Heterovariance migration network and particle filter-based battery life prediction method

Abstract

The invention provides a hetero-variance migration network and particle filter-based battery life prediction method, and relates to the technical field of battery health management, and the method comprises the steps: collecting a part of charging data fragments from a target battery under an actual operation condition, and extracting an incremental capacity feature; processing the incremental capacity characteristics by using a heterovariance migration network trained on a tagged source domain and an untagged target domain, outputting a mean value and a variance of a battery health state, and describing an evolution process of the battery health state by using a parameterized double-index degradation function; generating an initial parameter particle by taking a degradation function parameter of source domain fitting as a mean value, and updating the parameter particle by taking a state prediction value output by the heterovariance neural network as an observation value and a corresponding uncertainty variance as observation noise; and finally, predicting the remaining life of the battery based on the updated parameter result. According to the method, cross-working-condition and cross-battery-type residual life prediction is realized, and the prediction precision and robustness are improved.

Description

Technical Field

[0001] This invention relates to the field of battery health management technology, and specifically discloses a battery life prediction method based on heteroscedastic migration network and particle filtering. Background Technology

[0002] With the rapid development of new energy technologies, lithium-ion batteries are widely used as key energy storage devices. Their performance degradation and lifespan prediction directly affect the safety, reliability, and economy of the system. Accurate prediction of the battery's rated lifespan (RUL) is crucial for ensuring safe and efficient battery operation.

[0003] Due to its strong nonlinear fitting capability and the fact that it does not require consideration of complex electrochemical reaction mechanisms, the data-driven method (Q Si, S Matsuda, Y Yamaji, et al. Data-Driven Cycle Life Prediction of Lithium Metal-Based Rechargeable Battery Based on Discharge / Charge Capacity and Relaxation Features) is a powerful approach. Adv. Sci. (2024, 11, 2402608) has become the mainstream method for battery life prediction. This method usually requires complete charge and discharge data, and the model is trained using early capacity data. However, in real-world usage scenarios, batteries rarely achieve a complete charge and discharge process. In addition, different batteries have significantly different chemical systems and operating conditions, and their degradation trends show significant differences. This leads to a significant decrease in the prediction accuracy of purely data-driven methods when applied across batteries and operating conditions. At the same time, data-driven methods may produce prediction results that violate physical common sense, facing the dilemma of insufficient interpretability and unreliable results. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, such as strong dependence on complete charge and discharge data, insufficient generalization ability across operating conditions, and low reliability of results, the present invention aims to provide a battery life prediction method based on heteroscedasticity transfer network and particle filtering. This method can utilize cross-domain knowledge for transfer learning when only partial charge and discharge segments are available, and combine it with a physical degradation model to achieve constraints. It also has the ability to predict uncertainties in the output, thereby improving the reliability and practicality in engineering applications.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A battery lifetime prediction method based on heteroscedasticity transfer network and particle filtering is proposed. First, partial charging data fragments of the target battery under actual operating conditions are collected as source domain data. Incremental capacity (IC) features are extracted and correlation verification and normalization are performed. The processed labeled source domain and unlabeled target domain data are input into a CNN-LSTM feature extraction module, which combines a convolutional neural network (CNN) and a long short-term memory network (LSTM). Cross-domain knowledge transfer is achieved by incorporating maximum mean discrepancy (MMD) loss. A total loss function is constructed by introducing Gaussian negative log-likelihood loss and mean squared error loss to simultaneously learn the predicted mean and variance of the health status. In the lifetime inference stage, an initial particle set is generated using the parameters of the double exponential degradation function fitted from the source domain data. The health status predicted by the heteroscedasticity transfer network is used as the observation value and input into the particle filtering model to complete particle weight updates and resampling. When no new observation values ​​are available, the battery degradation trajectory is calculated based on the updated parameters. Finally, a battery lifetime prediction method with low data dependence, high cross-condition generalization ability, and high reliability is achieved.

