Lithium battery rul prediction method, system and device based on data-driven and particle filtering fusion model and medium

Abstract

The lithium battery RUL prediction method, system, device and medium based on data driving and particle filtering fusion model, the method is: combining convolutional neural network (CNN) and gated recurrent unit (GRU) network to form a series structure of CNN-GRU, which can capture spatial and temporal information simultaneously and simplify the complex time series problem in the prior art; a double exponential empirical model is used to establish a state equation, the predicted value output by the CNN-GRU is used as an observation value, a state space equation of the CNN-GRU-PF fusion prediction model is established, and the lithium battery RUL is predicted based on the CNN-GRU-PF fusion model; the system, device and method based on the above method can efficiently predict the lithium battery RUL, can reduce the noise and uncertainty of the data, ensure the reliability of the model method, and have good prediction accuracy and good universal applicability.

Description

Technical Field

[0001] This invention belongs to the field of power battery and energy storage technology, and in particular relates to a lithium battery RUL prediction method, system, device and medium based on a data-driven and particle filter fusion model. Background Technology

[0002] Lithium-ion batteries are widely used in power batteries and energy storage due to their advantages such as large capacity, no memory effect, high energy density, and strong adaptability to high and low temperatures. As lithium batteries undergo continuous charge-discharge cycles, their capacity gradually decreases, and their performance gradually weakens. It is generally considered that when the actual capacity of a battery decays to 70%–80% of its rated capacity, the battery can no longer meet usage requirements, i.e., it has reached the end of life (EOL). The threshold used to determine whether a lithium battery has reached its EOL is called the failure threshold. The remaining usable life (RUL) of a lithium-ion battery can be considered as the normal usable time before the battery reaches the failure threshold during its service life.

[0003] Lithium-ion battery range-extended aging (RUL) prediction mainly falls into three categories: one is establishing mathematical models to simulate the battery aging process; another is directly utilizing historical data to build data-driven models; and the third is a fusion model combining mathematical and data-driven models. Mathematical model-based methods mainly include electrochemical models, equivalent circuit models, and empirical degradation models. A drawback of electrochemical modeling is that it is difficult for electrochemical models to accurately reflect battery characteristics due to the dynamic and nonlinear nature of electrochemical processes. Sadabadi, KK, and others used electrochemical models to estimate the internal parameters of lithium-ion batteries, thereby achieving RUL prediction. However, the drawback of electrochemical modeling is that it is difficult to accurately reflect battery characteristics due to the dynamic and nonlinear nature of electrochemical processes. Equivalent circuit modeling requires establishing complex electrochemical models and determining model structure and parameters. Therefore, the drawback of equivalent circuit modeling is its complex structure, making process parameter selection and model optimization difficult. Furthermore, this method relies on research and understanding of the battery charging and discharging mechanism, requiring substantial prior electrochemical knowledge. Data-driven methods essentially extract features that reflect the system's regularity from raw data through various feature extraction methods, such as statistical features and frequency domain features. Data-driven algorithms establish a mapping relationship between input features and output targets, such as mapping historical data to future data in time series forecasting. A learning model is trained using statistical learning theory and algorithms, enabling it to accurately predict the target output based on input features. Therefore, a drawback of data-driven methods is the need for a large amount of training data; otherwise, it is difficult to guarantee prediction accuracy. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, the present invention aims to provide a lithium battery RUL prediction method, system, device, and medium based on a data-driven and particle filter fusion model. This method combines Convolutional Neural Networks (CNN) and Gated Recurrent Units (GRU) networks to form a CNN-GRU cascaded structure, which can simultaneously capture spatial and temporal information, simplifying complex time series problems in existing technologies. A dual-exponential empirical model is used to establish the state equation, and the predicted values ​​output by the CNN-GRU are used as observations to establish the state space equation of the CNN-GRU-PF fusion prediction model. The RUL of lithium batteries is then predicted based on this CNN-GRU-PF fusion model. This approach reduces data noise and uncertainty, ensures the reliability of the model method, provides accurate predictions, and has good general applicability.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A lithium battery RUL prediction method based on a data-driven and particle filter fusion model is proposed. First, a convolutional neural network (CNN) and a gated recurrent unit (GRU) network are combined to form a CNN-GRU cascaded structure, which can simultaneously capture spatial and temporal information. Then, a state equation is established using a double exponential empirical model. The predicted values ​​output by the CNN-GRU are used as observation values ​​to establish the state space equation of the CNN-GRU-PF fusion prediction model. The RUL of the lithium battery is then predicted based on the CNN-GRU-PF fusion model.

[0007] A lithium battery RUL prediction method based on a data-driven and particle filter fusion model includes the following steps:

[0008] Step 1: Divide the Convolutional Neural Network (CNN) into convolutional layers and fully connected layers, and eliminate the use of pooling layers. The fully connected layers integrate and map the data output by the convolutional layers to obtain the classification or regression results.

[0009] By combining convolutional neural networks (CNN) and gated recurrent units (GRU) networks, a CNN-GRU network model with a cascaded structure is formed.

[0010] Step 2: Using the predicted values ​​output by CNN-GRU as observation values, establish a CNN-GRU-PF fusion prediction model and perform RUL prediction.

