A hybrid perspective integrated learning method for improving the prediction accuracy of the remaining service life of lithium-ion batteries

Abstract

The application provides a hybrid perspective integrated learning method for improving the prediction accuracy of the residual service life of a lithium ion battery, and comprises the following steps: step 1, performing disorder processing on an original data set obtained from the HIs perspective of the battery, and performing capacity prediction training through an online sequential extreme learning machine (OS-ELM); step 2, decomposing data from the capacity degradation data perspective of the battery by using a completely adaptive noise set empirical mode decomposition (CEEMDAN) algorithm and reconstructing the data; step 3, performing time series prediction and post-processing through an autoregressive integrated moving average (ARIMA) model; and step 4, integrating the two perspectives, and proposing a meta-model data set construction method for training the OS-ELM, so that the RUL prediction from the hybrid perspective is realized. The application simultaneously considers the HIs and capacity degradation data two perspectives, solves the limitation that the traditional lithium ion battery RUL prediction method only depends on a single perspective, and is helpful to obtain more accurate and reliable prediction results.

Description

Technical Field

[0001] This invention proposes a hybrid perspective ensemble learning method to improve the accuracy of remaining life prediction for lithium-ion batteries, belonging to the field of RUL prediction technology for lithium-ion batteries. Background Technology

[0002] As an energy source for various electronic tools, lithium-ion batteries have become an indispensable part. However, due to the interference of internal response mechanisms and external environmental factors on battery charging and discharging, lithium-ion batteries will inevitably age, causing major safety accidents and serious consequences [1]. Therefore, it is extremely important to predict the remaining service life (RUL) of lithium-ion batteries.

[0003] The main methods for predicting the relative uptime (RUL) of lithium-ion batteries include model-based, data-driven, and hybrid methods [2]. Liu et al. [3] proposed a method to predict RUL by reconstructing the nonlinear Wiener process using dynamic latent variables, but the electrochemical reactions inside the battery are very complex and difficult to model accurately. Amir et al. [4] proposed an equivalent circuit model that can adapt to and capture the dynamics of the health state over time to predict RUL. However, in practical applications, the corresponding adjustment of model parameters may affect the accuracy of the model as the battery operating conditions change. Sangwan et al. [5] proposed an empirical model and a Kalman filter to predict RUL. However, using this method to predict RUL will face the problem of feature extraction difficulties.

[0004] Therefore, solving complex problems in research requires considering multiple perspectives. While traditional methods are effective in some cases, they may fail to fully capture the inherent complexity of modern challenges. To address this need for comprehensive solutions, the concept of a "hybrid perspective" has emerged as a promising paradigm. This approach advocates integrating multiple viewpoints, methods, or data sources to achieve a more comprehensive understanding and solution to complex problems. Existing lithium-ion battery RUL prediction methods typically employ a single perspective, either based on directly extracted HIs (Highly Indicative Resources) or on lithium-ion battery capacity degradation data. This can lead to a limited understanding of the complex dynamics of battery degradation, failing to capture the multifaceted characteristics of battery aging, resulting in inaccurate or incomplete predictions.

[0005] To employ a hybrid perspective approach, this paper utilizes ensemble learning to combine two perspectives to predict the relative uptime (RUL) of lithium-ion batteries. Ensemble learning, as a powerful technique, can integrate the advantages of various perspectives, reducing the bias and error of individual perspectives, thereby improving the accuracy and reliability of predictions. Applying ensemble learning to a hybrid perspective allows for the organic fusion of information from different viewpoints, deepening the analysis of each perspective and improving the effectiveness of the prediction model to address various research challenges.

[0006] References:

[0007] [1] Zhao J, Zhu Y, Zhang B, Liu M, Wang J, Liu C and Hao X2023Sustainability 15 5014.

[0008] [2]Jiahui Z,Liting T et al.2023Power Generation Technology 44 1.

[0009] [3] Liu Q, Zhang Y, Si X and Fan Z 2023IEEE Transactions on Instrumentation and Measurement 72 1–9.

[0010] [4] Amir S, Gulzar M, Tarar MO, Naqvi IH, Zaffar NA and Pecht M G2022IEEE Access 10 18279–18288.

[0011] [5]Jafari S and Byun YC 2022Sensors 22 9522. Summary of the Invention

[0012] This invention proposes a Hybrid Perspective Ensemble Learning (HyPELS) method to improve the prediction accuracy of the Remaining Life (RUL) of lithium-ion batteries. First, the raw dataset obtained from the battery's Hierarchical Hierarchical (HI) perspective is shuffled and trained for capacity prediction using an online Sequential Extreme Learning Machine (OS-ELM). Second, the data from the battery capacity degradation perspective is decomposed and reconstructed using the Fully Adaptive Noise Ensemble Empirical Mode Decomposition (CEEMDAN) algorithm, followed by time series prediction and post-processing using an Autoregressive Differential Moving Average (ARIMA) model. Finally, to integrate the two perspectives, a meta-model dataset construction method is proposed for training the OS-ELM, achieving RUL prediction from the hybrid perspective. Comprehensive experimental validation using two NASA datasets fully demonstrates the effectiveness of HyPELS and its advantages in improving prediction accuracy and reliability.