[0006] A battery lifetime prediction method based on heteroscedasticity transfer networks and particle filtering includes the following steps: 1) From the charging data stream collected during actual operation, a segment containing degradation information is extracted as the source domain data. The collected data is as follows: in , , They represent the first Voltage, current, and time series of each sample. , , These represent the sample number. Voltage, current, and time data at each sampling point , Indicates the total number of samples; 2) Extract IC features from the collected data segments. The calculation formula is as follows: in Represents a capacity difference sequence. Represents a voltage difference sequence. Represents the IC characteristic sequence, Indicates the first The first sample Individual IC characteristic values; The formula for calculating IC characteristics under constant current charging conditions can be written in the following form: The MA method is selected to smooth the IC characteristic curve, and its calculation formula is as follows: in It is the window length of the moving average; Spearman correlation analysis was performed on the source domain data IC characteristics and capacity data, and the calculation formula is as follows: in, Indicates the first The rank correlation coefficient between each feature and capacity It is the first In the nth sample The difference between the rank of an IC feature and the rank of its capacity. It is the total number of samples; The IC characteristics and capacity are normalized separately, and the calculation formulas are as follows: in, and These are the maximum and minimum values ​​of the original IC characteristics, respectively. It is the normalized value. This is the raw capacity data. This is the initial capacity data. It is normalized data; 3) Perform the same feature extraction on both labeled source domain data and unlabeled target domain data, and input them into the feature extraction modules of Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM) to extract deep features. CNN is used to extract local spatial features, and LSTM is used to extract temporal features. The deep features are then input into a fully connected layer, which outputs the predicted mean and variance. Given the feature sequence of the input samples... The calculation process of the convolutional layer is as follows: in, and These represent the weights and biases of the convolution kernel, respectively. This represents a one-dimensional convolution operation. This indicates batch normalization, used to accelerate model convergence and improve stability. Pool represents the ReLU activation function. This indicates max pooling. It is the local feature sequence obtained after the convolution module operation, which is used as the input of the next stage LSTM network; The LSTM network receives at each time step t eigenvectors in And update its internal state, the calculation equation is as follows: in , , These represent the activation vectors of the forget gate, input gate, and output gate, respectively. and These represent the cell state and hidden state at the current time step, respectively. and This refers to the state at the previous time step; Candidate cell state; , , , and , , , These are the learnable weight matrix and bias vector of the network; It is the Sigmoid activation function; It is the hyperbolic tangent activation function; This represents vector concatenation. This represents element-wise multiplication; the hidden state of the last layer of the LSTM is taken as the final depth feature vector. ; The extracted depth features A shared fully connected layer is input to perform a non-linear transformation on the features to generate features. The calculation formula is as follows: feature The data is then fed into two separate output layers, which are used to predict the mean of the target state. and variance The calculation formula is as follows: in, and These are the shared layer weight matrix and bias vector, respectively. For shared layer output features, , and , These are the weights and biases of the mean output layer and the variance output layer, respectively. ; To transfer degenerate knowledge learned from source domain data to the target domain, a domain adaptation mechanism is introduced, which utilizes deep features. The MMD loss is calculated to impose constraints on the feature extraction module, guiding it to learn domain-independent shared feature representations; let the set of deep features extracted from the source domain be... The set of deep features extracted from the target domain The MMD calculation formula is: in, and Let represent the total number of samples in the source domain and the target domain, respectively. and Representing the source domain's first... The nth sample and the target domain One sample, It is a feature mapping. It is the RKHS space; The MMD formula, expanded using the kernel trick, is as follows: in Gaussian kernels are commonly used. , It is the bandwidth hyperparameter of the Gaussian kernel, which controls the smoothness of the kernel function; The training uses the Gaussian negative log-likelihood loss function, the formula of which is: in, It is the overall negative log-likelihood loss. It is the first The true label value of each sample and These are the predictions of the model for the first time. The mean and variance of each sample; By introducing the MSE term as a regularization term into the loss function, the formula for the total loss function is: in, The coefficient of the MSE loss term. It is the coefficient of the MMD loss term; During the model application phase, the IC features obtained from actual working conditions will be input into the trained model, and the predicted mean and variance of the health status will be output: in, For IC characteristic sequences, and The predicted number of The health status and variance of each cycle; 4) A double-exponential degradation function is used. The first exponential term describes the rapid, nonlinear degradation process, and the second exponential term describes the slow degradation process. The combination of the two is used to accurately describe the dynamic evolution of the battery health state, as shown in the following formula: in, Index for time steps (loops), For particle serial numbers, , , and For the first Particles in each cycle The sequence of model parameters, Indicates the first A cycle is formed by particles The state values ​​calculated from the parameter sequence; 5) Fit the degradation function using source domain data, and generate initial particles using the parameter fitting results as the mean: in, , , and For particles The initial model parameter sequence; Substitute each set of parameter particle values ​​into the double exponential degradation function to calculate the corresponding health status value; 6) The predicted health status value is used as the observed value, and the predicted variance is used as the observation noise. The particle weights are updated using the observation function, which is: in, To predict the mean, This represents zero-mean Gaussian noise. To predict variance; The likelihood value is calculated using the observation function, and the particle weights at the current time are updated. The weight calculation formula is as follows: in, Represents particles In the The parameter weights of each loop; The weights of all particles are normalized, and it is determined whether the number of effective particles has reached a threshold. If it is lower than the threshold, resampling is performed. The formula for calculating the number of effective particles is: in, Represents the total number of particles. Represents particles In the Normalized weights for each loop; 7) Once there are no more observations, the particles will no longer be updated. Based on the updated parameters, the RUL of the battery and its distribution will be predicted. Assume the updated particle set is: in, Represents particles Parameter sequence, , , and These represent the final updated particles. The sequence of model parameters; Predict the battery's subsequent degradation trajectory based on each parameter particle. Calculate the corresponding failure time: Where FPT is the initial prediction point cycle, The failure threshold is set to 80% of the initial capacity. Represents particles Extrapolated predicted lifetime end cycle number; The remaining lifetime of the FPT cycle is expressed as: Based on the particle weighting results, the predicted posterior distribution of RUL is obtained as follows: in, Represents particles The final updated normalized weights, Indicates an indicator function, ; RUL expected values ​​are as follows: Step 1) The principle for extracting a segment of charging data containing degradation information is that it should cover the voltage range where the battery degradation characteristics are obvious and which is easy to obtain in the actual incomplete charging and discharging process.