[0011] The CNN-GRU network model structure in step 1 is as follows:

[0012] 1) Input layer, used to receive lithium battery capacity degradation data and output it to the CNN layer;

[0013] 2) The CNN layer includes a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function. The convolutional layer performs convolution operations on the lithium battery capacity data from the input layer and extracts local features. The batch normalization layer standardizes the input data in the network, reducing variations in data distribution and accelerating network convergence. The ReLU activation function increases the network's non-linearity and overcomes the gradient vanishing problem. Then, a flattening layer maps the data to the output, which serves as the input data for the GRU module.

[0014] 3) GRU layer: The GRU layer receives the output data of the convolutional layer, trains the data, predicts the capacity of the lithium battery based on the training results, and outputs the predicted data to the Dropout layer.

[0015] 4) Dropout layer: The Dropout layer is used to prevent overfitting in the GRU neural network, reduce the interdependence between neurons, and enhance the network's generalization ability and robustness. A fully connected layer is added after the Dropout layer to map the output of the Dropout layer to the desired output dimension.

[0016] 5) Output layer: The output data of the Dropout layer is the lithium battery capacity at the next moment.

[0017] The GRU layer uses the tanh function to scale and translate the training data, reducing the data distribution range of GRU training and improving the efficiency of GRU training.

[0018] The specific method for step 2 is as follows:

[0019] 2.1) A dual-exponential empirical model is used to establish the state equation. The capacity prediction value output by CNN-GRU is used as the observation value to establish the state space equation of the CNN-GRU-PF fusion prediction model:

[0020] x k =[a k b k ;c k ;d k (1)

[0021]

[0022] Q k =a k * exp(b k *k)+c k * exp(d k *k)+γ k ,γ k ~N(0,σQ (3)

[0023] In the formula: a, b, c, and d are the internal parameters of the battery, which are also state variables; the battery capacity is an observable quantity; k is the charge-discharge cycle period of the battery; Q k ω is the capacity observation value for k cycles, provided by the capacity prediction value of the CNN-GRU model; ω and γ are the process noise and measurement noise, respectively, and both process noise and measurement noise are Gaussian noise;

[0024] 2.2) The filtering method of PF is adopted, and the capacity prediction value of CNN-GRU network is used to update the parameters a, b, c, and d of the double exponential empirical model;

[0025] 2.3) Calculate the new capacity prediction value using the updated parameters from step 2.2);

[0026] 2.4) Update the CNN-GRU training dataset with the new capacity prediction values ​​from step 2.3) to achieve the fusion of CNN-GRU neural network prediction information and double exponential empirical model prediction information; predict the RUL of lithium battery based on the CNN-GRU-PF fusion model.

[0027] The specific process of step 2.4) is as follows:

[0028] 2.4.1) Noise Reduction Preprocessing: Extract battery capacity data from the battery dataset and perform noise reduction preprocessing;

[0029] 2.4.2) Initialize parameters: Set the prediction start point s, lifetime threshold, initial PF value, and CNN-GRU parameters;

[0030] 2.4.3) Divide the training and test sets: Data_train(s) is the training data vector for the CNN-GRU model; Data_test is the test data vector for the CNN-GRU model; Data_train(s) consists of the volumes C1, C2, C3, ..., C of the first s epochs. s composition;

[0031] 2.4.4) Train a CNN-GRU network model to predict the battery capacity at cycle s+1, and name it Q. Ps+1 ;

[0032] 2.4.5) PF-based state estimation: For the state-space model composed of equations (1), (2), and (3), Q is... Ps+1 As a measurement in PF, the model state is estimated using the PF algorithm [a] k b k ;c k ;d kThis allows us to obtain the optimized capacity prediction value Q. s+1 ;

[0033] 2.4.6) Constructing the mobile training dataset: Using the optimized capacity prediction value Q obtained in step 2.4.5) s+1 Update the training dataset Data_train(s) of the CNN-GRU model, and then retrain the CNN-GRU neural network using the updated training dataset. The specific method for updating the training dataset is as follows: [The text abruptly ends here, likely due to an incomplete translation or a formatting error.] s+1 Adding Data_train(s) to the CNN-GRU model generates the training dataset Data_train(s+1); the updated Data_train(s+1) consists of C2, C3, ..., C s Q s+1 composition;

[0034] 2.4.7) Repeat steps 2.4.4) to 2.4.6) until the prediction step size reaches the total length of the dataset, then the loop ends;

[0035] 2.4.8) Evaluate the prediction results based on the test set data and calculate the RUL.

[0036] A lithium battery RUL prediction system based on a data-driven and particle filter fusion model includes:

[0037] CNN-GRU Adjustment Module: Used in step 1 to adjust the relationship between the layers of CNN and GRU, select a one-dimensional CNN and a one-layer GRU to perform lithium battery capacity prediction and provide capacity observations for PF;

[0038] CNN-GRU-PF module: used in step 2 to combine CNN-GRU and PF to obtain optimized lithium battery capacity prediction values ​​and calculate lithium battery RUL.

[0039] The CNN-GRU-PF module includes a data preprocessing module, a CNN-GRU prediction module, a PF state estimation module, a moving training dataset module, and an error analysis module.

[0040] The preprocessing module comprises three parts: noise reduction, setting initialization parameters, and dataset partitioning. Noise reduction utilizes the Fully Adaptive Empirical Mode Decomposition (CEEMDAN) algorithm to decompose lithium battery capacity data and reconstruct it using Pearson correlation coefficients to reduce noise in the battery capacity data. Setting initialization parameters is used to configure the parameters of the RUL prediction system for system startup. The dataset is partitioned into training and testing datasets for the RUL prediction system's prediction work and comparison of prediction results.

