entropy and component classification. Clustering is used to classify components, and high-frequency components are then... Secondary decomposition;

[0445] The multi-model collaborative prediction module is used to predict the three components obtained by decomposition using the corresponding prediction models.

[0446] The hyperparameter optimization module is used to optimize all decomposition parameters and prediction parameters used in the prediction process. It adopts a hierarchical reinforcement learning strategy to optimize parameters in stages and at different levels.

[0447] The final prediction module is used to predict the final result based on the optimal parameters obtained by hyperparameter optimization.

[0448] The adaptive update module is used to determine whether the model needs to be retrained based on the current state of the battery.

[0449] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0450] In this embodiment of the invention, the lithium battery remaining life prediction method introduces a component decomposition mechanism and tailors prediction models for different frequency components, effectively capturing local fluctuations and long-term trends in the battery degradation process, significantly improving the accuracy and robustness of remaining life prediction. Appropriate prediction models are designed for the complexity differences of different components, avoiding redundant calculations caused by uniformly using complex models for all components, thus achieving a balance between prediction accuracy and computational efficiency. The hyperparameter optimization strategy based on hierarchical reinforcement learning can automatically and efficiently find the optimal parameter combination according to different battery data characteristics, overcoming the shortcomings of high computational cost and poor adaptability of traditional optimization methods.

[0451] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0452] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the remaining life of a lithium battery based on multi-frequency decomposition and hybrid neural networks, characterized in that, The method for predicting the remaining lifespan of lithium batteries includes: S1. Data Collection Historical capacity data of lithium-ion batteries during multiple charge-discharge cycles are collected. Each data sample includes the number of cycles and its corresponding discharge capacity value. The data is organized and managed in time series format to form a complete battery life dataset. S2. Data Preprocessing Original capacity sequence The standardized capacity sequence obtained after standardization ; S3. Data Decomposition and Component Classification For normalized capacity sequences Decompose the sample and calculate the sample entropy and Clustering is used to classify components into three categories: high-frequency, mid-frequency, and low-frequency. The high-frequency components are then further classified. Secondary decomposition; S4. Multi-model collaborative prediction The corresponding prediction models are used to predict the three components respectively; S5. Hyperparameter Optimization in Hierarchical Reinforcement Learning All decomposition and prediction parameters used in the prediction process are optimized by adopting a hierarchical reinforcement learning strategy to optimize parameters in stages and at different levels. S6. Final Prediction Using the optimal combination of hyperparameters obtained through hierarchical reinforcement learning, the entire process from S3 to S4 is re-executed to generate the final, accurate prediction of remaining battery life. S7. Adaptive Update Whether the model needs to be retrained depends on the current state of the battery. In S4, the high-frequency component prediction model is input after... The decomposed sub-components are used to capture the dependencies between time series data through a dual feature extraction encoder. After feature fusion, the data is input into the decoder and finally mapped through flattening layers, fully connected layers, etc., to output the predicted value. The intermediate frequency component prediction model includes: a simplified input mapping layer, position encoding, and a 1-layer... The encoder, the linear transformation layer for feature mapping, and the output layer; The low-frequency component prediction model employs an attention mechanism. Model; The prediction results of the three models for high, medium, and low frequency components are superimposed to obtain the final battery capacity prediction sequence: Based on this fused capacity prediction sequence, the remaining lifespan of the battery is calculated according to the preset failure threshold.

2. The lithium battery remaining life prediction method based on multi-frequency decomposition and hybrid neural network as described in claim 1, characterized in that, In S1, the complete battery life dataset is represented as follows: in, For the first One cycle; For the first The battery discharge capacity measured over one cycle; This represents the total number of cycles.

3. The lithium battery remaining life prediction method based on multi-frequency decomposition and hybrid neural network as described in claim 1, characterized in that, In S2, the normalized capacity sequence for: Among them, the Standardized capacity value per cycle for: in, For the first The original capacity value for each cycle; The mean of the original capacity sequence; denoted as the standard deviation of the original capacity sequence.

4. The lithium battery remaining life prediction method based on multi-frequency decomposition and hybrid neural network as described in claim 1, characterized in that, S3 specifically includes: S3.

1. Primary decomposition The complete set empirical mode decomposition method is used to preprocess the standardized capacity sequence. Decomposed into a series of intrinsic mode functions and residual terms: in, For the first One intrinsic mode function; This is the final residual term; S3.

2. Component Classification Calculate the sample entropy of all intrinsic mode function components obtained in the above steps, and then use... Clustering algorithm, preset number of clusters All intrinsic mode functions are classified into three categories: high frequency, mid frequency, and low frequency. S3.

3. Secondary decomposition For the high-frequency components obtained from the above clustering, further processing is performed... The method performs secondary noise reduction.

