Lithium battery early-stage RUL prediction method based on transfer learning and unscented particle filter (UPF) fusion model

Abstract

The invention provides a lithium battery early-stage RUL prediction method based on a transfer learning and unscented particle filter (UPF) fusion model, and belongs to the technical field of power batteries and energy storage. The method comprises the following steps of: smoothing and denoising an original data set by using a moving average method and a residual analysis method, and dividing the processed data into source domain data and target domain data; inputting the source domain data into the Mogrimer LSTM neural network for training to obtain a pre-training model, and performing fine tuning on the pre-training model by using a small amount of target domain data so as to realize RUL prediction of other batteries in the same type but in different experimental environments; in order to further improve the prediction precision, a double-exponential capacity attenuation empirical model is sampled to establish a state equation; and taking an RUL prediction value output by the fine-tuned model as an observation value of the UPF, obtaining a posterior capacity prediction value through each iteration update, and judging whether the posterior capacity prediction value reaches a capacity threshold value or not, thereby realizing prediction of the remaining service life of the lithium battery.

Description

Technical Field

[0001] The present invention belongs to the technical field of power batteries and energy storage, and in particular, relates to an early RUL prediction method for lithium batteries based on a transfer learning and unscented particle filter (UPF) fusion model. Background Art

[0002] In recent years, lithium batteries have been widely used in the fields of power batteries and energy storage due to their advantages such as high energy density, long life and low self-discharge rate. However, as the use time increases, the performance of lithium batteries gradually declines and the probability of failure increases, which may lead to huge economic losses and even safety hazards. Therefore, the study of advanced battery prognostic and health management (PHM) methods, especially the remaining useful life (RUL) prediction, has become an important topic to ensure battery safety and reliability. However, the capacity decay of lithium batteries is affected by many factors, such as lithium ion loss in electrochemical reactions, structural changes in positive and negative electrode materials, and use conditions such as charge and discharge cycles, current, voltage and temperature. These complex factors make it a challenge to accurately predict the RUL of lithium batteries.

[0003] Existing RUL estimation methods can be divided into model-based and data-driven methods; model-based methods make predictions by establishing mathematical models of chemical reactions or degradation mechanisms inside batteries, including electrochemical models, semi-empirical models, and statistical filters; electrochemical models can accurately describe the chemical reactions inside lithium batteries and have relatively clear physical meanings, but they require a large number of battery parameters, which are difficult to obtain in practice; semi-empirical models are modeled by combining battery attenuation laws and measurement data. The advantage is that the construction and implementation are relatively simple and do not require a lot of prior knowledge; but because the model relies on simple mathematical formulas, it is difficult to accurately capture complex battery dynamics processes; the statistical filter method first models the battery and estimates parameters or corrects the state through filtering methods; its advantage is that it can update the system state and RUL estimation in real time, which is suitable for applications in dynamic environments. The disadvantage is that it relies on accurate system models;

[0004] Unlike model-based methods, data-driven methods do not rely on detailed analysis of the internal reactions of the battery, but instead use data mining to reveal the intrinsic correlation between input and output, thereby learning and capturing the law of lithium battery capacity decay. In the field of battery capacity prediction, a large number of studies have been devoted to exploring various machine learning methods to infer capacity change trends and predict future capacity curves. These methods include support vector machines (SVM), Gaussian process regression (GPR), long short-term memory networks (LSTM), and gated recurrent unit networks (GRU). These models usually rely on large-scale data training to capture the complex nonlinear mapping relationship between input and output variables, so as to accurately fit the dynamic changes of battery capacity in time series prediction.

[0005] Different batteries have significant differences in the aging process. However, traditional machine learning methods usually assume that the data distribution of the training set and the test set should be consistent. In practical applications, this assumption is often difficult to hold due to the wide variety of lithium batteries and different working environments. In order to obtain cycle data for new batteries, a large number of experiments are usually required, which is time-consuming and costly. Fortunately, different batteries tend to follow similar degradation patterns. During the charge and discharge cycle, the structural changes and wear of the positive and negative electrode materials will gradually reduce their efficiency in embedding and de-embedding lithium ions, thereby causing battery capacity decay. Therefore, how to effectively transfer the degradation characteristics of one battery to another becomes the key to improving the accuracy of lithium battery aging prediction, and transfer learning provides a practical solution to this problem. Summary of the invention

[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a lithium battery early RUL prediction method based on transfer learning and unscented particle filter (UPF) fusion model, which solves the problems raised in the above background technology.