[0041] The CNN-GRU prediction module is trained based on the pre-divided training dataset to predict the capacity of lithium batteries and provide capacity observations for power PF.

[0042] Among them, the state estimation module of PF is used to correct the capacity prediction error of CNN-GRU and obtain the optimized capacity prediction value;

[0043] Among them, the moving training dataset module is used to add the capacity-optimized value of PF to the training set of the CNN-GRU module to improve the long-term prediction ability of CNN-GRU.

[0044] The error analysis module is used to analyze the accuracy of the prediction results and to calculate the RUL of the lithium battery.

[0045] A lithium battery RUL prediction device based on a data-driven and particle filter fusion model includes:

[0046] Memory, used to store computer programs;

[0047] A processor is used to implement the lithium battery RUL prediction method based on the data-driven and particle filter fusion model described in steps 1 to 2 when executing the computer program.

[0048] A computer-readable storage medium for storing a computer program that, when executed by a processor, is capable of performing a lithium battery RUL prediction based on a data-driven and particle filter fusion model according to the method described in steps 1 to 2.

[0049] Compared with the prior art, the present invention has the following advantages:

[0050] 1. To address the issue of noise in battery capacity data, which affects the accuracy of lithium battery RUL prediction, this invention utilizes the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) algorithm to decompose the lithium battery capacity data and reconstruct it using Pearson correlation coefficients, thereby reducing noise in the battery capacity data. Initialization parameters are set to configure the parameters of the RUL prediction system for system startup. The dataset is divided into training and testing datasets for comparing the prediction performance of the RUL prediction system. This reduces data noise and uncertainty, ensuring the reliability of the model method. This invention uses the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) algorithm to decompose the battery capacity sequence and reconstructs it using Pearson correlation coefficients, effectively reducing the noise present in the original data.

[0051] 2. This invention employs a comprehensive fusion method, combining Convolutional Neural Networks (CNN), Gated Recurrent Unit (GRU) neural networks, and Particle Filtering (PF). The advantage of this multi-model fusion method is that it fully leverages the strengths of different models, using them separately for feature extraction, temporal feature learning, and state prediction, thereby improving the accuracy and stability of RUL prediction. Secondly, to better adapt to long-term battery performance changes, this invention introduces an iterative training mechanism with a moving window. By dynamically updating the training dataset, the fusion model can more accurately capture the evolution trend of battery capacity and accurately predict its RUL during battery charge-discharge cycles, thus improving long-term prediction performance and achieving accurate predictions even with limited training data.

[0052] 3. Experimental tests were conducted on different datasets, demonstrating good prediction results on both the NASA and the Center for Advanced Life Cycle Engineering (CALCE) datasets at the University of Maryland. This fully proves that the battery RUL prediction method of this invention has sufficient versatility to perform predictions on different battery datasets, and is not limited to a specific dataset. This invention has a very good advantage of universal applicability and is suitable for various situations. Attached Figure Description

[0053] Figure 1 This is a structural diagram of the CNN-GRU network model of this invention.

[0054] Figure 2 This is a flowchart of the lithium battery RUL prediction using the CNN-GRU-PF fusion model of this invention.

[0055] Figure 3 This is the battery capacity degradation curve of the present invention, wherein, Figure 3 (a) shows the battery capacity degradation curve for dataset A. Figure 3 (b) shows the battery capacity degradation curve for dataset B.

[0056] Figure 4 This is an exploded view of the B5 battery CEEMDAN of the present invention.

[0057] Figure 5 These are the battery capacity prediction results of the four models of this invention, wherein, Figure 5 (a) is a B5 battery. Figure 5 (b) is a B6 battery. Figure 5 (c) is a B7 battery. Figure 5 (d) is a B18 battery. Figure 5 (e) is a CS35 battery. Figure 5 (f) is a CS36 battery. Figure 5(g) is a CS37 battery. Figure 5 (h) is a CS38 battery.

[0058] Figure 6 These are the prediction results from different prediction starting points of this invention, wherein, Figure 6 (a) is a B5 battery. Figure 6 (b) is a B6 battery. Figure 6 (c) is a B7 battery. Figure 6 (d) is a B18 battery. Figure 6 (e) is a CS35 battery. Figure 6 (f) is a CS36 battery. Figure 6 (g) is a CS37 battery. Figure 6 (h) is a CS38 battery. Detailed Implementation

[0059] The present invention will now be described in further detail with reference to the accompanying drawings.

[0060] A lithium battery RUL prediction method based on a data-driven and particle filter fusion model is proposed. First, a convolutional neural network (CNN) and a gated recurrent unit (GRU) network are combined to form a CNN-GRU cascaded structure, which can simultaneously capture spatial and temporal information, simplifying the complex time series problems in existing technologies. Then, a state equation is established using a double exponential empirical model. The predicted values ​​output by the CNN-GRU are used as observation values ​​to establish the state space equation of the CNN-GRU-PF fusion prediction model. The RUL of the lithium battery is then predicted based on the CNN-GRU-PF fusion model, achieving efficient prediction of lithium battery RUL.

