Generalized lithium ion battery state prediction method and system

Abstract

The invention discloses a generalized lithium ion battery state prediction method and system, and belongs to the technical field of lithium ion battery data analysis. Multi-source heterogeneous lithium ion battery data (including current, voltage and non-uniform timestamps) are segmented and subjected to zero padding according to a fixed time span, mixed mask reconstruction and coulomb integral regression tasks are jointly executed through a dual-scale encoder base model to complete pre-training, a regression head is cascaded, and base parameters are frozen; and finely tuning the regression head based on the battery state tag marked by the sequence. According to the method, by means of the pre-training task of physical perception and irregular time sequence modeling, the generalization ability and adaptability of battery data analysis are improved, and meanwhile irregular time sequence data are effectively processed. Experiments show that the method shows excellent performance in a plurality of downstream tasks, and an advanced modeling framework is provided for performance optimization, safety management and sustainable life cycle management of the lithium ion battery.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery data analysis technology, and in particular to a generalizable lithium-ion battery state prediction method and system. Background Technology

[0002] Lithium-ion batteries, as one of the most important energy storage devices in modern society, are widely used in portable electronic devices, electric vehicles, and grid energy storage. With the acceleration of the global energy transition, the demand and application scenarios for lithium-ion batteries are constantly expanding, and the requirements for their performance, safety, and lifespan are becoming increasingly stringent. Accurately analyzing and predicting the battery's state of health (SOH), remaining service life (RUL), and charge / discharge performance is crucial for optimizing battery management systems, reducing operating costs, and improving safety.

[0003] Traditional lithium-ion battery modeling methods are mainly divided into two categories: physical models and data-driven models. Physical models, such as pseudo-two-dimensional models (P2D) and equivalent circuit models (ECM), describe the internal processes of the battery by establishing physical equations for the reactions of electrodes, electrolytes, and interfaces. These models have high theoretical accuracy, but due to the complexity of the internal processes of the battery, the identification and calibration of model parameters often require a large amount of experimental data, and the computational complexity is high, making them difficult to apply in real time.

[0004] Data-driven models utilize machine learning and deep learning algorithms to learn the input-output relationships of batteries from historical data. In recent years, with the development of deep learning technology, neural network-based battery modeling methods have made significant progress. These methods can automatically extract features from large amounts of data, construct nonlinear mapping relationships, and have high prediction accuracy and adaptability. However, most existing data-driven models are optimized for specific battery types or specific operating conditions, resulting in limited generalization ability. When faced with batteries with different chemical systems (such as NMC-811 and LFP) or different operating scenarios (such as fast charging, slow discharging, or varying temperature environments), the model's performance often degrades significantly.

[0005] Furthermore, battery data typically exhibits non-uniform sampling frequencies and irregular time series characteristics. In practical applications, due to the dynamic characteristics of batteries, the sampling rate varies with changes in battery state; for example, the sampling frequency is lower at the beginning of charging and higher during the rapid voltage drop in the discharging phase. This irregular time series poses a challenge to traditional time series analysis methods. Conventional time series models usually assume uniform data sampling, which cannot effectively handle information loss and time aliasing issues caused by variations in sampling frequency. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a generalizable lithium-ion battery state prediction method and system. By employing physical sensing pre-training tasks and irregular time series modeling, it achieves unified pre-training on multiple lithium-ion battery datasets, enhancing the generalization and adaptability of battery data analysis while effectively handling irregular time series data.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] In a first aspect, the present invention proposes a generalizable method for predicting the state of lithium-ion batteries, comprising the following steps:

[0009] (1) Acquire multi-source heterogeneous lithium-ion battery data with different battery chemical systems and operating conditions. Each battery data is a continuous time series, consisting of current value, voltage value and non-uniform timestamp, and the battery state corresponding to the last timestamp is marked.

[0010] (2) Divide each battery data sequence into n time segment sequences according to the time span. The time span in each time segment sequence is the same. Pad the end of the time segment sequence with fewer data points with 0 to unify the length of all time segment sequences. The original timestamp distribution of data points in each segment sequence is preserved, and a natural sampling mask vector is introduced to mark the valid data points.

[0011] (3) The final representation of each battery data sequence is extracted using a lithium-ion battery base model based on a dual-scale encoder. The hybrid mask reconstruction task and the Coulomb integral regression task are jointly executed to pre-train the lithium-ion battery base model.

[0012] (4) After the lithium-ion battery base model, the regression head is spliced, the parameters of the lithium-ion battery base model are frozen, and the battery state corresponding to the last timestamp in each battery data sequence is used as the label. The regression head is fine-tuned by combining the loss between the battery state predicted by the regression head and the label.

[0013] (5) Combine the pre-trained lithium-ion battery base model and the fine-tuned regression head to predict the state of lithium-ion batteries.

[0014] Furthermore, the battery chemistry system corresponding to the multi-source heterogeneous lithium-ion battery data in step (1) includes LiCoO2, NMC, LFP, and NCA, the operating conditions include fast charging and slow discharging, and the battery status includes battery health status, remaining service life, and internal resistance.

[0015] Furthermore, in step (2), the range of the data stamp corresponding to the i-th time segment is [(i-1)×t]. s ,i×t s The natural sampling mask vector is represented as t. m= [1,…,1,0,…,0], where t s Indicates the time span, t m In this context, 1 represents a valid bit and 0 represents an invalid bit. By introducing a natural sampling mask vector, each time segment is filled with the same sequence length.

[0016] Furthermore, the lithium-ion battery base model includes:

[0017] The intra-block irregularity processing module takes the preprocessed time segment sequence, timestamp sequence, and natural sampling mask vector as input, and generates the intra-block embedding vector for each time segment sequence based on rotation position encoding and multi-head attention mechanism; wherein the timestamp is used as the rotation angle parameter in the rotation position encoding.

[0018] The inter-block rule processing module takes the concatenation result of the intra-block embedding vectors of each segment corresponding to each battery data sequence as input, extracts the final representation of each segment based on the coding network, and concatenates them to obtain the final representation of each battery data sequence.