[0013] This invention discloses a hybrid perspective ensemble learning method for improving the accuracy of predicting the remaining lifespan of lithium-ion batteries, comprising the following steps:

[0014] Step 1: The original dataset obtained from the perspective of battery HIs is shuffled and trained for capacity prediction using an online sequential extreme learning machine (OS-ELM).

[0015] Step 2: Decompose and reconstruct the data from the perspective of battery capacity degradation using the Fully Adaptive Noise Ensemble Empirical Mode Decomposition (CEEMDAN) algorithm.

[0016] Step 3: Perform time series forecasting and post-processing using the Autoregressive Differential Moving Average (ARIMA) model.

[0017] Step 4: Integrating the two perspectives, a meta-model dataset construction method is proposed for training OS-ELM, realizing RUL prediction from the hybrid perspective.

[0018] The advantages and beneficial effects of this invention are as follows:

[0019] This invention aims to improve the accuracy and robustness of lithium-ion battery raw uptime (RUL) prediction by combining battery health indicators and capacity degradation data using an ensemble learning method. Specifically, the main contributions of this invention are as follows:

[0020] (1) Perspective integration: HyPELS effectively integrates two perspectives, battery HIs and capacity degradation data, through ensemble learning, making full use of the advantages of both to improve the accuracy and robustness of RUL prediction.

[0021] (2) Accuracy enhancement techniques: HyPELS employs advanced techniques, such as shuffling the dataset during OS-ELM training to eliminate the effects of time, and using sophisticated methods to predict data with capacity degradation to ensure the accuracy and reliability of the prediction results.

[0022] (3) Meta-model approach: HyPELS introduces a novel meta-model dataset construction method, which enables the OS-ELM model to be trained using merged datasets from multiple perspectives, promoting seamless integration and enhancing the overall RUL prediction performance.

[0023] (4) Comprehensive experimental verification: Comprehensive experimental verification demonstrated the effectiveness of HyPELS and proved its advantages in improving prediction accuracy and reliability. Attached Figure Description

[0024] Figure 1a This is a capacity degradation curve for lithium-ion battery data B5 and B18 from the NASA constant voltage and constant current dataset.

[0025] Figure 1b Capacity degradation curves for NASA's random walk charge / discharge datasets RW9 and RW10.

[0026] Figure 2 This is a descriptive diagram of the hybrid perspectives (Perspective 1 is the HIs perspective based on lithium-ion batteries, and Perspective 2 is the capacity degradation data perspective based on lithium-ion batteries).

[0027] Figure 3a The flowchart is for the fully adaptive noise set empirical mode decomposition algorithm (CEEMDAN).

[0028] Figure 3b This diagram illustrates the reconstruction process of lithium-ion battery capacity degradation data after decomposition using CEEMDAN.

[0029] Figure 4 The graph shows the four intrinsic mode components (IMFs) and one residual (Res) obtained after decomposing NASA's B5 battery using the CEEMDAN algorithm.

[0030] Figure 5 This is a flowchart of the method proposed in this invention.

[0031] Figure 6a The RUL prediction curve and actual curve of lithium-ion battery data B5 from the NASA constant voltage and constant current dataset are shown in the algorithm proposed in this invention.

[0032] Figure 6b The RUL prediction curve and actual curve of the B18 lithium-ion battery data in the NASA constant voltage and constant current dataset are shown in the algorithm proposed in this invention.

[0033] Figure 6c The RUL prediction curve and actual curve of lithium-ion battery RW9 in the NASA random walk charge-discharge dataset are shown in the algorithm proposed in this invention.

[0034] Figure 6d The RUL prediction curves and actual curves of lithium-ion battery RW10 in the NASA random walk charge-discharge dataset are shown in the algorithm proposed in this invention.

[0035] Figure 7a To compare the RUL prediction curves and actual curves of lithium-ion battery data B5 from the NASA constant voltage and constant current dataset (HIs perspective) in the experiment, the RUL prediction curves of the algorithm proposed in this invention and other algorithms are compared.

[0036] Figure 7b To compare the RUL prediction curves and actual curves of the NASA constant voltage and constant current dataset B18 lithium-ion battery data from the HIs perspective in the experiment, the RUL prediction curves of the algorithm proposed in this invention and other algorithms are shown.

[0037] Figure 7cTo compare the RUL prediction curves and actual curves of the NASA random walk charge-discharge dataset RW9 from the HIs perspective in the experiment, the RUL prediction curves of the algorithm proposed in this invention and other algorithms are shown.

