Lithium-ion battery state-of-health estimation method integrating auxiliary information and multi-operating condition data

Abstract

The application provides a lithium ion battery health state estimation method integrating auxiliary information and multi-working condition data, first loads multiple battery data, extracts battery auxiliary information of each sample of each battery, and obtains voltage, current and temperature sequence data corresponding to a preset voltage interval, then carries out monotone non-decreasing, interpolation and normalization processing on the sequence data to obtain sample data of multiple batteries and corresponding auxiliary feature information, then divides the data into a test set and a training set, inputs the training set into a constructed hybrid neural network model for processing to complete battery health state estimation of the training set, and finally optimizes the hybrid neural network model through back propagation combined with a SHAP value algorithm and pruning and quantization operations. The application scheme greatly improves the prediction ability and precision of the battery health state, reduces the calculation complexity of the model, and enables the model to have stronger robustness and generalization ability in a complex environment.

Description

Technical Field

[0001] This invention relates to the field of lithium battery technology, and in particular to a method for estimating the state of health of lithium-ion batteries that integrates auxiliary information and multi-condition data. Background Technology

[0002] State of Health (SOH) estimation of lithium-ion batteries is a core indicator for measuring battery degradation and is crucial for ensuring safety and lifespan management in fields such as electric vehicles, grid energy storage, and portable electronic devices. When SOH drops to 70%-80%, the battery needs to be replaced; therefore, accurate SOH estimation is key to the development of Battery Management Systems (BMS). Currently, publicly available SOH research is mainly based on open-source laboratory data, including NASA's lithium battery dataset, experimental data from the University of Maryland (CALCE), and battery data from Sandia National Laboratories (SNL). While these experimental data mimic real-world battery data to some extent, they differ significantly from the complex operating conditions of real vehicles (such as random charging and discharging, temperature fluctuations, and differences in driving habits), leading to large errors in practical applications. Furthermore, most existing SOH estimation methods use a single feature for estimation, failing to comprehensively consider multi-dimensional battery characteristics and lacking validation with real-world battery operating data.

[0003] Existing SOH estimation methods mainly fall into two categories: model-based methods and data-driven methods. Model-based methods can be further divided into equivalent circuit models, electrochemical models, and mathematical models. Equivalent circuit models are based on the electrical characteristics of lithium-ion batteries, using basic electrical components such as constant voltage sources, resistors, and capacitors to establish models that approximate or represent the dynamic characteristics of the battery system. Common models include the Rint model, RC network model, PNGV model, and Thevenin model. These models have well-defined structures and good scalability, but their computational complexity is relatively high, making them difficult to apply in end-user scenarios. Electrochemical models, on the other hand, establish partial differential equations based on the internal electrochemical reaction mechanisms of the battery to describe its internal operation and aging state. For example, a SP-based SOH estimation model has been proposed based on the chemical and mechanical degradation mechanisms of the battery. These methods have fewer parameters and can quickly and accurately predict battery capacity decay. However, the models are highly dependent on the battery reaction mechanisms, thus requiring modelers to have strong professional skills, making them difficult to widely promote and popularize. Mathematical modeling methods mainly estimate battery aging by establishing the relationship between aging characterization parameters and SOH. Incremental Capacity Analysis (ICA) is a common mathematical modeling method. This type of method can improve the estimation accuracy and robustness of the model while reducing computational complexity, and has certain practical application needs. However, this method requires manually establishing battery aging characteristic parameters and has high requirements for model selection. Therefore, it still places high demands on modelers and is difficult to promote universally.

[0004] Against the backdrop of rapid development in artificial intelligence technology, data-driven SOH estimation methods are being widely used and have become the mainstream estimation method for lithium-ion batteries. Data-driven methods do not require consideration of the complex electrochemical reaction mechanisms within the battery; they only need to establish a mapping relationship between aging data (such as current, voltage, and temperature) and the battery's SOH, and then use a trained estimation model to estimate the target battery. Common methods include Support Vector Machines (SVM), Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM). These methods have advantages such as strong generalization performance, high estimation accuracy, and fast estimation speed, but the models typically require high data quality, have a large number of parameters, and demand significant computational resources.

[0005] With the acceleration of electrification and the continuous expansion of renewable energy storage, developing high-precision, real-time, and adaptable SOH estimation methods has become an inevitable requirement and important direction for technological progress and application in the industry. However, existing methods are mainly based on experimental data, resulting in large computational loads, limited data types, and low model generalization and robustness, making them difficult to apply in real-world scenarios. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a lithium-ion battery health state estimation method that integrates auxiliary information and multi-condition data. The aim is to construct a SOH estimation method that integrates battery auxiliary information, multiple operating condition data, and deep learning technology, enabling it to achieve high-precision, highly generalizable, interpretable, and real-time health state estimation under various types of lithium-ion batteries and complex operating conditions. This provides a safe, reliable, and efficient battery management technology for electric vehicles, energy storage systems, and other applications.

[0007] A method for estimating the health status of a lithium-ion battery that integrates auxiliary information and multi-condition data, the method comprising:

[0008] Step S1: Load the hybrid battery dataset. Within a preset voltage range, extract charging segment data to obtain the charging voltage sequence and its corresponding charging current and charging temperature sequences. Perform monotonically non-decreasing processing on the charging voltage, charging current, and charging temperature sequences, and resample these three sequences at fixed intervals using linear interpolation. Then, extract the battery capacity value corresponding to each sample data as a label to form a battery capacity label sequence. Next, extract four features from each sample: battery type, nominal battery capacity, battery operating temperature, and battery discharge rate to obtain an auxiliary information vector sequence. Finally, input the resampled charging voltage sequence, charging current sequence, charging temperature sequence, battery capacity label sequence, and auxiliary information vector sequence into the features.

[0009] Step S2: Divide the input features into training set and test set, and construct a hybrid neural network model to process the training set data in order to estimate the battery health status of the training set. The hybrid neural network model includes convolutional neural network, long short-term memory network and attention mechanism.