[0007] In order to solve the above technical problems, the basic concept of the technical solution adopted by the present invention is:

[0008] A lithium battery early RUL prediction method based on transfer learning and unscented particle filter (UPF) fusion model includes the following steps:

[0009] Step 1: Data preprocessing: First, the lithium battery is subjected to cycle charge and discharge experiments at different temperatures and charge and discharge rates, and samples are taken regularly to obtain the original capacity data set; second, the original data is smoothed using the moving average method to reduce random fluctuations; then, residual analysis is performed by calculating the standard residual between the original data and the smoothed data (i.e. (original data - smoothed data) / standard deviation of the original data). The specific formula for calculating the standard residual is:

[0010] The specific formula for calculating the standardized residual is:

[0011]

[0012] Where Q i and They represent the original capacity data and the data after smoothing, respectively. RMSE represents the root mean square error of the original data and the smoothed data. The sample values ​​outside the 99% confidence interval are marked as outliers. The outliers are then corrected using linear interpolation, and the outliers are estimated and replaced based on the adjacent normal values. In order to eliminate data bias and improve model training efficiency, the Z-Score normalization method is used to process the capacity value after removing the outliers. The specific normalization formula is:

[0013]

[0014] in μ represents the average value of the processed data. σ represents the standard deviation of the processed data, X Represents the processed data, Z Represents normalized capacity data;

[0015] Step 2: The neural network implements RUL prediction, including:

[0016] Step 2.1: Single-step prediction: Train the Mogrifier-LSTM network to predict the capacity value of the next cycle based on the capacity values ​​of the first m cycles; the sliding window method is used to construct the training samples of the Mogrifier-LSTM network. The specific operation is: set the window length to m+1 (the first m cycles as input, the m+1th cycle as output), and slide along the entire capacity value sequence step by step to obtain input and output pairs. Each training sample is constructed accordingly, which can be specifically expressed as:

[0017] Q=[Q1,…Q t , …Q N ]

[0018] Where Q t represents the capacity value of the tth cycle, N represents the sequence length, and the entire battery capacity value sequence, from any continuous sequence of m+1 capacity values ​​[Q i , Q i+1 , …Q i+m ] to obtain an input-output pair, the input of the network is represented as x i =[Q i , Q j , …Q i+m-1 ], the output is y i =Q i+m , the relationship is written as i =f(x i ), namely Q i+m =f([Q i , …Q j , …Q i+m-1 ]), where f(·) is the mapping relationship learned by the neural network through training;

[0019] Step 2.2: Iterative prediction: Based on the single-step prediction, the prediction results of the previous step are continuously used as input to gradually calculate the capacity change trajectory of the lithium battery until it reaches 75% of the rated capacity. The specific implementation method is as follows:

[0020] Given input x1 = [Q1, ...Q i, …Q m ], the capacity value predicted by the model is

[0021]

[0022] The predicted results To update x1 to get the new input x2:

[0023] x2=[Q2,…Q i ,… m+1 ]

[0024] Then the predicted capacity value of x2 is:

[0025]

[0026] By analogy, the predicted value of the kth step is expressed as:

[0027]

[0028] By performing iterative prediction according to the above recursive steps, the capacity change trajectory of the lithium battery can be obtained until the capacity reaches the set threshold, that is, 75% of the rated capacity;

[0029] Step 2.3: Transfer learning: Divide the preprocessed data in step 1 into source domain data and target domain data, use the source domain data as the input of the Mogrifier-LSTM model for pre-training, and use the target domain data to fine-tune the model to achieve transfer learning, and obtain the fine-tuned transfer learning model TL-Mogrifier-LSTM; in the fine-tuning process, freeze the Mogrifier-LSTM layer, and only use the target domain data to optimize the parameters of the fully connected layer;

[0030] Step 3: RUL prediction method based on fusion model: The prediction value of the neural network is dynamically corrected by introducing an unscented particle filter (UPF), and the capacity value predicted by the TL-Mogrifier-LSTM neural network obtained in step 2 is used as the observation value of the UPF. The capacity prediction value is updated in each iteration of the prediction stage; the specific steps of using the capacity value predicted by the TL-Mogrifier-LSTM neural network as the observation value of the UPF include:

[0031] Step 3.1: Use the double exponential capacity decay model to construct the state space equation; the specific formula for constructing the state space equation includes:

[0032]