[0019] Furthermore, the preprocessing of the time segment sequence, timestamp sequence, and natural sampling mask vector includes:

[0020] An embedding token is introduced at the beginning of each time segment sequence and initialized to zero. An additional zero bit is introduced at the beginning of the timestamp sequence and marked as 0.

[0021] For each timestamp sequence, subtract (i-1)*t from each of the remaining bits except the first bit. s Perform reinitialization, where i represents the time segment sequence number, t s Indicates the time span;

[0022] An extra zero bit is introduced at the beginning of each natural sampling mask vector and marked as 1.

[0023] Furthermore, the intra-block irregularity processing module performs calculations separately for each time segment sequence. The calculation process includes:

[0024] Each data point within the time segment sequence is linearly mapped to obtain the query, key, and value, denoted as q, k, and v, respectively. For each query value q at position j... j and key value k j Perform position encoding:

[0025]

[0026] Where θ is a predefined attenuation coefficient, t j Let q' be the timestamp at position j, i be the number of the time segment sequence, and q' be the timestamp at position j. j 、k′ jThe encoded query value and key value;

[0027] The multi-head self-attention value is calculated based on the original value, the encoded query value, and the key value. During the calculation, only the valid bits are calculated based on the natural sampling mask vector. The result corresponding to the embedded token position in the multi-head self-attention value is further input into the feedforward neural network to generate the intra-block embedding vector.

[0028] Furthermore, the calculation process of the inter-block rule processing module includes:

[0029] The block embedding vectors of each segment corresponding to each battery data sequence are concatenated and input into the stacked Transformer encoder layer;

[0030] The output of the last Transformer encoder layer is passed through a linear layer to generate the final representation of each segment;

[0031] The final representations of each segment are spliced ​​together to obtain the final representation of each battery data sequence.

[0032] Furthermore, the hybrid mask reconstruction task is as follows:

[0033] For each segment corresponding to each battery data sequence, cross-channel masking and cross-time masking are performed. Cross-channel masking refers to randomly masking the current or voltage channel, retaining only one channel of data. Cross-time masking refers to randomly masking multiple segments.

[0034] The masked time segment sequence, timestamp sequence, and natural sampling mask vector are input into the lithium-ion battery pedestal model to generate the final masked representation. The original battery data sequence is then reconstructed using this final masked representation. The reconstruction loss is calculated based on the difference between the reconstructed values ​​and the actual values ​​in the masked portion. The masked portion for which the reconstruction loss needs to be calculated for each segment is represented as follows: Among them, t m The 1 in p represents the valid bit. m The 0 in the text is the cross-channel mask part, c m The 0 in the text represents the time-crossing mask portion. They are respectively for p m c m Invert the 0s and 1s in the text;

[0035] The process of reconstructing the original battery data sequence in reverse adopts a network structure symmetrical to the lithium-ion battery base model. First, the final representation after masking is input into a new inter-block rule processing module, and the generated result is regarded as an intra-block embedding vector. Then, the processed intra-block embedding vector, timestamp sequence and natural sampling mask vector are input into a new intra-block irregularity processing module, and the generated result is regarded as the reconstruction result of the segment.

[0036] Furthermore, the Coulomb integral regression task is as follows:

[0037] For each battery data sequence, calculate the actual charge accumulation for each bit in the sequence based on the current data:

[0038] The final unmasked representation of the sequence is used to predict the charge accumulation in reverse. Only the effective bits are taken and concatenated to form the predicted charge accumulation result of the complete sequence. The difference between the predicted charge accumulation of each bit in the sequence and the actual charge accumulation is used to calculate the regression prediction loss.

[0039] The process of reverse prediction of charge accumulation is implemented using a network structure symmetrical to the lithium-ion battery base model. First, the unmasked final representation is input into a new inter-block rule processing module, and the generated result is regarded as an intra-block embedding vector. Then, the processed intra-block embedding vector, timestamp sequence and natural sampling mask vector are input into a new intra-block irregularity processing module, and the generated result is regarded as the predicted charge accumulation result of the segment.

[0040] Secondly, this invention proposes a generalizable lithium-ion battery state prediction system for implementing the aforementioned lithium-ion battery state prediction method.

[0041] The beneficial effects of this invention are:

[0042] This invention significantly improves the accuracy and generalization ability of lithium-ion battery data analysis by introducing a pre-training task based on physical perception and irregular time series modeling. Applications in multiple downstream tasks demonstrate that the lithium-ion battery pedestal model can effectively handle the complexity and irregularity of battery data, providing strong modeling support for battery performance optimization, safety management, and sustainable lifecycle management. Attached Figure Description

[0043] Figure 1 This is a structural block diagram of a lithium-ion battery base model shown in an embodiment of the present invention.

[0044] Figure 2 This is a schematic diagram of patch by time shown in an embodiment of the present invention, wherein (a) is the length segmentation method for comparison, and (b) is the timestamp segmentation method of the present invention. Detailed Implementation

[0045] The present invention will be further described and illustrated below with reference to specific embodiments. The embodiments described are merely examples of the content of this disclosure and do not limit the scope of the invention. The technical features of each embodiment in the present invention can be combined accordingly, provided that there is no mutual conflict.

[0046] The accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. Some of the block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0047] The flowchart shown in the attached diagram is merely an illustrative example and does not necessarily include all steps. For example, some steps may be broken down, while others may be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0048] Figure 1 The diagram illustrates the overall framework of a lithium-ion battery substrate modeling method according to the present invention, mainly comprising four core modules: data preprocessing, physical sensing pre-training, irregular time series modeling, and downstream task adaptation. The specific implementation steps are as follows:

[0049] I. Data Preprocessing

[0050] (1) Collection of multi-source heterogeneous battery datasets: In this embodiment, raw data such as current, voltage, and timestamps of lithium-ion batteries were obtained from eight public datasets, including NASA, CALCE, and HUST. Each dataset contains samples of different battery chemical systems (LiCoO2, NMC, LFP, NCA) and operating conditions (such as fast charging and slow discharging), with a total data volume of 351 million records.

[0051] (2) Data standardization processing: The current I and voltage V data are normalized and the timestamps are uniformly converted to second-level timestamps. At the same time, the present invention retains the original timestamp information to maintain the irregular characteristics of the time series.