[0010] Step S3: Use the SHAP value algorithm to perform interpretability analysis on the hybrid neural network model and obtain the global importance of each feature. ;

[0011] Step S4, based on the global importance of features Pruning and quantization operations are performed on the hybrid neural network model.

[0012] Further, in step S1, for the first The voltage sequence is extracted from the data from the second cycle. V i = [ v 1 i , v 2 i ， ⋯ , v m i ] Current sequence I i = [ a 1 i , a 2 i , ⋯ , a m i ] and temperature sequence T i = [ t 1 i , t 2 i , ⋯ , t m i ] , The total number of sampling points within a fixed voltage range is used, with each cycle's data treated as a single sample. A i = [ V i , I i , T i ] Simultaneously calculate the battery capacity within a fixed voltage range for this cycle. As the health status label values ​​for the samples, all cyclic samples are finally aggregated to obtain the battery sequence dataset. X = [ A 1 , A 2 , ⋯ , A n ] and its corresponding set of health status label values Y = [ y 1 , y 2 , ⋯ , y n ] , Total number of cycles, battery capacity It is obtained by the following formula for the integral of ampere-hours:

[0013] ;

[0014] in, For the first During the second iteration Instantaneous current at each sampling point For the first During the second iteration Instantaneous current at each sampling point For the first The time of each sampling point Indicates the first One sampling point, ;

[0015] Then, the voltage, current, and temperature sequences were processed to be monotonically non-decreasing, and linear interpolation was used to resample these three sequences at fixed intervals to obtain new voltage sequences. U i = [ u 1 i , u 2 i ， ⋯ , u r i ] Current sequence Cur i = [ cur 1 i , cur 2 i , ⋯ , cur r i ] and temperature sequence Tem i = [ time 1 i , time 2 i , ⋯ , time r i ] The three new sequence data obtained will be combined to form a new sample. A ¯ i = [ U i , Cur i , Tem i ] The processed battery sequence data is obtained through n iterations. X ¯ = [ A ¯ 1 , A ¯ 2 , ⋯ , A ¯ n ] .

[0016] Further, in step S1, an auxiliary information vector is defined. AUX = [ T type ， C a , T eh , R eat ] Where Type indicates the battery type. This refers to the battery's nominal capacity. The ambient temperature of the battery. This refers to the battery discharge rate. With battery sequence data Together with the label Y, they serve as the final input features. ,Right now Z = [ X ¯ , Y , AUX ] .

[0017] Further, step S2 specifically includes:

[0018] The input features are divided into training and test sets. The training set is fed into a convolutional neural network to extract local features, and then a long short-term memory network is used to capture global temporal dependencies. Subsequently, a self-attention mechanism is used to assign differential weights to these features, and the features weighted by the self-attention mechanism are concatenated with the auxiliary information vectors in the training set. Finally, the concatenated comprehensive features are fed into a fully connected layer to complete the estimation of the battery health status of the training set.

[0019] Furthermore, step S3 specifically includes:

[0020] set up For a well-trained hybrid neural network model, on the training set Randomly select M samples from the middle, for the first M samples... Sample A ¯ i = [ U i , Cur i , Tem i ] = [ u 1 i , u 2 i ， ⋯ , u r i , cur r + 1 i , cur r + 2 i , ⋯ , cur 2 r i , time 2 r + 1 i , time 2 r + 2 i , ⋯ , time 3 r i ] ,Will The characteristic elements are uniformly used It means that it has been obtained. A ¯ i = [ u 1 i , u 2 i ， ⋯ , u 3 r i ] The SHAP value of each feature in the sample is calculated using the SHAP algorithm, as follows:

[0021] For the Each sample has a SHAP value of ,in, For hybrid neural network models The average of the predicted values ​​for each sample. For the first The predicted value of each sample under the hybrid neural network model For the first In the nth sample The SHAP value of each feature, ;

[0022] The SHAP value of a single feature is calculated as follows:

[0023] f g i = ∑ S | S |!( 3 r − | S | − 1 )! ( 3 r )! [ f ( S ∪ { u g i }) − f ( S )] ;

[0024] in, For features not included All feature subsets, , For set The number of elements, For factorial operation, For the hybrid neural network model obtained through training, For this neural network model, a subset of features The predicted value, Indicates the inclusion of features Feature set, This represents the union operation;

[0025] No. Global importance of each feature Defined as:

[0026] ;

[0027] The larger the value, the greater the positive effect of the feature on the model. Arranging the values ​​in descending order can identify the key factors that have the greatest impact on SOH prediction.

[0028] Furthermore, the pruning in step S4 includes weight pruning and neuron pruning, applying the pruning to 50% of the sorted neurons. The absolute value of the weight corresponding to the feature is calculated. If the absolute value is lower than the preset threshold, the weight is determined to be unimportant and set to zero. For neuron pruning, the average absolute value of the activation value of each neuron on the training set is calculated. If the average value is lower than the set threshold, the output of the neuron is forcibly set to zero. Then, a linear quantization method is used to map the full-precision floating-point weight parameters to integer parameters.

[0029] Furthermore, the weighted pruning is specifically as follows:

[0030] Let the fully connected layer in a convolutional neural network be the... The weight matrix of the layer is The output after learning through this layer is:

[0031] ;

[0032] ;

[0033] in, and These represent the input and output features of a fully connected layer in a convolutional neural network, respectively. and These represent the number of input features and the number of output features in a fully connected layer of a convolutional neural network, respectively. For bias terms;

[0034] For the weight matrix Each weight in If its absolute value is less than a preset threshold If the value is zero, then set it to zero; otherwise, retain the original value. The specific definition is as follows:

[0035] .

[0036] Furthermore, neuron pruning specifically involves:

[0037] For the The first in the layer The importance of a neuron is determined by calculating its overall activation value. To measure this, the calculation process is as follows:

[0038] ;

[0039] ;

[0040] in, It is an activation function. For the first The first in the layer The activation value of each neuron;

[0041] if Less than the set threshold If the neuron's contribution to the network output is small, it should be pruned by forcing its activation value to zero. .