[0038] Figure 7d To compare the RUL prediction curves and actual curves of the NASA random walk charge-discharge dataset RW10 from the HIs perspective in the experiment, the RUL prediction curves of the algorithm proposed in this invention and other algorithms are shown.

[0039] Figure 8a To compare the RUL prediction curves and actual curves of the lithium-ion battery data B5 from the NASA constant voltage and constant current dataset in the battery capacity degradation data perspective of the experiment, the RUL prediction curves of the algorithm proposed in this invention and other algorithms are shown.

[0040] Figure 8b To compare the RUL prediction curves and actual curves of the proposed algorithm and other algorithms, based on the NASA constant voltage and constant current dataset for lithium-ion battery data B18 from the perspective of battery capacity degradation data in the experiment.

[0041] Figure 8c To compare the RUL prediction curves and actual curves of the algorithm proposed in this invention and other algorithms from the perspective of battery capacity degradation data in the experiment, this paper presents the data on the NASA group's random walk charge-discharge dataset RW9.

[0042] Figure 8d To compare the RUL prediction curves and actual curves of the algorithm proposed in this invention and other algorithms from the perspective of battery capacity degradation data in the experiment, this paper presents the data from the NASA group's random walk charge-discharge dataset RW10.

[0043] Figure 9a This is a graph showing the predicted and actual RUL curves of lithium-ion battery data B5 from the NASA constant voltage and constant current dataset in the ablation experiment, under different integration methods (SAE, HYPELS) in a single perspective (Perspective1, Perspective2) and a mixed perspective. Figure 9b This is a graph showing the predicted and actual RUL curves of lithium-ion battery data B18 from the NASA constant voltage and constant current dataset for ablation experiments, in different ensemble methods (SAE, HYPELS) from single perspectives (Perspective 1, Perspective 2) and mixed perspectives. Figure 9c The RUL prediction curves and actual curves of NASA's random walk charge-discharge dataset RW9 in ablation experiments are shown in different ensemble methods (SAE, HYPELS) for single-view (Perspective1, Perspective2) and mixed-view (SAE, HYPELS) perspectives. Figure 9dThe RUL prediction curves and actual curves of NASA's random walk charge-discharge dataset RW10 in ablation experiments are shown in different ensemble methods (SAE, HYPELS) for single-view (Perspective1, Perspective2) and mixed-view (SAE, HYPELS) perspectives. Detailed Implementation

[0044] 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.

[0045] This invention discloses a hybrid perspective ensemble learning method for improving the accuracy of predicting the remaining lifespan of lithium-ion batteries, comprising the following steps:

[0046] Step 1: The original dataset obtained from the perspective of battery HIs is shuffled and trained for capacity prediction using an online sequential extreme learning machine (OS-ELM).

[0047] Step 2: Decompose and reconstruct the data from the perspective of battery capacity degradation using the Fully Adaptive Noise Ensemble Empirical Mode Decomposition (CEEMDAN) algorithm;

[0048] Step 3: Perform time series forecasting and post-processing using the Autoregressive Differential Moving Average (ARIMA) model.

[0049] Step 4: Integrating the two perspectives, a meta-model dataset construction method is proposed for training OS-ELM, realizing RUL prediction from the hybrid perspective.

[0050] Specifically, in step 1:

[0051] The two datasets used in the experiment came from NASA and were collected from batteries with graphite anodes and lithium nickel cobalt manganese oxygen cathodes, respectively, under two different experimental conditions. The first dataset was obtained by charging and discharging the batteries under constant voltage and constant current conditions provided by NASA's Ames Prediction Center. The battery capacity degradation process is as follows: Figure 1a The figures shown are B5 and B18, respectively. The 18650 lithium-ion batteries were tested at 25°C, including three different operating modes: charging, discharging, and impedance. The fault threshold was set at 70% of the rated capacity.

[0052] The second dataset comes from the Random Walk (RW) charge and discharge dataset in NASA's open-source database. Each 18650 lithium-ion battery was run with a continuous series of charge and discharge currents (ranging from -4.5A to 4.5A). This type of charge and discharge operation is called a Random Walk (RW) operation. The batteries used in this experiment were RW9 and RW10, with a failure threshold set at 70% of rated capacity. Their capacity degradation process is as follows... Figure 1b As shown.

[0053] like Figure 2 As shown, Figure 2 The front Perspective 1 represents the HIs perspective, and the bottom Perspective 2 represents the capacity degradation data perspective. Figure 2 The side is Figure 1a The capacity degradation curves of B5 and B18 batteries are shown. OS-ELM is an improved Extreme Learning Machine (ELM) algorithm that overcomes the limitation of ELM's inability to learn online. OS-ELM outperforms ELM because it adjusts the weights in each iteration, allowing the network to approach the optimal solution faster. However, current methods for applying OS-ELM to lithium-ion battery RUL prediction typically divide the original dataset into multiple groups according to time series and then feed them into the network for training sequentially. This approach causes the network to learn only from the data distribution of the first few cycles in the initial stage, while remaining unaware of the data distribution of other groups, thus affecting the quality of the network's learning of the overall dataset. In other words, simply dividing the dataset according to time series to train OS-ELM makes the network susceptible to the influence of time-series data distribution, preventing it from purely considering the HIs (Heat, Values, and Influences) for learning, thus introducing unnecessary interference terms and making it difficult for the network to converge to a better solution. To ensure that OS-ELM training is only related to HIs and not to time series, it is necessary to shuffle the dataset. The formula is shown below.