[0033] Where a, b, c, d represent model parameters, k represents the kth cycle of the lithium battery, and Q represents the estimated capacity of the battery. N(0, σ) represents Gaussian white noise with a mean of 0 and a standard deviation of σ, w represents process noise, and v represents observation noise;

[0034] Step 3.2: Use the UPF algorithm to track and determine the parameters of the double exponential capacity decay model in step 3.1; where the double exponential capacity decay model parameter is a k , b k , c k , d k , the UPF algorithm specifically includes the following steps:

[0035] Step 3.21: Initialize UPF related parameters: initial value of state vector, initial covariance matrix, number of particles N;

[0036] Step 3.22: Importance sampling stage: Use the UKF algorithm combined with the state space equation to calculate the state and observation distribution of each particle;

[0037] Step 3.23: Weight correction: Calculate the weight update factor of each particle and normalize the particle weight so that the sum of the weights is 1.

[0038] Step 3.24: Resampling: Resampling according to particle weight;

[0039] Step 3.25: Estimate output: Use the updated particles and weights to get the estimated value of the system state;

[0040] Step 3.26: Repeat steps 3.21 to 3.25 until the prediction starting point T is reached;

[0041] Step 3.27: Output the double exponential capacity decay model parameter a k , b k , c k , d k ;

[0042] Step 3.3: Starting from the prediction starting point, use the equation in step 3.1 as the state equation, take the capacity prediction value of the TL-Mogrifier-LSTM model as the observation value, apply the UPF algorithm to dynamically update the model parameters, obtain the posterior capacity estimation value, and use the posterior capacity estimation value as the input of the neural network model to predict the capacity value of the next cycle. Repeat this cycle iteratively to gradually predict the battery capacity change until the set capacity threshold is reached.

[0043] After adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art. Of course, any product implementing the present invention does not necessarily need to achieve all the advantages described below at the same time:

[0044] 1. By applying transfer learning and improving the long short-term memory network (LSTM), the TL-Mogrifier-LSTM model constructed by the present invention has stronger feature capture and nonlinear modeling capabilities than traditional models. The model can flexibly adapt to the characteristics of different batteries and still achieve good RUL prediction accuracy when data is insufficient, greatly reducing the demand for the amount of target battery data.

[0045] 2. The introduction of the unscented particle filter (UPF) algorithm performs real-time dynamic correction on the prediction value of the neural network, significantly improving the prediction accuracy. UPF updates the capacity prediction value in each iteration, effectively suppressing the error accumulation problem that may occur when relying solely on neural networks for long-term predictions, thereby ensuring the accuracy and reliability of RUL predictions.

[0046] The specific implementation modes of the present invention are further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The drawings described below are only some embodiments. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0048] Figure 1 It is a schematic diagram of the data set preprocessing process of the present invention;

[0049] Figure 2 It is a schematic diagram of the structure of the neural network Mogrifier-LSTM used in the present invention;

[0050] Figure 3 It is a schematic diagram of the overall process of the method for predicting the remaining service life of a lithium battery based on transfer learning and UPF fusion model provided by the present invention;

[0051] Figure 4 It is a battery capacity curve diagram predicted by different sampling methods of the present invention;

[0052] Figure 5 It is a curve diagram of battery capacity error predicted by different sampling methods of the present invention.

[0053] It should be noted that these drawings and textual descriptions are not intended to limit the conceptual scope of the present invention in any way, but are intended to illustrate the concept of the present invention for those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0054] The present invention will now be described in further detail with reference to the accompanying drawings.

[0055] Example 1: Please refer to Figure 1-5 As shown, in this embodiment, a lithium battery early RUL prediction method based on transfer learning and unscented particle filter (UPF) fusion model is provided, comprising the following steps:

[0056] Step 1: Data preprocessing: First, the lithium battery is subjected to cycle charge and discharge experiments at different temperatures and charge and discharge rates, and samples are taken regularly to obtain the original capacity data set; second, the original data is smoothed using the moving average method to reduce random fluctuations; then, residual analysis is performed by calculating the standard residual between the original data and the smoothed data (i.e. (original data - smoothed data) / standard deviation of the original data). The specific formula for calculating the standard residual is:

[0057]

[0058] Where Q i and Q i They represent the original capacity data and the data after smoothing, respectively. RMSE represents the root mean square error of the original data and the smoothed data. The sample values ​​outside the 99% confidence interval are marked as outliers. The outliers are then corrected using linear interpolation, and the outliers are estimated and replaced based on the adjacent normal values. In order to eliminate data bias and improve model training efficiency, the Z-Score normalization method is used to process the capacity value after removing the outliers. The specific normalization formula is:

[0059]

[0060] Where μ represents the mean value of the processed data, σ represents the standard deviation of the processed data, X represents the processed data, and Z represents the normalized capacity data;

[0061] Step 2: The neural network implements RUL prediction, including:

[0062] Step 2.1: Single-step prediction: Train the Mogrifier-LSTM network to predict the capacity value of the next cycle based on the capacity values ​​of the first m cycles; the sliding window method is used to construct the training samples of the Mogrifier-LSTM network. The specific operation is: set the window length to m+1 (the first m cycles as input, the m+1th cycle as output), and slide along the entire capacity value sequence step by step to obtain input and output pairs. Each training sample is constructed accordingly, which can be specifically expressed as:

[0063] Q=[Q1,…Q t , …Q N ]

[0064] Where Q t represents the capacity value of the tth cycle, N represents the sequence length, and the entire battery capacity value sequence, from any continuous sequence of m+1 capacity values ​​[Q i , Q i+1 , …Q i+m ] to obtain an input-output pair, the input of the network is represented as x i =[Q i , …Q j , …Q i+m-1 ], the output is y i =Q i+m , the relationship is written as i =f(x i ), namely Q i+m =f([Q i , …Q j , …Q i+m-1 ]), where f(·) is the mapping relationship learned by the neural network through training;

[0065] Step 2.2: Iterative prediction: Based on the single-step prediction, the prediction results of the previous step are continuously used as input to gradually calculate the capacity change trajectory of the lithium battery until it reaches 75% of the rated capacity. The specific implementation method is as follows:

[0066] Given input x1 = [Q1,…Q i , …Q m ], the capacity value predicted by the model is

[0067]

[0068] The predicted results To update x1 to get the new input x2:

[0069] x2=[Q2,…Q i , …Q m+1 ]

[0070] Then the predicted capacity value of x2 is:

[0071]

[0072] By analogy, the predicted value of the kth step is expressed as:

[0073]

[0074] By performing iterative prediction according to the above recursive steps, the capacity change trajectory of the lithium battery can be obtained until the capacity reaches the set threshold, that is, 75% of the rated capacity;

[0075] Step 2.3: Transfer learning: Divide the preprocessed data in step 1 into source domain data and target domain data, use the source domain data as the input of the Mogrifier-LSTM model for pre-training, and use the target domain data to fine-tune the model to achieve transfer learning, and obtain the fine-tuned transfer learning model TL-Mogrifier-LSTM; in the fine-tuning process, freeze the Mogrifier-LSTM layer, and only use the target domain data to optimize the parameters of the fully connected layer;

[0076] Step 3: RUL prediction method based on fusion model: The prediction value of the neural network is dynamically corrected by introducing an unscented particle filter (UPF), and the capacity value predicted by the TL-Mogrifier-LSTM neural network obtained in step 2 is used as the observation value of the UPF. The capacity prediction value is updated in each iteration of the prediction stage; the specific steps of using the capacity value predicted by the TL-Mogrifier-LSTM neural network as the observation value of the UPF include:

[0077] Step 3.1: Use the double exponential capacity decay model to construct the state space equation; the specific formula for constructing the state space equation includes:

[0078]

[0079] Where a, b, c, d represent model parameters, k represents the kth cycle of the lithium battery, and Q represents the estimated capacity of the battery. N(0, σ) represents Gaussian white noise with a mean of 0 and a standard deviation of σ, ω represents process noise, and ν represents observation noise;

[0080] Step 3.2: Use the UPF algorithm to track and determine the parameters of the double exponential capacity decay model in step 3.1; where the double exponential capacity decay model parameter is a k , b k , c k , d k , the UPF algorithm specifically includes the following steps:

[0081] Step 3.21: Initialize UPF related parameters: initial value of state vector, initial covariance matrix, number of particles N;

[0082] Step 3.22: Importance sampling stage: Use the UKF algorithm combined with the state space equation to calculate the state and observation distribution of each particle;

[0083] Step 3.23: Weight correction: Calculate the weight update factor of each particle and normalize the particle weight so that the sum of the weights is 1.