[0042] Furthermore, the quantitative operation specifically involves:

[0043] Set the weight matrix in the fully connected layer network. for:

[0044] ;

[0045] Each weight value is a high-precision floating-point number. The original floating-point weights are linearly mapped to an 8-bit integer range, i.e., 0~255. The quantification formula is:

[0046] ;

[0047] in, These are the original high-precision floating-point number weight parameters. For the calculated low-precision fixed-point numbers, and These are the minimum and maximum values ​​in the original weighted sample. This is the rounding operation.

[0048] To accurately estimate the state of health (SOH) of a battery, this invention has systematically optimized both data acquisition and model selection, thereby effectively improving the accuracy of prediction and the robustness of the model. Compared with the prior art, the beneficial technical effects of this invention are mainly reflected in the following aspects:

[0049] Extracting charging segment data and constructing auxiliary information vectors: At the data level, this invention significantly improves data quality through refined processing of charging segment data, thereby improving the accuracy of SOH prediction. First, by setting high-voltage segment data as the core input feature, this processing method is more in line with the actual operating conditions of the battery. Monotonic non-decreasing processing effectively eliminates noise interference and avoids possible errors during the acquisition process. Compared with the manual feature extraction of existing methods, this method better preserves the original characteristics of the data and ensures data accuracy. Second, in the high-voltage range, this invention uses interpolation to transform densely sampled voltage data... Transforming points into discrete points as input to the model not only preserves the most critical degradation information during battery degradation but also significantly reduces the computational complexity of the model. This process greatly improves data utilization efficiency and avoids the impact of redundant data on computation. Finally, to ensure data consistency and balance, this invention fixes the number of data samples for each battery at 1000 and merges data samples from multiple battery types to form a unified dataset. This method can effectively improve the model's generalization ability and avoid the problem of training only on data under a single operating condition in traditional methods, thereby improving the model's robustness under different usage conditions.

[0050] Constructing Auxiliary Information and Multi-Condition Hybrid Data: This invention constructs battery auxiliary information and introduces it into the model training process. Considering the diverse types of batteries and complex operating conditions in practical applications, this invention improves the accuracy and applicability of SOH prediction by introducing battery auxiliary information, thereby achieving precise guidance of the model and optimizing its predictive ability. Specifically, this invention extracts four important battery characterization information: battery type (different types of batteries exhibit different behaviors during degradation), battery standard capacity (the nominal capacity of the battery is a benchmark reflecting its health status), operating temperature (temperature has a profound impact on battery performance and is one of the key factors in battery degradation), and discharge rate (high-rate discharge accelerates battery degradation and therefore plays an important role in predicting SOH). This information is concatenated to the features learned by the model in the final stage of model training, through fusion... The combined features are further used for SOH prediction. In this way, the model can make more accurate predictions based on the specific battery type and usage environment. To enhance the model's adaptability and generalization ability, this invention fuses data from different battery types and operating conditions to form a training set containing 23,000 samples. These samples come from battery data of 23 different types and operating conditions. The introduction of this multi-condition mixed dataset enables the model to better adapt to the health status prediction needs of various batteries. Compared with traditional methods, this invention introduces more battery characterization information and diverse datasets, making the prediction results not only highly accurate in experimental environments, but also better able to cope with the changing operating conditions in real-world applications. This fusion method significantly enhances the robustness of the model, enabling it to exhibit stronger adaptability and stability in a wide range of practical applications.

[0051] This invention constructs a CNN-LSTM-ATTENTION hybrid neural network model: By integrating multiple deep learning methods, it significantly improves the predictive ability of battery state of health (SOH). First, a convolutional neural network (CNN) is used to extract local features from battery data. CNNs can automatically identify local features during battery degradation, such as voltage fluctuations and abnormal changes during charging and discharging. This process effectively captures subtle changes during battery aging, especially short-term performance degradation signals, avoiding the limitations of traditional methods that rely on manual feature extraction. Next, a long short-term memory (LSTM) network is introduced to capture long-term dependencies in battery data. Battery degradation often exhibits a time-series nature, and LSTM can effectively remember important information from historical data, helping the model understand the long-term trend of battery degradation and make more accurate predictions about future states. This approach can capture changes in battery performance over time. The dynamic information is incorporated into the prediction, further improving the model's accuracy. Finally, the Attention mechanism assigns different weights to the features at each time step, thus focusing more on the most critical moments in the battery degradation process. The introduction of Attention enables the model to automatically learn which time periods are most important for SOH prediction, thereby enhancing the model's flexibility and accuracy. In addition, by combining this auxiliary information with the deep features of battery degradation, the model can more accurately reflect the health status of different battery types and operating conditions. This combination enables the model to not only accurately capture local features and long-term trends, but also to make fine adjustments for the actual operating conditions of different batteries, thereby obtaining better prediction results. Through this multi-dimensional and multi-technology combination, this invention achieves comprehensive and accurate estimation of SOH, which can better cope with the degradation problems of various battery types and their different operating conditions, and greatly enhances the practical application value of battery SOH prediction.

[0052] Model Optimization and Interpretability Analysis Based on SHAP Value Algorithm: This invention uses the SHAP value algorithm to perform model interpretability analysis and dynamically adjusts the weight parameters of each feature during training based on the analysis results to improve model performance.

[0053] Pruning and Quantization: This invention employs pruning and quantization as key technologies for model compression, providing significant advantages for deploying the model in resource-constrained environments. The core advantage of pruning lies in its effective improvement of model efficiency. By removing unimportant weights (weight pruning) or neurons with low activation values ​​(neuron pruning), it significantly reduces model complexity and the number of parameters. This not only reduces storage usage but also greatly improves computation speed, while mitigating overfitting to some extent and enhancing the model's generalization ability. The core advantage of quantization lies in its extreme optimization of storage and computation. It converts model weights and activation values ​​from 32-bit floating-point numbers to low-precision formats such as 8-bit integers, making the model... The model size is significantly reduced (typically by 75%). This reduction in precision means that computing speed can be significantly improved on hardware that supports low-precision operations (such as mobile chips and edge computing devices), while memory bandwidth pressure is also effectively alleviated. More importantly, when pruning and quantization are used together, they can produce a synergistic optimization multiplier effect. The sparse model structure after pruning provides a more stable numerical distribution for quantization, while quantization further amplifies the storage and bandwidth benefits brought by pruning. The two methods work together to enable complex models to achieve efficient, low-latency real-time inference on edge devices such as embedded devices and mobile terminals while maintaining acceptable precision loss, which greatly expands the application boundaries of artificial intelligence. Attached Figure Description

[0054] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart of a lithium-ion battery health status estimation method that integrates auxiliary information and multi-condition fusion data, provided by an embodiment of the present invention.