[0007] Compared with the prior art, the present invention has the following beneficial effects: This invention effectively addresses the problems of strong dependence on complete charge-discharge data, insufficient generalization ability across operating conditions, and low reliability of prediction results. This method utilizes only charging data fragments with sufficient degradation characteristics for modeling and prediction, significantly reducing the difficulty of data acquisition and better reflecting the availability of battery data under actual operating conditions. Simultaneously, the combination of heteroscedasticity transfer networks (HIFL) and particle filtering achieves an organic fusion of data-driven and physically constrained approaches. During cross-domain training, the HIFL can simultaneously output health state estimates and their uncertainties, used to construct an observation noise model, thereby improving the reliability and stability of the prediction results. The bi-exponential degradation function can uniformly describe the degradation trend of the battery at different stages, making the prediction results consistent with physical mechanisms. The dynamic update mechanism of particle filtering further ensures iterative correction of model parameters as available observation data arrives, enhancing the robustness and online adaptability of the method. In summary, the method of this invention can achieve migration prediction of remaining life under different battery types and operating conditions, while simultaneously outputting confidence intervals, balancing prediction accuracy and risk assessment, and has good engineering application prospects. Attached Figure Description

[0008] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.

[0009] Figure 2 This is a schematic diagram of a heteroscedastic migration network according to an embodiment of the present invention.

[0010] Figure 3 The graph shows the battery life prediction results under different temperature conditions according to the embodiments of the present invention.

[0011] Figure 4 The graph shows the battery life prediction results under different temperature conditions and active materials in the embodiments of the present invention. Detailed Implementation

[0012] The present invention will now be described in detail with reference to the embodiments and accompanying drawings.