[0052] Because the types and number of variables vary across different battery datasets, this invention focuses on current, voltage, and timestamps, which are present in all datasets. Consider a lithium-ion battery current dataset I = [I1, I2, ..., I...] N ], where I>0 represents charging, I<0 represents discharging, and voltage V=[V1,V2,…,V N Here, I i and V i At time t i Recorded. For simplicity, use X = [x1, x2, ..., x...]. N ] represents I and V, where x j =(V j ,I j The timestamp sequence T = [t1, t2, ..., t...] N] is irregular, and N is the total number of records.

[0053] (3) Timestamp Segmentation (patch by time): Define personalized time spans t based on the characteristics of different battery datasets. s This involves dividing a continuous time series into irregular time segments with the same time span. Taking the time series X above as an example, it is divided into several time segments x. i The superscript i represents the i-th time segment, and each time segment contains all the data points recorded within that time period, such as... Figure 2 As shown, Figure (a) on the left illustrates a traditional sliding window length segmentation method. These time segments are generated by sliding a fixed-length window, which fails to accurately preserve the temporal information of irregular battery time series. Furthermore, positional encoding is crucial for the capture order and relative relationships. Segments covering longer time intervals should exhibit more significant differences in positional encoding than those covering shorter intervals. However, traditional segmentation techniques based on a fixed number of timestamps fail to consider these differences. Figure (b) on the right illustrates the timestamp segmentation strategy proposed in this invention. Specifically, based on a fixed time span t... s Define the data points within each segment, where the i-th segment contains data points from (i-1)×t. s to i×t s All data points recorded within a time range can be represented by the i-th segment as x. i =[x j ,x j+1 ,…,x j+k The corresponding timestamp is t. i =[t j ,t j+1 ,…,t j+k ], where (i-1)×t s ≤t j <t j+1 <··· <t j+k <i×t s The sequence X is divided into n time segments, each with a length of k+1. Data points within each time segment are arranged according to their actual time distribution, rather than using a fixed-length sliding window. However, the length of each segment may vary, making parallel computation difficult. To overcome this, a mask vector t is used... m =[1,…,1,0,…,0] marks the length of valid data, t m Element 1 corresponds to the valid bit, and element 0 corresponds to the invalid bit, ensuring the feasibility of parallel computing.

[0054] II. Pre-training tasks for physical perception

[0055] This invention enhances the model's understanding of battery data by introducing two physical sensing pre-training tasks—Mixed Masked Autoencoder (MMAE) and Coulomb Integral Regression (CIR)—ensuring its accuracy in transvariate electrochemical relationships and time dynamics.

[0056] S21 Hybrid Mask Autoencoder (MMAE)

[0057] Hybrid masking autoencoders aim to learn the physical relationships within a battery by reconstructing data points that have been randomly masked. The specific steps are as follows:

[0058] (1) Cross-channel masking: The current or voltage channel is randomly masked. In this invention, the masking ratio is set to 30%. For example, in a small segment of charge and discharge data, only the voltage data may be retained while the current information is masked, forcing the model to reconstruct the missing variables through physical relationships such as V = E(SOC) - IR.

[0059] (2) Cross-time masking: Randomly mask time segments. In this invention, the masking ratio is set to 30%. The model needs to infer the original data of the masked area based on the unmasked time segments, so that the model can obtain the time dependence in the captured battery data.

[0060] (3) Reconstruction Loss Function: The masked sequence (including a set of masked time segment sequences, timestamp sequences, and mask vectors) is sequentially processed with intra-block irregularity handling and inter-block regularity handling to generate the representation Z = (z 1 ,z 2 ,…,z n Then, the generated representation is used to reconstruct the mask part in reverse, and the reconstruction loss is calculated based on the difference between the reconstructed value and the actual value of the mask part.

[0061] It should be noted that the mask part used to calculate the reconstruction loss is based on the natural sampling mask vector t. m , time-transition mask vector p m and cross-channel mask vector c m The calculated dual masking strategy, involving cross-temporal and cross-channel masking vectors, forces the model to reconstruct missing segments through temporal continuity and physical consistency constraints. The final effective masking vector is represented as follows: Among them, t m The 1 in p represents the valid bit. m The 0 in the text represents the part that needs to be rebuilt, and c... m The 0 in the text represents the part that needs to be rebuilt. They are respectively for p m c mInverting the 0s and 1s in the vector t, the part with a value of 1 after inversion is the part that needs to be reconstructed. m , First, to and Take the union, then combine it with t m Taking the intersection yields the result. As the part that needs to be reconstructed for each time segment sequence, the m of all segments are spliced ​​together to obtain the mask part that needs to be reconstructed for the complete sequence.

[0062] The reconstruction process for each segment can be represented as follows: z m This represents the masked time segment sequence, and MMAE(.) represents the reconstruction process; the reconstruction result of each time segment sequence is... The prediction results for each time segment are concatenated, taking only the valid bits, to form the complete sequence reconstruction result. The reconstruction loss is then calculated based on the difference between the actual values ​​of the masked portions before and after reconstruction and the reconstructed values. S22 Coulomb Integral Regression (CIR)

[0063] Coulomb integral regression uses Coulomb's law to calculate the cumulative charge transfer, thereby constraining the model's understanding of charge conservation. Specific steps include:

[0064] (1) Charge conservation modeling: For the complete sequence, the actual charge accumulation of each bit in the sequence is calculated based on the current data:

[0065]

[0066] Among them, I j t represents the current value at position j in the sequence. j Q represents the actual timestamp of the j-th position in the sequence. j Represents the timestamp t in the sequence j The corresponding actual charge accumulation.

[0067] Use the trapezoidal rule to handle irregular sampling intervals and ensure that timestamps are converted to second-level units.

[0068] (2) Regression task design: The original unmasked representation z corresponding to each time segment sequence is used to predict the charge accumulation in reverse. This process is represented as: CIR(.) represents the regression prediction process, which concatenates the prediction results of each time segment, taking only the valid bits, to form the predicted charge accumulation result of the complete sequence. The regression prediction loss is calculated based on the difference between the predicted charge accumulation of each bit in the sequence and the actual charge accumulation. S23 Irregular Time Series Modeling

[0069] To address the non-uniform sampling characteristics of lithium-ion battery data, this invention proposes a dual-scale time encoder for processing irregular time series. First, each segment is mapped to an embedding representation h of equal length. Then, the embedding representations h of all segments are concatenated, and the final representation is generated through inter-block rule processing.

