Lithium battery remaining life prediction method fusing projection operator and conditional variational autoencoder

Abstract

The application discloses a lithium battery residual life prediction method fusing a projection operator and a conditional variational autoencoder, and comprises the following steps: acquiring lithium battery data to be measured; wherein the lithium battery data to be measured comprises the terminal voltage, the charging and discharging current and the battery shell temperature of the lithium battery to be measured at each moment within a sampling duration; inputting the lithium battery data to be measured into m base learner models which are pre-trained to obtain first prediction results for the residual life of the lithium battery to be measured; obtaining m second prediction results for the residual life of the lithium battery to be measured based on the m first prediction results and a multi-layer conditional variational autoencoder model which is pre-trained; and performing weighted summation on all the first prediction results and the second prediction results by using 2m projection operator coefficients which are pre-trained to obtain the final prediction result of the residual life of the lithium battery to be measured. The application can obtain stable and optimal integrated prediction results for the residual life of the lithium battery.

Description

Technical Field

[0001] This invention belongs to the field of lithium battery RUL (Remaining useful life) prediction, specifically involving a lithium battery remaining useful life prediction method that integrates projection operators and conditional variational autoencoders. Background Technology

[0002] Lithium batteries, with their high energy density, relatively stable electrochemical characteristics, low pollution, and long cycle life, have become the core energy supply component in various systems and devices. They are widely used in various fields such as transportation, mobile communication, aerospace and military, and consumer electronics, and have shown outstanding advantages. For example, lithium batteries are currently commonly used in electric vehicles.

[0003] Because the performance of lithium batteries degrades to some extent during the charging and discharging process, and this degradation can lead to system malfunctions or even catastrophic consequences, predicting the health status of lithium batteries is essential.

[0004] Lithium-ion battery health status prediction estimates the battery's health status by monitoring changes in performance indicators such as voltage, current, and temperature during battery cycling, thereby reducing system risks. As the number of battery cycles increases, various performance indicators of the lithium-ion battery also change dynamically. The remaining lifespan of a lithium-ion battery is the number of cycles required for its capacity to decline to the failure threshold under charge-discharge conditions. Predicting the remaining lifespan of a lithium-ion battery is an important aspect of battery health status management. It can, to a certain extent, make the dynamic changes in lithium-ion battery performance controllable, thereby improving the safety and availability of the lithium-ion battery system and providing highly valuable reference suggestions for lithium-ion battery system managers and users.

[0005] In recent years, deep learning-based lithium battery health status management has developed rapidly. Among them, Convolutional Neural Networks (CNN) models, trained using parameters from a partial complete charge-discharge cycle of the lithium battery, achieve end-to-end remaining life prediction and perform robustly on large-scale lithium battery datasets. However, this model does not consider the impact of battery operating temperature on its degradation rate. Recurrent Neural Networks (RNNs) based on Long Short-Term Memory (LSTM) estimate the health status of electric vehicles under driving conditions using only voltage and current curves. LSTM-based RNNs alleviate gradient vanishing problems and can retain information for longer periods, thus capturing long-term data correlations. Compared to LSTM, Gated Recurrent Unit (GRU) neural networks have slightly higher prediction errors but require fewer parameters to learn. Deep neural network technology avoids complex mathematical modeling processes, overcomes the limitations of manual feature extraction in data-driven approaches, reduces the influence of human factors on prediction results, and achieves end-to-end remaining life prediction, possessing significant technological advantages. Based on this, some scholars have used multi-neural network ensemble learning methods to train models for predicting the remaining lifespan of lithium batteries. However, these methods suffer from drawbacks such as high training difficulty and low accuracy of ensemble prediction. Currently, this problem has become a critical issue that urgently needs to be addressed in this field. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this invention provides a method for predicting the remaining lifetime of lithium batteries by fusing a projection operator and a conditional variational autoencoder. The technical problem to be solved by this invention is achieved through the following technical solution: A method for predicting the remaining lifetime of lithium batteries by fusing projection operators and conditional variational autoencoders includes: Acquire data of the lithium battery under test; wherein, the data of the lithium battery under test includes the terminal voltage, charging and discharging current and battery casing temperature of the lithium battery under test at each time point within the sampling period; The data of the lithium battery under test is input into the pre-trained... Each base learner model yields the first prediction result for the remaining lifespan of the lithium battery under test. based on Based on the first prediction result and the pre-trained multi-layer conditional variational autoencoder model, a result is obtained for the remaining lifespan of the lithium battery under test. A second prediction result; Using the 2 obtained from pre-training The projection operator coefficients are used to weight and sum all the first and second prediction results to obtain the final prediction result of the remaining life of the lithium battery under test.

[0007] In one embodiment of the present invention, the pre-training yields 2 The process of generating projection operator coefficients includes: Obtain a sample dataset of lithium batteries; wherein the sample dataset includes sample data of multiple lithium batteries, and the sample data of each lithium battery includes the terminal voltage, charging and discharging current and battery casing temperature of the lithium battery in various sampling periods, and the tag information corresponding to a sample data of each lithium battery represents the remaining life of the lithium battery. Based on the aforementioned sample dataset, for the preset Each base learner performs The cross-folding experiment yielded the trained result. Each base learner is trained and obtained. The initial prediction results of each base learner on the corresponding test set in the sample dataset are combined into a set of initial prediction results. Construct an encoder and decoder for a conditional variational autoencoder, and stack the encoder and decoder to form a multilayer conditional variational autoencoder; Set the loss function of the multi-layer conditional variational autoencoder; The initial prediction result set is smoothed using a locally weighted linear regression method to obtain a smoothed initial prediction result set. Using the smoothed initial prediction result set as input data, and the label information of the sample data corresponding to the initial prediction result set as conditions, the multilayer conditional variational autoencoder is trained using the loss function of the multilayer conditional variational autoencoder and the backpropagation algorithm to obtain the trained multilayer conditional variational autoencoder model. Based on the trained multi-layer conditional variational autoencoder model, the generated data corresponding to each initial prediction result is used as the new prediction result. The newly added prediction results constitute a collection of new prediction results for the remaining life of lithium batteries. The initial set of prediction results and the newly added set of prediction results are concatenated, and a projection operator is trained based on the concatenation result to obtain 2. Projection operator coefficients.

