Lithium battery remaining life prediction method and system based on multiple conditional variational autoencoder

Abstract

This invention discloses a method and system for predicting the remaining life of lithium batteries using a multi-conditional variational autoencoder, belonging to the field of lithium battery remaining life prediction technology. The method first uses a pre-trained Transformer model to perform preliminary predictions on historical operating data, generating a preliminary prediction sequence. Second, it dynamically constructs guiding vectors based on clean samples, builds joint input features, and inputs them into the multi-conditional variational autoencoder model in conjunction with conditional information. Feature extraction and progressive correction are achieved through alternating cascades of multi-level composite encoders and attention networks, outputting deep latent variables. Subsequently, a locally weighted linear regression module smooths the latent variables, suppressing abnormal fluctuations. Finally, the model is reconstructed by a multi-level composite decoder and attention network to output a high-precision remaining life prediction result. This invention effectively solves the problem of traditional models easily learning false degradation patterns in high-noise environments, significantly improving prediction robustness and accuracy.

Description

Technical Field

[0001] This invention belongs to the field of lithium battery remaining useful life (RUL) prediction technology, and relates to a method and system for predicting the remaining useful life of lithium batteries using a multi-condition variational autoencoder. Background Technology

[0002] Lithium-ion batteries are the core power source for electric vehicles, aerospace, portable electronic devices, and large-scale energy storage systems. Their performance directly determines the reliability, safety, and lifespan of the entire system. With the rapid development of new energy technologies, precise management of lithium-ion battery health status and reliable prediction of remaining lifespan have become core technologies for ensuring the continuous and stable operation of critical equipment, optimizing maintenance strategies, and reducing total lifespan costs, possessing significant economic and safety value. Against this backdrop, data-driven deep learning methods, with their powerful nonlinear fitting and feature learning capabilities, are gradually becoming the mainstream technology in the field of lithium-ion battery remaining lifespan prediction.

[0003] However, there is a significant difference between the ideal laboratory environment and the complex real-world operating scenarios. Under real-world conditions, lithium battery operating data is continuously affected by various factors such as sensor accuracy drift, occasional electromagnetic interference, and unstable communication links. This results in the time-series data often containing a large number of contaminated and abnormal samples, while pure samples that clearly represent the battery's internal health degradation path are relatively rare. This constitutes a highly challenging data reality of "mostly contaminated, a few pure," severely limiting the ability of existing predictive models to extract true degradation patterns from noisy observations.

[0004] Current mainstream prediction models all face limitations imposed by their inherent architecture when addressing this common challenge. While convolutional neural network-based methods can effectively extract local features, their local connectivity and weight sharing characteristics make it difficult to model the long-range temporal dependencies crucial to battery degradation. They are also insensitive to contamination noise scattered throughout long sequences and tend to learn partial local correlations rather than global degradation trends. Recurrent neural networks, represented by Long Short-Term Memory networks, are designed specifically for sequence modeling, but their sequence state propagation mechanism, when faced with a high proportion of outlier data points, propagates and amplifies noise along time steps, ultimately "contaminating" the hidden states of the entire sequence and forcing the model to learn degradation patterns that deviate from the true trajectory.

[0005] Generative models, such as variational autoencoders, provide a probabilistic framework for handling data uncertainty, but their standard form lacks an explicit conditional guidance mechanism. While conditional variational autoencoders can introduce conditional information, their encoder-decoder structure remains insufficient for global temporal modeling of complex long sequences. More critically, existing generative models generally lack an explicit, structured mechanism to proactively identify and retrieve key information from a small number of clean samples in historical long sequences during the model inference phase, in order to correct the representation and prediction of contaminated data at the current moment. Existing technologies struggle to achieve accurate long-term temporal dependency modeling and efficient utilization of sparse clean information in complex noisy environments, leading to a significant risk of performance degradation in real-world, complex operating scenarios. This deficiency has become a key bottleneck restricting the practical application of lithium battery remaining life prediction technology with high reliability. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for predicting the remaining life of lithium batteries using a multi-condition variational autoencoder, so as to solve the technical problem that the prediction accuracy of lithium batteries is insufficient under complex noise backgrounds in the prior art.