[0070] (1) Handling irregularities within blocks

[0071] A point-level multi-head self-attention mechanism (MHSA) is used to obtain the embedding representation. First, an embedding token is introduced and initialized to zero, and the corresponding first timestamp is also set to zero. Therefore, the input representation for each segment is: x = [0, x...]. j ,x j+1 ,…,x j+k ], t=[0,t j ,t j+1 ,…,t j+k ] and t m = [1,…,1,0,…,0]. For each fragment x, the timestamp within the fragment (excluding the embedding token) is obtained by subtracting (i-1)*t. s Reinitialize and update the timestamp sequence t. This is because each segment is processed independently, and within each segment, the difference between timestamps represents a relative time difference, which remains constant regardless of the segment's start time.

[0072] Position Encoding: A method is used to adjust Rotary Position Embeddings (RoPE), replacing the discrete position index values ​​in the original RoPE with the timestamp as the rotation angle parameter. Specifically, each data point within a segment is linearly mapped to obtain a query, key, and value, denoted as q, k, and v, respectively. For each position j, the query q... j and key k j Position encoding is performed, and the encoding formula is as follows:

[0073]

[0074] Where θ is a predefined attenuation coefficient, t j For timestamps.

[0075] Multi-head self-attention: Based on the q', k', and v obtained above, a multi-head self-attention (MHSA) mechanism is used within the block to capture local physical associations through point-level interactions, as shown below:

[0076]

[0077] The token representations embedded in the MHSA output are further input into the FeedForward neural network to generate intra-block embedding vectors h, which are used for subsequent inter-block rule processing; here, the set of intra-block embedding vectors for all segments is H = (h 1 ,h 2 ,…,h n ).

[0078] (2) Inter-block rule processing

[0079] The segments are modeled in a regular manner, using standard Rotational Position Encoding (RoPE) to encode the relative temporal distance between blocks. After stacking multiple Transformer encoder layers, the final representation Z = (z...) is generated through a linear layer. 1 ,z 2 ,…,z n ).express

[0080] Here, the inter-block rule processing uses a standard transformer coding network.

[0081] III. Overall Description

[0082] The specific implementation process of the generalizable lithium-ion battery state prediction method proposed in this invention includes:

[0083] (1) Data acquisition of multi-source heterogeneous lithium-ion batteries

[0084] In this embodiment, raw data was collected and standardized from multiple publicly available battery datasets. The battery chemistry systems corresponding to the multi-source heterogeneous lithium-ion battery data include LiCoO2, NMC, LFP, and NCA; the operating conditions include fast charging and slow discharging; and the battery states include battery health, remaining lifespan, and internal resistance.

[0085] (2) Timestamp segmentation

[0086] In this step, timestamps are uniformly converted to second-level timestamps, while retaining the original timestamp information to maintain the irregular characteristics of the time series. A patch-by-time strategy is adopted to divide the continuous time series into irregular time segments.

[0087] Specifically, each battery data sequence is divided into n time segment sequences according to the time span. The time span within each time segment sequence is the same. For time segment sequences with fewer data points, zeros are padded at the end to unify the length of all time segment sequences. The original timestamp distribution of data points within each segment sequence is preserved, and a natural sampling mask vector is introduced to mark valid data points.

[0088] (2) Physical perception pre-training

[0089] It is used to extract the final representation of each battery data sequence using a lithium-ion battery base model based on a dual-scale encoder, and jointly perform the hybrid mask reconstruction task and the Coulomb integral regression task to pre-train the lithium-ion battery base model.

[0090] In the two pre-training tasks, the hybrid masked autoencoder forces the model to learn physical relationships through cross-channel and cross-temporal masks, while Coulomb integral regression constrains charge accumulation predictions through the law of charge conservation. These two tasks work together to enable the model to understand the electrochemical behavior of the battery at different time scales.

[0091] The lithium-ion battery base model includes:

[0092] The intra-block irregularity processing module takes the preprocessed time segment sequence, timestamp sequence, and natural sampling mask vector as input, and generates an intra-block embedding vector for each time segment sequence based on rotational position encoding and multi-head attention mechanism; in the rotational position encoding, the timestamp is used as the rotation angle parameter; the intra-block irregularity processing module calculates separately for each time segment sequence, the calculation process includes: linearly mapping each data point in the time segment sequence to obtain the query, key, and value, represented as q, k, and v respectively, and for each query value q at position j... j and key value k j Position encoding is performed, and multi-head self-attention values ​​are calculated based on the original value, the encoded query value, and the key value. During the calculation, only the valid bits are calculated based on the natural sampling mask vector. The results of the multi-head self-attention values ​​corresponding to the embedded token positions are further input into the feedforward neural network to generate intra-block embedding vectors.

[0093] The inter-block rule processing module takes the concatenated result of the intra-block embedding vectors of each segment corresponding to each battery data sequence as input, extracts the final representation of each segment based on the encoding network, and concatenates them to obtain the final representation of each battery data sequence. The computation process of the inter-block rule processing module includes: concatenating the intra-block embedding vectors of each segment corresponding to each battery data sequence and inputting it into a stacked Transformer encoder layer; passing the output of the last Transformer encoder layer through a linear layer to generate the final representation of each segment; and concatenating the final representations of each segment to obtain the final representation of each battery data sequence.

[0094] In this step, the hybrid mask reconstruction task is as follows:

[0095] For each segment corresponding to each battery data sequence, cross-channel masking and cross-time masking are performed. Cross-channel masking refers to randomly masking the current or voltage channel, retaining only one channel of data. Cross-time masking refers to randomly masking multiple segments.