[0008] In one embodiment of the present invention, obtaining a sample dataset of lithium batteries includes: Obtain the MIT-ST dataset for lithium batteries; The battery packs in the MIT-ST dataset are filtered to obtain the filtered MIT-ST dataset; Based on the field information of lithium battery terminal voltage, charging and discharging current and battery casing temperature, the field information of the filtered MIT-ST dataset is extracted to obtain the field-extracted MIT-ST dataset. For the MIT-ST dataset after the fields are extracted, the charge and discharge data of each lithium battery, represented as a complete time series, are generated as samples by overlapping slices in a sliding window, resulting in multiple sample data for that lithium battery. Each sample data for the lithium battery includes the terminal voltage, charge and discharge current, and battery casing temperature of the lithium battery within a sampling period matched with the size of the sliding window. The normalized result of the ratio of the lithium battery's fully charged capacity to its rated capacity within the sampling period is determined as the tag information corresponding to the sample data of the lithium battery to characterize the remaining life of the lithium battery. The lithium battery sample dataset is composed of sample data from all lithium batteries.

[0009] In one embodiment of the present invention, the Each base learner includes: These are distinct neural network models, including CNN, GRU, ResNet, Transformer, and combinations of different neural network models in series and parallel.

[0010] In one embodiment of the present invention, the construction of the encoder and decoder of the conditional variational autoencoder, and the stacking of the encoder and the decoder to form a multilayer conditional variational autoencoder, includes: An encoder for constructing a conditional variational autoencoder includes: a first hidden layer, a first fully connected layer, a second fully connected layer, and a latent variable calculation output layer; wherein, the first hidden layer accepts input data and conditions in a fully connected manner, and uses ReLU activation function; the first fully connected layer and the second fully connected layer are used to calculate the mean and variance of the latent variables based on the output of the first hidden layer, respectively; the latent variable calculation output layer is used to generate latent variables through reparameterization based on the calculated mean and variance of the latent variables; A decoder for a conditional variational autoencoder is constructed, comprising a second hidden layer and a third hidden layer; wherein the second hidden layer receives the latent variables and the conditions using a fully connected method; and the third hidden layer uses a sigmoid activation function to calculate the generated data. A preset number of encoders are stacked sequentially to form a whole encoder, and a preset number of decoders are stacked sequentially to form a whole decoder. Using the whole encoder and the whole decoder, a multilayer conditional variational autoencoder is constructed.

[0011] In one embodiment of the present invention, the loss function of the multilayer conditional variational autoencoder is expressed as: ; in, This represents the input data of the multilayer conditional variational autoencoder; This represents the generated data output by the multi-layer conditional variational autoencoder; Indicates the quantity of the input data; This represents the standard deviation of the latent variables generated by the encoder; The preset number represents the number of layers in the multi-layer conditional variational autoencoder; and Let represent the mean of the latent variables generated by the encoder and the decoder, respectively, and assume that the variances of the latent variables generated by the encoder and the decoder are equal. This indicates the calculation of the L2 norm.

[0012] In one embodiment of the present invention, the initial prediction result set is smoothed using a locally weighted linear regression method to obtain a smoothed initial prediction result set, including: For each original data point in the initial prediction result set, select the data point in the initial prediction result set that is closest to that original data point. The set of nearest neighbors of each original data point is obtained from the given data points, denoted as: ;in, ; Based on the difference between the original data point and the corresponding neighboring data points, the weights of each data point in the neighboring data points are calculated using a preset weight calculation formula. Based on the initial prediction result set, the label information set corresponding to the initial prediction result set, and the weights calculated for each data point in the initial prediction result set, a weighted regression model is used to calculate the estimated value corresponding to each data point in the initial prediction result set, and the results are merged to obtain a smoothed initial prediction result set.

[0013] In one embodiment of the present invention, the preset weight calculation formula is expressed as: ; in, This represents the original data points in the initial prediction result set. The corresponding nearest neighbor set The Middle The weights corresponding to each data point; ; ; Represents the original data points and the corresponding neighbor set Medium data points The absolute value of the difference between them; The formula used to calculate the estimated value corresponding to each data point in the initial prediction result set using a weighted regression model is expressed as follows: ; in, This represents the set of initial prediction results; This represents the set of initial prediction results after smoothing. This represents the set of weights corresponding to the data points in the initial prediction result set; This represents the set of tag information corresponding to the initial prediction result set.

[0014] In one embodiment of the present invention, the initial set of prediction results and the newly added set of prediction results are concatenated, and a projection operator is trained based on the concatenation result to obtain 2. The projection operator coefficients include: The initial prediction results are combined into a set. The prediction results of the battery cells are spliced ​​together in sequence to obtain the initial set of prediction results. The newly added prediction results are combined into a single set. The prediction results of the battery cells are spliced ​​together in sequence to obtain a new set of prediction results after splicing. The initial prediction result set and the newly added prediction result set after splicing are merged, and the corresponding lithium battery tag information is also merged to obtain a result containing 2 A merging set of vectors; Initialize the coefficients of each vector in the merge set; Based on the merged set and the preset learning rate, a coefficient update operation is performed to obtain the coefficients updated in the current iteration. One coefficient; Determine whether the current iteration has reached the preset total number of iterations; If so, end the iteration and update the current iteration's value to 2. If one coefficient is determined to be the projection operator coefficient, then return to perform a coefficient update operation based on the merge set and the preset learning rate to obtain the coefficients updated in the current iteration. The steps for each coefficient.

[0015] In one embodiment of the present invention, a coefficient update operation is performed based on the merged set and a preset learning rate to obtain the coefficients updated in the current iteration. The coefficients include: For each vector in the merged set, calculate the product of the vector and the current coefficient to obtain the weighted result of the vector; The difference between the weighted result of the calculated vector and the lithium battery tag data is used to obtain the corresponding difference vector; Calculate the inner product between the difference vector and each vector, and multiply the inner product result by the preset learning rate to obtain the coefficient update value of the corresponding vector; The difference between the current coefficient of each vector and the updated coefficient value is determined as the updated coefficient of the corresponding vector in the current iteration.