[0084] Step 3.24: Resampling: Resampling according to particle weight;

[0085] Step 3.25: Estimate output: Use the updated particles and weights to get the estimated value of the system state;

[0086] Step 3.26: Repeat steps 3.21 to 3.25 until the prediction starting point T is reached;

[0087] Step 3.27: Output the double exponential capacity decay model parameter a k , b k , c k , d k ;

[0088] Step 3.3: Starting from the prediction starting point, use the equation in step 3.1 as the state equation, take the capacity prediction value of the TL-Mogrifier-LSTM model as the observation value, apply the UPF algorithm to dynamically update the model parameters, obtain the posterior capacity estimation value, and use the posterior capacity estimation value as the input of the neural network model to predict the capacity value of the next cycle. Repeat this cycle iteratively to gradually predict the battery capacity change until the set capacity threshold is reached.

[0089] 1. By applying transfer learning and improving the long short-term memory network (LSTM), the TL-Mogrifier-LSTM model constructed by the present invention has stronger feature capture and nonlinear modeling capabilities than traditional models. The model can flexibly adapt to the characteristics of different batteries and still achieve good RUL prediction accuracy when data is insufficient, greatly reducing the demand for the amount of target battery data.

[0090] 2. The introduction of the unscented particle filter (UPF) algorithm performs real-time dynamic correction on the prediction value of the neural network, significantly improving the prediction accuracy. UPF updates the capacity prediction value in each iteration, effectively suppressing the error accumulation problem that may occur when relying solely on neural networks for long-term predictions, thereby ensuring the accuracy and reliability of RUL predictions.

[0091] Embodiment 2: Figure 1 The data set preprocessing process of the present invention is shown. Figure 1 It can be seen that there is a certain amount of noise in the original data. After smoothing and outlier processing, data suitable for further analysis are obtained. Figure 2 The neural network structure model used in the present invention is shown. Figure 2 It can be seen that compared with the traditional LSTM structure, Mogrifier-LSTM enhances the information interaction ability of the input gate and the forget gate, thus having stronger nonlinear expression performance. Figure 3Describes the lithium battery RUL prediction process based on transfer learning and UPF fusion model. Figure 3 It can be seen that the preprocessed data is divided into two parts: one part is input into the Mogrifier-LSTM network for training to obtain a pre-trained model. Then the model is fine-tuned using the target domain data to generate the target model TL-Mogrifier-LSTM, and the capacity prediction value is obtained; the other part is used to obtain the initial parameter values ​​of the double exponential model. The capacity prediction value of the transfer learning model is used as the observation value of the UPF prediction model, and is updated during each iteration of the UPF algorithm to obtain the posterior capacity prediction value. Finally, the RUL prediction is achieved by comparing whether the posterior capacity prediction value reaches the capacity prediction threshold. The specific steps are as follows:

[0092] Step 1, data preprocessing: First, the lithium battery is subjected to cycle charge and discharge experiments at different temperatures and charge and discharge rates, and samples are taken regularly to obtain the original capacity data set; secondly, the original data is smoothed using the moving average method to reduce random fluctuations; then, residual analysis is performed by calculating the standard residual between the original data and the smoothed data (i.e. (original data - smoothed data) / standard deviation of the original data). The specific formula for calculating the standard residual is:

[0093]

[0094] Where Q i and Represent the original capacity data and the smoothed data, respectively, and RMSE represents the root mean square error of the original data and the smoothed data. The sample values ​​outside the 99% confidence interval are marked as outliers. Then the linear interpolation method is used to correct the outliers, and the outliers are estimated and replaced according to the adjacent normal values.

[0095] In addition, in order to eliminate data bias and improve model training efficiency, the present invention uses the Z-Score normalization method to process the capacity value after removing outliers, so as to obtain preprocessed data. The normalization strategy is as follows:

[0096]

[0097] Where μ represents the mean value of the processed data, σ represents the standard deviation of the processed data, X represents the processed data, and Z represents the normalized capacity data.

[0098] Step 2: Neural network realizes RUL prediction

[0099] Step 2.1, single-step prediction. In the present invention, we train the Mogrifier-LSTM network to predict the capacity value of the next cycle based on the capacity values ​​of the previous m cycles. The sample is constructed by a sampling sliding window method, with the window length set to m+1 (where the first m cycles are used as input and the m+1th cycle is used as output), and the entire capacity value sequence is scanned step by step to obtain input-output pairs. Each training sample can be expressed as, Q = [Q1, ...Q t , …Q N ], where Q t represents the capacity value of the tth cycle, and N represents the sequence length. Extending to the entire battery capacity value sequence, we can obtain the capacity value of any continuous m+1 capacity value sequence [Q i , Q i+1 , …Q i+m ] to obtain an input-output pair. The input of the network can be represented as x i =[Q i , …Q j , …Q i+m-1 ], the output is y i =Q i+m , this relationship can be written as i =f(x i ), namely Q i+m =f([Q i , …Q j , …Q i+m-1 ]), where f(·) is the mapping relationship learned by the neural network through training. In this way, the function of predicting the capacity value of the next cycle based on the capacity value of the previous m cycles is realized.