[0096] The masked time segment sequence, timestamp sequence, and natural sampling mask vector are input into the lithium-ion battery pedestal model to generate the final masked representation. The original battery data sequence is then reconstructed using this final masked representation. The reconstruction loss is calculated based on the difference between the reconstructed values ​​and the actual values ​​in the masked portion. The masked portion for which the reconstruction loss needs to be calculated for each segment is represented as follows: Among them, t m The 1 in p represents the valid bit. m The 0 in the text is the cross-channel mask part, c m The 0 in the text represents the time-crossing mask portion. They are respectively for p m c m Invert the 0s and 1s in the text;

[0097] The process of reconstructing the original battery data sequence employs a network structure symmetrical to the lithium-ion battery substrate model. First, the masked final representation is input into a new inter-block rule processing module, and the generated result is considered as an intra-block embedding vector. This intra-block embedding vector is then processed, for example, linearly mapped to twice the original segment length through a fully connected layer. Then, a reshape operation transforms it into two vectors of the original segment length, which are stacked to form a reconstruction matrix of shape (2, original segment length). These two vectors of the original segment length are used to reconstruct the original current and voltage, respectively. Finally, the reconstruction matrix (i.e., the processed intra-block embedding vector), the timestamp sequence, and the natural sampling mask vector are input into a new intra-block irregularity processing module, and the generated result is considered as the segment reconstruction result.

[0098] In this step, the Coulomb integral regression task is as follows:

[0099] For each battery data sequence, calculate the actual charge accumulation for each bit in the sequence based on the current data:

[0100] The final unmasked representation of the sequence is used to predict the charge accumulation in reverse. Only the effective bits are taken and concatenated to form the predicted charge accumulation result of the complete sequence. The difference between the predicted charge accumulation of each bit in the sequence and the actual charge accumulation is used to calculate the regression prediction loss.

[0101] The process of reverse prediction of charge accumulation is implemented using a network structure symmetrical to the lithium-ion battery base model. First, the unmasked final representation is input into a new inter-block rule processing module, and the generated result is regarded as an intra-block embedding vector. Then, the intra-block embedding vector is processed, for example, linearly mapped to the reconstruction vector of the original segment length through a fully connected layer. Then, the reconstruction vector (i.e. the processed intra-block embedding vector), the timestamp sequence, and the natural sampling mask vector are input into a new intra-block irregularity processing module, and the generated result is regarded as the predicted charge accumulation result of the segment.

[0102] (4) After the lithium-ion battery base model, the regression head is spliced, the parameters of the lithium-ion battery base model are frozen, and the battery state corresponding to the last timestamp in each battery data sequence is used as the label. The regression head is fine-tuned by combining the loss between the battery state predicted by the regression head and the label.

[0103] For example, the regression heads used to predict battery health status, remaining lifespan, and internal resistance are implemented using independent multilayer sensing machines.

[0104] (5) Combine the pre-trained lithium-ion battery base model and the fine-tuned regression head to predict the state of lithium-ion batteries.

[0105] Specifically, the historical battery data sequence of the lithium-ion battery to be predicted is obtained and divided into n time segment sequences according to the timestamp segmentation method. The time span within each time segment sequence is the same. The time segment sequence with fewer data points is padded with 0s at the end to unify the length of all time segment sequences. The original timestamp distribution of data points within each segment sequence is preserved, and a natural sampling mask vector is introduced to mark the valid data points. The final representation of each battery data sequence is extracted using a lithium-ion battery base model based on a dual-scale encoder. The last result in the final representation is taken as input into the regression head to obtain the prediction result.

[0106] This embodiment uses a specific experiment to verify the implementation effect of the present invention.

[0107] I. Data Description

[0108] In a specific embodiment of the present invention, to verify its effectiveness, we used lithium-ion battery datasets from various sources for experiments. The following is a detailed description of each dataset:

[0109] The NASA dataset originates from lithium-ion battery charge-discharge experiments conducted by NASA. It records 7,203,084 data points from 34 battery cells, each with a rated capacity of 2.0 Ah. The charging process employed constant current (CC) mode at 1.5 A until the battery voltage reached 4.2 V, then switched to constant voltage (CV) mode until the charging current dropped to 20 mA. The positive electrode material was lithium cobalt oxide.

[0110] The CALCE dataset, sourced from the University of Maryland's Center for Advanced Life Cycle Engineering (CALCE), includes two battery types: CS2 and CX2. The CS2 battery has a nominal capacity of 1.1 Ah, with a lithium cobalt oxide cathode containing trace amounts of manganese. The charging process follows a standard constant current / constant voltage (CC / CV) protocol, where the CC phase charges at a current rate of 0.5C until the voltage reaches 4.2V, then maintains 4.2V until the charging current drops below 0.05A. Unless otherwise specified, the discharge cutoff voltage for these batteries is 2.7V. The CX2 battery has a nominal capacity of 1.35 Ah, also with a lithium cobalt oxide cathode (containing trace amounts of manganese). The charging profiles for CX2 batteries are similar to those for CS2, but some minor variations are introduced on the standard charging profiles for certain CX2 batteries. This dataset contains a total of 27 cells from both battery types, comprising 16,852,982 records.

[0111] The HUST dataset contains 77 LFP / graphite battery cells (nominal capacity 1.1Ah, nominal voltage 3.3V). All cells use the same charging protocol, but different multi-stage discharge protocols are employed under a constant temperature of 30°C. The dataset contains 91,735,853 records.

[0112] The MATR dataset consists of 185 commercial lithium-ion batteries cycled under fast-charging conditions until failure. All battery cells employ a one- or two-step fast-charging strategy, formatted as "C1(Q1)-C2", where C1 and C2 are the first and second constant-current steps, respectively, and Q1 is the state of charge (SOC, %) at current transition. The second current step ends after reaching 80% SOC, followed by charging at 1C via CC-CV. Upper and lower voltage limits are 3.6V and 2.0V, respectively, and all battery cells are discharged at a rate of 4C. This dataset contains four batches of lithium iron phosphate (LFP) / graphite battery cells, totaling 143,576,451 records.

[0113] The RWTH dataset contains time-series data (time, current, voltage, temperature) for 48 lithium-ion battery cells used for cyclic aging tests. During the experiment, 48 identical battery cells were aged under the same conditions. The batteries used carbon anodes and NMC as cathode materials. The dataset contains 48,625,944 records.