[0016] The beneficial effects of this invention are: The present invention proposes a novel multi-neural network ensemble method. Utilizing the projection theorem, it projects label information onto an invariant linear subspace composed of predictions from multiple neural networks, thereby obtaining a stable and optimally performing ensemble prediction result. To further improve the prediction performance of the ensemble method and reduce training complexity, the present invention employs a conditional variational autoencoder (CDAE) to generate more prediction results based on existing neural network predictions. Since the CDAE is nonlinear, the generated prediction results are linearly independent of the original prediction results, thus expanding the invariant subspace. When the label information is projected onto this expanded invariant subspace, the performance of the resulting ensemble prediction is further improved. The method proposed in this invention integrates more neural network prediction results using a projection operator, not only further improving prediction performance but also enhancing the stability of the prediction results due to the stability of the projection operator. This invention enables more accurate and stable prediction of the remaining lifespan of lithium batteries, overcoming the shortcomings of existing neural network ensemble methods in predicting the remaining lifespan of lithium batteries, such as high training difficulty and the need for further improvement in the ensemble prediction performance. It provides technical support for the safe operation of lithium battery swapping modes. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a lithium battery remaining lifetime prediction method that integrates a projection operator and a conditional variational autoencoder, as provided in an embodiment of the present invention. Figure 2 2 were pre-trained for the embodiments of the present invention. A flowchart illustrating the coefficients of the projection operator; Figure 3 This is a lithium battery capacity degradation curve from the MIT-ST dataset in an embodiment of the present invention; Figures 4(a) and 4(b) are schematic diagrams of the encoder and decoder of the conditional variational autoencoder constructed according to an embodiment of the present invention; Figure 5 A schematic diagram of the structure of a 4-layer conditional variational autoencoder constructed according to an embodiment of the present invention; Figure 6 shows the comparison results between the method of the present invention and a single neural network; Figure 7 shows the comparison results of the method of the present invention with other neural network integration methods. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0019] First, a brief explanation of the relevant background knowledge and the purpose of this invention will be given.

[0020] Taking the electric vehicle sector as an example, with the large-scale popularization of electric vehicles, the long charging time of traditional charging methods has become a pain point in the market.

[0021] Battery swapping refers to a mode where, when an electric vehicle's battery is low, its power battery can be quickly removed from the vehicle and a new power battery installed at a battery swapping station, achieving vehicle-battery separation. The removed batteries are then charged, stored, and allocated by the swapping station. A battery swapping station mainly comprises five modules: a positioning system, a battery swapping system, an operation and maintenance system, a safety system, and a logistics system. The battery swapping system is the core of the station, consisting of a battery swapping platform, a palletizer, a lifting mechanism, a locking mechanism, and connectors. The quick-swap locking mechanism is a key component connecting the power battery box and the vehicle body, and is crucial for enabling battery swapping services. It not only ensures a reliable connection between the electric vehicle's battery box and the vehicle body but also meets the need for rapid battery pack replacement. For the replaced batteries, their remaining lifespan needs to be predicted to ensure no safety issues arise during subsequent installation.

[0022] Because lithium batteries have a very complex structure and degradation mechanism, and electric vehicles operate under various conditions, the remaining lifespan of a battery is not a technical parameter that can be directly measured. It is necessary to establish a corresponding model for estimation, which is one of the major challenges currently facing the health status management of lithium batteries in electric vehicles.

[0023] Because the degradation process of lithium batteries is characterized by nonlinearity and uncertainty, a single neural network, due to its structural limitations, cannot capture all data features. Therefore, it is necessary to employ multiple heterogeneous neural networks for predicting the remaining life of lithium batteries, and to synthesize the prediction results of multiple network models through ensemble learning techniques to obtain a better prediction result.

[0024] Ensemble learning refers to the process of generating multiple learners with independent decision-making capabilities according to a certain strategy, and then combining them to solve the same problem. Generally, ensemble learners have stronger generalization capabilities than individual learners, and their predictive performance is significantly better, thus this technology has promising applications. For example, random forests are a classic bagging ensemble method based on decision trees. Multiple decision trees are trained in parallel, and all prediction results are combined using majority voting (classification) or averaging (regression). In random forests, each decision tree undergoes sample and feature sampling during training, which helps prevent overfitting. AdaBoost is a serially trained model where the results of each weak learner influence the weighting of the next learner and the weighting of training samples, thereby improving the generalization ability of the ensemble model. Stacking, as the name suggests, completes the ensemble process by "stacking" models. It combines the outputs of multiple base learners in a certain way, or selects the optimal model from them. The working process of this type of method can be roughly summarized into two stages: the first stage is to train the base learner to obtain the preliminary prediction results of the base learner; the second stage is to use the outputs of multiple base learners as training data to train another model in order to achieve the purpose of ensemble.

[0025] Existing ensemble methods, when applied to multi-neural network ensembles for predicting the remaining lifespan of lithium batteries, suffer from complex training, reliance on majority rules, and no significant performance improvement. This invention aims to address these issues, thereby achieving more accurate and stable predictions of lithium battery remaining lifespan and providing technical support for the safe operation of lithium battery swapping systems.

[0026] This invention provides a method for predicting the remaining lifetime of lithium batteries by fusing projection operators and conditional variational autoencoders, such as... Figure 1 As shown, the method may include the following steps: S1, acquire data of the lithium battery under test; The data of the lithium battery under test includes the terminal voltage, charging and discharging current and battery casing temperature of the lithium battery under test at each time point within the sampling period. S2, input the lithium battery data to be tested into the pre-trained... Each base learner model yields the first prediction result for the remaining lifespan of the lithium battery under test. S3, based on Based on the first prediction result and the pre-trained multi-layer conditional variational autoencoder model, a result is obtained for the remaining lifespan of the lithium battery under test. A second prediction result; S4, using pre-trained 2 The projection operator coefficients are used to weight and sum all the first and second prediction results to obtain the final prediction result of the remaining life of the lithium battery under test.

