Battery Health Estimation Method and Electronic Device Based on Trend Filter Decomposition

Abstract

This invention discloses a battery health estimation method and electronic device based on trend filtering decomposition, belonging to the field of new energy power battery health management. The method includes preprocessing lithium-ion battery aging data; decomposing the battery degradation sequence using trend filtering decomposition technology, and dynamically optimizing the filtering parameters by introducing reasonable granularity criteria from granular computing to construct a feature case library reflecting the physical meaning of battery aging; constructing a long short-term memory network model integrating a Dropout regularization module, and suppressing overfitting by randomly discarding neuron outputs during network training; inputting the case library into the optimized network model for training and iteration, ultimately achieving high-precision estimation of the lithium-ion battery health state. This invention can adaptively extract the essential trend of battery degradation, improve feature interpretability, and enhance model generalization ability, thereby significantly improving the accuracy and robustness of battery health state prediction.

Description

Technical Field

[0001] This invention belongs to the field of health management of new energy power batteries, specifically involving a battery health estimation method and electronic equipment based on trend filtering decomposition. Background Technology

[0002] Against the backdrop of continuously deteriorating environmental quality and increasingly strained energy supply, my country is actively promoting the development of new energy sources to facilitate economic restructuring and achieve sustainable social development. Lithium-ion batteries, with their advantages such as high energy density and long cycle life, have become a key component of the new energy system. Among these, the prediction of remaining useful life (RUL) and the assessment of state of health (SOH) are core research areas in battery health management and reliability engineering. These aspects are crucial for optimizing equipment maintenance strategies, extending service life, and reducing total lifespan costs, and have therefore become a research hotspot in the field of electronic system health management and fault diagnosis.

[0003] As a crucial link in the new energy industry, data-driven lithium-ion battery health prediction methods have attracted widespread attention due to their flexibility. However, the core premise of such methods is to accurately and robustly extract trend features reflecting the inherent aging patterns from battery operating data. Currently, there are still significant shortcomings in the trend feature extraction stage. Existing trend extraction methods (such as moving averages and fixed-parameter filtering algorithms) often rely on empirically setting smoothness parameters, making it difficult to adapt to the complex non-stationarity and stage-specific changes in battery degradation data. Inappropriate parameter selection can easily lead to over-smoothing, resulting in the loss of key abrupt changes, or under-smoothing, retaining too much noise, making the extracted "trend" unable to accurately characterize the aging physical process, thus limiting the upper limit of the accuracy of subsequent prediction models.

[0004] Furthermore, even if a smooth curve is obtained, interpreting and structurally representing the trend with clear physical meaning remains a challenge. Many methods treat feature extraction and trend interpretation as independent steps, resulting in extracted features lacking direct correlation with the internal state evolution of the battery (such as SEI film growth and lithium deposition), leading to poor model interpretability and hindering state diagnosis and risk warning in practical management systems.

[0005] Finally, at the predictive model level, while Long Short-Term Memory (LSTM) networks can effectively handle time series data, they are prone to overfitting in scenarios with limited training data and complex, variable operating conditions, leading to a decline in model generalization ability. Although regularization techniques are widely used, how to more effectively combine them with the structural characteristics of LSTM to enhance its ability to remember and learn long-term health evolution trends while suppressing overfitting still requires further exploration.

[0006] In summary, existing technologies have systemic shortcomings in effectively extracting data trend features and improving the generalization ability of prediction models. Therefore, there is an urgent need for a new method that can synergistically address these issues to achieve a more accurate, reliable, and interpretable estimate of the health status of lithium-ion batteries. Summary of the Invention

[0007] The purpose of this invention is to provide a battery health estimation method and electronic device based on trend filtering decomposition. This method addresses the significant shortcomings of existing technologies in processing high-dimensional battery data, detecting structural changes in the data, effectively extracting data trend features, model generalization, overfitting risk, and practicality. It adopts a deep fusion of granular computing and trend filtering, and automatically separates noise and interference by dynamically optimizing the key parameters of trend filtering, extracting the smooth trend component that best represents the battery's inherent aging pattern. Through structured regularization, it effectively suppresses overfitting and improves the model's adaptability and robustness to unseen battery data and different charging and discharging scenarios, thereby achieving high-precision, high-reliability, and highly interpretable battery health state estimation.

[0008] To achieve the above objectives, the technical solution adopted by this invention is: a battery health estimation method based on trend filter decomposition, comprising the following steps:

[0009] S1. Preprocess the raw lithium-ion battery aging data;

[0010] S2. The preprocessed data is decomposed using a trend filtering algorithm to extract aging features from the battery data. The regularization parameter of the trend filtering is dynamically optimized by the reasonable granularity criterion in granular computing, and the sequence is iteratively decomposed based on the optimized parameter until the entropy value of the component reaches the set condition. The decomposed components are then stored in the feature case library.