5. The lithium battery remaining life prediction method based on multi-frequency decomposition and hybrid neural network as described in claim 1, characterized in that, In the high-frequency component prediction model, the feature extraction encoder includes: a sensor encoder and a time step encoder; The sensor encoder is composed of a multi-head parallel computation mechanism and a feedforward neural network, combined with layer normalization and residual connections. The formula for calculating the single-head attention score in the self-attention mechanism is as follows: in, For query matrix; The key matrix; It is a value matrix; The dimension of the key is used for scaling to avoid gradient anomalies; mask To block invalid locations as needed; T This indicates matrix transpose; The results of multiple single-head attention scores are concatenated and then transformed linearly to generate the final attention features; The specific formula for a feedforward neural network is: in, A This is the output of the attention mechanism; This is the weight matrix for the first layer of linear transformation; This is the bias for the first-level linear transformation; GELU is the activation function for Gaussian error linear units, and its formula is: in, It is the cumulative distribution function of the standard normal distribution; This is the weight matrix for the second-level linear transformation; This is the bias for the second-level linear transformation; Layer normalization is applied to the attention output and the feedforward network output, using the following formula: in, The mean; Standard deviation; It is the minimum value; Learnable scaling factor As a bias factor, it can stabilize training and accelerate convergence; The logic of time-step encoders and sensor encoders is similar. The main difference is that time-step encoders add position encoding before multi-head attention. The specific formula is as follows: in, For location index; For dimension indexing; For model dimensions; The data obtained by fusing the features from the two encoders is input into the decoder, which generates a hidden representation; the final output of the decoder at the last time step is: in, This is the final output of the entire high-frequency prediction model decoder module; L The length of the input sequence; It is the domain; The function of the flattening layer is to flatten this two-dimensional tensor It is compressed into a one-dimensional vector to connect standard fully connected layers, and its formula is: in, Flatten To be The tensor is reshaped into a length of ; The fully connected layer maps high-dimensional features to specific predicted values, and the formula is as follows: in, This represents the final predicted value for the high-frequency components; This is the weight matrix of the fully connected layer; This is the bias vector of the fully connected layer; Represents matrix multiplication; The calculation formula for the intermediate frequency component prediction model is as follows: in, The input sequence is... The tensor is mapped and then its dimensions are adjusted to fit. Model; Weights of the linear layer For the bias of the linear layer; Used for each position in time series data The positional coding is generated using a simplified periodic positional coding form, which assigns different periodic coding features to even-numbered positions using a sine function; It is through including Layer encoder The intermediate feature representation obtained after processing; The encoder adds position-encoded input sequences Encode the sequence, extract its feature information, and output the result. It includes features processed by mechanisms such as self-attention, which are used for subsequent decoding and prediction; This refers to the predicted value of the mid-frequency prediction model, passed through the weights of the fully connected layer. and Features of encoder output Calculated; The calculation formula for the low-frequency component prediction model is as follows: in, It means The layer in the first The hidden state of each cycle contains information up to the current moment. The layer stores information about the input sequence for subsequent computation and context transmission. The low-frequency component time series is in the first... Input samples for each cycle; yes Layer at time The hidden state is the memory information from the previous moment; It is the attention weight, used to measure the attention weight. Hidden state of each cycle The importance of the overall characteristics, through After normalization, the values ​​range from 0 to 1, and the sum of the attention weights at all times is 1; This refers to the context vector, which is processed through attention weights. Hidden state at each time point The weighted summation yields an integrated set of important information about the hidden states at different times, highlighting the characteristics of key time points; It is the prediction output of the low-frequency prediction model; Used for context vectors Perform a linear transformation; It is the rectified linear unit activation function, for Perform a nonlinear transformation, setting values ​​less than 0 to 0 and keeping values ​​greater than 0 unchanged, thereby increasing the model's nonlinear expressive power. It is a regularization operation; Used for passing Activation function calculation and The data after regularization is then linearly transformed again to obtain the final predicted value.

6. The lithium battery remaining life prediction method based on multi-frequency decomposition and hybrid neural network as described in claim 1, characterized in that, S5 specifically includes: S5.

1. State Awareness The system obtains the current status; S5.

2. High-level decision-making The PPO agent selects a high-level action to execute based on its policy network according to its current state. S5.

3. Low-level execution Based on the selected high-level action, execute the corresponding low-level action and adjust the relevant hyperparameters accordingly. S5.

4. Performance Evaluation and Reward Calculation Run the prediction framework using the new hyperparameters to calculate the reward. The reward function is calculated using the following formula: in, For the current round compared to the baseline Improvement amount; for The historical standard deviation; This represents the average accuracy of the predicted direction. This is a stability penalty term, which measures the imbalance in the prediction performance of each component; S5.

5. State Update and Strategy Learning New status and rewards R Feedback is given to the PPO agent to update its policy network and value network; This process is repeated a preset number of times, and the final output is the hyperparameter combination that performs best in the optimization process.