[0054] D'=σ(D) (1)

[0055] Where D represents the original dataset, and D' represents the shuffled dataset. σ represents the permutation function that rearranges the samples in dataset D.

[0056] OS-ELM consists of two phases: an initialization phase and an online sequential learning phase. In the initialization phase, OS-ELM is equivalent to ELM, where OS-ELM uses the first N0 training samples. To train the network. i It is the input vector of the i-th sample, t i This corresponds to the target output vector. The number of hidden layer nodes is set to k, the activation function is of type g(x), and the process is performed by minimizing... The optimal output weight matrix β0 is obtained by using this method. Here, w is the input weight, b is the hidden layer bias, β is the weight from the hidden layer to the output layer, and T0 is the target output matrix. Based on the above analysis, the processing procedure from the HIs perspective is as follows.

[0057] 1) Randomly generate weights w i and bias b i ,in,

[0058] 2) Calculate the hidden layer output matrix H0.

[0059]

[0060] 3) Calculate the output weight matrix β (0) .

[0061]

[0062] in

[0063] 4) In the online sequential learning phase, N1 sample data are used for learning, and the learning is achieved by minimizing... Obtain the most suitable output weight matrix β (1) .

[0064]

[0065] in,

[0066] 5) Substitute K1 into formula (4), and let The recursive formula for the online learning phase is as follows.

[0067]

[0068] Through the above process, the shuffled dataset is used to obtain prediction results based on the HIs perspective through the trained OS-ELM.

[0069] Specifically, in step 2:

[0070] To better understand and analyze signal characteristics, CEEMDAN is widely used in data preprocessing. The basic principle of CEEMDAN is to decompose a complex signal into a set of intrinsic mode functions (IMFs) and a residual signal (Res). By introducing adaptive noise, CEEMDAN can more effectively suppress mode aliasing in each iteration, which not only improves the stability of signal decomposition but also better handles nonlinear and non-stationary signals. The algorithm flowchart is shown below. Figure 3aAs shown. Assume x(t) is the original signal to be decomposed, and δ(t) is Gaussian white noise. Repeatedly add zero-mean Gaussian white noise to the original signal for decomposition. Use Empirical Mode Decomposition (EMD) to obtain the first Intrinsic Mode Function (IMF) and residual (Res). Add specific noise to the residual of the first IMF for further decomposition, obtaining the second IMF and residual. The EMD decomposition process continues until it can no longer be decomposed. The formulas for the final IMF and residual are as follows.

[0071]

[0072] r d (t)=r d-1 (t)-IMF d (t) (7)

[0073] In formulas (6) and (7), the IMF d (t) represents the d-th modal component obtained through CEEMDAN decomposition, E d-1 (.) represents the (d-1)th modal component obtained from the EMD decomposition sequence, ε d-1 This represents the weighting coefficient of the residual noise added by CEEMDAN to stage d-1. d (t) represents the residual at stage d. Taking B5 as an example, the components obtained through decomposition are as follows: Figure 4 As shown.

[0074] This invention uses CEEMDAN decomposition to obtain multiple IMFs and one Res. Then, the decomposed portion is divided into high-frequency and low-frequency components based on the zero-crossing rate. For a signal sequence, a zero-crossing point occurs when there are two adjacent sample values, one positive and the other negative. The formula for the zero-crossing rate is shown below.

[0075]

[0076] Where, n zero s represents the number of zero-crossing points of the signal. len Rate represents the length of the signal. zero-crossing This represents the zero-crossing rate. If the zero-crossing rate is greater than or equal to 0.01, it is classified as a high-frequency component; if the zero-crossing rate is less than 0.01, it is classified as a low-frequency component.

[0077] The components obtained from CEEMDAN decomposition and zero-crossing rate need to be properly reconstructed before being used for subsequent predictions. Without reconstruction, redundant information will exist, consuming unnecessary memory. Therefore, before prediction, we reconstruct the obtained high-frequency and low-frequency components according to different reconstruction rules to eliminate redundant information and improve prediction efficiency, such as... Figure 3bAs shown.

[0078] (i) Low-frequency components. Low-frequency components refer to components that change slowly and have longer periods, which are related to the basic trend or main characteristics of the signal. They represent the long-term trend and periodic components in the original signal. To prevent the loss of important information and to obtain low-frequency characteristics, only a simple summation process is needed for the low-frequency components.