[0114] The HNEI dataset consists of 14 commercially available 18650 battery cells with graphite anodes and cathodes composed of a mixture of NMC and LCO. These cells were cycled more than 1000 times at 1.5C at room temperature with a depth of discharge (DOD) of 100%. The dataset contains 7,497,493 records.

[0115] The SNLLFP dataset consists of 18 commercially available 18650LFP battery cells, acquired from Sandia National Laboratories, cycled to 80% capacity (although cycling is still ongoing). The study examined the effects of temperature, depth of discharge (DOD), and discharge current on the long-term degradation of these commercial batteries. The dataset contains 6,904,086 records.

[0116] The ULPurdue dataset consists of 22 commercially available 18650 battery cells with graphite anodes and NCA cathodes. These cells were cycled at 0.5C at room temperature between 2.7–4.2V (0–100% SOC) or 2.9–4.0V (2.5–96.5% SOC, when the cells are fresh) until different levels of capacity decay (10%, 15%, 20%) were achieved. The dataset contains 6,246,486 records.

[0117] II. Baseline Model

[0118] To comprehensively verify the superiority of the lithium-ion battery pedestal model proposed in this invention in the field of lithium-ion battery analysis, this experiment selected several representative baseline models for comparison. These baseline models cover non-pre-trained models, time-series pre-trained models, and irregular time-series models to ensure the comprehensiveness and reliability of the experimental results. The specific selection of baseline models is based on their wide applicability and technical representativeness in battery modeling tasks.

[0119] Non-pretrained models include Crossformer and iTransformer. Crossformer is a time series modeling method based on a cross-attention mechanism, which improves prediction performance by capturing long-range dependencies; iTransformer optimizes time series feature extraction capabilities through an improved Transformer architecture.

[0120] Time series pre-trained models include GPT4TS, PatchTST, TS2Vec, and TFC. GPT4TS is based on a generative pre-training paradigm, learning a general representation through large-scale time series data; PatchTST uses a patch segmentation strategy, combined with a Transformer encoder to extract local and global features; TS2Vec generates a general embedding representation of time series through contrastive learning; and TFC utilizes time-frequency consistency for pre-training to enhance the model's adaptability to time series patterns.

[0121] Irregular time series models include mTAN, Contiformer, and tPatchGNN. mTAN models irregular dependencies through a multi-temporal attention mechanism; Contiformer uses a continuous-time Transformer to handle irregularly sampled data; and tPatchGNN combines graph neural networks and adaptive patch segmentation to address the asynchronous nature of time series data. Furthermore, experiments also incorporated recently popular hybrid models (such as mTAN and Contiformer) to verify the robustness of the lithium-ion battery pedestal model in complex scenarios.

[0122] By comparing these baseline models, the innovations of the lithium-ion battery pedestal model in terms of physical consistency constraints, irregular time modeling, and cross-task generalization ability can be comprehensively evaluated. The selection of baseline models covers the full spectrum from traditional time series modeling to cutting-edge pre-training methods, ensuring that the experimental results reflect the technological breakthroughs and practicality of the lithium-ion battery pedestal model.

[0123] III. Experimental Setup

[0124] The experimental data came from eight publicly available lithium-ion battery datasets, including NASA, CALCE, MATR, HUST, HNEI, UL Purdue, SNLLFP, and RWTH, covering mainstream cathode materials (such as LiCoO2, NMC, LFP, and NCA) and 351 million data points. These datasets include different battery chemistry systems, charge / discharge protocols (such as fast charging, slow discharging, and constant current / constant voltage), and environmental conditions (such as temperature variations), ensuring the broad applicability of the experimental results. The experiments employed a dataset-level partitioning strategy: 85% of the batteries were used for pre-training, and 15% were used for downstream task evaluation, with strict avoidance of overlap between the pre-training and testing phases for the same battery. This partitioning method not only reflects the fragmented nature of battery data in real-world applications but also effectively prevents overfitting.

[0125] The experimental task focuses on the core state estimation problem of lithium-ion batteries, including state of health (SOH), remaining useful life (RUL), internal resistance (IR), and charge / discharge capacity (ΔQ). SOH is defined as the ratio of the battery's maximum usable capacity to its nominal capacity, used to quantify the degree of battery degradation; RUL predicts the battery's end-of-life based on the decreasing trend of SOH; IR reflects the internal electrochemical changes of the battery through dynamic impedance measurement; and ΔQ calculates the battery's state of charge (SOC) and energy distribution through cumulative charge transfer. Two evaluation modes are employed: battery-split and cycle-split. In battery-split mode, the training and test sets contain completely different battery samples, simulating the need for the model to generalize to unknown batteries in real-world scenarios. In cycle-split mode, different charge / discharge cycles of the same battery are used for training and testing, evaluating the model's ability to capture long-term degradation trends.

[0126] The evaluation metrics used were mean squared error (MSE) and mean absolute error (MAE), which measure the squared error and absolute error between the predicted and actual values, respectively. MSE is sensitive to large errors and is suitable for tasks requiring high accuracy (such as capacity estimation); MAE, on the other hand, focuses more on the average distribution of errors and is suitable for scenarios with lower tolerance for extreme errors (such as RUL prediction). The experiment used 10 independent repeated experiments (with different random seeds) to calculate the mean and standard deviation, ensuring the statistical significance of the results. Furthermore, the experiment used paired t-tests (p<0.05) to verify whether the improvement in the lithium-ion battery base model was statistically significant, further strengthening the credibility of the conclusions.

[0127] IV. Comparative Experiment

[0128] The experimental results are shown in Tables 1-1 and 1-2 (Battery-split) and Tables 2-1 and 2-2 (Cycle-split).

[0129] Table 1-1

[0130] Table 1-2

[0131]

[0132] Table 2-1

[0133]

[0134] Table 2-2

[0135]

[0136] Experimental results demonstrate that the lithium-ion battery pedestal model proposed in this invention exhibits significant performance advantages in lithium-ion battery analysis tasks. In the battery segmentation mode, the mean mean square error (MSE) and mean absolute error (MAE) of the lithium-ion battery pedestal model are reduced by 28.3% and 27.3%, respectively, compared to the best baseline model. In the periodic segmentation mode, these figures are further improved to 49.3% and 59.8%. This performance improvement is attributed to three core innovations of the lithium-ion battery pedestal model: the hybrid masked autoencoder (MMAE), coulomb integral regression (CIR), and the dual-scale time encoder.