[0100] Step 2.2 Iterative prediction: In step 2.1, we used the Mogrifier-LSTM network to implement single-step prediction. Next, we will introduce how to use this model for long-term iterative prediction to obtain the capacity change trajectory of the battery before reaching the threshold capacity. The specific steps are as follows:

[0101] Given input x1 = [Q1, ...Q i , …Q m ], the capacity value predicted by the model is

[0102]

[0103] Next, the prediction results To update x1 to get the new input x2:

[0104] x2=[Q2,…Q i …Q m+1 ]

[0105] Then the predicted capacity value of x2 is:

[0106]

[0107] By analogy, the predicted value of the kth step can be expressed as:

[0108]

[0109] By performing iterative prediction according to the above recursive steps, the capacity change trajectory of the lithium battery can be obtained until the capacity reaches the set threshold, that is, 75% of the rated capacity.

[0110] Step 2.3 transfer learning, the neural network model used in the present invention is composed of a Mogrifier LSTM layer and a fully connected layer. First, the preprocessed data in step 1 is divided into source domain data and target domain data. Then, the source domain data is used as the input of the Mogrifier-LSTM model for pre-training. Subsequently, the model is fine-tuned using the data of the target domain to achieve transfer learning. During the fine-tuning process, the Mogrifier-LSTM layer is frozen, and only the target domain data is used to optimize and adjust the parameters of the fully connected layer, thereby obtaining a fine-tuned transfer learning model TL-Mogrifier-LSTM.

[0111] Step 3, RUL prediction method based on fusion model. The prediction value of the neural network is dynamically corrected by introducing the unscented particle filter (UPF). The capacity value predicted by the TL-Mogrifier-LSTM neural network obtained in step 2 is used as the observation value of the UPF, and the capacity prediction value is updated in each iteration of the prediction stage. Furthermore, the posterior capacity prediction value output by the UPF algorithm is input into the TL-Mogrifier-LSTM model to iteratively predict the capacity value of the next sampling point, thereby significantly improving the accuracy of the prediction.

[0112] Step 3.1 Construct the state space equation using the double exponential capacity decay model

[0113]

[0114] Where a, b, c, d represent model parameters, k represents the kth cycle of the lithium battery, and Q represents the estimated capacity of the battery. N(0, σ) represents Gaussian white noise with a mean of 0 and a standard deviation of σ, ω represents process noise, and ν represents observation noise;

[0115] Step 3.2 Use the UPF algorithm to track and determine the double exponential capacity decay model parameter a in step 3.1 k , b k , c k , d k, specifically the following steps:

[0116] a. Initialize UPF related parameters: initial value of state vector, initial covariance matrix, number of particles N;

[0117] b. Importance sampling stage: Use the UKF algorithm combined with the state space equation to calculate the state and observation distribution of each particle;

[0118] c. Weight correction: Calculate the weight update factor of each particle and normalize the particle weight so that the sum of the weights is 1, that is

[0119] d. Resampling: Resampling according to particle weight;

[0120] e. Estimated output: Use the updated particles and weights to get the estimated value of the system state;

[0121] f. Repeat steps ae until the prediction starting point T;

[0122] g. Output double exponential capacity decay model parameter a k , b k , c k , d k .

[0123] Step 3.3 starts from the prediction starting point T, takes the equation in step 3.1 as the state equation, takes the capacity prediction value of the TL-Mogrifier-LSTM model as the observation value, and applies the UPF algorithm to dynamically update the model parameters to obtain the posterior capacity estimation value. Subsequently, the posterior capacity estimation value is used as the input of the neural network model to predict the capacity value of the next cycle. This cycle is repeated iteratively to gradually predict the battery capacity change until the set capacity threshold is reached.

[0124] In order to verify the effect of the present invention, experiments were conducted.