[0079] (ii) The high-frequency component represents minute fluctuations and transient events in the signal, which can help us understand the instantaneous state and local characteristics of the signal. During reconstruction, we must consider not only global features but also local features. Therefore, this invention employs two reconstruction rules to reconstruct the high-frequency component. One rule uses the maximum value of each period of each high-frequency component as the value of the high-frequency feature for that period, called the maximum value reconstruction rule. The other rule fully considers the combined influence of a point and its surrounding points on the data change trend and uses the window energy maximization rule to reconstruct the high-frequency component. The detailed steps of the window energy maximization reconstruction rule are as follows.

[0080] Assume the decomposition produces m high-frequency components. Where e is the position of the data point in each high-frequency component.

[0081] 1) Select a window of size 3*1 and calculate the energy value of each high-frequency component within the window area.

[0082]

[0083] In the formula, This represents the energy value of each high-frequency component.

[0084] 2) Take the part at the center point where the window energy reaches its maximum value as i and take it as the high-frequency part at position i.

[0085]

[0086] In the formula, max_value represents the maximum value of the window energy, and index represents the index of the position of the maximum value in the high-frequency part.

[0087] Specifically, in step 3:

[0088] In the preprocessing of capacity degradation data, we obtained low-frequency and high-frequency features through decomposition and reconstruction. Due to the impact of capacity regeneration, low-frequency features often exhibit abrupt changes, while high-frequency features show many short-term fluctuations and obvious non-stationarity. ARIMA can capture short-term fluctuations and abrupt changes in the data, and can also convert non-stationary signals into stationary signals through differencing operations, thus effectively enabling modeling and prediction. Therefore, ARIMA is used to analyze high- and low-frequency features. A brief description of ARIMA is as follows.

[0089] 1) Perform g-order differencing on a non-stationary sequence until it becomes a stationary sequence. The formula for g-order differencing is as follows.

[0090] Δ g y t =Δ(Δ g-1 y t (12)

[0091] Wherein, the time series to be decomposed is y t Δ represents the difference operation.

[0092] 2) Use AR to process the autoregressive part of the time series.

[0093]

[0094] In the formula, y t This represents the current value, μ is a constant term, and γ... g It is the autocorrelation coefficient, ε t It represents the error. p is the order of AR, which is determined by the partial autocorrelation coefficient function.

[0095] 3) Use MA to process the moving average portion of the time series.

[0096]

[0097] In the formula, μ is a constant term, and θ g It is the autocorrelation coefficient, ε t It represents the error. q is the order of MA, which is determined by the autocorrelation coefficient function.

[0098] A baseline value can be obtained by analyzing the low-frequency features describing capacity regeneration phenomena and overall trends using ARIMA analysis. Further analysis of two high-frequency features representing non-stationary changes yields a calibration value. Adding the baseline value to the two calibration values ​​provides two more accurate final results from the perspective of capacity degradation data.

[0099] Specifically, in step 4:

[0100] We recognize that the two different perspectives differ not only in information features and model selection, but also in data preprocessing and prediction methods. Integrating these two perspectives into ensemble learning is a challenge. Therefore, we propose a method for constructing a meta-model dataset for hybrid perspective ensemble learning, with the following steps. Assume DR represents the original dataset of length L, and DS represents the shuffled dataset.

[0101] 1) Obtain DS by shuffling DR. Randomly extract r samples from DS as the training dataset DS. trainAnd record the index of these samples in DR as Idx DR (z), where z = 1, ..., s × r.

[0102] 2) Applying the idea of ​​k-fold cross-validation, the DS... train Divide the dataset into r subsets of equal size. Each subset serves as the validation set, and the remaining r-1 subsets are combined to form the training set. In this way, we obtain the combined training and validation sets, denoted as C. o , where o=1,..,r.

[0103] 3) From the perspective of HIs, traverse C o The model is trained and predicted to obtain s×r×1 sets of prediction results.

[0104] 4) From the perspective of capacity decay data, when traversing C o At that time, for each sample data, we in DS train Find its index z, and thus locate its index Idx in DR. DR (z). If Based on DR, we extract Idx DR (z) Perform ARIMA time series forecasting on several points prior to the current point; conversely, if... Based on DR, we extract Idx in reverse. DR ARIMA time series predictions were performed on several points following (z). After post-processing, we obtained s×r×2 sets of prediction results.

[0105] 5) By merging the prediction results obtained from the two perspectives, we obtained a meta-model training dataset DM of size s×r×3. train .

[0106] Through the above steps, we integrate the perspectives of HIs and capacity degradation data into ensemble learning, and ultimately obtain the DM for training the OS-ELM meta-model. train Finally, the prediction results are obtained through OS-ELM.