[0011] S3. Construct a regularized long short-term memory network model, embed the Dropout regularization operation into the internal computation process of the LSTM unit, use the result of the Dropout operation on the hidden state of the previous time step as the gating computation of the current time step, and randomly discard some neuron connections during training to improve the generalization ability of the model.

[0012] S4. Input the data from the feature case library into the regularized long short-term memory network model for training. Optimize the model by adjusting its hyperparameters to obtain the trained optimal prediction model. Use the optimal prediction model to estimate the health status and remaining lifespan of the lithium battery.

[0013] Further, step S2 includes:

[0014] S21. Calculate the regularization parameter in the trend filtering algorithm. Selection range ,in for And gradually increase the critical value corresponding to dividing the battery decay sequence into an information granule;

[0015] S22. Based on the reasonable granularity criterion in granular computing, determine the optimal parameter values ​​within the selected range. At this point, the comprehensive evaluation function of the reasonable granularity criterion reaches its maximum value;

[0016] S23. Obtain the optimal value of the parameters. The trend filtering algorithm decomposes the battery degradation sequence X to obtain the components. When a certain component When the entropy value exceeds a set threshold, the decomposition of the X sequence ends, yielding m components. and residual ;

[0017] S24. Place each component obtained from the decomposition into the feature case library for model training and testing; the residual is white noise and is not stored in the feature case library.

[0018] Further, in step S21, the regularization parameter is calculated. maximum value The specific process is as follows:

[0019] (1) Construct Gaussian linear fuzzy information particles corresponding to the battery decay sequence X ,in:

[0020] The core line represents the Gaussian linear fuzzy information particle. These are the slope and intercept of the coreline, respectively. This indicates the battery degradation duration corresponding to this information particle. The variance of the core line and the original data;

[0021] (2) Calculate the decay sequence of the galvanic cell Each data point Membership degree of the Gaussian linear fuzzy information granules:

[0022]

[0023] (3) Calculate the sum of the membership degrees of all data points in the galvanic cell decay sequence to the information particle. :

[0024]

[0025] (4) Gradually increase The value of , when When the value increases to a certain critical value, the battery degradation sequence X is divided into a complete information granule, and the sum of membership degrees F converges to a constant value. This critical value is the critical value. .

[0026] Further, in step S22, the comprehensive evaluation function is:

[0027]

[0028] in, The coverage rate is inversely proportional to the number of information particles s generated. ;

[0029] Due to its special characteristics, the discreteness of information particles Related, ;

[0030] exist Within the range of choices, when When the maximum value is obtained, the corresponding That is, the optimal value of the parameter. .

[0031] Furthermore, in step S2, the objective function of the trend filtering algorithm is:

[0032]

[0033] In the formula, n represents the total number of data points in the battery degradation sequence. Represents the battery degradation sequence. To decompose into components, where and Let X and Y be the values ​​of the i-th data points in sequences X and Y, respectively; the first term in the formula is the loss term, which measures the degradation sequence X of the galvanic cell and the decomposed components. The error between them; the second term is the smoothing penalty term, This is a regularization parameter that controls the smoothness of the fitted curve.

[0034] Furthermore, in step S3, the calculation process of the regularized long short-term memory network model includes:

[0035] (1) Based on the input battery aging feature data and combined with the hidden state of the previous time step after the Dropout operation, the input gate is calculated. Forgotten Gate Output gate and candidate cell status :

[0036]

[0037] In the formula, D is the Dropout operator, and its parameter is set to a random subset of 0; This represents the sigmoid function. This represents the hyperbolic tangent activation function. For affine transformation operators, This is the hidden state at the previous time step t. This represents the hidden state of the previous time step t-1 in the current layer;

[0038] (2) Update cell state :

[0039]

[0040] in, for Cell state at time step t of layer. for Cell state at time step t-1 on the layer. This represents the element-wise multiplication operation;

[0041] (3) Calculate the current hidden state : Finally, after processing by the fully connected layer and the output layer, the estimated result of the battery health status is obtained.

[0042] Furthermore, step S2 also includes the step of dividing the battery aging stage: based on the components after trend filtering decomposition. The sequence characteristics are used to divide the battery degradation process into multiple information particles, and the components corresponding to the information particles are used as the basis for further analysis. The entropy, decay rate change and residual The variance of the information particle is used to assign a semantic label to each information particle, which represents the physical stage of battery aging. The semantic label includes the formation of the adaptation stage, the stable health stage, the transition stage, or the accelerated degradation stage.

[0043] Furthermore, the rules for assigning semantic tags include:

[0044] If the information particle is in the early stage of battery cycling, and its residual Large variance and components If the changes are gradual, then it is labeled as an adaptation phase;

[0045] If the component corresponding to the information granule If the decay rate is stable and the entropy value is below the first threshold, and the overall sequence exhibits low variance and linearity, then it is labeled as a stable and healthy stage.