[0007] To achieve the above objectives, the present invention employs the following technical solution: In a first aspect, the present invention provides a method for predicting the remaining life of a lithium battery using a multi-condition variational autoencoder, comprising the following steps: Acquire historical operating data of the lithium battery to be tested, and divide the historical operating data into contaminated samples and clean samples; The contaminated sample is input into a pre-trained Transformer model for preliminary prediction to obtain a preliminary prediction sequence; K guiding vectors are constructed based on the pure samples, and the historical running data is fused with the preliminary prediction sequence to construct joint input features; The joint input features are used as the first input, combined with the historical running data and the guiding vector as conditional information, and input into the multi-conditional variational autoencoder model with fused attention network for feature extraction and correction, and mapping to generate deep latent variables. The deep latent variables are smoothed to obtain the smoothed latent variable features; The smoothed latent variable features are used as the primary input, combined with the historical running data and the guiding vector as conditional information, and then input into the multi-conditional variational autoencoder model of the fused attention network for reconstruction and decoding, outputting the final remaining lifetime prediction result.

[0008] Furthermore, the multi-conditional variational autoencoder model of the fused attention network includes multiple composite encoders, a final encoder, an intermediate connection module, multiple composite decoders, and a final decoder connected in series. Each of the composite encoders includes a base encoder and an attention network module; The multiple composite encoders are connected in series with the final encoder to perform step-by-step feature extraction and correction using the joint input features as the first-level input, combined with the historical running data as conditional information and the guidance information, and the deep latent variables are output by the final encoder. The intermediate connection module is a locally weighted linear regression module, which is used to receive the deep latent variables output by the final encoder, perform regression smoothing based on local weighting, and output the smoothed latent variable features. Each of the composite decoders includes a basic decoder and an attention network module; The multiple composite decoders are connected in series with the final-stage decoder. They are used to perform step-by-step reconstruction decoding by taking the smoothed latent variable features as the first-stage input and combining the historical running data as conditional information and the guiding information. Finally, the final-stage decoder outputs the final remaining lifetime prediction result.

[0009] Furthermore, the workflow of the composite encoder is as follows: The input of the basic encoder receives a vector formed by concatenating the input feature matrix of the composite encoder with the historical running data as conditional information. The basic encoder uses a first fully connected layer with a ReLU activation function to perform a nonlinear transformation on the concatenated vector to obtain fused features; the fused features are processed in parallel using two fully connected layers to calculate the mean vector and variance vector of the latent variable distribution, respectively; based on the mean vector and variance vector, a reparameterization technique is used to sample and generate latent variables; The latent variables output by the basic encoder and the K guiding vectors are input to the attention network module. After feature correction by the attention network module, the corrected latent variable features are output as the output of the composite encoder. Wherein, when the composite encoder is the first composite encoder in the series, its input feature matrix is ​​the joint input feature; when the composite encoder is not the first composite encoder, its input feature matrix is ​​the corrected latent variable feature output by the previous composite encoder.

[0010] Furthermore, the workflow of the composite decoder is as follows: The input of the basic decoder receives a vector formed by concatenating the input feature matrix of the composite decoder with the historical running data as conditional information. The basic decoder uses a second fully connected layer to linearly map the concatenated vector, mapping the feature dimension to the data space; and uses a third fully connected layer configured with a Sigmoid activation function to perform a nonlinear transformation on the mapped features to generate decoded features. The decoded features output by the basic decoder and the K guiding vectors are input into the attention network module. After feature correction by the attention network module, the corrected decoded features are output as the output of the composite decoder. Wherein, when the composite decoder is the first composite decoder in the series, its input feature matrix is ​​the smoothed latent variable feature output by the intermediate connection module; when the composite decoder is not the first composite decoder, its input feature matrix is ​​the corrected decoding feature output by the previous composite decoder.