[0107] HyPELS leverages ensemble learning to fuse the perspectives of high-intensity (HI) and capacity-degraded data, fully utilizing the advantages of both to mitigate the limitations of a single perspective and improve the accuracy and robustness of RUL predictions. To achieve accurate predictions from the HI perspective, we shuffled the dataset to eliminate the influence of time factors during OS-ELM training. To achieve accurate predictions from the capacity-degraded data perspective, we performed data decomposition and reconstruction, ARIMA time series prediction, and post-processing. To integrate the two perspectives, we propose a meta-model dataset construction method to train OS-ELM, ultimately achieving the final integration of hybrid perspective RUL predictions. The overall algorithm flowchart of HyPELS is shown below. Figure 5 As shown.

[0108] To demonstrate the effectiveness and reliability of HyPELS, comparative and ablation experiments were conducted using two datasets. In the comparative experiments, algorithms from two different perspectives were compared. The actual and predicted curves of battery capacity degradation data, along with evaluation metrics for each algorithm, are shown below, demonstrating the superiority of HyPELS. In the ablation experiments, both the global to simple average ensemble (SAE) method and the two single-perspective stepwise ablation were performed. The actual and predicted curves of battery capacity degradation data, along with evaluation metrics for each part, are shown below, demonstrating the effectiveness of each perspective and ensemble method, thus demonstrating the rationale for HyPELS.

[0109] Mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) are introduced to measure the performance of HyPELS.

[0110]

[0111]

[0112] Among them, y k This refers to the actual capacity of the lithium-ion battery. This is the predicted capacity of lithium-ion batteries. is the average actual capacity of the lithium-ion battery, and N is the number of predicted data points.

[0113] To verify the feasibility of HyPELS, experiments were conducted with 70% of the initial capacity of the two battery groups as the failure threshold. Table 1 shows the predicted remaining lifespan of each lithium-ion battery group and compares the predicted and actual remaining lifespan values. MAE, RMSE, and R2 were used to evaluate the performance of different algorithms. Figure 6a , 6b Figures 6c and 6d demonstrate the fit between the predicted and actual curves of battery capacity degradation data.

[0114] Table 1. Prediction results of the two battery groups using HyPELS

[0115]

[0116] As shown in Figure 6, the prediction curve of HyPELS for lithium-ion battery capacity degradation data is highly consistent with the actual curve, and the predicted lithium-ion battery failure threshold (referred to as RULpre) is close to the actual failure threshold (referred to as RUL). Table 1 shows that RMSE and MAE are very small, while R2 is close to 1, indicating that HyPELS can predict RUL well, thus proving the rationality of HyPELS.

[0117] To demonstrate the advantages of HyPELS, this section compares a series of algorithm evaluation metrics based on capacity degradation data and the HIs perspective. Figure 7a , Figure 7b , Figure 7c and Figure 7d This demonstrates the degree of fit between the predicted curve and the actual curve for each algorithm from the perspective of HIs. Figure 8a , Figure 8b , Figure 8c and Figure 8d The table shows the goodness of fit between the predicted and actual curves for each algorithm from the perspective of capacity degradation data. Detailed evaluation metrics for each algorithm are shown in Table 2. Here, Perspective 1 represents the HIs perspective, and Perspective 2 represents the capacity degradation data perspective shown in Table 3.

[0118] Table 2 Comparison of prediction results from different algorithms and HyPELS for each perspective

[0119]

[0120]

[0121] From HIs's perspective, such as Figure 7a , Figure 7b , Figure 7c , Figure 7d As shown in Table 2, OS-ELM exhibits the lowest MAE and RMSE among methods viewed from the same perspective, indicating minimal error. Furthermore, the predicted capacity curve for each battery group shows the best and most stable fit to the actual capacity curve, demonstrating the rationality of the OS-ELM method used. Figure 8a , Figure 8b , Figure 8c , Figure 8dAs shown in Table 2, from the perspective of capacity degradation data, CEEMDAN-ARIMA used in this experiment had the lowest predicted MAE and RMSE for each battery group, indicating that its error was much smaller than other algorithms at the same perspective, and its prediction accuracy was higher. Simultaneously, the RUL prediction curve for each battery group showed the best fit with the actual capacity degradation data curve at this perspective, demonstrating the significant advantage of this method in predicting lithium-ion battery capacity. Therefore, CEEMDAN-ARIMA was chosen for capacity prediction from the perspective of capacity degradation data. Experimental data show that, for both perspectives, the RUL capacity degradation data prediction curve obtained using HyPELS had a higher fit with the actual curve, and the prediction error was significantly smaller than other methods, proving the advantages of HyPELS.

[0122] To further demonstrate the effectiveness of HyPELS, we progressively ablate the system from a global perspective to a simple average ensemble (SAE) method and two single-viewpoint approaches. In our experiments, we compared single-viewpoint approaches Perspective1 and Perspective2 with the two proposed ensemble methods, SAE and HyPELS, using the evaluation metrics MAE, RMSE, and R². Figure 9a , Figure 9b , Figure 9c , Figure 9d Table 3 provides a comparison of capacity degradation data prediction curves and evaluation metrics for each method.