[0061] A lithium battery RUL prediction method based on a data-driven and particle filter fusion model includes the following steps:

[0062] Step 1: Divide the Convolutional Neural Network (CNN) into convolutional layers and fully connected layers, and eliminate the use of pooling layers. The fully connected layers can integrate and map the data output by the convolutional layers to obtain classification or regression results.

[0063] The convolutional neural network (CNN) and gated recurrent unit (GRU) network are combined to form a cascaded CNN-GRU network model, and the layers are adjusted to select a one-dimensional CNN and a one-layer GRU.

[0064] Step 2: Using the predicted values ​​output by CNN-GRU as observations, establish a CNN-GRU-PF fusion prediction model and perform RUL prediction:

[0065] 2.1) A dual-exponential empirical model is used to establish the state equation. The capacity prediction value output by CNN-GRU is used as the observation value to establish the state space equation of the CNN-GRU-PF fusion prediction model:

[0066] x k =[a k b k ;c k ;d k (1)

[0067]

[0068] Q k =a k * exp(b k *k)+c k * exp(d k *k)+γ k ,γ k ~N(0,σ Q (3)

[0069] In the formula: a, b, c, and d are the internal parameters of the battery, which are also state variables; the battery capacity is an observable quantity; k is the charge-discharge cycle period of the battery; Q k ω is the capacity observation value for k cycles, provided by the capacity prediction value of the CNN-GRU model; ω and γ are the process noise and measurement noise, respectively, and both process noise and measurement noise are Gaussian noise;

[0070] 2.2) The filtering method of PF is adopted, and the capacity prediction value of CNN-GRU network is used to update the parameters a, b, c, and d of the double exponential empirical model;

[0071] 2.3) Calculate the new capacity prediction value using the updated parameters from step 2.2);

[0072] 2.4) Update the CNN-GRU training dataset with the new capacity prediction values ​​from step 2.3) to achieve the fusion of CNN-GRU neural network prediction information and dual-exponential empirical model prediction information; predict the RUL of lithium batteries based on the CNN-GRU-PF fusion model;

[0073] 2.4.1) Noise Reduction Preprocessing: Extract battery capacity data from the battery dataset and perform noise reduction preprocessing;

[0074] 2.4.2) Initialize parameters: Set the prediction start point s, lifetime threshold, initial PF value, and CNN-GRU parameters;

[0075] 2.4.3) Divide the training and test sets: Data_train(s) is the training data vector for the CNN-GRU model; Data_test is the test data vector for the CNN-GRU model; Data_train(s) consists of the volumes C1, C2, C3, ..., C of the first s epochs. s composition;

[0076] 2.4.4) Train a CNN-GRU network model to predict the battery capacity at cycle s+1, and name it Q. Ps+1 ;

[0077] 2.4.5) PF-based state estimation: For the state-space model composed of equations (1), (2), and (3), Q is... Ps+1 As a measurement in PF, the model state is estimated using the PF algorithm [a] k b k ;c k ;d k This allows us to obtain the optimized capacity prediction value Q. s+1 ;

[0078] 2.4.6) Constructing the mobile training dataset: Using the optimized capacity prediction value Q obtained in step 2.4.5) s+1 Update the training dataset Data_train(s) of the CNN-GRU model, and then retrain the CNN-GRU neural network using the updated training dataset. The specific method for updating the training dataset is as follows: [The text abruptly ends here, likely due to an incomplete translation or a formatting error.] s+1 Adding Data_train(s) to the CNN-GRU model generates the training dataset Data_train(s+1); the updated Data_train(s+1) consists of C2, C3, ..., C s Q s+1 composition;

[0079] 2.4.7) Repeat steps 2.4.4) to 2.4.6) until the prediction step size reaches the total length of the dataset, then the loop ends;

[0080] 2.4.8) Evaluate the prediction results based on the test set data and calculate the RUL.

[0081] The CNN-GRU network model structure in step 1 is as follows:

[0082] 1) Input layer, used to receive lithium battery capacity degradation data and output it to the CNN layer;

[0083] 2) The CNN layer includes a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function. The convolutional layer performs convolution operations on the lithium battery capacity data from the input layer and extracts local features. The batch normalization layer standardizes the input data in the network, reducing variations in data distribution and accelerating network convergence. The ReLU activation function increases the network's non-linearity and overcomes the gradient vanishing problem. Then, a flattening layer maps the data to the output, which serves as the input data for the GRU module.

[0084] 3) GRU layer: The GRU layer receives the output data from the convolutional layer and trains the data. It then predicts the capacity of the lithium battery based on the trained data. The tanh function is used to scale and translate the GRU training data to distribute it within a smaller range, thereby improving the efficiency of GRU training.

[0085] 4) Dropout layer: The Dropout layer is used to prevent overfitting in the GRU neural network, reduce the interdependence between neurons, and enhance the network's generalization ability and robustness. A fully connected layer is added after the Dropout layer to map the output of the Dropout layer to the desired output dimension.

[0086] 5) Output layer: The output data of the Dropout layer is the lithium battery capacity at the next moment.