[0056] Figure 2 This is a schematic diagram of the battery capacity and voltage aging with the number of cycles in the SNL dataset provided in this embodiment of the invention, wherein (a) is an NCA battery, (b) is an NMC battery, and (c) is an LFP battery;

[0057] Figure 3 This is a schematic diagram of the data preprocessing process provided in an embodiment of the present invention;

[0058] Figure 4 This is a schematic diagram of the process for constructing auxiliary information and mixed data provided in the embodiments of the present invention;

[0059] Figure 5 This is a schematic diagram of the structure of the hybrid neural network model provided in an embodiment of the present invention;

[0060] Figure 6 This is a comparative schematic diagram of pruning before and after the pruning process provided in an embodiment of the present invention;

[0061] Figure 7 This is a comparative diagram of the quantization operation before and after, provided in an embodiment of the present invention. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0063] This invention proposes a method for estimating the state of health of lithium-ion batteries that integrates auxiliary information and multi-condition data, such as... Figure 1 As shown, the method includes the following steps:

[0064] S1: Extracting charging segment data and constructing auxiliary information vectors

[0065] S11: Load the hybrid battery dataset. Within a preset voltage range, extract the charging voltage segment and its corresponding charging current and charging temperature segments for each battery in each cycle. Iterate through all batteries to obtain the charging voltage sequence and its corresponding charging current and charging temperature sequences. Subsequently, perform monotonically non-decreasing processing on the charging voltage, charging current, and charging temperature sequences. Finally, use a linear interpolation method to resample these three sequences at fixed intervals to generate new sequences of uniform length and equal intervals.

[0066] S12: Treat each loop as a sample data, extract its corresponding battery capacity value as a label, traverse the battery dataset to obtain the battery capacity values ​​of all samples, and form a battery capacity label sequence.

[0067] S13: Extract four features from each sample: battery type, nominal battery capacity, battery operating temperature, and battery discharge rate, and construct them as auxiliary information vectors; traverse the battery dataset to obtain the auxiliary information vector sequence corresponding to all samples; finally, use the three new sequences after resampling (i.e., charging voltage, charging current, and charging temperature sequences), the battery capacity label sequence, and the auxiliary information vector sequence as input to the model.

[0068] Hybrid battery dataset used in this invention The data comes from battery data from Sandia National Laboratories (SNL) and two real-world energy storage charge-discharge datasets. The SNL dataset consists of three types of batteries: lithium nickel cobalt aluminum oxide (NCA), lithium nickel cobalt manganese oxide (NMC), and lithium iron phosphate (LFP). These batteries have different discharge rates, depths of discharge, and ambient discharge temperatures, as detailed in Table 1. During battery use, one charge and one discharge cycle constitute one battery cycle. As the number of cycles increases, the battery gradually begins to age, with the most obvious change being a decrease in battery capacity. Therefore, the change in battery capacity is commonly used to characterize the battery's State of Health (SOH), which is defined as follows:

[0069] ;

[0070] in, This represents the battery's state of equilibrium (SOH) under the capacity definition. Indicates the current available capacity of the battery. This indicates the available capacity of the new battery.

[0071] Table 1 SNL Battery Data Types

[0072]

[0073] As the number of battery cycles increases, its capacity shows a clear downward trend. Figure 2 This paper presents the relationship between battery voltage and capacity as a function of cycle number in the SNL battery dataset. (a) represents an NCA battery, (b) an NMC battery, and (c) an LFP battery. The graph shows that battery capacity gradually decreases with cycling, reflecting the battery aging process. The graph labels two battery capacity values, e1 and e2, at the same voltage position for different cycle numbers. The color depth corresponds to the cycle number; darker colors indicate more cycles and a higher degree of battery aging. Clearly, the value of e2 is significantly smaller than e1, indicating that battery aging intensifies and capacity significantly decreases with increasing cycle number. Furthermore, in practical applications, battery capacity is also affected by ambient temperature; temperature changes also have a certain impact on its performance.

[0074] Therefore, by extracting charging current, charging voltage, and charging temperature data from the hybrid battery dataset and calculating the rechargeable battery capacity, this information can effectively characterize the battery's health status and degradation trend.

[0075] according to Figure 2 As shown, the standard voltage of NCA and NMC batteries is 4.3V, while the standard voltage of LFP batteries is 3.7V. In actual use, most users, concerned about battery capacity, tend to operate the batteries in a higher capacity range, corresponding to a higher voltage range. Based on this, this invention, drawing on practical application experience, selects the voltage range of NCA and NMC batteries as [4.0, 4.198] and the voltage range of LFP batteries as [3.4, 3.598], and collects corresponding voltage, current, and temperature data within this range.

[0076] For the The data from the next cycle yields the voltage sequence. V i = [ v 1 i , v 2 i ， ⋯ , v m i ] Current sequence I i = [ a 1 i , a 2 i , ⋯ , a m i ] and temperature sequence T i = [ t 1 i , t 2 i , ⋯ , t m i ] , The total number of sampling points within a fixed voltage range is used, with each cycle's data treated as a single sample. A i = [ V i , I i , T i ] Simultaneously calculate the battery capacity within a fixed voltage range for this cycle. As the health status label values ​​for the samples, all cyclic samples are finally aggregated to obtain the battery sequence dataset. X = [ A 1 , A 2 , ⋯ , A n ] and its corresponding set of health status label values Y = [ y 1 , y 2 , ⋯ , y n ] , Total number of cycles, battery capacity It is calculated using the following formula for the integral of ampere-hours:

[0077] ;

[0078] in, For the first During the second iteration Instantaneous current at each sampling point For the first During the second iteration Instantaneous current at each sampling point For the first The time of each sampling point Indicates the first One sampling point, .