[0027] To facilitate understanding of the embodiments of the present invention, the pre-completed training process will first be described in detail. Specifically: Please see Figure 2 The pre-training yielded 2 The process of generating projection operator coefficients may include the following steps: Step a1: Obtain a sample dataset of lithium batteries; The sample dataset includes sample data of multiple lithium batteries. The sample data of each lithium battery includes the terminal voltage, charging and discharging current and battery casing temperature of the lithium battery in each sampling period. The tag information corresponding to a sample data of each lithium battery represents the remaining life of the lithium battery. Step a1 may include the following steps: Step a11: Obtain the MIT-ST dataset for lithium batteries; Specifically, this embodiment of the invention uses the lithium-ion battery capacity degradation dataset provided by the MIT-Stanford-Toyota Research Center (MIT-ST), referred to as the MIT-ST dataset.

[0028] Step a12: Filter the battery packs in the MIT-ST dataset to obtain the filtered MIT-ST dataset; The MIT-ST dataset was collected from 124 sets of cycle charge-discharge experiments of commercial lithium iron phosphate (LiFePO4, LFP) / graphite batteries manufactured by A123 Systems, model APR18650M1A. The batch from June 30, 2017 showed more stable cycle counts than other batches, and this batch of lithium battery data can be selected to construct the training and test sets in the experiments of this invention. Furthermore, Figure 3 The capacity degradation curves of all lithium batteries in this batch are shown. It can be seen that when the battery capacity drops to 75% of the rated capacity, most batteries have between 450 and 550 charge-discharge cycles. In addition, the capacity decay process of all batteries is non-linear, that is, the capacity decay is slow in the early stage and relatively fast in the later stage.

[0029] The selected batch of lithium batteries from June 30, 2017 contained 48 APR18650M1A batteries. First, all batteries exhibiting anomalies during data acquisition were removed. These removed batteries included: lithium battery data collected from channels 1, 2, 3, 5, and 6 of a previous batch (which had already been used in the May 12, 2017 batch and whose performance had degraded); lithium batteries from channels 7 and 21 with abnormal thermocouple detachment, resulting in abnormal temperature data; batteries from channel 10 with inherent defects, exhibiting extremely rapid capacity decay during the experiment; and thermocouples from channels 15 and 16 exhibiting anomalies. Therefore, 35 sets of the remaining batteries were selected as the filtered MIT-ST dataset for 5 rounds of cross-validation. In the cross-folding experiment It can be 5. Each fold includes 28 training sets and 7 test sets, and detailed grouping information is shown in Table 1.

[0030] Table 1 Grouping of the MIT-ST dataset

[0031] Step a13: Based on the field information of lithium battery terminal voltage, charging and discharging current and battery casing temperature, the field of the filtered MIT-ST dataset is extracted to obtain the field-extracted MIT-ST dataset. Since the internal resistance of lithium batteries does not change significantly in the early stages of degradation, but rises sharply as degradation progresses, the change in internal resistance over a single charge-discharge cycle is not significant for predicting the health status of lithium batteries. For other fields in the filtered MIT-ST dataset, the state information of the lithium battery's input power, load power, or stored energy can be calculated using the integral of terminal voltage, charge / discharge current, and charging time. Furthermore, the battery's operating temperature also affects the rate of its internal electrochemical reactions. Therefore, the MIT-ST dataset with extracted fields such as lithium battery terminal voltage, charge / discharge current, and battery casing temperature was ultimately selected to train the base learner.

[0032] Step a14: For the MIT-ST dataset after field extraction, the charge and discharge data of each lithium battery, represented as a complete time series, are generated as samples using a sliding window overlapping slicing method to obtain multiple sample data for the lithium battery; each sample data of the lithium battery includes the terminal voltage, charge and discharge current, and battery casing temperature of the lithium battery within a sampling duration matched with the size of the sliding window; and the normalized result of the ratio of the lithium battery's fully charged capacity to its rated capacity within the sampling duration is determined as the tag information corresponding to the sample data of the lithium battery to characterize the remaining life of the lithium battery; Typically, time series data can be processed using sliding windows to generate a dataset that the network model can read. Sliding windows are categorized into continuous and discontinuous time windows based on their sliding method, and can also be classified as fixed or variable time windows by changing their size. For the MIT-ST dataset after field extraction, the original data consists of continuous charge and discharge data of lithium batteries, including terminal voltage, charge and discharge current, and battery casing temperature. Therefore, a fixed-size discontinuous sliding window is used to generate samples; that is, the window slides by jumping a certain step size. First, the charge and discharge data of a single lithium battery are merged into a complete time series, resulting in the charge and discharge data of each lithium battery represented by a complete time series. This is the charge and discharge data sequence from the battery's factory state to when its capacity decays to 75% of its rated capacity. The sliding window is then used to overlap and slice the charge and discharge data of each lithium battery represented by its complete time series to generate samples. Since the time step required for the lithium battery to charge from the 2.0V discharge cutoff voltage to 100% State of Charge (SoC) during the acquisition of this dataset is approximately 500, the sliding window size can be set to 500 time steps in the experiment. This means that the sampling duration matched by the sliding window size is 500 time steps; the sliding window slides by 100 time steps each time. For each generated sample, the lithium battery's health status, i.e., remaining lifetime, refers to the ratio of the lithium battery's capacity when fully charged to its rated capacity under the current window state, after normalization.

[0033] Step a15: The sample dataset of lithium batteries is constructed from the sample data of all lithium batteries.

[0034] For ease of understanding, the sample dataset of lithium batteries can be represented as follows: ,in, This represents the number of lithium batteries and is a natural number greater than 0. For the first The sample data of the lithium battery includes multiple sample data, each of which includes the terminal voltage, charging and discharging current and battery casing temperature of the lithium battery within a sampling period that matches the size of the sliding window.