[0013] Reference Figure 1 A battery lifetime prediction method based on heteroscedasticity transfer networks and particle filtering includes the following steps: 1) From the charging data stream collected during actual operation, a segment containing degradation information is extracted as the source domain data. The collected data is as follows: in , , They represent the first Voltage, current, and time series of each sample. , , These represent the sample number. Voltage, current, and time data at each sampling point , Indicates the total number of samples; 2) Extract IC features from the collected data segments. The calculation formula is as follows: in Represents a capacity difference sequence. Represents a voltage difference sequence. Represents the IC characteristic sequence, Indicates the first The first sample Individual IC characteristic values; The formula for calculating IC characteristics under constant current charging conditions can be written in the following form: The MA method is selected to smooth the IC characteristic curve, and its calculation formula is as follows: in It is the window length of the moving average; Spearman correlation analysis was performed on the source domain data IC characteristics and capacity data, and the calculation formula is as follows: in, Indicates the first The rank correlation coefficient between each feature and capacity It is the first In the nth sample The difference between the rank of an IC feature and the rank of its capacity. It is the total number of samples; The IC characteristics and capacity are normalized separately, and the calculation formulas are as follows: in, and These are the maximum and minimum values ​​of the original IC characteristics, respectively. It is the normalized value. This is the raw capacity data. This is the initial capacity data. It is normalized data; 3) Perform the same feature extraction on both labeled source domain data and unlabeled target domain data, and input them into the feature extraction modules of Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM) to extract deep features. CNN is used to extract local spatial features, and LSTM is used to extract temporal features. The deep features are then input into a fully connected layer, which outputs the predicted mean and variance. Given the feature sequence of the input samples... The calculation process of the convolutional layer is as follows: in, and These represent the weights and biases of the convolution kernel, respectively. This represents a one-dimensional convolution operation. This indicates batch normalization, used to accelerate model convergence and improve stability. Pool represents the ReLU activation function. This indicates max pooling. It is the local feature sequence obtained after the convolution module operation, which is used as the input of the next stage LSTM network; The LSTM network receives at each time step t eigenvectors in And update its internal state, the calculation equation is as follows: in , , These represent the activation vectors of the forget gate, input gate, and output gate, respectively. and These represent the cell state and hidden state at the current time step, respectively. and This refers to the state at the previous time step; Candidate cell state; , , , and , , , These are the learnable weight matrix and bias vector of the network; It is the Sigmoid activation function; It is the hyperbolic tangent activation function; This represents vector concatenation. This represents element-wise multiplication; the hidden state of the last layer of the LSTM is taken as the final depth feature vector. ; The extracted depth features A shared fully connected layer is input to perform a non-linear transformation on the features to generate features. The calculation formula is as follows: feature The data is then fed into two separate output layers, which are used to predict the mean of the target state. and variance The calculation formula is as follows: in, and These are the shared layer weight matrix and bias vector, respectively. For shared layer output features, , and , These are the weights and biases of the mean output layer and the variance output layer, respectively. ; To transfer degenerate knowledge learned from source domain data to the target domain, a domain adaptation mechanism is introduced, which utilizes deep features. The MMD loss is calculated to impose constraints on the feature extraction module, guiding it to learn domain-independent shared feature representations; let the set of deep features extracted from the source domain be... The set of deep features extracted from the target domain The MMD calculation formula is: in, and Let represent the total number of samples in the source domain and the target domain, respectively. and Representing the source domain's first... The nth sample and the target domain One sample, It is a feature mapping. It is the RKHS space; The MMD formula, expanded using the kernel trick, is as follows: in Gaussian kernels are commonly used. , It is the bandwidth hyperparameter of the Gaussian kernel, which controls the smoothness of the kernel function; The training uses the Gaussian negative log-likelihood loss function, the formula of which is: in, It is the overall negative log-likelihood loss. It is the first The true label value of each sample and These are the predictions of the model for the first time. The mean and variance of each sample; By introducing the MSE term as a regularization term into the loss function, the formula for the total loss function is: in, The coefficient of the MSE loss term. It is the coefficient of the MMD loss term; During the model application phase, the IC features obtained from actual working conditions will be input into the trained model, and the predicted mean and variance of the health status will be output: in, For IC characteristic sequences, and The predicted number of The health status and variance of each cycle; 4) A double-exponential degradation function is used. The first exponential term describes the rapid, nonlinear degradation process, and the second exponential term describes the slow degradation process. The combination of the two is used to accurately describe the dynamic evolution of the battery health state, as shown in the following formula: in, Index for time steps (loops), For particle serial numbers, , , and For the first Particles in each cycle The sequence of model parameters, Indicates the first A cycle is formed by particles The state values ​​calculated from the parameter sequence; 5) Fit the degradation function using source domain data, and generate initial particles using the parameter fitting results as the mean: in, , , and For particles The initial model parameter sequence; Substitute each set of parameter particle values ​​into the double exponential degradation function to calculate the corresponding health status value; 6) The predicted health status value is used as the observed value, and the predicted variance is used as the observation noise. The particle weights are updated using the observation function, which is: in, To predict the mean, This represents zero-mean Gaussian noise. To predict variance; The likelihood value is calculated using the observation function, and the particle weights at the current time are updated. The weight calculation formula is as follows: in, Represents particles In the The parameter weights of each loop; The weights of all particles are normalized, and it is determined whether the number of effective particles has reached a threshold. If it is lower than the threshold, resampling is performed. The formula for calculating the number of effective particles is: in, Represents the total number of particles. Represents particles In the Normalized weights for each loop; 7) Once there are no more observations, the particles will no longer be updated. Based on the updated parameters, the RUL of the battery and its distribution will be predicted. Assume the updated particle set is: in, Represents particles Parameter sequence, , , and These represent the final updated particles. The sequence of model parameters; Predict the battery's subsequent degradation trajectory based on each parameter particle. Calculate the corresponding failure time: Where FPT is the initial prediction point cycle, The failure threshold is set to 80% of the initial capacity. Represents particles Extrapolated predicted lifetime end cycle number; The remaining lifetime of the FPT cycle is expressed as: Based on the particle weighting results, the predicted posterior distribution of RUL is obtained as follows: in, Represents particles The final updated normalized weights, Indicates an indicator function, ; RUL expected values ​​are as follows: This embodiment uses the NCA and NCM battery datasets from Tongji University to verify the effectiveness of the method of the present invention.