[0011] Furthermore, the attention network module is configured to process the input feature matrix. and stated by The guiding feature matrix consists of several guiding vectors. Perform the following operations: Using the first trainable weight matrix obtained through pre-training For the input feature matrix Perform a linear transformation to map its feature dimensions to the target feature dimensions. And a query vector is generated using the Sigmoid activation function. ; Using the second trainable weight matrix obtained through pre-training and the third trainable weight matrix obtained through pre-training For the guiding feature matrix respectively Perform a linear transformation, where Map the feature dimension to the target feature dimension The key matrices are generated by applying the Sigmoid activation function. Sum matrix ; The query vector Copy in the row direction Next, construct an initial query matrix. ,in The length of the input sequence; then a preset position encoding matrix is ​​used. With the initial query matrix Add them together to get the final query matrix. ,Right now ; According to the formula Calculate the middle score matrix ; then The attention weight matrix is ​​obtained by normalizing using the Softmax function. ; The attention weight matrix With the value matrix Perform matrix multiplication to obtain the weighted modulated feature matrix. ; The weighted modulated feature matrix The feature dimensions are processed through a fully connected layer and mapped to the input feature matrix. With the same feature dimensions, the processed feature matrix is ​​obtained; a residual connection operation is performed to connect the input feature matrix. Copy in the same direction. After performing dimension alignment, the result is added to the processed feature matrix to output the final processed feature matrix. ; Wherein, when the attention network module is applied to the composite encoder, the input feature matrix The latent variables output by the basic encoder When the attention network module is applied to the composite decoder, the input feature matrix Decoding features output by the basic decoder .

[0012] Furthermore, the workflow of the intermediate connection module is as follows: Using the deep latent variables output by the final encoder as query points, the distance between them and a set of reference points learned during model training is calculated, where the distance is calculated using the Euclidean distance formula:

[0013] in, This represents the deep hidden variable. Represents the first of the set of reference points A reference point, Represents the deep hidden variables With the Reference points The Euclidean distance between them; Based on the distance, a set of local weights is calculated using a kernel function, wherein the kernel function is a Gaussian kernel, and the calculation formula is as follows:

[0014] in, Indicates that for the first Local weights calculated from reference points Indicates the distance, Indicates bandwidth parameter; The deep latent variables are subjected to weighted linear regression using the local weights. The weighted least squares problem is solved to obtain the regression coefficients, and then the transformed features after regression fitting are output as the smoothed latent variable features. The smoothed latent variable features serve as the input feature matrix of the first composite decoder in the composite decoder cascade structure.

[0015] Furthermore, the step of constructing K guiding vectors based on the pure samples specifically includes: Complete charge-discharge cycles are extracted from historical operating data as query timing segments, which include synchronization sequences of three channels: terminal voltage, charge-discharge current, and battery casing temperature. Extract the corresponding segments of each cycle from the complete historical running data of the clean sample to form a time-series segment library of the clean sample; The multivariate dynamic time warping (DTW) algorithm is used to calculate the similarity between the query time series segment and each segment in the time series segment library, and the K clean sample cyclic segments that are most similar to the query time series segment are selected as the K nearest neighbors. Based on the optimal curved path obtained by the DTW algorithm, each nearest pure sample loop segment is aligned with the query time sequence segment point by point on the time axis. The terminal voltage, charging and discharging current and temperature values ​​corresponding to each alignment point are extracted to construct K guiding vectors.

[0016] In a second aspect, the present invention provides a lithium battery remaining life prediction system based on a multi-condition variational autoencoder, comprising: The data acquisition module is used to acquire historical operating data of the lithium battery under test and divide the historical operating data into contaminated samples and pure samples. The preliminary prediction module is used to input the contaminated sample into a pre-trained Transformer model for preliminary prediction to obtain a preliminary prediction sequence; The feature construction module is used to construct K guiding vectors based on the clean samples and to fuse the historical running data with the preliminary prediction sequence to construct joint input features; The latent variable generation module is used to take the joint input features as the first input, combine them with the historical running data as conditional information and the guiding vector, and input them into the multi-conditional variational autoencoder model with fused attention network for feature extraction and correction, and map to generate deep latent variables. The smoothing module is used to smooth the deep latent variables and obtain the smoothed latent variable features; The prediction output module is used to take the smoothed latent variable features as the first input, combine them with the historical running data as conditional information and the guiding vector, and input them into the multi-conditional variational autoencoder model of the fused attention network for reconstruction and decoding, and output the final remaining lifetime prediction result.