[0079] Subsequently, based on the physical law that voltage, current, and temperature all exhibit monotonicity during battery charging, the extracted sequence fragments underwent corresponding monotonicity processing. The processing flow is as follows: Figure 3 As shown.

[0080] In this invention, monotonicity processing is achieved by iteratively comparing each element in a loop for voltage sequences. V i = [ v 1 i , v 2 i ， ⋯ , v m i ] Current sequence I i = [ a 1 i , a 2 i , ⋯ , a m i ] and temperature sequence T i = [ t 1 i , t 2 i , ⋯ , t m i ] The processing procedure is as follows:

[0081] ;

[0082] in, , , The first The first value in the voltage, current, and temperature sequences in the next cycle. , , The first In the second cycle, the voltage sequence, current sequence, and temperature sequence are in the first... One value, The loop count is given. After the above processing, a monotonically non-decreasing voltage sequence is obtained. V ¯ i = [ v ¯ 1 i , v ¯ 2 i ， ⋯ , v ¯ m i ] This yields a monotonically non-decreasing current sequence. I ¯ i = [ a ¯ 1 i , a ¯ 2 i , ⋯ , a ¯ m i ] and temperature sequence T ¯ i = [ t ¯ 1 i , t ¯ 2 i , ⋯ , t ¯ m i ] .

[0083] To address the issues of unequal time intervals and a large number of sampling points in the original voltage, current, and temperature sequence data, this invention employs linear interpolation to interpolate the original voltage, current, and temperature sequence data. Data from each sampling point is resampled at fixed intervals to obtain... A new sequence of data from each sampling point. The specific method is as follows:

[0084] ;

[0085] in, The first During the second iteration The voltage, current, and temperature values ​​at each sampling point; and They are the first During the second iteration The voltage value of the sampling point and the first sampling point Voltage values ​​at each sampling point; and They are the first During the second iteration The current value at the sampling point and the first sampling point Current values ​​at each sampling point; and They are the first During the second iteration The temperature value of the sampling point and the first sampling point Temperature values ​​at each sampling point. For the first During the second iteration The time of each sampling point j ∈ [ 1 , 2 , ⋯ , r ] , k ∈ [ 1 , 2 , ⋯ , m ] △t is the time interval, which is 5 seconds in this embodiment.

[0086] This method generates new voltage data points at fixed intervals by performing linear interpolation between adjacent data points, thereby transforming a monotonically decreasing voltage, current, and temperature sequence into a uniform and dense voltage sequence on the time axis. U i = [ u 1 i , u 2 i ， ⋯ , u r i ] Current sequence Cur i = [ cur 1 i , cur 2 i , ⋯ , cur r i ] and temperature sequence Tem i = [ time 1 i , time 2 i , ⋯ , time r i ] This provides a regular data foundation for subsequent model training and effectively eliminates the misalignment problem of periodic data caused by the randomness of sampling time. The three obtained sequence data are then used to form a new sample. A ¯ i = [ U i , Cur i , Tem i ] The processed battery sequence data is obtained through n iterations. X ¯ = [ A ¯ 1 , A ¯ 2 , ⋯ , A ¯ n ] .

[0087] Finally, the battery auxiliary information and multi-condition hybrid data are constructed, and the process is as follows: Figure 4 As shown.

[0088] To improve the accuracy and generalization ability of lithium-ion battery state of health (SOH) estimation, an auxiliary information vector (AUX) is introduced as prior guidance information for the model. This auxiliary information vector is used to characterize the key attributes and operating conditions of the battery, thereby providing guidance and differentiation during model training and inference.

[0089] First, define the auxiliary information vector. AUX = [ T type ， C a , T eh , R eat ] Where Type indicates the battery type. , The nominal capacity of the battery is represented by the following values:

[0090] ;

[0091] Battery ambient temperature Battery discharge rate C is the unit of discharge rate, where 1C represents the current required to discharge the battery to its nominal capacity in one hour. Therefore, for an NMC battery with a nominal capacity of 3.0 Ah, an ambient temperature of 15℃, and a discharge rate of 0.5C, its auxiliary information vector is: C ell aux = [ 0 , 3 . 0 , 15 , 0 . 5 ] In the model's input, With input battery sequence data Together with the label Y, they serve as the final input features. ,Right now Z = [ X ¯ , Y , AUX ] .

[0092] S2: Divide the input features into training and testing sets, and construct a hybrid neural network model to process the training set data.

[0093] S21: The above input features A rolling window approach was used for partitioning, with 10 consecutive samples forming a group, and each group divided into test sets at a ratio of 1:9. Z test = [ X ¯ test ， Y test ， AUX test ] With training set Z train = [ X ¯ train ， Y train ， AUX train ] Subsequently, the training set samples are shuffled and saved, along with the corresponding auxiliary information vectors of the shuffled samples. and health status label sequence ;

[0094] S22: The training set is input into a convolutional neural network to extract local features, and then a long short-term memory network is used to capture global temporal dependencies. Subsequently, a self-attention mechanism is used to assign differential weights to these features, and the weighted features are then compared with the auxiliary information vector in the training set. The data is then stitched together, and finally, the combined features are input into a fully connected layer to estimate the health status of the training set cells.

[0095] Specifically, this invention directly inputs voltage, current, temperature, and auxiliary information data from the training set as features into a hybrid neural network model constructed from convolutional neural networks, long short-term memory networks, and attention mechanisms, without the need for separately designed statistical features. The process is as follows: Figure 5 As shown. The above input features First, a Convolutional Neural Network (CNN) is used to extract local information contained in the battery data. The CNN then processes the input features... Battery sequence data To perform convolution operations, use a convolution kernel. ( (where 3 is the kernel size and 3 is the dimension of the input features). The convolution operation is represented as:

[0096] ;

[0097] in, This is the output after passing through the convolutional neural network. It is a convolutional neural network model. This represents the convolution operation. is the bias term of the convolutional layer, and ReLU is the activation function.