[0035] Step a2, based on the sample dataset, perform the preset... Each base learner performs The cross-folding experiment yielded the trained result. Each base learner is trained and obtained. The initial prediction results of each base learner on the corresponding test set in the sample dataset are combined into a set of initial prediction results. Among them, the Each base learner includes: These are distinct neural network models, including CNN, GRU, ResNet, Transformer, and combinations of different neural network models in series and parallel.

[0036] The cascading and parallel combinations of different neural network models can be, for example, CNN and GRU in cascading, CNN and GRU in parallel, ResNet and GRU in cascading, ResNet and GRU in parallel, etc., which will not be listed in detail here.

[0037] For ease of understanding, Each base learner can be represented as ; Individual base learners on a sample dataset of lithium batteries Upward The cross-validation experiment, as shown in Table 1, can be performed 5 times. Each fold includes 28 training sets and 7 test sets. The training sets are used to train the base learner, resulting in the trained... Each base learner can be represented as The trained base learner makes predictions on the test set to obtain initial prediction results. All initial predictions obtained from the crossover experiment are combined into a set of initial predictions, which can be used as... express, , respectively corresponding The prediction results of each base learner. Corresponding to the The prediction results of each base learner include One portion, This refers to the number of lithium batteries.

[0038] Taking a CNN neural network as the first base learner as an example, in a crossover experiment, the model is first trained on the training set (sample data from 28 lithium batteries). Then, the trained model is used to predict the remaining lifespan of the remaining 7 batteries on the test set. This crossover experiment is repeated 5 times to obtain the initial prediction results for all batteries, which can be represented as follows: The prediction results of other neural networks are the same, and will not be elaborated here.

[0039] Step a3: Construct the encoder and decoder of the conditional variational autoencoder, and stack the encoder and decoder to form a multilayer conditional variational autoencoder; The Conditional Variational Autoencoder (CVAE) is a generative model that has been widely used in the field of dialogue generation. CVAE combines the ideas of autoencoders and variational inference, enabling the model to generate dialogue content that conforms to the context and speech by learning latent semantic space representations.

[0040] This embodiment of the invention applies a variational autoencoder to a preset scenario of remaining lithium battery life. Specifically, step a3 may include: Step a31: Construct the encoder of the conditional variational autoencoder, including: a first hidden layer, a first fully connected layer, a second fully connected layer, and a latent variable computation output layer; The first hidden layer accepts input data and conditions using a fully connected approach, and the activation function used is ReLU. The first fully connected layer and the second fully connected layer are used to calculate the mean and variance of the latent variables based on the output of the first hidden layer, respectively. The latent variable calculation output layer is used to generate latent variables based on the calculated mean and variance of the latent variables through a reparameterization method. The structure of the encoder is shown in Figure 4(a), where the hidden layer is the first hidden layer, which uses a fully connected method to receive input data. and conditions In the first hidden layer The components (neurons) representing latent variables are linear combinations of input data and conditions. The mean and variance of the latent variables are expressed as follows: as well as The latent variables generated by the encoder are ,in, The number of input data.

[0041] Step a32: Construct the decoder of the conditional variational autoencoder, including: a second hidden layer and a third hidden layer; The second hidden layer uses a fully connected method to receive the hidden variables and the conditions; the third hidden layer uses the sigmoid activation function to calculate and generate data. The structure of the decoder is shown in Figure 4(b), where the two hidden layers are the second hidden layer and the third hidden layer; and This refers to the width of these two hidden layers. The generated data (i.e., the generated data) is represented as... The generated data is an estimate of the input data.

[0042] Step a33: Stack a preset number of encoders sequentially to form a whole encoder, and stack the preset number of decoders sequentially to form a whole decoder. Use the whole encoder and the whole decoder to form a multilayer conditional variational autoencoder.

[0043] The preset number can be determined according to needs or experimental results. In this embodiment of the invention, the preset number can specifically be 4, that is, a 4-layer conditional variational autoencoder is constructed. The structure is as follows: Figure 5 As shown.

[0044] Where Encoder represents the encoder; Decoder represents the decoder; and the Latent following Encoder... This represents the implicit variables in its output; the Latent after the Decoder. This represents the implicit variables in its output.

[0045] Step a4: Set the loss function of the multi-layer conditional variational autoencoder; Wherein, assuming the conditional variational autoencoder is Markovian, the loss function of the multilayer conditional variational autoencoder is expressed as: ; in, This represents the input data of the multilayer conditional variational autoencoder; This represents the generated data output by the multi-layer conditional variational autoencoder; Indicates the quantity of the input data; This represents the standard deviation of the latent variables generated by the encoder; The preset number represents the number of layers in the multi-layer conditional variational autoencoder; and Let represent the mean of the latent variables generated by the encoder and the decoder, respectively, and assume that the variances of the latent variables generated by the encoder and the decoder are equal. This indicates the calculation of the L2 norm.

[0046] Step a5: The initial prediction result set is smoothed using a locally weighted linear regression method to obtain a smoothed initial prediction result set. In this embodiment of the invention, the initial set of prediction results is used as the training set of the multilevel conditional variational autoencoder. First, the training set is smoothed by local weighted linear regression to prevent gradient explosion or vanishing when constructing the hierarchical conditional variational autoencoder.

[0047] In one optional implementation, step a5 may include: Step a51: For each original data point in the initial prediction result set, select the data point in the initial prediction result set that is closest to that original data point. The set of nearest neighbors of each original data point is obtained from the given data points, denoted as: ;in, ; The quantity can be set as needed, for example, it can be 7, etc.

[0048] Step a52: Based on the difference between the original data point and the corresponding data points in the neighboring set, calculate the weights of each data point in the neighboring set of the original data point using a preset weight calculation formula. The preset weight calculation formula is expressed as follows: ; in, This represents the original data points in the initial prediction result set. The corresponding nearest neighbor set The Middle The weights corresponding to each data point; ; ; Represents the original data points and the corresponding neighbor set Medium data points The absolute value of the difference between them; It is understandable that for any original data point in the set of initial prediction results, a corresponding weight can be calculated for each data point in its neighboring set.

[0049] Step a53: Based on the initial prediction result set, the label information set corresponding to the initial prediction result set, and the weights calculated for each data point in the initial prediction result set, a weighted regression model is used to calculate the estimated value corresponding to each data point in the initial prediction result set, and the results are merged to obtain a smoothed initial prediction result set.