[0125] The experimental data of the present invention uses the NCA battery data set, and the effectiveness of different methods in RUL prediction is evaluated by three indicators: RMSE (root mean square error), MAE (mean absolute error) and PE (prediction error). The experimental results are detailed in Table 1. Figure 4 and Figure 5Among them, Mogrifier-LSTM represents the experimental results of iterative prediction using the Mogrifier-LSTM network; TL-Mogrifier-LSTM represents the experimental results of the TL-Mogrifier-LSTM model obtained by first training the pre-trained model with source domain data and then fine-tuning the model with target domain data; TL-Mogrifier-LSTM-UPF is the experimental result of the transfer learning and UPF fusion model proposed in the present invention.

[0126] Table 1 Comparison of prediction results

[0127] Real RUL Predicting RUL PE RMSE MAE Mogrifier-LSTM 527 597 70 0.05947 0.05231 TL-Mogrifier-LSTM 527 548 21 0.3265 0.03239 TL-Mogrifier-LSTM-UPF 527 523 4 0.01800 0.01775

[0128] Table 1 shows the comparison of long-term prediction results of different methods. It can be seen that the RMSE, MAE, and PE errors of the method using only a single neural network model are large, and the TL-Mogrifier-LSTM method using transfer learning has significantly reduced the error, which shows that the use of transfer learning can effectively improve the RUL prediction effect of the battery. The last row of Table 1 is the fusion model method proposed by the present invention. Obviously, this method can obtain better RUL prediction results, indicating the effectiveness of this method.

[0129] Figure 4 and Figure 5 The capacity curves and prediction error values ​​of the lithium battery RUL prediction results of different methods used in the present invention are shown respectively. It can be observed from the figure that the method proposed in the present invention has a smaller capacity prediction error, and the prediction curve is more consistent with the actual value, indicating that this method has a significant advantage in accuracy.

[0130] The present invention is not limited to the above-mentioned embodiments. Anyone should be aware that any structural changes made under the enlightenment of the present invention, and any technical solutions that are the same or similar to the present invention, fall within the protection scope of the present invention. The technology, shape, and structural parts not described in detail in the present invention are all well-known technologies.

Claims

1. A lithium battery early RUL prediction method based on transfer learning and unscented particle filter (UPF) fusion model, characterized in that: The steps include: Step 1: Data preprocessing: Perform cyclic charge and discharge experiments on lithium batteries at different temperatures and different charge and discharge rates, and regularly sample lithium batteries to obtain the original data set of lithium batteries. Use the moving average method and residual analysis method to smooth and reduce noise on the original data set to reduce the impact of noise on subsequent model training; Step 2: The neural network implements RUL prediction, including: Step 2.1: Single-step prediction: Train the Mogrifier-LSTM network to predict the capacity value of the next cycle based on the capacity values ​​of the previous m cycles; Step 2.2: Iterative prediction: Based on the single-step prediction, the prediction result of the previous step is continuously used as input to gradually calculate the capacity change trajectory of the lithium battery until it reaches 75% of the rated capacity; Step 2.3: Transfer learning: Divide the preprocessed data in step 1 into source domain data and target domain data, use the source domain data as the input of the Mogrifier-LSTM model for pre-training, and use the target domain data to fine-tune the model to achieve transfer learning, and obtain the fine-tuned transfer learning model TL-Mogrifier-LSTM; Step 3: RUL prediction method based on fusion model: The prediction value of the neural network is dynamically corrected by introducing the unscented particle filter (UPF). The capacity value predicted by the TL-Mogrifier-LSTM neural network obtained in step 2 is used as the observation value of the UPF, and the capacity prediction value is updated in each iteration of the prediction stage.

2. According to claim 1, a lithium battery early RUL prediction method based on transfer learning and unscented particle filter (UPF) fusion model is characterized in that: In step 1, the data is preprocessed by first smoothing the initial capacity data using a moving average method, and then performing a residual analysis to calculate the standard residual using the smoothed data and the original capacity data.

3. According to claim 2, a lithium battery early RUL prediction method based on transfer learning and unscented particle filter (UPF) fusion model is characterized in that: The specific formula for calculating the standardized residual is: Where Q i and They represent the original capacity data and the smoothed data respectively, RMSE represents the root mean square error of the original data and the smoothed data, and the sample values ​​outside the 99% confidence interval are marked as outliers, which are then corrected using linear interpolation and replaced based on adjacent normal values.

4. The early RUL prediction method of lithium battery based on transfer learning and unscented particle filter (UPF) fusion model according to claim 3 is characterized in that: In order to eliminate data bias and improve model training efficiency, the Z-Score normalization method is used to process the capacity value after removing outliers. The specific normalization formula is: Where μ represents the mean value of the processed data, σ represents the standard deviation of the processed data, X represents the processed data, and Z represents the normalized capacity data.