[0123] Table 3 Comparison of prediction results from algorithms, SAE, and HyPELS based on different quantity perspectives.

[0124]

[0125]

[0126] We averaged the prediction results of Perspective 1 and Perspective 2 using SAE. We then integrated the prediction results of Perspective 1 and Perspective 2 into the HyPELS prediction result using OS-ELM. By comparing evaluation metrics, it can be seen that the fit between the SAE-based capacity degradation data prediction curve and the actual curve is better than that of the single-view prediction results. The prediction error is significantly smaller than that of other single-view methods, thus proving that HyPELS is indeed superior to other prediction methods. HyPELS integrates two perspectives through OS-ELM, and its capacity degradation data prediction curve has a better fit with the actual curve than SAE, with smaller error, thus fully demonstrating the effectiveness and superiority of HyPELS.

[0127] The HyPELS proposed in this invention considers both HIs (Heat Intake) and capacity degradation data, overcoming the limitation of traditional lithium-ion battery RUL prediction methods that rely on only a single perspective. This method can extract more comprehensive features, contributing to more accurate and reliable prediction results.

[0128] From the perspective of HIS, considering the sensitivity of OS-ELM to time data, the dataset is shuffled to eliminate the influence of time series on model training and prevent unnecessary interference, thereby promoting the network to develop towards a better solution. Subsequently, the prediction results of Perspective1 are obtained by using OS-ELM on the shuffled dataset.

[0129] For capacity degradation data, CEEMDAN is used to decompose the data, and then the data is divided into high-frequency and low-frequency components based on the zero-crossing rate. The resulting components are then reconstructed according to reconstruction rules to obtain the reconstructed high-frequency and low-frequency features. Finally, ARIMA analysis and post-processing are used to obtain the prediction results for Perspective2.

[0130] To integrate predictions from both HIs (High-Intensity Loss) and capacity degradation (HUL) data, a stacking approach from ensemble learning was employed. OS-ELM was trained as the meta-model. To facilitate the training of the meta-model, a set of metadata construction rules was designed to integrate the two perspectives of lithium-ion batteries. Two NASA datasets were used to evaluate the performance of HyPELS. Comparative and ablation experiments demonstrate that HyPELS exhibits lower error and higher accuracy compared to other methods, highlighting its significant advantages in predicting the RUL of lithium-ion batteries.

Claims

1. A hybrid perspective ensemble learning method for improving the prediction accuracy of remaining lifespan of lithium-ion batteries, characterized in that: Includes the following steps: Step 1: The original dataset obtained from the perspective of battery HIs is shuffled and trained for capacity prediction using an online sequential extreme learning machine (OS-ELM). To ensure that the training of OS-ELM is only related to HIs and not to time series, it is necessary to shuffle the dataset. The formula is shown below. （1） in, Represents the original dataset. This represents the shuffled dataset; This indicates a rearrangement of the dataset. The permutation function of the middle sample; Step 2: Decompose and reconstruct the data from the perspective of battery capacity degradation using the fully adaptive noise ensemble empirical mode decomposition (CEEMDAN) algorithm; Step 3: Perform time series forecasting and post-processing using the Autoregressive Differential Moving Average (ARIMA) model; Step 4: Integrating the two perspectives, a meta-model dataset construction method is proposed for training OS-ELM, achieving RUL prediction from the hybrid perspective; the specific steps are as follows; let... The representative length is The original dataset, This represents the shuffled dataset; Step 4.1 By... To obtain by mixing and washing ;from Random extraction 1 sample as training dataset and these samples in The index records in the data are ,in ; Step 4.2 Application Cross-validation, Divided into Each subset is of equal size; each subset is used as the validation set in turn, while the remaining subsets are used as the validation set. The subsets are combined to form the training set; the combination of the training and validation sets is obtained, denoted as . ,in ; Step 4.3 From the perspective of HIs, traverse Perform model training and prediction to obtain Group prediction results; Step 4.4 From the perspective of capacity decay data, during the traversal... At that time, for each sample data, in Find its index And therefore in Locating its index ;if Based on ,extract ARIMA time series forecasting is performed on several previous points; conversely, if... Then, based on DR, reverse extraction is performed. ARIMA time series forecasting was then performed on several subsequent points; after post-processing, the following results were obtained. Group prediction results; Step 4.5 merges the prediction results obtained from the two perspectives to obtain a result of size [size missing]. Metamodel training dataset ; Through the above steps, the perspectives of HIs and capacity degradation data are integrated into ensemble learning, ultimately yielding a meta-model for OS-ELM training. Finally, the prediction results are obtained through OS-ELM.