[0098] Next, the output features obtained from the convolutional neural network are... It serves as the input to a Long Short-Term Memory (LSTM) network model. LSTM possesses memory capabilities, effectively capturing the temporal features between nodes and enhancing the accuracy of the model's State of Interest (SOH) estimation. The core idea of ​​LSTM is to introduce three gates, namely the input gates. Forgotten Gate and output gate This is used to control the transmission and updating of information. Its main operations process the input data using the following formula:

[0099] { i t = s ( oh i ⋅ [ h t − 1 , X CNN ] + b i ) f t = s ( oh f ⋅ [ h t − 1 , X CNN ] + b f ) o t = s ( oh o ⋅ [ h t − 1 , X CNN ] + b o ) e t = tanh( oh e ⋅ [ h t − 1 , X CNN ] + b e ) c t = f t ⊙ c t − 1 + i t ⊙ e t h t = o t ⋅ tanh ( c t ) ;

[0100] Where σ is the sigmoid activation function, which outputs a value between 0 and 1; These are the weight matrices for the input gate, forget gate, and output gate, respectively. These are the bias terms for the input gate, forget gate, and output gate, respectively. It represents the features after the convolution operation; h t-1 This indicates the hidden state at the previous moment. It is a candidate state. It is the weight matrix in the candidate state. It is the bias vector, and tanh is the hyperbolic tangent activation function. It represents the current state of the memory unit update. It is the updated state of the memory unit in the previous moment. This indicates element-wise multiplication. This indicates the output status at the current moment.

[0101] The convolutional features After computation by the LSTM network, the current output state features are obtained. ,make = .

[0102] The features obtained at this time The aging local features and correlations between aging data in battery data have been learned. However, because this invention directly inputs voltage, current, and temperature data into the network model, the computation becomes complex, and the importance of each feature cannot be measured. Therefore, the output of the LSTM is... The input is fed into the attention mechanism network, and the operational principle of the attention mechanism is as follows:

[0103] ;

[0104] in, The feature output after the attention mechanism operation. These are query, key, and value matrix. Represents a matrix space with 100 rows and 3 columns. , , ,in , , These are the weight parameters in the query, key, and value matrices, respectively, and their initial values ​​are all randomly generated values ​​in the range [-0.1, 0.1]. It is the dimension of the key, here Softmax is a function used for normalization, ensuring that the weights sum to 1.

[0105] Through attention mechanisms Through computation, the model can dynamically assign appropriate weights to input features, thereby highlighting key information and downplaying secondary or redundant content. This process allows the model to autonomously identify and focus on the most relevant parts of the input, assigning them appropriate attention based on their importance, ultimately enhancing the model's representational ability and generalization performance, and achieving better training results.

[0106] At this time To learn the features of all information and to perform SOH estimation more accurately, this invention introduces auxiliary information as prior information to guide the model in SOH estimation. This is done by using the auxiliary information vectors from the training set. and The features are concatenated and then passed through a fully connected network. The output of this network... This is the predicted value of the battery's health status, and its mathematical calculation process is as follows:

[0107] ,in, For the weight parameters of a fully connected network, This is the bias term for a fully connected network.

[0108] S3: Interpretability analysis of hybrid neural network models using the SHAP value algorithm.

[0109] S31: On the training set, first calculate the SHAP value of each feature in each sample, and then at the global level, summarize the absolute values ​​of the SHAP values ​​of each feature in all samples to evaluate the overall impact of each feature on the model's prediction.

[0110] Specifically, in the above-mentioned hybrid neural network model, in order to deeply analyze the contribution of each feature in the voltage sequence, current sequence and temperature sequence to the prediction results, this invention uses the Shapley Additive Explanatory (SHAP) algorithm to perform model interpretability analysis. It mainly analyzes the importance of the extracted voltage, current and temperature values ​​within the fixed voltage range in the model, and dynamically adjusts the weight parameters of each feature during training based on the analysis results to improve model performance.

[0111] The Shapley Value Algorithm (SHAP), based on Shapley values ​​in cooperative game theory, is used to fairly allocate the contribution of each feature in a model to the prediction results. As a cutting-edge method for model interpretability, SHAP offers several significant advantages over other interpretation tools. First, its theoretical foundation is solid, originating from Shapley values ​​in cooperative game theory, providing a feature contribution allocation scheme supported by robust mathematical axioms, ensuring fairness and rationality in the interpretation. Second, SHAP satisfies ideal interpretability properties, including local accuracy (ensuring consistency between the interpretation of a single prediction and the actual model output) and consistency (i.e., if a model change increases the contribution of a feature, its SHAP value should also increase accordingly), a theoretical guarantee lacking in many other methods. Furthermore, the SHAP algorithm can provide both local and global perspectives simultaneously. It can not only accurately quantify the contribution of each feature to the prediction of a single sample, enabling detailed case analysis, but also assess the global importance of features by aggregating sample-level results, thus comprehensively revealing the model's decision-making mechanism. In addition, the SHAP algorithm has excellent model versatility. As a model-independent post-interpretation method, it can be widely applied to various black-box models (such as neural networks, random forests, and gradient boosting trees), greatly expanding its application scenarios.

[0112] The process of interpretability analysis of hybrid neural network models using the SHAP value algorithm is as follows:

[0113] Assumption For the above-trained hybrid neural network model, in the training set Randomly select M samples from the middle, for the first M samples... Sample A ¯ i = [ U i , Cur i , Tem i ] = [ u 1 i , u 2 i ， ⋯ , u r i , cur r + 1 i , cur r + 2 i , ⋯ , cur 2 r i , time 2 r + 1 i , time 2 r + 2 i , ⋯ , time 3 r i ] ,Will The characteristic elements are uniformly used It means that it has been obtained. A ¯ i = [ u 1 i , u 2 i ， ⋯ , u 3 r i ] The SHAP value of each feature in the sample is calculated using the SHAP value algorithm. The process is as follows:

[0114] For the Each sample has a SHAP value of ,in, For hybrid neural network models The average of the predicted values ​​for each sample. For the first The predicted value of each sample under the hybrid neural network model For the first In the nth sample The SHAP value of each feature, .