[0017] Thirdly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the lithium battery remaining life prediction method of a multi-condition variational autoencoder as described above.

[0018] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the lithium battery remaining life prediction method of a multi-condition variational autoencoder as described above.

[0019] Compared with the prior art, the present invention has the following beneficial effects: This invention discloses a method and system for predicting the remaining life of lithium batteries using a multi-conditional variational autoencoder. By embedding an attention network based on clean sample guidance vectors into the encoder and decoder, cross-level feature guidance enhancement is achieved. This enables the model to actively retrieve and utilize degenerate priors from a small number of clean samples during inference, fundamentally suppressing the progressive amplification of noise during temporal propagation. This solves the problems of traditional RNN models being easily disturbed by outliers and CNN models struggling to capture long-range dependencies. Simultaneously, deep latent variables are smoothed to suppress abnormal fluctuations. After reconstruction by a multi-level decoder and attention network, a high-precision remaining life prediction result is finally output. This invention achieves robust and accurate lithium battery life prediction in real industrial scenarios dominated by contaminated data, providing reliable technical support for battery health status assessment, preventative maintenance scheduling, and full lifecycle cost management in energy storage power stations, electric vehicles, and other fields. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the system of the present invention; Figure 3This is a schematic diagram of the structural framework of the multiconditional variational autoencoder model that integrates an attention network in an embodiment of the present invention. Figure 4 This is a schematic diagram of the composite encoder structure of the multi-conditional variational autoencoder model that integrates an attention network in an embodiment of the present invention. Figure 5 This is a schematic diagram of the composite decoder structure of the multi-conditional variational autoencoder model that integrates an attention network in an embodiment of the present invention. Figure 6 This is a schematic diagram of the intermediate connection part of the multiconditional variational autoencoder model that integrates an attention network in an embodiment of the present invention. Figure 7 This is a detailed structural diagram of the attention network module in an embodiment of the present invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application.

[0023] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0024] It should be understood that in the embodiments of this application, "at least one" means one or more, and "more than one" means two or more. "And / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. The character " / " generally indicates that the related objects before and after it are in an "or" relationship. "Contains A, B and / or C" means containing any one, two, or three of A, B, and C.

[0025] It should be understood that in the embodiments of this application, "B corresponding to A", "B corresponding to A", "A corresponds to B" or "B corresponds to A" means that B is associated with A, and B can be determined based on A. Determining B based on A does not mean that B is determined solely based on A; B can also be determined based on A and / or other information.

[0026] The lithium battery remaining life prediction method based on a multi-condition variational autoencoder provided by this invention can be executed by an electronic device, such as a terminal or a server. The terminal can be a smartphone, tablet, laptop, or other similar device. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. It is understood that this invention does not specifically limit the executing entity of this lithium battery remaining life prediction method based on a multi-condition variational autoencoder.

[0027] The technical solution of this application will be described in detail below through specific embodiments. It should be noted that the following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments described below are used to explain the technical solution of this application and are not intended to limit actual use.

[0028] First, a brief explanation of the relevant background knowledge and the purpose of this invention will be given.

[0029] With the rapid popularization of electric vehicles and the large-scale deployment of new electrochemical energy storage in the power grid, lithium batteries, as the core energy storage unit, have entered a stage of large-scale cluster operation. Their widespread application has also brought severe data governance challenges. In actual operating environments, due to differences in the accuracy of distributed battery pack sensors, and inherent defects in vehicle battery management systems, power grids, and building energy management systems during complex electromagnetic environments and long-distance data transmission, massive amounts of battery operating data inevitably contain contaminated data including outliers, missing data, and random noise. Directly using such data for battery health status assessment will lead to serious misjudgments in remaining lifespan predictions, potentially causing safety risks due to failure to promptly warn of battery failure, as well as economic losses due to premature battery replacement or imbalanced energy storage system scheduling. Although existing prediction models perform excellently on clean laboratory data, their prediction accuracy often deteriorates sharply when faced with real industrial data scenarios characterized by "most contaminated, a few clean" data. This technical bottleneck has become a key obstacle restricting the refined management and digital transformation of battery lifecycle operations.