[0050] Specifically, the formula used to calculate the estimated value corresponding to each data point in the initial prediction result set using a weighted regression model is expressed as follows: ; in, This represents the set of initial prediction results; This represents the set of initial prediction results after smoothing. This represents the set of weights corresponding to the data points in the initial prediction result set; This represents the set of tag information corresponding to the initial prediction result set.

[0051] It is understandable that, through the above processing, each data point in the initial prediction result set can obtain a corresponding estimated value. Step a6: Using the smoothed initial prediction result set as input data, and the label information of the sample data corresponding to the initial prediction result set as conditions, the multilayer conditional variational autoencoder is trained using the loss function of the multilayer conditional variational autoencoder and the backpropagation algorithm to obtain the trained multilayer conditional variational autoencoder model. against Figure 5 The smoothed initial prediction result set is the input data set of the multilayer conditional variational autoencoder (MLCE), and the label information of the sample data corresponding to the initial prediction result set is the conditional set. The MLCE is trained using the backpropagation algorithm with the Adam optimizer. When the loss function of the MLCE converges, the trained MLCE model is obtained. The specific training process can be understood by referring to existing conditional variational autoencoder training processes, and will not be described in detail here.

[0052] Step a7: Based on the trained multi-layer conditional variational autoencoder model, the generated data corresponding to each initial prediction result is obtained as the new prediction result, and then... The newly added prediction results constitute a collection of new prediction results for the remaining life of lithium batteries. For each initial prediction result, the trained multi-layer conditional variational autoencoder model can be used to generate corresponding generated data, resulting in... The newly added prediction results can achieve the modification of the original ones. The expansion of the initial predictions obtained by each base learner results in a larger number of predictions, thus creating a larger search space.

[0053] Step a8: Concatenate the initial prediction result set and the newly added prediction result set; train the projection operator based on the concatenation result to obtain 2. Projection operator coefficients.

[0054] In one optional implementation, step a8 may include: Step a81, combine the initial prediction results into the set. The prediction results of the battery cells are spliced ​​together in sequence to obtain the initial set of prediction results. The initial prediction results are aggregated. In The prediction results of each battery cell are sequentially concatenated to obtain an initial set of prediction results. (The sentence is incomplete and requires further context.) Taking one model as an example , ;in This indicates the concatenation of multidimensional arrays; Indicates the first In the model, the first The prediction results for the block battery correspond to a total of Block lithium battery; Indicates the first The concatenation result of the models is represented as follows: .

[0055] Step a82, combine the newly added prediction results into the set. The prediction results of the battery cells are spliced ​​together in sequence to obtain a new set of prediction results after splicing. The newly added set of prediction results can be represented as: ,in This represents the number of generated prediction results; the newly added prediction results are then aggregated. Using the same method, the newly added prediction results were spliced ​​together to obtain a collection of spliced ​​prediction results. . Step a83: Merge the initial prediction result set and the newly added prediction result set after splicing, and merge the corresponding lithium battery tag information to obtain a result containing 2 A merging set of vectors; The merge set can be represented as: .in, Indicates the first A vector.

[0056] Step a84: Initialize the coefficients of each vector in the merged set; initialization The coefficient of each vector in the vector is ; Step a85: Based on the merged set and the preset learning rate, perform a coefficient update operation to obtain the coefficients updated in the current iteration. One coefficient; Specifically, step a85 includes: 1) For each vector in the merged set, calculate the product of the vector and the current coefficient to obtain the weighted result of the vector; This step is calculation. ; 2) The difference between the weighted result of the calculated vector and the lithium battery tag data is used to obtain the corresponding difference vector; This step involves calculating the difference vector. Lithium battery tag data refers to the tag information corresponding to the sample data of lithium batteries.

[0057] 3) Calculate the inner product between the difference vector and each vector, and multiply the inner product result by the preset learning rate to obtain the coefficient update value of the corresponding vector; This step is calculation. ,in, This represents the result of the inner product.

[0058] 4) The difference between the current coefficient of each vector and the updated coefficient value is determined as the updated coefficient of the corresponding vector in the current iteration.

[0059] This step can be represented as: ; The coefficients of each vector can be updated through the above processing.

[0060] Step a86: Determine whether the current iteration has reached the preset total number of iterations; The preset total number of iterations can be set as needed and is not limited here.

[0061] If so, execute step a87 to end the iteration and update the current iteration's value to 2. The coefficients are determined as projection operator coefficients; If not, return to step a85, that is, perform a coefficient update operation based on the merged set and the preset learning rate to obtain the updated coefficients for the current iteration. The steps for each coefficient.

[0062] In this embodiment of the invention, multiple neural networks are selected as base learners, trained on the training set of the sample dataset, and the trained base learners are used to predict the test set. The trained dataset is then saved. A base learner and an initial set of prediction results are used. Then, a multilayer conditional variational autoencoder (MLCE) is constructed. The prediction results of the base learners on the test set (i.e., the initial set of prediction results) are used as the training set to train the constructed MCE. The trained MCE model is then used to generate a new set of prediction results from the training data, and the trained MCE model is saved. The purpose is to use the MCE model as a generative model to generate new lithium battery prediction results, thereby reducing the number of base learners to train and thus reducing the training difficulty. Next, the prediction results of the base learners on the test set are concatenated with the results generated by the MCE to train the projection operator and obtain projection coefficients. The purpose is to integrate the obtained lithium battery prediction results to obtain a more accurate and stable final prediction result.

[0063] The above is the training process of an embodiment of the present invention. After obtaining 2... After obtaining the projection operator coefficients, various trained models can be used to predict the health status of the lithium battery under test. That is, for the actual lithium battery remaining life prediction process, steps S1~S4 are executed.

[0064] Specifically, in S1, the form of the lithium battery data to be tested is the same as that of the sample data of any lithium battery during the training process. The difference is that it does not contain label data that characterizes the remaining life of the battery, but rather aims to predict the remaining life of the lithium battery.