[0087] Experimental Dataset Preprocessing and Evaluation Criteria

[0088] (1) Experimental Dataset and Preprocessing

[0089] This invention uses two different datasets. Dataset A is a lithium-ion battery aging dataset from NASA's PCoE laboratory. Dataset B is a lithium-ion battery dataset from the Center for Advanced Life Cycle Engineering (CALCE) at the University of Maryland. Dataset A contains data from four batteries: B5, B6, B7, and B18, all of which underwent charge-discharge aging tests using 18650 lithium-ion batteries with a nominal capacity of 2Ah. This dataset was obtained by performing CCCV cycle tests on the batteries under the same room temperature (24°C). First, the batteries were charged with a constant current of 1.5A until the battery voltage reached 4.2V. Then, they were charged in constant voltage mode until the charging current dropped below 20mA. The discharge test was performed with a constant current of 2A until the voltages of B5, B6, B7, and B18 dropped to 2.7V, 2.5V, 2.2V, and 2.5V, respectively.

[0090] Dataset B includes four batteries with lithium cobalt oxide (LiCoO2) as the cathode material and a nominal capacity of 1.1 Ah: CS35, CS36, CS37, and CS38. Under the same room temperature (24°C) environment, they were charged at a constant current rate of 0.5C until the battery voltage reached 4.2V. Charging was then continued in constant voltage mode until the charging current dropped to 50mA. Discharge tests were conducted in 1C constant current output mode until the voltage dropped to 2.7V.

[0091] The batteries were cycle-tested for charging and discharging. In dataset A, the failure threshold was considered to be 70% of the rated capacity (1.4 Ah). Since the B7 battery in dataset A did not reach end-of-life (EOL), the failure threshold was considered to be 72% of the rated capacity (1.44 Ah) in dataset A. In dataset B, the failure threshold was considered to be 80% of the rated capacity (0.88 Ah). The battery capacity degradation curves for the two datasets are shown below. Figure 3 As shown.

[0092] from Figure 3 As can be seen, the battery capacity degradation curve is a nonlinear curve, with its nonlinear components mainly consisting of capacity regeneration and noise components. Capacity regeneration is caused by the composition and characteristics of lithium batteries and must be considered in RUL prediction. However, the noise component is caused by equipment errors and external factors, and eliminating its impact on lithium battery RUL prediction is a key issue. Therefore, to eliminate the influence of noise, this invention uses the CEEMDAN algorithm to decompose the original data, calculates the correlation coefficient of each component based on the Pearson correlation coefficient, removes components with low correlation coefficients, and sums the remaining components to obtain the reconstructed data.

[0093] CEEMDAN is a signal decomposition and denoising method. Its main function is to decompose complex signals into a series of intrinsic mode functions (IMFs) with inherently different frequency and amplitude characteristics, thereby achieving the purpose of processing noise and interference in the signal. This invention uses CEEMDAN to decompose the capacity degradation data of all batteries. The decomposed signal of the B5 battery is as follows: Figure 4 As shown, the same applies to other batteries.

[0094] from Figure 4As can be seen, the residual is the main component of the original capacity, IMF4 to IMF6 are the low-frequency fluctuation components of the original capacity, and IMF1 to IMF3 are the high-frequency fluctuation components of the original capacity. Noise typically exists in the high-frequency components and requires further analysis using the Pearson correlation coefficient. The closer the correlation coefficient is to 1, the stronger the correlation between the component and the original signal; the closer the correlation coefficient is to 0, the weaker the correlation. It is worth noting that discarding too many components will filter out both capacity regeneration and noise components. Therefore, to preserve the capacity regeneration components while discarding the noise components, this invention, through repeated experimental simulations, determined the following criteria: In dataset A, IMF components with a correlation of less than 5% with the original capacity data are discarded, and the remaining IMF components and residuals are reconstructed; in dataset B, IMF components with a correlation of less than 10% with the original capacity data are discarded, and the remaining components are reconstructed. The correlation coefficients of datasets A and B are shown in Table 1.

[0095] As shown in Table 1, the correlation coefficient of the IMF3 component of battery B5 is less than 5%, therefore the IMF3 component is discarded, and the remaining IMF components and residual components are reconstructed. Similarly, the IMF3 and IMF6 components of batteries B6 and B7, and the IMF1 and IMF2 components of battery B18 are discarded. For dataset B, the IMF1 to IMF4 components of batteries CS35, CS37, and CS38, and the IMF1 to IMF5 components of battery CS36 are discarded. Subsequent experimental simulation data are all noise-reduced data.

[0096] Table 1. Pearson correlation coefficients for each battery

[0097]

[0098] (2) Evaluation criteria

[0099] 1) Mean Absolute Percentage Error (MAPE):

[0100]

[0101] 2) Root Mean Square Error (RMSE):

[0102]

[0103] 3) Coefficient of determination (R squared, R 2 ):

[0104]

[0105] In the formula, To predict SOH; y i This represents the actual SOH.

[0106] 4) RUL prediction error E r :

[0107] E r =|P RUL -R RUL | (7)

[0108] In the formula, P RUL The RUL prediction value is calculated as the difference between the predicted number of cycles to reach the lifetime threshold and the predicted starting point; R RUL The true RUL value is calculated as the difference between the number of cycles required to reach the lifetime threshold and the predicted starting point.

[0109] This invention uses the aforementioned metrics to evaluate the predictive performance of the proposed model. Among them, MAPE, RMSE, and E... r The smaller the value of R, the more accurate the prediction result. 2 The closer the value is to 1, the better the model's predictive performance.