[0030] The purpose of this invention is to overcome the aforementioned bottlenecks and provide a method for predicting the remaining life of lithium batteries using a multi-conditional variational autoencoder model that integrates an attention network. This invention is not limited to pursuing theoretical accuracy limits in ideal laboratory data, but rather aims to achieve and maintain high accuracy and robust predictive performance in noisy real-world industrial data environments. Specific technical objectives include: First, it significantly improves the model's robustness to data contamination, ensuring that even on challenging datasets with "most contaminated and a few clean" datasets, its prediction error metrics (such as root mean square error RMSE) remain stably maintained at an acceptable high level of accuracy, avoiding a catastrophic decline in model performance. This is achieved by using a Transformer model for initial predictions and then constructing a subsequent multi-conditional variational autoencoder model with an attention-fused network for feature correction.

[0031] Second, it enables efficient use of information from a few clean samples for guided learning. By designing an attention-based explicit guidance mechanism, the model can actively refer to and integrate guidance vectors constructed from clean samples during the inference process, thereby fundamentally improving its prediction accuracy and generalization ability in real noisy scenarios.

[0032] Third, a hybrid prediction framework is constructed that combines a multi-level encoding-decoding structure with local weighted linear regression correction capabilities. This framework can extract deep features from noisy preliminary predictions and smooth out abnormal fluctuations through local regression, providing a robust feature representation foundation for the final high-precision prediction.

[0033] Fourth, an adaptive information fusion and transmission strategy is provided. Through the collaborative work of a multi-level attention network and encoder-decoder, the transmission and fusion of features in the latent space are optimized, effectively reducing the overall error of the prediction results and further improving the reliability and practicality of the prediction system.

[0034] See Figure 1 This invention discloses a method for predicting the remaining life of a lithium battery using a multi-condition variational autoencoder, comprising the following steps: S1, acquire the historical operating data of the lithium battery to be tested, and divide the historical operating data into contaminated samples and pure samples; S2, The contaminated sample is input into a pre-trained Transformer model for preliminary prediction to obtain a preliminary prediction sequence; S3, construct K guiding vectors based on the pure samples, and fuse the historical running data with the preliminary prediction sequence to construct joint input features; S4, the joint input features are used as the first-level input, combined with the historical running data and the guiding vector as conditional information, and input into the multi-conditional variational autoencoder model of the fusion attention network for feature extraction and correction, and mapping to generate deep latent variables; S5, Smooth the deep latent variables to obtain the smoothed latent variable features; S6. The smoothed latent variable features are used as the primary input, combined with the historical running data and the guiding vector as conditional information, and input into the multi-conditional variational autoencoder model of the fusion attention network for reconstruction and decoding, and the final remaining lifetime prediction result is output.

[0035] To facilitate understanding of the embodiments of the present invention, each step is described in detail. Specifically: In one feasible embodiment of the present invention, step S1 specifically includes the following sub-steps: Step S11: Obtain and filter the experimental dataset. In one embodiment of the present invention, a publicly available lithium battery accelerated aging experimental dataset (such as the MIT-ST dataset) is used as the data source. Based on the principles of data integrity and performance consistency, a five-fold cross-validation scheme is employed to filter 35 battery cells from the dataset, forming the final experimental dataset. The five-fold cross-validation scheme randomly divides the dataset into five mutually exclusive subsets. In each round of cross-validation, the data from four subsets (a total of 28 battery cells) are used as the training set, and the data from the remaining subset (7 battery cells) is used as the test set.

[0036] Step S12: Construct a training dataset simulating real-world contamination. During each round of cross-validation training, to simulate the "most contaminated, a few clean" data scenario in real-world applications, a controllable noise injection strategy is adopted: a pre-defined, relatively small number of cells are randomly selected from the battery cells in the current training set. (like For each cell in the training set, all sample data are retained as clean samples. For the remaining cell cells in the training set, Gaussian white noise is injected into their time series data to simulate actual noise such as sensor drift, thereby generating contaminated samples.