[0137] MMAE forces the model to learn the physical correlation between voltage, current, and time by jointly reconstructing time mask patches and cross-channel mask variables. Experiments show that MMAE reduces the error to 0.002 (MSE) and 0.015 (MAE) in the charge / discharge capacity calculation task, representing improvements of 6 times and 45 times compared to non-pre-trained models (such as Crossformer). CIR further improves the accuracy of battery state estimation by explicitly modeling the charge conservation law. For example, in the SOH estimation task, the MSE of the lithium-ion battery pedestal model reaches 0.002 in battery segmentation mode, which is 50% lower than the best baseline model (Crossformer); in cycle segmentation mode, the MSE further decreases to 0.001, which is 75% lower than PatchTST. The dual-scale time encoder addresses the problem of dynamic changes in battery data sampling frequency by combining intra-patch processing for irregularities and inter-patch attention for regularities. Experiments show that the MSE of the lithium-ion battery pedestal model decreases from 0.45e-04 to 7.25e-07 in irregular time modeling tasks (such as internal resistance estimation), verifying its strong modeling ability for physical constraints.

[0138] Furthermore, the pre-training paradigm of the lithium-ion battery pedestal model does not rely on complete charge-discharge cycle annotations, enabling it to directly handle fragmented real-world application data (such as partial charge-discharge segments in electric vehicles and smartphones). Experiments show that the lithium-ion battery pedestal model outperforms all baseline models in battery segmentation mode, demonstrating its applicability in real-world scenarios. By comparing pre-trained models (such as PatchTST) and non-pre-trained models (such as Crossformer), it was found that the physical perception objective functions (MMAE and CIR) of the lithium-ion battery pedestal model significantly improve the model's understanding of battery electrochemical characteristics, while traditional pre-training methods (such as TS2Vec) rely solely on statistical correlation, resulting in poor performance in tasks with strong physical correlations. This result demonstrates that the innovative design of the lithium-ion battery pedestal model can effectively bridge the gap between data-driven methods and electrochemical principles, providing a more reliable solution for battery modeling.

[0139] This embodiment also provides a generalizable lithium-ion battery state prediction system, including:

[0140] The multi-source heterogeneous lithium-ion battery data acquisition module is used to acquire multi-source heterogeneous lithium-ion battery data containing different battery chemical systems and operating conditions. Each battery data is a continuous time series, consisting of current value, voltage value and non-uniform timestamp, marking the battery state corresponding to the last timestamp.

[0141] The timestamp segmentation module is used to divide each battery data sequence into n time segment sequences according to the time span. The time span within each time segment sequence is the same. The time segment sequence with fewer data points is padded with 0s at the end to unify the length of all time segment sequences. The data points within each segment sequence retain the original timestamp distribution, and a natural sampling mask vector is introduced to mark the valid data points.

[0142] The physical perception pre-training module is used to extract the final representation of each battery data sequence using a lithium-ion battery base model based on a dual-scale encoder, and jointly execute the hybrid mask reconstruction task and the Coulomb integral regression task to pre-train the lithium-ion battery base model.

[0143] The regression head training module is used to stitch the regression head after the lithium-ion battery base model, freeze the lithium-ion battery base model parameters, use the battery state corresponding to the last timestamp in each battery data sequence as the label, and fine-tune the regression head by combining the loss between the battery state predicted by the regression head and the label.

[0144] The lithium-ion battery state prediction module is used to predict the state of lithium-ion batteries by combining a pre-trained lithium-ion battery pedestal model and a fine-tuned regression head.

[0145] For the system embodiments, since they basically correspond to the method embodiments, relevant details can be found in the descriptions of the method embodiments; the implementation methods of the remaining modules will not be repeated here. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0146] The system embodiments of the present invention can be applied to any device with data processing capabilities, such as a computer or other similar device. The system embodiments can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution.

[0147] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. Those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A generalizable method for predicting the state of lithium-ion batteries, characterized in that, Includes the following steps: (1) Acquire multi-source heterogeneous lithium-ion battery data with different battery chemical systems and operating conditions. Each battery data is a continuous time series, consisting of current value, voltage value and non-uniform timestamp, and the battery state corresponding to the last timestamp is marked. (2) Divide each battery data sequence into n time segment sequences according to the time span. The time span in each time segment sequence is the same. Pad the end of the time segment sequence with fewer data points with 0 to unify the length of all time segment sequences. The original timestamp distribution of data points in each segment sequence is preserved, and a natural sampling mask vector is introduced to mark the valid data points. (3) The final representation of each battery data sequence is extracted using a lithium-ion battery base model based on a dual-scale encoder. The hybrid mask reconstruction task and the Coulomb integral regression task are jointly executed to pre-train the lithium-ion battery base model. (4) After the lithium-ion battery base model, the regression head is spliced, the parameters of the lithium-ion battery base model are frozen, and the battery state corresponding to the last timestamp in each battery data sequence is used as the label. The regression head is fine-tuned by combining the loss between the battery state predicted by the regression head and the label. (5) Combine the pre-trained lithium-ion battery base model and the fine-tuned regression head to predict the state of lithium-ion batteries.

2. The generalizable lithium-ion battery state prediction method according to claim 1, characterized in that, The battery chemistry system corresponding to the multi-source heterogeneous lithium-ion battery data in step (1) includes LiCoO2, NMC, LFP, and NCA. The operating conditions include fast charging and slow discharging. The battery status includes battery health status, remaining service life, and internal resistance.

3. The generalizable lithium-ion battery state prediction method according to claim 1, characterized in that, In step (2), the range of data stamps corresponding to the i-th time segment is [(i-1)×t]. s ,i×t s The natural sampling mask vector is represented as t. m = [1,…,1,0,…,0], where t s Indicates the time span, t m In this context, 1 represents a valid bit and 0 represents an invalid bit. By introducing a natural sampling mask vector, each time segment is filled with the same sequence length.