[0046] If the component corresponding to the information granule The decay rate accelerates and there is a trend inflection point, while the residual If the variance increases, a transition stage label is assigned;

[0047] If the component corresponding to the information granule A significantly accelerated and nonlinear decay rate is labeled as an accelerated degradation stage.

[0048] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the battery health estimation method based on trend filter decomposition.

[0049] The beneficial effects of the above scheme are as follows:

[0050] 1. This invention achieves high-fidelity and adaptive extraction of battery degradation trends. Traditional filtering methods using fixed parameters struggle to adapt to the non-stationary nature of battery data, often resulting in a trade-off between smoothing noise and preserving key abrupt changes. This invention innovatively utilizes reasonable granularity criteria from granular computing to dynamically optimize key parameters of trend filtering. This adaptively removes random noise from the original data while accurately extracting the essential trends reflecting the inherent aging patterns of the battery, significantly improving the feature quality of the input model.

[0051] 2. This invention endows the extracted data features with clear physical meaning, achieving interpretable assessment of battery health status. Existing methods obtain abstract mathematical features that are difficult to directly correlate with the internal aging mechanisms of the battery. This invention automatically maps the mathematical components decomposed by trend filtering into information particles with clear aging semantics based on their statistical properties, such as dividing them into physical stages like adaptation, stable health, transitional stages, and accelerated degradation. This makes the prediction results no longer black-box numerical values, but can intuitively reflect the battery's health status and its evolutionary trajectory, greatly enhancing the method's engineering applicability.

[0052] 3. This invention constructs a prediction model that is resistant to overfitting and has strong generalization ability. Addressing the problem that Long Short-Term Memory (LSTM) networks are prone to overfitting with limited data and that their generalization ability declines with changing operating conditions, this invention deeply embeds Dropout technology into the LSTM gating structure. This structured regularization method enables the model to learn more robust feature representations, effectively reducing dependence on noise and random patterns in the training data, thereby ensuring the model's robustness and accuracy in predictions under different individual batteries and complex, variable charging and discharging conditions.

[0053] 4. This invention ultimately forms a complete technical closed loop from accurate feature extraction and physical interpretation to reliable prediction. High-quality trend features are the foundation for accurate prediction, interpretable feature representation enhances the reliability of the results, and a highly generalizable model ensures the stable performance of the method in real-world scenarios. The synergistic effect of these three elements jointly promotes a systematic improvement in the accuracy, reliability, and practicality of lithium-ion battery health state estimation, providing strong technical support for battery health management. Attached Figure Description

[0054] Figure 1 This is a flowchart of the battery health estimation method of the present invention;

[0055] Figure 2 This invention is based on Flowchart of the trend filter decomposition method;

[0056] Figure 3 These are the trend filter parameters based on the convergence of the sum of membership degrees of information granules in Example 1. A schematic diagram of the solution process;

[0057] Figure 4 In Example 1, the optimal values ​​of the trend filter parameters are obtained by using a reasonable granularity criterion. A schematic diagram of the process;

[0058] Figure 5 This is a schematic diagram of the process of multi-level decomposition of the battery signal by trend filtering in Example 1;

[0059] Figure 6 This is the structure of the deep learning model using Dropout regularization in Example 1;

[0060] Figure 7 This is the specific process of step S4 in Example 1;

[0061] Figure 8 This is a comparison of the regularized long short-term memory network model of this invention with other prediction methods;

[0062] Figure 9 This is a comparison chart of the effectiveness of trend filtering for different individual cells in this invention;

[0063] Figure 10 This is a comparison chart of the verification of different Dropout ratios in this invention. Detailed Implementation

[0064] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0065] It should be noted that, unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0066] Example 1:

[0067] like Figure 1 As shown, a battery health estimation method based on trend filter decomposition includes the following steps:

[0068] S1. Preprocess the raw lithium-ion battery aging data;

[0069] S2. The preprocessed data is decomposed using a trend filtering algorithm to extract aging features from the battery data. The regularization parameter of the trend filtering is dynamically optimized by the reasonable granularity criterion in granular computing, and the sequence is iteratively decomposed based on the optimized parameter until the entropy value of the component reaches the set condition. The decomposed components are then stored in the feature case library.

[0070] S3. Construct a regularized long short-term memory network model, embed the Dropout regularization operation into the internal computation process of the LSTM unit, use the result of the Dropout operation on the hidden state of the previous time step as the gating computation of the current time step, and randomly discard some neuron connections during training to improve the generalization ability of the model.

[0071] S4. Input the data from the feature case library into the regularized long short-term memory network model for training. Optimize the model by adjusting its hyperparameters to obtain the trained optimal prediction model. Use the optimal prediction model to estimate the health status and remaining lifespan of the lithium battery.