7. The lithium battery remaining life prediction method based on multi-frequency decomposition and hybrid neural network as described in claim 1, characterized in that, Specifically, S7 includes: S7.

1. Triggering Mechanism and Data Collection The system presets a trigger interval, which can be based on time or a cycle. Once the trigger condition is met, the module automatically starts the update process. The system collects the latest battery operation data from the battery management system since the last update, mainly including the latest cycle capacity measurement value and its corresponding cycle number; S7.

2. Data Buffering and Sequence Reconstruction The system maintains a fixed-length first-in-first-out (FIFO) data buffer to store historical data for a recent period of time. Newly collected data is added to the buffer while the oldest data points are removed to keep the total length of the buffer constant; thus, an updated time series reflecting the latest state of the battery is formed. S7.

3. Update Strategy Decisions Based on the significance of data changes, the system intelligently selects between full-process updates or rapid fine-tuning to balance computational costs and prediction accuracy. S7.3.

1. Change Detection Calculate the statistical difference between the new data and recent historical data; calculate a mean dissimilarity based on a sliding window, using the following formula: in, The average capacity of the newly added data points; To update the average capacity of windows of the same length in the front buffer; To update the standard deviation of the contents of windows of the same length in the front buffer; S7.3.

2. Strategy Selection Two thresholds are set, and the threshold for determining no significant change is: The threshold for determining a significant change is: ; When the battery degradation status is determined to have changed significantly, the system triggers a full-process update mode. When this is determined to be a normal degradation fluctuation, the system triggers the rapid fine-tuning mode; If the status does not change significantly, no update will be made this time, and the system will wait for the next trigger cycle. The full-process update mode starts from S1 and executes all steps completely. The rapid fine-tuning mode, in order to save computational resources, only performs online fine-tuning on the already trained prediction model; specifically, it includes: Fine-tuning data Use all or part of the latest data in the data buffer as the fine-tuning dataset; Fine-tuning method A small learning rate is used to update the weights of the three prediction models with a small number of iterations of gradient descent; the loss function is typically the mean squared error. in, To fine-tune the loss value; For each sample in the fine-tuning dataset; This represents the total number of data samples for fine-tuning. This is the actual capacity; For predicted capacity; The gradient descent algorithm is used during fine-tuning to continuously and incrementally adjust the internal parameters of the three prediction models in order to minimize the value of the loss function. Dynamic learning rate The learning rate decays dynamically with each fine-tuning epoch, and the calculation formula is as follows: in, For fine-tuning the rounds; To initially fine-tune the learning rate; This represents the attenuation rate, with a value greater than 0 and less than 1. S7.

4. Prediction Result Output and Model Version Management Output results: Regardless of the update mode used, the system will eventually output an updated prediction of remaining useful life based on the latest data and model. Version management: The system archives the model parameters, hyperparameters, and key performance indicators after this update and adds a timestamp; this helps track the evolution of model performance and allows for a quick rollback to a previous stable version if problems occur during the update.

8. A lithium battery remaining life prediction system based on multi-frequency decomposition and hybrid neural network, characterized in that, The lithium battery remaining life prediction system includes: a data collection module, a data preprocessing module, a data decomposition and component classification module, a multi-model collaborative prediction module, a hyperparameter optimization module, a final prediction module, and an adaptive update module. The data collection module is used to collect historical operating data of lithium-ion batteries, including capacity data and the corresponding number of cycles, and store them in a certain format. The data preprocessing module is used to standardize the collected lithium battery data in preparation for subsequent input into the prediction model. The data decomposition and component classification module is used to decompose the capacity data of lithium batteries and calculate the sample entropy and component classification. Clustering is used to classify components, and high-frequency components are then... Secondary decomposition; The multi-model collaborative prediction module is used to predict the three components obtained by decomposition using the corresponding prediction models. The hyperparameter optimization module is used to optimize all decomposition parameters and prediction parameters used in the prediction process. It adopts a hierarchical reinforcement learning strategy to optimize parameters in stages and at different levels. The final prediction module is used to predict the final result based on the optimal parameters obtained by hyperparameter optimization. The adaptive update module is used to determine whether the model needs to be retrained based on the current state of the battery. In the multi-model collaborative prediction module, the high-frequency component prediction model is input after... The decomposed sub-components are used to capture the dependencies between time series data through a dual feature extraction encoder. After feature fusion, the data is input into the decoder and finally mapped through flattening layers, fully connected layers, etc., to output the predicted value. The intermediate frequency component prediction model includes: a simplified input mapping layer, position encoding, and a 1-layer... The encoder, the linear transformation layer for feature mapping, and the output layer; The low-frequency component prediction model employs an attention mechanism. Model; The prediction results of the three models for high, medium, and low frequency components are superimposed to obtain the final battery capacity prediction sequence: Based on this fused capacity prediction sequence, the remaining lifespan of the battery is calculated according to the preset failure threshold.

Images