[0110] 3. Experimental Results and Analysis

[0111] To verify the accuracy and stability of the proposed CNN-GRU-PF fusion model, this invention first divides the first 50% of the battery dataset into a training set and the last 50% into a test set. The proposed model is then compared with PF, GRU, and GRU-PF models through simulation. Secondly, the proposed CNN-GRU-PF fusion model is validated under different training sets, with simulation tests conducted when the training data constitutes 30%, 50%, and 60% of the total dataset. Finally, the running time and accuracy of each model are compared and analyzed.

[0112] (1) Parameter settings of the present invention

[0113] The proposed CNN-GRU-PF fusion model requires initial parameter settings for both the CNN-GRU data-driven model and the PF model. The CNN-GRU model parameters include the number of filters in the CNN convolutional layers, the kernel length, the activation function, the number of hidden neurons in the GRU, and the dropout layer coefficients. Parameter adjustments affect the model's learning efficiency and performance. After extensive experimental comparisons, the selected CNN-GRU model parameters are shown in Table 2.

[0114] Table 2 CNN-GRU Model Parameter Settings

[0115]

[0116]

[0117] This invention uses the Adaptive Moment Estimation (Adam) optimizer to optimize the hyperparameters of the CNN-GRU model, setting the initial learning rate to 0.005 and the learning rate reduction factor to 0.2.

[0118] The PF model particle size was set to 200. The capacity data of eight batteries in datasets A and B were fitted using the MATLAB curve fitting toolbox. The objective function was selected as a double exponential function to obtain the parameter values ​​of the double exponential model. These estimated values ​​were used as the initial values ​​for PF estimation. The obtained initial parameter fitting values ​​are shown in Table 3.

[0119] Table 3 Initial Parameters of PF

[0120]

[0121] (2) RUL prediction and comparative analysis based on CNN-GRU-PF

[0122] Based on the length of the existing data, the first 50% of the total data is divided into a training set and the last 50% into a test set. For batteries B5, B6, and B7 in dataset A, the prediction start point is set to 84 iterations; for battery B18, the prediction start point is set to 66 iterations. For batteries CS35 in dataset B, the prediction start point is set to 359 iterations; for CS36, it is set to 368 iterations; for CS37, it is set to 422 iterations; and for CS38, it is set to 379 iterations. The prediction results of the proposed model compared to PF, GRU, and GRU-PF models are as follows: Figure 5 As shown in Table 1. To ensure fairness, all models are initialized with the same parameters. For example, the number of hidden neurons in the GRU model is set as in Table 1, and a Dropout layer is also added.

[0123] from Figure 5It can be seen that the PF and GRU single models performed poorly. The GRU model exhibited divergence in the later stages of predicting the capacity of all eight batteries, resulting in poor prediction accuracy. This is because the GRU model suffers from accumulated error; without error compensation, the prediction becomes increasingly larger, eventually leading to divergence at the end of the prediction process. The PF model also performed poorly, with varying prediction results across different datasets; its prediction performance on dataset B was significantly better than that on dataset A. This demonstrates the poor applicability of the single PF model, making it difficult to guarantee its accuracy in practical applications. Compared to the two single models, GRU-PF showed better prediction performance. This is because the PF method corrects the accumulated error of the GRU model through its state prediction capabilities, thus significantly improving its prediction performance. However, due to the limited feature extraction capabilities of the GRU model, its prediction curve still shows considerable fluctuation, and its prediction performance remains insufficient. The CNN-GRU-PF model proposed in this invention demonstrates excellent prediction performance for the capacity prediction of all batteries in both datasets A and B, with the prediction curve showing a very high degree of fit to the true values. This demonstrates that the model method proposed in this invention has good accuracy and applicability, and has significant advantages in practical applications. To more intuitively illustrate the prediction accuracy and other information of each model, the MAPE, RMSE, and R-values ​​of the four models are compared. 2 and E r The errors are presented in a table, as shown in Table 4. Those that did not reach the failure threshold are indicated by “—”.

[0124] Table 4. Prediction errors of the four prediction models

[0125]

[0126]

[0127] As shown in Table 4, taking battery B5 as an example, the RMSE of the GRU, PF, and GRU-PF models are 0.0644, 0.0479, and 0.0184, respectively, while the RMSE of the CNN-GRU-PF model is 0.0082. Therefore, the proposed model method improves prediction accuracy by 87.27% compared to the GRU model, 82.88% compared to the PF model, and 55.43% compared to the GRU-PF model. For other batteries in both datasets, the prediction accuracy also shows a significant improvement. This fully demonstrates that the CNN-GRU-PF model has good prediction accuracy, prediction stability, and general applicability.

[0128] It is worth noting that the GRU model did not predict the RUL of battery B7, while the prediction error for the RUL of battery CS38 was E. rThe prediction stability is low, reaching up to 102 iterations. While the PF model improves RUL prediction, its stability remains low due to its reliance on data type. The RUL prediction error E of GRU-PF is... r Smaller, but its E for CS38 batteries r However, it reached 11 iterations, which is due to the poor feature extraction capability of GRU, resulting in poor prediction performance of the combined model GRU-PF. In contrast, the CNN-GRU-PF model achieved a lower RUL prediction error E for the four batteries in dataset A. r All loops are performed in order of 0, in dataset B, four batteries, E r The maximum value is only 4 cycles. It exhibits good stability and accuracy in predicting the RUL of lithium batteries.