[0072] This invention employs trend filtering decomposition technology combined with granular computing to decompose battery degradation sequences and extract relevant features to construct a case library. Simultaneously, addressing the issues of high dimensionality and strong nonlinearity in the feature space of battery degradation data, a regularized long short-term memory network model (deep prediction model) constructed using dropout technology is introduced. Data from the case library is input into the regularized long short-term memory network model for network training and prediction. This effectively avoids overfitting while improving prediction accuracy and robustness, enabling the estimation of the health status of lithium batteries under different operating conditions.

[0073] The steps of this invention are described in detail below:

[0074] Step S1: Collect and preprocess the raw lithium-ion battery aging data.

[0075] Data source: Commercial rechargeable APR18650M1A lithium-ion battery pack using the A123 system, rated capacity 2.0Ah, positive electrode material is nickel cobalt manganese oxide, negative electrode material is graphite, charge / discharge temperature is set at 24℃. Battery capacity decay and battery cycle life data, including charge / discharge current, voltage, temperature, and capacity decay curves, are collected through high-precision sensors.

[0076] Battery data with too few cycle counts were removed (considered as incomplete aging process); battery aging cycles with too few effective cycle counts were also removed; missing values ​​in the experiment were supplemented using the average method; and outliers were removed using the three-standard-deviation method; voltage and current data were normalized using z-score. Battery health was defined as...

[0077]

[0078] In the formula, This represents the battery's current maximum usable capacity. This refers to the battery's rated capacity. The battery is considered to have reached the end of its lifespan when its capacity decays to 80% of its rated capacity. After data cleaning, interference between the battery capacity decay data and the battery cycle life is eliminated.

[0079] In this embodiment, 100 sets of valid data are retained after cleaning, of which 80 sets are used for training and 20 sets are used for testing.

[0080] The preprocessed data is decomposed and processed using trend filtering decomposition technology to extract battery data aging characteristics and construct a case library. The trend filtering decomposition technology uses granular computing to solve the parameters optimally, obtains the maximum value of the parameters based on the construction of linear information granules, and obtains the maximum value of the parameters according to the reasonable granularity criterion.

[0081] Step S2: Use trend filtering decomposition technology to extract aging features from battery data and build a case library.

[0082] Figure 2 This is a flowchart of trend filtering decomposition. The trend filtering decomposes the battery decay sequence, and the optimal value is obtained using granular computation to build a case library.

[0083] The specific implementation process of trend filtering decomposition of the preprocessed data in this embodiment is as follows:

[0084] S21, Trend filtering algorithm for battery degradation sequence Decompose into linear trends and residual series The sum. Satisfying two principles: linear trend. To achieve smoothness and residuals as much as possible It should be as small as possible. Converting these two principles into mathematical expressions and combining them, we get the following formula:

[0085]

[0086] In the formula, n represents the total number of data points in the battery degradation sequence. Represents the battery degradation sequence. To decompose into components, where and Let X and Y be the values ​​of the i-th data points in sequences X and Y, respectively; the first term in the formula is the loss term, which measures the degradation sequence X of the galvanic cell and the decomposed components. The error between them; the second term is the smoothing penalty term, This is a regularization parameter that controls the smoothness of the fitted curve.

[0087] Furthermore, in step S21, in order to perform trend filtering decomposition on the battery degradation sequence X, a regularization parameter must be established. The selection criteria. First, for any given battery degradation sequence X, the regularization parameter needs to be determined. Selection range .Sure The process requires granular computation to construct fuzzy information granules. Fuzzy information granules are sets constructed from elements that are functionally similar or related, and are described using a fuzzy set-based representation method. The expression for a Gaussian linear fuzzy information granule is defined as follows:

[0088]

[0089] in, The core line represents the Gaussian linear fuzzy information particle. These are the slope and intercept of the coreline, respectively; parameters This indicates the length of the information particle, which here represents the duration of battery degradation contained within that information particle. The variance between the core line and the original data represents the degree of fit between the information particles and the decay sequence of the galvanic cell.

[0090] Then, the decay sequence of the galvanic cell is calculated. Each point in the and The membership degree of the information particles formed by the trend filtering algorithm is expressed as follows:

[0091]

[0092] In the trend filtering algorithm formula, As the error gradually increases, the error weights become smaller, and the battery degradation sequence is divided into fewer and fewer linear information particles. At a certain critical value, the battery degradation sequence X is divided into entire information particles. After this point, Further increasing the value will not affect the decomposition of the battery degradation sequence X. The critical value is marked as That is, the battery degradation sequence is divided into a whole information granule, and the total membership value of each point in the sequence to that information granule is... and no longer change, that is

[0093]

[0094] It converges to a constant value; at this point... The critical value is .

[0095] In practice, each... The sum of the membership degrees corresponding to the values ,when When the value no longer changes, this is when the value is obtained. An example of the process is shown below. Figure 3 As shown.