[0115] The formula for calculating the SHAP value of a single feature is:

[0116] f g i = ∑ S | S |!( 3 r − | S | − 1 )! ( 3 r )! [ f ( S ∪ { u g i }) − f ( S )] ;

[0117] in, For features not included All feature subsets, ; For set The number of elements, For factorial operation, For the hybrid neural network model obtained through training, For this neural network model, a subset of features The predicted value, Indicates the inclusion of features Feature set, This indicates the union operation.

[0118] At the global level, by summing the absolute values ​​of the SHAP values ​​of each feature across all samples, the overall impact of each feature on the model's predictions can be assessed. Global importance of each feature Defined as: , The larger the value, the greater the positive effect of the feature on the model. By sorting the values ​​in descending order, the key factors that have the greatest impact on SOH prediction can be identified.

[0119] Based on the global importance of each feature Introducing feature weight parameters during model training to make them correlate with... Proportional to the model, the weighted loss function allows the model to focus more on important features and suppress the interference of noisy features, thereby improving the model's prediction accuracy and robustness. This method not only enhances the interpretability of the model but also provides a quantitative basis for feature selection and model optimization, making it suitable for applications requiring high reliability, such as battery health status prediction.

[0120] S4: Global Importance of Features Pruning and quantization operations on hybrid neural network models

[0121] The global importance of each feature calculated above For the 50% after sorting The absolute value of the weight corresponding to the feature is calculated. If the absolute value is lower than the preset threshold, the weight is determined to be unimportant and set to zero. For neuron pruning, the average absolute value of the activation value of each neuron on the training set is calculated. If the average value is lower than the set threshold, the output of the neuron is forcibly set to zero. Then, a linear quantization method is used to map the full-precision floating-point weight parameters to 8-bit integer parameters.

[0122] To reduce model complexity and improve inference efficiency, this invention is based on the global importance of features. Pruning and quantization operations are performed on the hybrid neural network model. The pruning process is divided into two levels: weight pruning and neuron pruning, such as... Figure 6 As shown, the details are as follows:

[0123] Weight pruning aims to identify and remove unimportant weight parameters in hybrid neural networks, i.e., setting some weights to zero, thereby sparsifying the network. This invention focuses on the fully connected layers (assuming to be the first) in convolutional neural networks. Perform operations on the first layer, let the first layer be the first layer. The weight matrix of the layer is The output after learning through this layer is:

[0124] ;

[0125] ;

[0126] in, and These represent the input and output features of a fully connected layer in a convolutional neural network, respectively. and These represent the number of input features and the number of output features in a fully connected layer of a convolutional neural network, respectively. This is a bias term.

[0127] This invention employs a threshold-based pruning strategy: for the weight matrix Each weight in If its absolute value is less than a preset threshold If the value is zero, then set it to zero; otherwise, retain the original value. The specific definition is as follows:

[0128] ;

[0129] This process can effectively remove weights that have little impact on the output, thus achieving model compression.

[0130] Neuron pruning: In addition to weight-level pruning, this invention also implements neuron-level pruning to further simplify the network structure. After completing the above weight pruning, neuron pruning is further performed on the fully connected layers in the convolutional neural network. For the first... The first in the layer The importance of a neuron is determined by calculating its overall activation value. To measure this, the calculation process is as follows:

[0131] ;

[0132] ;

[0133] in, It is an activation function. For the first The first in the layer The activation value of each neuron.

[0134] if Less than the set threshold If the neuron's contribution to the network output is considered small, it is pruned, meaning its activation value is forcibly set to zero. This operation removes redundant neurons, reducing subsequent computation. By traversing all neurons, neurons that contribute less to the output are removed, thus reducing the complexity of the hybrid neural network.

[0135] Next, based on the pruned model, to further reduce model storage usage and improve inference speed, the weight matrix in the fully connected layer network was adjusted. After quantization, due to neuron pruning, the number of inputs and outputs of the fully connected layer features (neurons) are now respectively... and Quantization aims to map the weight values, represented by high-precision floating-point numbers (such as FP32) in the weight matrix of a fully connected layer network, to low-precision fixed-point numbers (such as INT8). The process is as follows: Figure 7 As shown, the specific operation is as follows:

[0136] ;

[0137] Each weight value is a high-precision floating-point number. The original floating-point weights are linearly mapped to an 8-bit integer range (0~255). The quantification formula is: ,in, These are the original high-precision floating-point number weight parameters. For the calculated low-precision fixed-point numbers, and These are the minimum and maximum values ​​in the original weighted sample. This is the rounding operation.

[0138] After quantization, the original floating-point weights are mapped to... Figure 7 The integer values ​​shown (such as 68, 19, 58, etc.) realize the compression from 32-bit floating-point numbers to 8-bit integers. The quantized model can perform efficient inference on hardware that supports integer operations, greatly improving the computational efficiency of the model.

[0139] This invention is the first to propose combining battery auxiliary information (battery type, standard capacity, operating temperature, and discharge rate) with convolutional neural networks (CNN), long short-term memory networks (LSTM), and self-attention mechanisms to construct a novel battery state of health (SOH) estimation model. Furthermore, it integrates data from different battery types and operating conditions into a unified dataset for training, overcoming the limitations of single datasets in traditional methods. In addition, it standardizes the training samples to a fixed length, resolving data inconsistency and imbalance issues, enhancing the model's generalization ability and robustness, thus enabling the model to exhibit high prediction accuracy in a wider range of application scenarios. This invention also calculates the SHAP value of features to measure the contribution of each feature. Based on the different contributions of features, targeted pruning and quantification operations are performed, not only providing interpretability analysis of the model but also significantly optimizing the model's network structure, effectively reducing the computational load, and optimizing computational efficiency and storage resource utilization during training. By successfully applying the hybrid neural network model integrating battery auxiliary information to the test set and two other real-world energy storage datasets, accurate prediction results were achieved, further validating the model's strong generalization ability and adaptability.