[0129] (2) RUL prediction based on different prediction starting points of CNN-GRU-PF

[0130] Considering that the available lithium battery capacity data is limited in some real-world situations, this may lead to poor prediction performance of the prediction model. To verify that the model proposed in this invention still has good prediction performance even with insufficient training data, this invention selects training data accounting for 30%, 50%, and 60% of the total data, respectively, and uses the CNN-GRU-PF model to predict the RUL of lithium batteries. The prediction results are as follows: Figure 6 As shown in Table 5, the error results are shown in Table 5, where ST is the prediction starting point.

[0131] Table 5 Prediction errors at different prediction starting points

[0132]

[0133]

[0134] from Figure 6 As shown in Table 5, both 50% and 60% training sets demonstrate excellent prediction performance, while 30% training set still maintains good prediction accuracy, essentially predicting the trend of actual capacity degradation. Therefore, in practical applications, even with limited available capacity data, the CNN-GRU-PF fusion model proposed in this invention still achieves good prediction results. It is noteworthy that, except for batteries B5 and B6, the MAPE and RMSE of other batteries are quite close under both 50% and 60% training set conditions. This indicates that the fusion model can achieve very high prediction accuracy even with only 50% training set data.

[0135] (3) Comparison of running times of different model methods

[0136] This section uses a B5 battery as an example, with 50% of the data as the training set. Simulation experiments are conducted using four models: PF, GRU, GRU-PF, and CNN-GRU-PF. The model's runtime, RMSE, and RUL prediction error (E) are compared. r A comparative analysis was conducted. Both the GRU-PF and CNN-GRU-PF models employed a moving window iterative update method to improve the prediction accuracy in each iteration. To verify the advantages of the moving window dynamic update method, a comparative analysis was also performed on the GRU-PF and CNN-GRU-PF models without the moving window. The running time and prediction accuracy of the different models are shown in Table 6.

[0137] Table 6 Running time and prediction accuracy of different models

[0138]

[0139] As shown in Table 6, firstly, while the individual GRU and PF models have short running times, their prediction accuracy is very low, making them unsuitable for practical RUL prediction tasks. Secondly, both GRU-PF and CNN-GRU-PF, which use a moving window to iteratively update the training set, have longer running times than methods without a moving window, but their prediction accuracy is significantly improved. GRU-PF improves prediction speed by 46.2% compared to GRU-PF without a moving window; CNN-GRU-PF improves prediction accuracy by 69.06% compared to CNN-GRU-PF without a moving window. In practice, RUL predictions are typically performed 10 to 20 times, which far exceeds the running time of the CNN-GRU-PF fusion model proposed in this invention. Therefore, the method proposed in this invention is practically feasible.

[0140] In summary, based on the experimental simulations, the proposed fusion model demonstrates the following advantages: First, it achieves very high prediction accuracy. By combining the strengths of CNN, GRU, and PF, and employing a moving window to iteratively update the training set, it significantly improves prediction accuracy. Second, it can achieve long-term predictions and maintain accuracy even with limited data samples.

Claims

1. A lithium battery RUL prediction method based on a data-driven and particle filter fusion model, characterized in that, Specifically, the following steps are included: Step 1: Divide the Convolutional Neural Network (CNN) into convolutional layers and fully connected layers, and eliminate the use of pooling layers. The fully connected layers integrate and map the data output from the convolutional layers to obtain classification or regression results. By combining convolutional neural networks (CNN) and gated recurrent units (GRU) networks, a CNN-GRU network model with a cascaded structure is formed. Step 2: Establish a CNN-GRU-PF fusion prediction model to predict the RUL of lithium batteries; The PF filtering method is adopted, and the parameters of the double exponential empirical model are updated using the capacity prediction value of the CNN-GRU network. The updated parameters are used to calculate the new capacity prediction value. The new capacity prediction value is used to update the CNN-GRU training dataset, realizing the fusion of the CNN-GRU neural network prediction information and the double exponential empirical model prediction information. The RUL of lithium battery is predicted based on the CNN-GRU-PF fusion model. 2.1) A dual-exponential empirical model is used to establish the state equation. The capacity prediction value output by CNN-GRU is used as the observation value to establish the state space equation of the CNN-GRU-PF fusion prediction model: x k =[a k ;b k ;c k ;d k ] (1) Q k =a k * exp(b k *k)+c k * exp(d k *k)+γ k ,γ k ~N(0,σ Q ) (3) In the formula: a, b, c, and d are the internal parameters of the battery, which are also state variables; the battery capacity is the observed quantity; k is the charge-discharge cycle period of the battery; Q k ω is the capacity observation value for k cycles, provided by the capacity prediction value of the CNN-GRU model; ω and γ are the process noise and measurement noise, respectively, and both process noise and measurement noise are Gaussian noise; 2.2) The filtering method of PF is adopted, and the capacity prediction value of CNN-GRU network is used to update the parameters a, b, c, and d of the double exponential empirical model; 2.3) Calculate the new capacity prediction value using the updated parameters from step 2.2); 2.4) Update the CNN-GRU training dataset with the new capacity prediction values ​​from step 2.3) to achieve the fusion of CNN-GRU neural network prediction information and dual-exponential empirical model prediction information; predict the RUL of lithium batteries based on the CNN-GRU-PF fusion model; 2.4.1) Noise Reduction Preprocessing: Extract battery capacity data from the battery dataset and perform noise reduction preprocessing; 2.4.2) Initialize parameters: Set the prediction start point s, lifetime threshold, initial PF value, and CNN-GRU parameters; 2.4.3) Divide the training and test sets: Data_train(s) is the training data vector for the CNN-GRU model; Data_test is the test data vector for the CNN-GRU model; Data_train(s) consists of the volumes C1, C2, C3, ..., C of the first s epochs. s composition; 2.4.4) Train and predict the battery capacity at cycle s+1 using a CNN-GRU network model, named 2.4.5) PF-based state estimation: For the state-space model composed of equations (1), (2), and (3), the PF-based state estimation is performed as follows: As a measurement in PF, the model state is estimated using the PF algorithm [a] k b k c k ;d k This allows us to obtain the optimized capacity prediction value Q. s+1 ; 2.4.6) Constructing the mobile training dataset: Using the optimized capacity prediction value Q obtained in step 2.4.5) s+1 Update the training dataset Data_train(s) of the CNN-GRU model, and then retrain the CNN-GRU neural network using the updated training dataset. The specific method for updating the training dataset is as follows: [The text abruptly ends here, likely due to an incomplete translation or a formatting error.] s+1 Adding Data_train(s) to the CNN-GRU model generates the training dataset Data_train(s+1); the updated Data_train(s+1) consists of C2, C3, ..., C s Q s+1 composition; 2.4.7) Repeat steps 2.4.4) to 2.4.6) until the prediction step size reaches the total length of the dataset, then the loop ends; 2.4.8) Evaluate the prediction results based on the test set data and calculate the RUL.