[0096] S22, Optimal value of parameters The specific implementation process is as follows:

[0097] First, trend filtering divides the battery degradation sequence X into several segments, resulting in a piecewise constant function (information granules) as the output is made, where the function value within each segment is as close as possible to the dividing point of that interval. Then, in the above trend filtering objective formula, the decomposed components... ,like Then we get They are not on the same information granule (line segment).

[0098] As can be seen from the above trend filtering objective formula, with As the error weight increases, the error weight decreases, while the smoothing penalty weight increases. Consequently, fewer and fewer information particles are formed during the attenuation sequence filtering process, until the entire attenuation sequence is divided into a single information particle. With... During the ascent process, combined with a reasonable granularity criterion, the average number of data points contained in the generated s information granules is represented as follows: The proportion is In the reasonable granularity criterion, as... Increase its coverage Represented as:

[0099]

[0100] In the formula, The value range is [0,1], and it is inversely proportional to the number of information particles s generated. In the reasonable granularity criterion, as... As it rises, its particularity decreases accordingly. During the process, its particularity decreases monotonically, and its particularity... Represented as:

[0101]

[0102] In the above formula, This represents the dispersion of information particles during the filtering process (the difference between adjacent information particles). Based on the reasonable granularity criterion, its... The comprehensive evaluation function for the value is:

[0103]

[0104] In specific implementation, as Ascending, calculate the monotonically increasing functions respectively. and monotonically decreasing function The value of the product of the two. When the maximum value is obtained, at this point, we can find... The process of obtaining the result is as follows: Figure 4 As shown, in each round of trend filtering decomposition, , The entropy values ​​of the corresponding components are shown in Table 1.

[0105] Table 1. Parameters and entropy values ​​of each round of trend filter decomposition.

[0106]

[0107] Table 1 , , All are effective feature components obtained by trend filtering decomposition of the battery decay sequence X; the decomposition process also produces residuals. (White noise), residual Not listed in Table 1.

[0108] S23, Obtain parameters , The trend filtering algorithm decomposes the battery degradation sequence X to obtain the components. . When the amount When the entropy value exceeds a set threshold, the decomposition of the X sequence ends, yielding m components. and residual The residual sequence is considered to be the first of the battery decay sequence X. The trend filter decomposes the components, which are not stored in the feature case library.

[0109] Combining in the pursuit The basic principle of reasonable particle size in the process is to analyze the corresponding physical meaning of battery aging through component decomposition.

[0110] The specific implementation process of this embodiment is as follows:

[0111] (1) For the battery decay sequence Solve its corresponding And using the reasonable granularity criterion, the optimal value of the parameter is solved. Solve for the trend filter decomposition components at this time. And calculate its entropy value as 0.212;

[0112] (2) For components Using trend filtering decomposition, the corresponding And using the reasonable granularity criterion, the optimal value of the parameter is solved. Solve for the trend filter decomposition components at this time. And calculate its entropy value as 0.602;

[0113] (3) Also for components Using trend filtering decomposition, the corresponding And using a reasonable granularity criterion, the solution is obtained. Solve for the trend filter decomposition components at this time. The entropy value of 0.87 is calculated. At this point, the entropy value is greater than the set threshold (0.8), so the trend filtering decomposition process ends.

[0114] (4) Find the residual series , It should be noted that the residual column can be considered as white noise, therefore it is not used as a feature case library.

[0115] Figure 5 The curves are the characteristic curves after trend filtering decomposition. It is clear from the figure that the battery capacity decay curve after noise removal is smooth and the abrupt change points are completely preserved, which strongly proves the effectiveness and superiority of this method in data processing.

[0116] The entire lifecycle of the battery is semantically divided into information granules, and then trend filtering decomposition is performed on these information granules:

[0117] Information particle ①, according to the reasonable granularity criterion, the information particle is located in 5-15% of the entire battery cycle. Based on its trend filtering decomposition characteristics, it can be seen that the information particle fluctuates greatly in the short term during this stage. (The variance is large), and the capacity fluctuates slightly. Its semantics are "forming an adaptation stage", indicating that the information particle represents the adaptation stage information particle of the battery transitioning from the manufacturing state to the stable working state.

[0118] Information particle ②, according to the reasonable granularity criterion, this information particle is located in the 15-60% range of the entire cycle, and its trend filtering decomposition characteristics are: the decay rate is relatively stable (decomposition components) The value is smaller compared to other stages, and the entropy value is the smallest at 0.212), and the capacity decreases slowly in an approximately linear manner (decomposition components). (The fluctuation range is small), and the overall trend filtering decomposition features are low variance and linear. The semantics represented by its information granule is "stable and healthy stage". This information granule represents a healthy operating stage in which the battery performance degrades slowly and predictably.