[0140] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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 or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for lithium-ion battery state-of-health estimation integrating ancillary information and multi-operating condition data, characterized in that, The method includes: Step S1: Load the hybrid battery dataset. Within a preset voltage range, extract charging segment data to obtain the charging voltage sequence and its corresponding charging current and charging temperature sequences. Perform monotonically non-decreasing processing on the charging voltage, charging current, and charging temperature sequences, and resample these three sequences at fixed intervals using linear interpolation. Then, extract the battery capacity value corresponding to each sample data as a label to form a battery capacity label sequence. Next, extract four features from each sample: battery type, nominal battery capacity, battery operating temperature, and battery discharge rate to obtain an auxiliary information vector sequence. Finally, input the resampled charging voltage sequence, charging current sequence, charging temperature sequence, battery capacity label sequence, and auxiliary information vector sequence into the features. Step S2: Divide the input features into training set and test set, and construct a hybrid neural network model to process the training set data in order to estimate the battery health status of the training set. The hybrid neural network model includes convolutional neural network, long short-term memory network and attention mechanism. Step S3, using SHAP value algorithm to analyze the explainability of the mixed neural network model, and obtaining the global importance of each feature ; Step S4, according to the feature global importance Pruning and quantization operations are performed on the mixed neural network model. Step S2 further includes: The input features are divided into training and test sets. The training set is fed into a convolutional neural network to extract local features, and then a long short-term memory network is used to capture global temporal dependencies. Subsequently, a self-attention mechanism is used to assign differential weights to these features, and the features weighted by the self-attention mechanism are concatenated with the auxiliary information vectors in the training set. Finally, the concatenated comprehensive features are fed into a fully connected layer to complete the estimation of the battery health status of the training set.

2. The method according to claim 1, characterized in that, In step S1, for the first cycle data, the voltage sequence , the current sequence and the temperature sequence are extracted. , The total sampling points in the fixed voltage range are taken as a sample for each cycle data, and the battery capacity in the fixed voltage interval of the cycle is calculated as a sample health state label value. Finally, all cycle samples are aggregated to obtain the battery sequence data set and the corresponding health state label value set , The total number of cycles is The battery capacity is calculated by the ampere-hour integral formula as follows: ； wherein, is the instantaneous current at the sample point of the cycle, is the instantaneous current at the sample point of the cycle, is the time of the sample point, ; Then monotone non-decreasing processing and fixed interval resampling are performed on the voltage sequence, the current sequence and the temperature sequence respectively by using linear interpolation method to obtain new voltage sequence , current sequence and temperature sequence , and the obtained new three sequence data are combined to form new sample , and the processed battery sequence data are obtained through n times of circulation .

3. The method according to claim 2, characterized in that, In step S1, an auxiliary information vector is defined where Type denotes the battery type, is the nominal capacity of the battery, is the ambient temperature of the battery, is the discharge rate of the battery, is the battery sequence data and the label Y together as final input features i.e. .

4. The method of claim 3, wherein, Step S3 further includes: set up For a well-trained hybrid neural network model, on the training set Randomly select M samples from the middle, for the first M samples... Sample ,Will The characteristic elements are uniformly used It means that it has been obtained. The SHAP value of each feature in the sample is calculated using the SHAP algorithm, as follows: For the Each sample has a SHAP value of ,in, For hybrid neural network models The average of the predicted values ​​for each sample. For the first The predicted value of a sample under a hybrid neural network model For the first In the nth sample The SHAP value of each feature, ; The SHAP value of a single feature is calculated as follows: ； wherein, is a subset of all features that do not contain feature , is the number of elements of the set , is the factorial operation, is the trained hybrid neural network model, is the predicted value of the neural network model for the feature subset , denotes the feature set containing feature , denotes the union operation; The global importance of the first characteristic is defined as: ； The larger the value, the greater the positive effect of the feature on the model. Arranging the values ​​in descending order can identify the key factors that have the greatest impact on SOH prediction.

5. The method according to claim 4, characterized in that, The pruning in the step S4 includes weight pruning and neuron pruning, the weights of the top 50% of the neurons are pruned , the absolute values of the weights corresponding to the features are calculated, if the absolute value is lower than a preset threshold, it is determined that the weight is not important and is set to zero, for neuron pruning, the average absolute values of the activation values of each neuron on the training set are counted, if the average value is lower than a set threshold, the output of the neuron is forced to zero, then, a linear quantization method is used to map the full-precision floating-point weight parameters to integer parameters.

6. The method of claim 5, wherein, Weighted pruning specifically involves: Let the weight matrix of the full connection layer, i.e. the last layer in the convolutional neural network, be Then the output after learning through this layer is: ​ ； ； wherein, and are input and output features of a fully connected layer in a convolutional neural network, respectively, and are the number of input and output features of a fully connected layer in a convolutional neural network, respectively, is a bias term. For each weight in the weight matrix if its absolute value is less than a pre-set threshold then it is set to zero, otherwise the original value is preserved, defined as follows:​ 。 7. The method of claim 6, wherein, Neuron pruning specifically involves: For the first layer, the importance of the th neuron is measured by computing the integrated activation value of this neuron as follows: ； ； in, It is an activation function. For the first The first in the layer The activation value of each neuron; If is less than a set threshold , the neuron has a small contribution to the network output and is pruned, i.e. its activation is forced to zero: .

8. The method according to claim 7, characterized in that, The specific quantitative operation is as follows: Set the weight matrix in the fully connected layer network. for: ； Each weight value is a high-precision floating-point number. The original floating-point weights are linearly mapped to an 8-bit integer range, i.e., 0~255. The quantification formula is: ； in, These are the original high-precision floating-point number weight parameters. For the calculated low-precision fixed-point numbers, and These are the minimum and maximum values ​​in the original weighted sample. This is the rounding operation.

Images