5. The early RUL prediction method of lithium battery based on transfer learning and unscented particle filter (UPF) fusion model according to claim 1 is characterized in that: In step 2.1, the sliding window method is used to construct the training samples of the Mogrifier-LSTM network. The specific operation is: set the window length to m+1 (the first m cycles are used as input, and the m+1th cycle is used as output), and gradually slide along the entire capacity value sequence to obtain input and output pairs. Each training sample is constructed accordingly, which can be specifically expressed as: Q=[Q1,…Q t ,…Q N ] Where Q t represents the capacity value of the tth cycle, N represents the sequence length, and the entire battery capacity value sequence, from any continuous sequence of m+1 capacity values ​​[Q i , Q i+1 , …Q i+m ] to obtain an input-output pair, the input of the network is represented as x i =[Q i , …Q j , …Q i+m-1 ], the output is y i =Q i+m , the relationship is written as i =f(x i ), namely Q i+m =f([Q i , …Q j , …Q i+m-1 ]), where f(·) is the mapping relationship learned by the neural network through training.

6. The early RUL prediction method of lithium battery based on transfer learning and unscented particle filter (UPF) fusion model according to claim 1 is characterized in that: In step 2.2, the prediction formula is: Given input x1 = [Q1, ...Q i , …Q m ], the capacity value predicted by the model is The predicted results To update x1 to get the new input x2: x2=[Q2,…Q i ,…Q m+1 ] Then the predicted capacity value of x2 is: By analogy, the predicted value of the kth step is expressed as: By performing iterative prediction according to the above recursive steps, the capacity change trajectory of the lithium battery can be obtained until the capacity reaches the set threshold, that is, 75% of the rated capacity.

7. The early RUL prediction method of lithium battery based on transfer learning and unscented particle filter (UPF) fusion model according to claim 1 is characterized in that: In step 2.3, during the fine-tuning process, the Mogrifier-LSTM layer is frozen and only the target domain data is used to optimize the parameters of the fully connected layer.

8. The early RUL prediction method of lithium battery based on transfer learning and unscented particle filter (UPF) fusion model according to claim 1 is characterized in that: In step 3, the specific steps of using the capacity value predicted by the obtained TL-Mogrifier-LSTM neural network as the observed value of UPF include: Step 3.1: Construct the state space equation using the double exponential capacity decay model; Step 3.2: Use the UPF algorithm to track and determine the parameters of the double exponential capacity decay model in step 3.1; Step 3.3: Starting from the prediction starting point, use the equation in step 3.1 as the state equation, take the capacity prediction value of the TL-Mogrifier-LSTM model as the observation value, apply the UPF algorithm to dynamically update the model parameters, obtain the posterior capacity estimation value, and use the posterior capacity estimation value as the input of the neural network model to predict the capacity value of the next cycle. Repeat this cycle iteratively to gradually predict the battery capacity change until the set capacity threshold is reached.

9. The early RUL prediction method of lithium battery based on transfer learning and unscented particle filter (UPF) fusion model according to claim 8, characterized in that: In step 3.1, the specific formula for constructing the state space equation includes: Where a, b, c, d represent model parameters, k represents the kth cycle of the lithium battery, and Q represents the estimated capacity of the battery. N(0, σ) represents Gaussian white noise with a mean of 0 and a standard deviation of σ, ω represents process noise, and v represents observation noise.

10. The early RUL prediction method of lithium battery based on transfer learning and unscented particle filter (UPF) fusion model according to claim 8, characterized in that: In step 3.2, the double exponential capacity decay model parameter is a k , b k , c k , d k , the UPF algorithm specifically includes the following steps: Step 3.21: Initialize UPF related parameters: initial value of state vector, initial covariance matrix, number of particles N; Step 3.22: Importance sampling stage: Use the UKF algorithm combined with the state space equation to calculate the state and observation distribution of each particle; Step 3.23: Weight correction: Calculate the weight update factor of each particle and normalize the particle weight so that the sum of the weights is 1. Step 3.24: Resampling: Resampling according to particle weight; Step 3.25: Estimate output: Use the updated particles and weights to get the estimated value of the system state; Step 3.26: Repeat steps 3.21 to 3.25 until the prediction starting point T is reached; Step 3.27: Output the double exponential capacity decay model parameter a k , b k , c k , d k .

Images