2. The lithium battery RUL prediction method based on a data-driven and particle filter fusion model according to claim 1, characterized in that, The CNN-GRU network model structure in step 1 is as follows: 1) Input layer, used to receive lithium battery capacity degradation data and output it to the CNN layer; 2) The CNN layer includes a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function. The convolutional layer performs convolution operations on the lithium battery capacity data from the input layer and extracts local features. The batch normalization layer standardizes the input data in the network, reducing variations in data distribution and accelerating network convergence. The ReLU activation function increases the network's non-linearity and overcomes the gradient vanishing problem. Then, a flattening layer maps the data to the output, which serves as the input data for the GRU module. 3) GRU layer: The GRU layer receives the output data of the convolutional layer, trains the data, predicts the capacity of the lithium battery based on the training results, and outputs the predicted data to the Dropout layer. 4) Dropout layer: The Dropout layer is used to prevent overfitting in the GRU neural network, reduce the interdependence between neurons, and enhance the network's generalization ability and robustness. A fully connected layer is added after the Dropout layer to map the output of the Dropout layer to the desired output dimension. 5) Output layer: The output data of the Dropout layer is the lithium battery capacity at the next moment.

3. The lithium battery RUL prediction method based on a data-driven and particle filter fusion model according to claim 2, characterized in that, The GRU layer uses the tanh function to scale and translate the training data, reducing the data distribution range of GRU training and improving the efficiency of GRU training.

4. A prediction system for the lithium battery RUL prediction method based on a data-driven and particle filter fusion model according to any one of claims 1 to 3, characterized in that, include: CNN-GRU Adjustment Module: Used in step 1 to adjust the relationship between the layers of CNN and GRU, select a one-dimensional CNN and a one-layer GRU to perform lithium battery capacity prediction and provide capacity observations for PF; CNN-GRU-PF module: used in step 2 to combine CNN-GRU and PF to obtain optimized lithium battery capacity prediction values ​​and calculate lithium battery RUL.

5. The prediction system for the lithium battery RUL prediction method based on the data-driven and particle filter fusion model according to claim 4, characterized in that, The CNN-GRU-PF module includes a data preprocessing module, a CNN-GRU prediction module, a PF state estimation module, a moving training dataset module, and an error analysis module. The preprocessing module comprises three parts: noise reduction, initialization parameter setting, and dataset partitioning. Noise reduction utilizes the Fully Adaptive Empirical Mode Decomposition (CEEMDAN) algorithm to decompose lithium battery capacity data and reconstruct it using Pearson correlation coefficients to reduce noise in the battery capacity data. Initialization parameter setting is used to configure the parameters of the RUL prediction system for system startup. Data partitioning is used to divide the dataset into training and testing datasets for the RUL prediction system's prediction work and comparison of prediction results. The CNN-GRU prediction module is trained based on the pre-divided training dataset to predict the capacity of lithium batteries and provide capacity observations for power PF. Among them, the state estimation module of PF is used to correct the capacity prediction error of CNN-GRU and obtain the optimized capacity prediction value. Among them, the moving training dataset module is used to add the capacity-optimized value of PF to the training set of the CNN-GRU module to improve the long-term prediction ability of CNN-GRU. The error analysis module is used to analyze the accuracy of the prediction results and to calculate the RUL of the lithium battery.

6. A prediction device for a lithium battery RUL prediction method based on a data-driven and particle filter fusion model, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the lithium battery RUL prediction method based on a data-driven and particle filter fusion model as described in any one of claims 1 to 3 when executing the computer program.

7. A computer-readable storage medium for storing a computer program, characterized in that, When the computer program is executed by the processor, it is capable of performing a lithium battery RUL prediction based on a data-driven and particle filter fusion model according to any one of claims 1 to 3.

Images