[0119] Information particle ③, according to the reasonable granularity criterion, this information particle is located in 60-85% of the entire battery cycle, and its trend filtering decomposition characteristics are: the capacity decline rate is significantly accelerated (decomposition components) The absolute value of the slope of the corresponding information particles increases significantly, and volatility is enhanced (decomposition components). The amplitude increases. (The variance is large), the semantics of this information particle is "transition stage". The core features of this information particle are the existence of trend inflection point in trend filtering decomposition and the increase of residual variance value in trend filtering decomposition, which represent the key stage of the battery-dominated aging mechanism from the stable consumption stage to the structural damage stage.

[0120] Information granule ④, according to the reasonable granularity criterion, is located at 85% - EOL (End of Life) of the battery's entire life cycle, and its trend filtering decomposition characteristic is rapid capacity decline (decomposition component). The slope of the corresponding information particle increases significantly and is the largest relative to other information particles. This information particle is given the semantic meaning of "accelerated degradation (pre-failure) stage". Its core feature is the nonlinear enhancement of trend filtering decomposition. This information particle represents the dangerous stage where the battery performance deteriorates rapidly and approaches the end of its life.

[0121] By using trend filtering decomposition and combining it with reasonable granularity criteria, we can better assign semantic meaning to information particles and clarify the physical meaning of battery aging they represent. This makes it easier for staff to understand battery dynamics in a timely manner and better maintain the battery, ensuring the normal use of electrical equipment.

[0122] S24. Place the decomposed components (decay sequences) into the feature case library for model training and testing. After trend filtering decomposition and the formation of the case library, the decay features of the battery sequence are extracted. The decay trend is smoother and more distinctive as a feature. At the same time, the residual part is effectively isolated to remove or weaken invalid features, which can improve the model training efficiency.

[0123] Step S3: Construct a regularized long short-term memory network model.

[0124] A regularized long short-term memory network was used as the battery prediction model. This model, through Dropout regularization, effectively avoids the overfitting risk common in traditional deep learning models, while significantly improving the model's robustness and prediction accuracy. The parameter values ​​of the regularized long short-term memory network model are shown in Table 2.

[0125] Table 2. Parameters of the Regularized Long Short-Term Memory Network Model of this Invention

[0126]

[0127] The specific process of constructing a battery prediction model using a regularized optimized Long Short-Term Memory network is as follows:

[0128] S31. Set up the initial network and divide the case library into training and testing sets. Set the initial network parameters and the training cutoff conditions.

[0129] S32. Add a Dropout regularization module to the prediction model and use particle swarm optimization to select the optimal Dropout occurrence rate [0,1], so that the network can remember more historical information, avoid overfitting, and increase the generalization performance of the model.

[0130] In the traditional LSTM architecture,

[0131]

[0132]

[0133]

[0134]

[0135] In the formula, for A hidden state at a layer time step. The symbol for element-wise multiplication. This refers to the affine transformation operator. Dropout is a regularization method that randomly discards the outputs of some neurons during LSTM training, reducing the model's dependence on specific neurons and improving its generalization performance. With Dropout, each neuron has an equal probability of being dropped, resulting in a different network in each training iteration. This is equivalent to training different sub-networks each time, which helps improve the model's generalization ability and effectively prevents overfitting. Neurons that are not dropped are not part of the current network, but their work is resumed in the next training iteration, and then some neurons are randomly selected again for training. The Dropout regularization module allows the network to learn features more autonomously, without over-relying on a single neuron or group of neurons, thereby improving the model's robustness and adaptability to new data.

[0136] In this embodiment, the architecture of the LSTM model is as follows:

[0137]

[0138]

[0139]

[0140] In the formula, D is the Dropout operator, which sets a random subset of its argument to 0; This represents the sigmoid function. This represents the hyperbolic tangent activation function. For affine transformation operators, This is the hidden state at the previous time step t. This represents the hidden state at the previous time step t-1 of the current layer.

[0141] This method ensures that the LSTM model can remember more historical information during the learning process, while reducing overfitting, thereby improving its robustness and generalization performance. An LSTM with a Dropout layer, such as... Figure 6 As shown, the input layer receives battery aging-related features, which are then passed to the LSTM layer. While the LSTM units extract temporal features, a hyperparameter optimization layer (covering the number of units, learning rate, number of iterations, and Dropout value) dynamically adjusts the model parameters. Next, a fully connected layer integrates the extracted features, and Dropout randomly discards some neuron connections for regularization. Finally, the output layer outputs the prediction result. This network structure, through the combined design of LSTM temporal feature extraction and Dropout regularization, achieves accurate prediction of battery health status.

[0142] Subsequently, the particle swarm optimization method was used to obtain the optimal Dropout occurrence rate (0.2), and the optimal regularized long short-term memory network model was trained to predict the health status of lithium-ion batteries.

[0143] Step S4: Evaluate the performance of the prediction model, optimize the model hyperparameters, and output the optimal prediction model to estimate the health status and remaining lifespan of the lithium-ion battery. See [link to implementation details] for more information. Figure 7 .