[0014] Table 1. Details of the dataset used This dataset contains test data of batteries with different chemically active materials under different ambient temperature conditions. Detailed information about the batteries used is shown in Table 1.

[0015] To evaluate the migration prediction performance of this invention across different chemical materials and temperature conditions under incomplete charge-discharge conditions, two migration experiments were designed. First, the data was fragmented to simulate actual operating conditions, and data from CY35-05_1-#1, #2, #3, and #4 batteries with NCM cathode material at 35°C were selected as tagged source domains. Experiment 1 selected CY45-05_1-#21, #22, and #25 batteries with the same material but different operating conditions as untagged target domains; Experiment 2 selected CY45-05_1-#1, #2, and #12 batteries with NCA cathode material, which have different materials and operating conditions, as target domains to systematically evaluate the migration prediction performance of this invention across different chemical materials and temperature conditions under incomplete charge-discharge conditions.

[0016] In the data preprocessing stage, IC features were extracted from both the source and target domain data, MA smoothing was used, and Spearman correlation analysis was used to verify its correlation with capacity degradation. Subsequently, normalization was performed to ensure data scale uniformity.

[0017] The processed features are input into a heteroscedasticity transfer network for training. During training, source domain data undergoes supervised learning using capacity labels, while target domain data is aligned using MMD loss. This network integrates CNNs to extract local spatial features and utilizes LSTMs to capture temporal dependencies. Finally, it outputs the predicted mean and variance of the health status simultaneously through fully connected layers, such as... Figure 2 As shown.

[0018] In the particle filtering stage, the parameters of the double exponential degradation function obtained by fitting the source domain data are used as priors to generate 300 initial parameter particles to balance accuracy and efficiency. The mean and variance of the health state output by the neural network are used as the observed value and observation noise of PF, respectively, to achieve dynamic updating of the parameter particles.

[0019] To verify the prediction performance under different amounts of observation data, the experiment used incomplete observation data from the first 300 and first 400 cycles for parameter updates and lifetime prediction, respectively.

[0020] The experimental results of the model are shown in Table 2. Figure 3 and Figure 4 As shown. In Experiment 1, the proposed method can predict the actual degradation trajectory well and provide a narrow lifetime distribution range, such as... Figure 3 As shown, when the number of observation data increases from 300 cycles to 400 cycles, both the predicted MAE and RMSE decrease, the RULE (Remaining Useful Life Error) is smaller, and the RUL prediction is more accurate. This is due to the degradation information contained in more observation data, which makes the distribution of parametric particles more convergent.