[0065] For S2, it is completed and saved using training. Each base learner model The remaining lifespan of the lithium battery under test is predicted to obtain... The first prediction result.

[0066] For S3, it is to The first prediction result is input into a pre-trained multi-layer conditional variational autoencoder model to obtain generated data, which is the data for the remaining life of the lithium battery under test. The second prediction result.

[0067] For S4, the prediction results from the base learner are merged with the results generated by the multilayer variational autoencoder model. Specifically, the results are... The first prediction result and The second prediction result utilizes 2 The weighted summation of the projection operator coefficients yields the final prediction result of the remaining lifespan of the lithium battery under test, which serves as the final prediction result of the lithium battery's health status.

[0068] The innovation of this invention lies in combining a hierarchical conditional variational autoencoder with locally weighted linear regression. The model is trained using the initial predictions from the base learner, and the trained model generates predictions with better performance than the initial predictions. The final prediction result is then obtained by combining this with a projection operator. This invention proposes a novel multi-neural network ensemble method that utilizes the projection theorem to project label information onto an invariant linear subspace composed of predictions from multiple neural networks, thereby obtaining a stable and optimally performing ensemble prediction result. To further improve the prediction performance and reduce training complexity, this invention employs a conditional variational autoencoder to generate more predictions based on existing neural network predictions. Since the conditional variational autoencoder is nonlinear, the generated predictions are linearly independent of the original predictions, thus expanding the invariant subspace. Projecting label information into this expanded invariant subspace further enhances the performance of the ensemble prediction result. The method proposed in this invention integrates more neural network predictions using a projection operator. Compared to existing neural network ensemble methods, this invention not only improves prediction performance but also enhances the stability of the prediction results due to the stability of the projection operator. This invention can more accurately and stably predict the remaining lifespan of lithium batteries, overcoming the shortcomings of existing neural network integration methods in predicting the remaining lifespan of lithium batteries, such as high training difficulty and the need for further improvement in the prediction performance of the integration. It can provide technical support for the safe operation of lithium battery swapping mode.

[0069] To verify the effectiveness of the method provided in the embodiments of the present invention, the relevant simulation results are given below.

[0070] Figure 6 This invention provides a comparison between the method provided by the present invention and the prediction results of a single neural network. From... Figure 6 As can be seen, the root mean square error (RMSE) of the method provided by this invention is lower than the average value of a single neural network on all battery packs, indicating that the method provided by this invention is effective.

[0071] Figure 7 This invention provides a comparison between the method provided by this invention and other neural network integration methods. From... Figure 7 As can be seen, the root mean square error of the method provided by this invention is lower than that of other neural network ensemble methods on all battery packs, indicating that the prediction performance of the method provided by this invention is better than that of the compared neural network ensemble methods.

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for predicting the remaining life of a lithium battery by fusing a projection operator and a conditional variational autoencoder, characterized in that, include: Acquire data of the lithium battery under test; wherein, the data of the lithium battery under test includes the terminal voltage, charging and discharging current and battery casing temperature of the lithium battery under test at each time point within the sampling period; The data of the lithium battery under test is input into m pre-trained base learner models to obtain the first prediction result for the remaining life of the lithium battery under test. Based on m first prediction results and a pre-trained multi-layer conditional variational autoencoder model, m second prediction results are obtained for the remaining life of the lithium battery under test. Using the 2m projection operator coefficients obtained through pre-training, all the first and second prediction results are weighted and summed to obtain the final prediction result of the remaining life of the lithium battery under test.

2. The lithium battery remaining lifetime prediction method based on the fusion of projection operator and conditional variational autoencoder according to claim 1, characterized in that, The process of obtaining 2m projection operator coefficients through pre-training includes: Obtain a sample dataset of lithium batteries; wherein the sample dataset includes sample data of multiple lithium batteries, and the sample data of each lithium battery includes the terminal voltage, charging and discharging current and battery casing temperature of the lithium battery in various sampling periods, and the tag information corresponding to a sample data of each lithium battery represents the remaining life of the lithium battery. Based on the sample dataset, k-fold crossover experiments are performed on the preset m base learners to obtain the m base learners that have been trained. The initial prediction results of the m base learners on the corresponding test set in the sample dataset are obtained and merged as the initial prediction result set. Construct an encoder and decoder for a conditional variational autoencoder, and stack the encoder and decoder to form a multilayer conditional variational autoencoder; Set the loss function of the multi-layer conditional variational autoencoder; The initial prediction result set is smoothed using a locally weighted linear regression method to obtain a smoothed initial prediction result set. Using the smoothed initial prediction result set as input data, and the label information of the sample data corresponding to the initial prediction result set as conditions, the multilayer conditional variational autoencoder is trained using the loss function of the multilayer conditional variational autoencoder and the backpropagation algorithm to obtain the trained multilayer conditional variational autoencoder model. Based on the trained multi-layer conditional variational autoencoder model, the generated data corresponding to each initial prediction result is obtained as the new prediction result, and the m new prediction results constitute a set of new prediction results for the remaining life of lithium battery. The initial set of prediction results and the newly added set of prediction results are concatenated, and the projection operator is trained based on the concatenation result to obtain 2m projection operator coefficients.

3. The lithium battery remaining life prediction method based on the fusion of projection operator and conditional variational autoencoder according to claim 2, characterized in that, The process of obtaining the lithium battery sample dataset includes: Obtain the MIT-ST dataset for lithium batteries; The battery packs in the MIT-ST dataset are filtered to obtain the filtered MIT-ST dataset; Based on the field information of lithium battery terminal voltage, charging and discharging current and battery casing temperature, the field information of the filtered MIT-ST dataset is extracted to obtain the field-extracted MIT-ST dataset. For the MIT-ST dataset after the fields are extracted, the charge and discharge data of each lithium battery, represented as a complete time series, are generated as samples by overlapping slices in a sliding window, resulting in multiple sample data for that lithium battery. Each sample data for the lithium battery includes the terminal voltage, charge and discharge current, and battery casing temperature of the lithium battery within a sampling period matched with the size of the sliding window. The normalized result of the ratio of the lithium battery's fully charged capacity to its rated capacity within the sampling period is determined as the tag information corresponding to the sample data of the lithium battery to characterize the remaining life of the lithium battery. The lithium battery sample dataset is composed of sample data from all lithium batteries.