[0144] S41. Input the case library data (training set) after trend filtering and denoising into the regularized long short-term memory network model. The model calculates neuron parameters through backpropagation, performs battery health status prediction analysis, and uses mean-square error (MSE) to evaluate model performance. The calculation formula is as follows:

[0145]

[0146] In the formula, and y i These are the predicted and actual values ​​for the i-th battery cell, where n represents the number of battery cells.

[0147] S42. Following the method in step S41, the regularized long short-term memory network model is trained multiple times. The learning rate is dynamically adjusted using the learning rate warmup strategy. The Adam optimizer is used to improve the model parameters. The dropout rate is optimized using particle swarm optimization. The prediction results are recorded and evaluated. The model hyperparameters are continuously adjusted based on the results until the optimal performance is achieved. The trained optimal prediction model is then output.

[0148] S43. The trained optimal prediction model is used as the final model for estimating the battery health status to perform lithium-ion battery health status estimation.

[0149] It is important to emphasize that for different types of battery data, parameters must be adjusted using a grid search method to obtain the most suitable network model parameters; network model parameters cannot be generalized. The grid search method exhaustively searches through all candidate parameter choices, trying every possible combination of parameters. This method permutes and combines the possible values ​​of each parameter to generate a "grid," then uses each combination for model training, and evaluates performance using cross-validation. After trying all parameter combinations in the fitted function, the classifier is returned, automatically adjusted to the optimal parameter combination.

[0150] In this embodiment, five-fold cross-validation is used to ensure the robustness of the model results. The iteration criterion set in this embodiment is: iteration stops when MSE < 0.1, and the optimal prediction model is output.

[0151] Comparative experiment:

[0152] Figure 8 The comparison results between this invention and the prediction methods CNN-ATResNet and CNN-LSTM are shown. The horizontal axis represents the number of battery cycles, and the vertical axis represents the predicted battery health status. The predicted values ​​of this invention are represented by orange lines, while the predicted values ​​of CNN-ATResNet and CNN-LSTM are represented by yellow and purple lines, respectively. Figure 8 It can be seen that the predicted values ​​of this invention almost coincide with the actual values. Therefore, the regularized long short-term memory network model of this invention has better prediction accuracy and is significantly superior to the other two methods.

[0153] Figure 9 This paper compares the predictive effectiveness of deep learning models for batteries No. 18, No. 28, No. 44, and No. 62 using two different approaches: "without trend filtering decomposition" and "with trend filtering decomposition." Experimental results show that the prediction accuracy of the deep learning models is improved to varying degrees after applying trend filtering decomposition, specifically by 21%, 26%, 22%, and 24%, with an average improvement of 23.25%. These results demonstrate that the introduction of trend filtering decomposition technology significantly optimizes the predictive performance of the deep learning models in this invention.

[0154] Figure 10 To verify the effectiveness of integrating the Dropout module into the deep learning model, comparative experiments were conducted for four individual battery cells (No.18, No.28, No.44, and No.62) with different Dropout occurrence rates (0%, 10%, 20%, 30%, and 40%). Analysis of the experimental data shows that, with the integrated Dropout technology, the deep learning model achieves the best prediction performance when the Dropout occurrence rate is 20%, and the results of repeated experiments exhibit the smallest variance and the strongest robustness. This demonstrates that the integration of Dropout technology into the deep learning model in this invention has significant advantages, effectively improving model stability and prediction accuracy.

[0155] The batteries used in this invention are all derived from automotive power batteries. Their charging and discharging strategies simulate different usage scenarios for new energy power batteries (fast and slow charging, power battery discharging scenarios, etc.). Verified through 100 different charging and discharging scenarios, this invention demonstrates good prediction accuracy and robustness, and can adapt to the diversity and complexity of battery data under different operating conditions. It has very broad market and application value in the battery field, especially in the field of new energy vehicle power batteries.

[0156] Example 2:

[0157] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the battery health estimation method based on trend filtering decomposition described in Embodiment 1.