[0021] Figure 4 The evaluation results for Experiment 2 show that the predicted degradation trajectory is also consistent with reality, and the accuracy increases significantly with the addition of observational data. In comparison, the prediction error of Experiment 2 is higher than that of Experiment 1. This is because the source and target domains in Experiment 2 differ in temperature and active materials, leading to fundamental differences in electrochemical properties and degradation mechanisms, and a greater variation in data distribution, increasing the difficulty of knowledge transfer. Nevertheless, this method can still provide effective predictions in more complex scenarios, demonstrating its robustness.

[0022] Table 2 Life Prediction Assessment Results

Claims

1. A battery lifetime prediction method based on heteroscedasticity transfer networks and particle filtering, characterized in that: First, partial charging data fragments of the target battery under actual operating conditions are collected as source domain data. Incremental Capacity (IC) features are extracted and correlation verification and normalization are performed. The processed labeled source domain and unlabeled target domain data are input into a CNN-LSTM feature extraction module composed of a convolutional neural network (CNN) and a long short-term memory network (LSTM). Cross-domain knowledge transfer is achieved by combining the maximum mean discrepancy (MMD) loss. The total loss function is constructed by introducing Gaussian negative log-likelihood loss and mean squared error loss to simultaneously learn the predicted mean and variance of the health status. In the lifetime inference stage, the parameters of the double exponential degradation function fitted by the source domain data are used to generate an initial particle set. The health status predicted by the heteroscedastic migration network is used as the observation value input to the particle filter model to complete the particle weight update and resampling. When no new observation values ​​are available, the battery degradation trajectory is calculated based on the updated parameters, and finally, the battery RUL prediction with low data dependence, high cross-condition generalization ability and high reliability is achieved.