4. The generalizable lithium-ion battery state prediction method according to claim 1, characterized in that, The lithium-ion battery base model includes: The intra-block irregularity processing module takes the preprocessed time segment sequence, timestamp sequence, and natural sampling mask vector as input, and generates the intra-block embedding vector for each time segment sequence based on rotation position encoding and multi-head attention mechanism; wherein the timestamp is used as the rotation angle parameter in the rotation position encoding. The inter-block rule processing module takes the concatenation result of the intra-block embedding vectors of each segment corresponding to each battery data sequence as input, extracts the final representation of each segment based on the coding network, and concatenates them to obtain the final representation of each battery data sequence.

5. The generalizable lithium-ion battery state prediction method according to claim 4, characterized in that, The preprocessing of time segment sequences, timestamp sequences, and natural sampling mask vectors includes: An embedding token is introduced at the beginning of each time segment sequence and initialized to zero. An additional zero bit is introduced at the beginning of the timestamp sequence and marked as 0. For each timestamp sequence, subtract (i-1)*t from each of the remaining bits except the first bit. s Perform reinitialization, where i represents the time segment sequence number, t s Indicates the time span; An extra zero bit is introduced at the beginning of each natural sampling mask vector and marked as 1.

6. The generalizable lithium-ion battery state prediction method according to claim 5, characterized in that, The intra-block irregularity processing module performs calculations separately for each time segment sequence. The calculation process includes: Each data point within the time segment sequence is linearly mapped to obtain the query, key, and value, denoted as q, k, and v, respectively. For each query value q at position j... j and key value k j Perform position encoding: Where θ is a predefined attenuation coefficient, t j Let q be the timestamp at position j, i be the number of the time segment sequence, and q be the timestamp at position j. j ′、k j ′ represents the encoded query value and key value; The multi-head self-attention value is calculated based on the original value, the encoded query value, and the key value. During the calculation, only the valid bits are calculated based on the natural sampling mask vector. The result corresponding to the embedded token position in the multi-head self-attention value is further input into the feedforward neural network to generate the intra-block embedding vector.

7. The generalizable lithium-ion battery state prediction method according to claim 5, characterized in that, The calculation process of the inter-block rule processing module includes: The block embedding vectors of each segment corresponding to each battery data sequence are concatenated and input into the stacked Transformer encoder layer; The output of the last Transformer encoder layer is passed through a linear layer to generate the final representation of each segment; The final representations of each segment are spliced ​​together to obtain the final representation of each battery data sequence.

8. The generalizable lithium-ion battery state prediction method according to claim 5, characterized in that, The hybrid mask reconstruction task is as follows: For each segment corresponding to each battery data sequence, cross-channel masking and cross-time masking are performed. Cross-channel masking refers to randomly masking the current or voltage channel, retaining only one channel of data. Cross-time masking refers to randomly masking multiple segments. The masked time segment sequence, timestamp sequence, and natural sampling mask vector are input into the lithium-ion battery pedestal model to generate the final masked representation. The original battery data sequence is then reconstructed using this final masked representation. The reconstruction loss is calculated based on the difference between the reconstructed values ​​and the actual values ​​in the masked portion. The masked portion for which the reconstruction loss needs to be calculated for each segment is represented as follows: Among them, t m The 1 in p represents the valid bit. m The 0 in the text is the cross-channel mask part, c m The 0 in the text represents the time-crossing mask portion. They are respectively for p m c m Invert the 0s and 1s in the text; The process of reconstructing the original battery data sequence in reverse adopts a network structure symmetrical to the lithium-ion battery base model. First, the final representation after masking is input into a new inter-block rule processing module, and the generated result is regarded as an intra-block embedding vector. Then, the processed intra-block embedding vector, timestamp sequence and natural sampling mask vector are input into a new intra-block irregularity processing module, and the generated result is regarded as the reconstruction result of the segment.

9. The generalizable lithium-ion battery state prediction method according to claim 4, characterized in that, The Coulomb integral regression task is as follows: For each battery data sequence, calculate the actual charge accumulation for each bit in the sequence based on the current data: The final unmasked representation of the sequence is used to predict the charge accumulation in reverse. Only the effective bits are taken and concatenated to form the predicted charge accumulation result of the complete sequence. The difference between the predicted charge accumulation of each bit in the sequence and the actual charge accumulation is used to calculate the regression prediction loss. The process of reverse prediction of charge accumulation is implemented using a network structure symmetrical to the lithium-ion battery base model. First, the unmasked final representation is input into a new inter-block rule processing module, and the generated result is regarded as an intra-block embedding vector. Then, the processed intra-block embedding vector, timestamp sequence and natural sampling mask vector are input into a new intra-block irregularity processing module, and the generated result is regarded as the predicted charge accumulation result of the segment.

10. A generalizable lithium-ion battery state prediction system for implementing the lithium-ion battery state prediction method of claim 1, characterized in that, The system includes: The multi-source heterogeneous lithium-ion battery data acquisition module is used to acquire multi-source heterogeneous lithium-ion battery data containing different battery chemical systems and operating conditions. Each battery data is a continuous time series, consisting of current value, voltage value and non-uniform timestamp, marking the battery state corresponding to the last timestamp. The timestamp segmentation module is used to divide each battery data sequence into n time segment sequences according to the time span. The time span within each time segment sequence is the same. The time segment sequence with fewer data points is padded with 0s at the end to unify the length of all time segment sequences. The data points within each segment sequence retain the original timestamp distribution, and a natural sampling mask vector is introduced to mark the valid data points. The physical perception pre-training module is used to extract the final representation of each battery data sequence using a lithium-ion battery base model based on a dual-scale encoder, and jointly execute the hybrid mask reconstruction task and the Coulomb integral regression task to pre-train the lithium-ion battery base model. The regression head training module is used to stitch the regression head after the lithium-ion battery base model, freeze the lithium-ion battery base model parameters, use the battery state corresponding to the last timestamp in each battery data sequence as the label, and fine-tune the regression head by combining the loss between the battery state predicted by the regression head and the label. The lithium-ion battery state prediction module is used to predict the state of lithium-ion batteries by combining a pre-trained lithium-ion battery pedestal model and a fine-tuned regression head.

Images