[0158] Finally, it should be noted that any parts of this invention not described in detail are prior art. Those skilled in the art will understand that the above descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A battery health estimation method based on trend filter decomposition, characterized in that, Includes the following steps: S1. Preprocess the raw lithium-ion battery aging data; S2. The preprocessed data is decomposed using a trend filtering algorithm to extract aging features from the battery data. The regularization parameter of the trend filtering is dynamically optimized by the reasonable granularity criterion in granular computing, and the sequence is iteratively decomposed based on the optimized parameter until the entropy value of the component reaches the set condition. The decomposed components are then stored in the feature case library. S3. Construct a regularized long short-term memory network model, embed the Dropout regularization operation into the internal computation process of the LSTM unit, use the result of the Dropout operation on the hidden state of the previous time step as the gating computation of the current time step, and randomly discard some neuron connections during training to improve the generalization ability of the model. S4. Input the data from the feature case library into the regularized long short-term memory network model for training. Optimize the model by adjusting its hyperparameters to obtain the trained optimal prediction model. Use the optimal prediction model to estimate the health status and remaining lifespan of the lithium battery. Step S2 includes: S21. Calculate the regularization parameter in the trend filtering algorithm. Selection range ,in for And gradually increase the critical value corresponding to dividing the battery decay sequence into an information granule; S22. Based on the reasonable granularity criterion in granular computing, determine the optimal parameter values ​​within the selected range. At this point, the comprehensive evaluation function of the reasonable granularity criterion reaches its maximum value; S23. Obtain the optimal value of the parameters. The trend filtering algorithm decomposes the battery degradation sequence X to obtain the components. When a certain component When the entropy value exceeds a set threshold, the decomposition of the X sequence ends, yielding m components. and residual ; S24. Place each component obtained from the decomposition into the feature case library for model training and testing; the residual is white noise and is not stored in the feature case library. The comprehensive evaluation function is: in, Coverage, relative to the number of information particles generated. s Inversely proportional, ; Due to its special characteristics, the discreteness of information particles Related, ; exist Within the range of choices, when When the maximum value is obtained, the corresponding That is, the optimal value of the parameter. .

2. The battery health estimation method based on trend filtering decomposition according to claim 1, characterized in that, In step S21, the regularization parameter is calculated. maximum value The specific process is as follows: (1) Construct Gaussian linear fuzzy information particles corresponding to the battery decay sequence X ,in: The core line represents the Gaussian linear fuzzy information particle. These are the slope and intercept of the coreline, respectively. This indicates the battery degradation duration corresponding to this information particle. The variance of the core line and the original data; (2) Calculate the decay sequence of the galvanic cell Each data point Membership degree of the Gaussian linear fuzzy information granules: (3) Calculate the sum of the membership degrees of all data points in the galvanic cell decay sequence to the information particle. : (4) Gradually increase The value of , when When increased to a certain critical value, the battery degradation sequence X When the entire information is divided into a single complete information particle, the sum of membership degrees F converges to a constant value; this critical value is... .

3. The battery health estimation method based on trend filtering decomposition according to claim 1, characterized in that, In step S2, the objective function of the trend filtering algorithm is: In the formula, n represents the total number of data points in the battery degradation sequence. Represents the battery degradation sequence. To decompose into components, where and The first and second sequences in X and Y are respectively the first and second sequences in Y. i The values ​​of each data point; the first term in the formula is the loss term, which measures the difference between the galvanic cell decay sequence X and the decomposed components. The error between them; the second term is the smoothing penalty term, This is a regularization parameter that controls the smoothness of the fitted curve.

4. The battery health estimation method based on trend filtering decomposition according to claim 1, characterized in that, In step S3, the calculation process of the regularized long short-term memory network model includes: (1) Based on the input battery aging feature data and combined with the hidden state of the previous time step after the Dropout operation, the input gate is calculated. Forgotten Gate Output gate and candidate cell status : In the formula, D is the Dropout operator, and its parameter random subset is set to 0; This represents the sigmoid function. This represents the hyperbolic tangent activation function. For affine transformation operators, This represents the hidden state at the previous time step t. This represents the hidden state of the previous time step t-1 in the current layer; (2) Update cell state : in, for Cell state at time step t of layer. for Cell state at time step t-1 on the layer. This represents the element-wise multiplication operation; (3) Calculate the current hidden state : Finally, after processing by the fully connected layer and the output layer, the estimated result of the battery health status is obtained.

5. The battery health estimation method based on trend filtering decomposition according to claim 1, characterized in that, Step S2 also includes the step of dividing the battery aging stage: based on the components after trend filtering decomposition. The sequence characteristics are used to divide the battery degradation process into multiple information particles, and the components corresponding to the information particles are used as the basis for further analysis. The entropy, decay rate change and residual The variance of the information particle is used to assign a semantic label to each information particle, which represents the physical stage of battery aging. The semantic label includes the formation of the adaptation stage, the stable health stage, the transition stage, or the accelerated degradation stage.

6. The battery health estimation method based on trend filtering decomposition according to claim 5, characterized in that, The rules for assigning semantic tags include: If the information particle is in the early stage of battery cycling, and its residual Large variance and components If the changes are gradual, then it is labeled as an adaptation phase; If the component corresponding to the information granule If the decay rate is stable and the entropy value is below the first threshold, and the overall sequence exhibits low variance and linearity, then it is labeled as a stable and healthy stage. If the component corresponding to the information granule The decay rate accelerates and there is a trend inflection point, while the residual If the variance increases, a transition stage label is assigned; If the component corresponding to the information granule A significantly accelerated and nonlinear decay rate is labeled as an accelerated degradation stage.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the battery health estimation method based on trend filter decomposition as described in any one of claims 1 to 6.

Images