2. The hybrid perspective ensemble learning method for improving the prediction accuracy of remaining lifespan of lithium-ion batteries according to claim 1, characterized in that: In step 1, the two sets of datasets used in the experiment came from NASA and were collected from batteries with graphite anodes and lithium nickel cobalt manganese oxygen cathodes, respectively. The first dataset was obtained by charging and discharging batteries under constant voltage and constant current conditions provided by NASA's Ames Prediction Center; the second dataset comes from the random walk charging and discharging dataset in NASA's open-source database.

3. The hybrid perspective ensemble learning method for improving the prediction accuracy of remaining lifespan of lithium-ion batteries according to claim 1, characterized in that: OS-ELM consists of two phases: the initialization phase and the online sequential learning phase. During the initialization phase, OS-ELM is used before training samples To train the network; It is the input vector of the i-th sample. This corresponds to the target output vector; set the number of hidden layer nodes to... The type of activation function is And by minimizing To obtain the optimal output weight matrix ;in, These are the input weights. It is the hidden layer bias. These are the weights from the hidden layer to the output layer. It is the target output matrix.

4. The hybrid perspective ensemble learning method for improving the prediction accuracy of remaining lifespan of lithium-ion batteries according to claim 3, characterized in that: The processing procedure based on the HIs perspective is as follows; Step 1.1 Randomly generate weights and bias ,in, ; Step 1.2 Calculate the hidden layer output matrix ; （2） Step 1.3 Calculate the output weight matrix ; （3） in , ; Step 1.4 During the online sequential learning phase, use Learn from sample data and minimize Obtain the most suitable output weight matrix ; （4） in, ; Step 1.5 will Substitute into formula (4), let The recursive formula for the online learning phase is: （5） Through the above process, the shuffled dataset is used to obtain prediction results based on the HIs perspective through the trained OS-ELM.

5. The hybrid perspective ensemble learning method for improving the prediction accuracy of remaining lifespan of lithium-ion batteries according to claim 1, characterized in that: In step 2, specifically: let x(t) be the original signal to be decomposed, and δ(t) be Gaussian white noise; repeatedly add Gaussian white noise with zero mean to the original signal for decomposition; use Empirical Mode Decomposition (EMD) to decompose and obtain the first Intrinsic Mode Function (IMF) and residual Res; add specific noise to the residual of the first IMF for further decomposition to obtain the second IMF and residual; the EMD decomposition process continues until it can no longer be decomposed; the formulas for the IMF and residual in the final stage are as follows; （6） （7） In formulas (6) and (7), This represents the first decomposition obtained through CEEMDAN. One modal component, This represents the sequence obtained through EMD decomposition. One modal component, This indicates that CEEMDAN is sending a message to the first... Weighting coefficients for residual noise added at each stage; Indicates the first The residual of the stage.

6. The hybrid perspective ensemble learning method for improving the prediction accuracy of remaining lifespan of lithium-ion batteries according to claim 5, characterized in that: The original capacity degradation data is decomposed using CEEMDAN to obtain multiple IMFs and one Res; then, the decomposed part is divided into high frequency and low frequency according to the zero-crossing rate; if there are two adjacent sample values, one of which is positive and the other is negative, a zero-crossing point will be generated; the formula for the zero-crossing rate is as follows. （8） in, Indicates the number of zero-crossing points of the signal. Indicates the length of the signal. This indicates the zero-crossing rate; if the zero-crossing rate is greater than or equal to 0.01, it is classified as a high-frequency component; if the zero-crossing rate is less than 0.01, it is classified as a low-frequency component.

7. A hybrid perspective ensemble learning method for improving the prediction accuracy of remaining lifespan of lithium-ion batteries according to claim 5 or 6, characterized in that: The steps of the window energy maximization reconstruction rule are as follows; Suppose the decomposition produced High-frequency components ,in, It refers to the location of the data point in each high-frequency component; Step 2.1 Select size as The window is defined, and the energy value of each high-frequency component within the window region is calculated. （9） In the formula, This represents the energy value of each high-frequency component; Step 2.2: The center point where the window energy reaches its maximum value is taken as... Part as position The high-frequency part; （10） （11） In the formula, This represents the maximum value of the window energy. This indicates the index of the position of the maximum value in the high-frequency component.

8. The hybrid perspective ensemble learning method for improving the prediction accuracy of remaining lifespan of lithium-ion batteries according to claim 1, characterized in that: In step 3, ARIMA is used to analyze high and low frequency features, and the steps are as follows; Step 3.1 Perform processing on non-stationary sequences Differences are repeated until the sequence becomes stationary; The formula for the order difference is as follows; （12） The time series to be decomposed is: , Indicates the difference operation; Step 3.2 Use AR to process the autoregressive component of the time series; （13） In the formula, Indicates the current value. It is a constant term. It is the autocorrelation coefficient. It is an error; It is the order of AR, which is determined by the partial autocorrelation coefficient function; Step 3.3 Use MA to process the moving average portion of the time series; （14） In the formula, It is a constant term. It is the autocorrelation coefficient. It is an error; It is the order of MA, which is determined by the autocorrelation coefficient function.

Images