4. The lithium battery remaining life prediction method based on the fusion of projection operator and conditional variational autoencoder according to claim 3, characterized in that, The m base learners include m distinct neural network models, including CNN, GRU, ResNet, Transformer, and combinations of different neural network models in series or parallel.

5. The lithium battery remaining life prediction method based on the fusion of projection operator and conditional variational autoencoder according to claim 4, characterized in that, The construction of the conditional variational autoencoder includes an encoder and a decoder, and stacking the encoder and the decoder to form a multilayer conditional variational autoencoder, comprising: An encoder for constructing a conditional variational autoencoder includes: a first hidden layer, a first fully connected layer, a second fully connected layer, and a latent variable calculation output layer; wherein, the first hidden layer accepts input data and conditions in a fully connected manner, and uses ReLU activation function; the first fully connected layer and the second fully connected layer are used to calculate the mean and variance of the latent variables based on the output of the first hidden layer, respectively; the latent variable calculation output layer is used to generate latent variables through reparameterization based on the calculated mean and variance of the latent variables; A decoder for a conditional variational autoencoder is constructed, comprising a second hidden layer and a third hidden layer; wherein the second hidden layer receives the latent variables and the conditions using a fully connected method; and the third hidden layer uses a sigmoid activation function to calculate the generated data. A preset number of encoders are stacked sequentially to form a whole encoder, and a preset number of decoders are stacked sequentially to form a whole decoder. Using the whole encoder and the whole decoder, a multilayer conditional variational autoencoder is constructed.

6. The lithium battery remaining lifetime prediction method based on the fusion of projection operator and conditional variational autoencoder according to claim 5, characterized in that, The loss function of the multilayer conditional variational autoencoder is expressed as: Where, x i This represents the input data of the multilayer conditional variational autoencoder; This represents the generated data output by the multilayer conditional variational autoencoder; N represents the number of input data; σ(q) φ (z t )) represents the standard deviation of the latent variables generated by the encoder; T corresponds to the preset quantity, representing the number of layers of the multilayer conditional variational autoencoder; μ(q) φ (z t )) and μ(p θ (z t )) represent the means of the latent variables generated by the encoder and the decoder, respectively, and it is assumed that the variances of the latent variables generated by the encoder and the decoder are equal; ||·|| 2 This indicates the calculation of the L2 norm.

7. The lithium battery remaining lifetime prediction method based on the fusion of projection operator and conditional variational autoencoder according to claim 6, characterized in that, The initial prediction result set is smoothed using a locally weighted linear regression method to obtain a smoothed initial prediction result set, including: For each original data point in the initial prediction result set, the h nearest data points in the initial prediction result set are selected to obtain the nearest neighbor set of the original data point, denoted as {x1, x2, ..., x...}. h }; where h <N; Based on the difference between the original data point and the corresponding neighboring data points, the weights of each data point in the neighboring data points are calculated using a preset weight calculation formula. Based on the initial prediction result set, the label information set corresponding to the initial prediction result set, and the weights calculated for each data point in the initial prediction result set, a weighted regression model is used to calculate the estimated value corresponding to each data point in the initial prediction result set, and the results are merged to obtain a smoothed initial prediction result set.

8. The lithium battery remaining lifetime prediction method based on the fusion of projection operator and conditional variational autoencoder according to claim 7, characterized in that, The preset weight calculation formula is expressed as follows: Among them, w i This represents the set of nearest neighbors {x1, x2, ..., xn} corresponding to the original data point x in the initial prediction result set. h The weight corresponding to the i-th data point in}; |xx i | represents the original data point x and its corresponding neighbor set {x1, x2, ..., x}. h The absolute value of the difference between data points xi in the data set; The formula used to calculate the estimated value for each data point in the initial prediction result set using a weighted regression model is expressed as follows: Where X represents the set of initial prediction results; W represents the set of initial prediction results after smoothing; W represents the set of weights corresponding to the data points in the set of initial prediction results; Y represents the set of label information corresponding to the set of initial prediction results.

9. The lithium battery remaining lifetime prediction method based on the fusion of projection operator and conditional variational autoencoder according to claim 8, characterized in that, The initial set of prediction results and the newly added set of prediction results are concatenated. Based on the concatenation result, a projection operator is trained to obtain 2m projection operator coefficients, including: The prediction results of K batteries in the initial prediction result set are spliced ​​together in order to obtain the spliced ​​initial prediction result set. The prediction results of K batteries in the newly added prediction result set are spliced ​​together in order to obtain the spliced ​​newly added prediction result set. The initial prediction result set and the newly added prediction result set after splicing are merged, and the corresponding lithium battery tag information is also merged to obtain a merged set containing 2m vectors. Initialize the coefficients of each vector in the merge set; Based on the merged set and the preset learning rate, a coefficient update operation is performed to obtain 2m coefficients after the current iteration update; Determine whether the current iteration has reached the preset total number of iterations; If yes, end the iteration and determine the 2m coefficients updated in the current iteration as projection operator coefficients; if no, return to the step of performing coefficient update operation based on the merge set and the preset learning rate to obtain the 2m coefficients updated in the current iteration.

10. The lithium battery remaining lifetime prediction method based on the fusion of projection operator and conditional variational autoencoder according to claim 9, characterized in that, Based on the merged set and the preset learning rate, a coefficient update operation is performed to obtain 2m coefficients after the current iteration update, including: For each vector in the merged set, calculate the product of the vector and the current coefficient to obtain the weighted result of the vector; The difference between the weighted result of the calculated vector and the lithium battery tag data is used to obtain the corresponding difference vector; Calculate the inner product between the difference vector and each vector, and multiply the inner product result by the preset learning rate to obtain the coefficient update value of the corresponding vector; The difference between the current coefficient of each vector and the updated coefficient value is determined as the updated coefficient of the corresponding vector in the current iteration.

Images