2. The battery lifetime prediction method based on heteroscedasticity transfer network and particle filtering according to claim 1, characterized in that, Includes the following steps: 1) From the charging data stream collected during actual operation, a segment containing degradation information is extracted as the source domain data. The collected data is as follows: in , , They represent the first Voltage, current, and time series of each sample. , , These represent the sample number. Voltage, current, and time data at each sampling point , Indicates the total number of samples; 2) Extract IC features from the collected data segments. The calculation formula is as follows: in Represents a capacity difference sequence. Represents a voltage difference sequence. Represents the IC characteristic sequence, Indicates the first The first sample Individual IC characteristic values; The formula for calculating IC characteristics under constant current charging conditions can be written in the following form: The MA method is selected to smooth the IC characteristic curve, and its calculation formula is as follows: in It is the window length of the moving average; Spearman correlation analysis was performed on the source domain data IC characteristics and capacity data, and the calculation formula is as follows: in, Indicates the first The rank correlation coefficient between each feature and capacity It is the first In the nth sample The difference between the rank of an IC feature and the rank of its capacity. It is the total number of samples; The IC characteristics and capacity are normalized separately, and the calculation formulas are as follows: in, and These are the maximum and minimum values ​​of the original IC characteristics, respectively. It is the normalized value. This is the raw capacity data. This is the initial capacity data. It is normalized data; 3) Perform the same feature extraction on both labeled source domain data and unlabeled target domain data, and input them into the feature extraction modules of Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM) to extract deep features. CNN is used to extract local spatial features, and LSTM is used to extract temporal features. The deep features are then input into a fully connected layer, which outputs the predicted mean and variance. Given the feature sequence of the input samples... The calculation process of the convolutional layer is as follows: in, and These represent the weights and biases of the convolution kernel, respectively. This represents a one-dimensional convolution operation. This indicates batch normalization, used to accelerate model convergence and improve stability. Pool represents the ReLU activation function. This indicates max pooling. It is the local feature sequence obtained after the convolution module operation, which is used as the input of the next stage LSTM network; The LSTM network receives at each time step t eigenvectors in And update its internal state, the calculation equation is as follows: in , , These represent the activation vectors of the forget gate, input gate, and output gate, respectively. and These represent the cell state and hidden state at the current time step, respectively. and This refers to the state at the previous time step; Candidate cell state; , , , and , , , These are the learnable weight matrix and bias vector of the network; It is the Sigmoid activation function; It is the hyperbolic tangent activation function; This represents vector concatenation. This represents element-wise multiplication; the hidden state of the last layer of the LSTM is taken as the final depth feature vector. ; The extracted depth features A shared fully connected layer is input to perform a non-linear transformation on the features to generate features. The calculation formula is as follows: feature The data is then fed into two separate output layers, which are used to predict the mean of the target state. and variance The calculation formula is as follows: in, and These are the shared layer weight matrix and bias vector, respectively. For shared layer output features, , and , These are the weights and biases of the mean output layer and the variance output layer, respectively. ; To transfer degenerate knowledge learned from source domain data to the target domain, a domain adaptation mechanism is introduced, which utilizes deep features. The MMD loss is calculated to impose constraints on the feature extraction module, guiding it to learn domain-independent shared feature representations; let the set of deep features extracted from the source domain be... The set of deep features extracted from the target domain The MMD calculation formula is: in, and Let represent the total number of samples in the source domain and the target domain, respectively. and Representing the source domain's first... The nth sample and the target domain One sample, It is a feature mapping. It is the RKHS space; The MMD formula, expanded using the kernel trick, is as follows: in Gaussian kernels are commonly used. , It is the bandwidth hyperparameter of the Gaussian kernel, which controls the smoothness of the kernel function; The training uses the Gaussian negative log-likelihood loss function, the formula of which is: in, It is the overall negative log-likelihood loss. It is the first The true label value of each sample and These are the predictions of the model for the first time. The mean and variance of each sample; By introducing the MSE term as a regularization term into the loss function, the formula for the total loss function is: in, The coefficient of the MSE loss term. It is the coefficient of the MMD loss term; During the model application phase, the IC features obtained from actual working conditions will be input into the trained model, and the predicted mean and variance of the health status will be output: in, For IC characteristic sequences, and The predicted number of The health status and variance of each cycle; 4) A double-exponential degradation function is used. The first exponential term describes the rapid, nonlinear degradation process, and the second exponential term describes the slow degradation process. The combination of the two is used to accurately describe the dynamic evolution of the battery health state, as shown in the following formula: in, Index for time steps (loops), For particle serial numbers, , , and For the first Particles in each cycle The sequence of model parameters, Indicates the first A cycle is formed by particles The state values ​​calculated from the parameter sequence; 5) Fit the degradation function using source domain data, and generate initial particles using the parameter fitting results as the mean: in, , , and For particles The initial model parameter sequence; Substitute each set of parameter particle values ​​into the double exponential degradation function to calculate the corresponding health status value; 6) The predicted health status value is used as the observed value, and the predicted variance is used as the observation noise. The particle weights are updated using the observation function, which is: in, To predict the mean, This represents zero-mean Gaussian noise. To predict variance; The likelihood value is calculated using the observation function, and the particle weights at the current time are updated. The weight calculation formula is as follows: in, Represents particles In the The parameter weights of each loop; The weights of all particles are normalized, and it is determined whether the number of effective particles has reached a threshold. If it is lower than the threshold, resampling is performed. The formula for calculating the number of effective particles is: in, Represents the total number of particles. Represents particles In the Normalized weights for each loop; 7) Once there are no more observations, the particles will no longer be updated. Based on the updated parameters, the RUL of the battery and its distribution will be predicted. Assume the updated particle set is: in, Represents particles Parameter sequence, , , and These represent the final updated particles. The sequence of model parameters; Predict the battery's subsequent degradation trajectory based on each parameter particle. Calculate the corresponding failure time: Where FPT is the initial prediction point cycle, The failure threshold is set to 80% of the initial capacity. Represents particles Extrapolated predicted lifetime end cycle number; The remaining lifetime of the FPT cycle is expressed as: Based on the particle weighting results, the predicted posterior distribution of RUL is obtained as follows: in, Represents particles The final updated normalized weights, Indicates an indicator function, ; RUL expected values ​​are as follows: 。 3. The battery lifetime prediction method based on heteroscedasticity transfer network and particle filtering according to claim 2, characterized in that: Step 1) The principle for extracting a segment of charging data containing degradation information is that it should cover the voltage range where the battery degradation characteristics are obvious and which is easy to obtain in the actual incomplete charge and discharge process.

Images