[0037] Step S13 involves time-series data slicing and label generation. For all samples (including clean and contaminated samples), a sliding window method is used for overlapping sampling to transform the long sequence into multiple fixed-length sample segments. For each sample segment, the battery capacity degradation level at the end of the window is calculated and normalized to a remaining lifespan label, thus constructing a structured mapping sample from multivariate time-series input to remaining lifespan labels.

[0038] Step S14: The final heterogeneous training and testing sets are formed. All the processed samples described above are integrated to create a dataset for model training. This dataset includes both contaminated samples simulating real-world noise and some uncontaminated, clean samples to reflect the characteristics of a real industrial data environment. The testing set is constructed independently and does not participate in the noise injection process described above; it is used only for final evaluation.

[0039] Step S15: Obtain real-time prediction data for the lithium battery under test. During the actual application prediction phase of the model, the terminal voltage, charging / discharging current, and battery casing temperature of the lithium battery under test at various times within the sampling period are directly obtained through sensors or the battery management system, serving as historical operating data for prediction. Due to the complexity of actual operating conditions, this acquired data is by default considered as a contaminated sample with noise.

[0040] In one feasible embodiment of the present invention, step S2 specifically includes the following sub-steps: Step S21: Pre-training the Transformer model based on the training set. The Transformer model is pre-trained under supervision using the training set data containing contaminated samples constructed in Step S1. The Transformer model employs an encoder-decoder architecture based on a self-attention mechanism to effectively capture the global dependencies of the input sequences. The model's input consists of three time-series data variables—terminal voltage, charge / discharge current, and battery casing temperature—within a specific time window, and the output is the corresponding normalized remaining life prediction value. The core purpose of this pre-training is to enable the model to initially learn the mapping relationship from noisy data to remaining life, thereby obtaining an initial prediction model that reflects the basic trend of battery degradation, rather than pursuing high-precision prediction results.

[0041] During training, the prediction results are calculated using forward propagation, and the mean squared error (MSE) is used as the loss function to measure the difference between the predicted and the true values. Its mathematical expression is: ; in, This represents the actual remaining lifespan value. These are the model's predicted values. The sample size is denoted as . The gradient is calculated using the backpropagation algorithm, and the model parameters are updated using the Adam optimizer until the model converges.

[0042] Step S22: Perform preliminary predictions on the test set based on the pre-trained model. During the model evaluation and final prediction stage, the real-time operating data of the lithium battery under test (i.e., noisy contaminated samples) obtained in step S15 is input into the pre-trained Transformer model. The model performs forward computation on the test set and outputs the corresponding preliminary remaining lifetime prediction sequence. This noisy sequence This will serve as one of the core inputs to the subsequent multiconditional variational autoencoder model that integrates the attention network, providing it with an initial prediction target for further deep correction and optimization. This prediction sequence In subsequent steps, depending on the training and prediction phases, the data will be fused with real lifespan labels or historical operational data.

[0043] In one feasible embodiment of the present invention, step S3 specifically includes the following sub-steps: Step S31: Construct a guiding vector based on the Dynamic Time Warping (DTW) algorithm. For the current lithium battery sample to be processed, perform the following operations to construct... One guiding vector: From the historical operating data obtained from the current sample under test, the most recent complete charge-discharge cycle is extracted as the query time sequence segment. This segment contains the synchronization sequence of three channels: terminal voltage, charge-discharge current, and battery casing temperature. At the same time, from the complete historical operating data of each of the clean samples obtained in step S1, the corresponding segments of each cycle are extracted to form a time sequence segment library of clean samples.

[0044] In one specific embodiment of the present invention, a multivariate DTW algorithm is used to measure sequence similarity. The multivariate DTW distance between the queried time series segment and each segment in the time series segment library is calculated. This distance comprehensively measures the similarity in shape between two multivariate time series segments, overcoming the problem of nonlinear distortion of the time axis caused by small differences in charge and discharge rates. Based on this DTW distance, the segment most similar to the currently queried time series segment is selected. A clean sample cyclic fragment, as A close neighbor.

[0045] For each selected nearest-neighbor pure sample cyclic segment, the optimal curved path calculated using the DTW algorithm is used to align it point-by-point with the current query time series segment on the time axis. After alignment, a correspondence is established between each time point of the two segments.

[0046] Finally, based on this alignment relationship, the original values ​​(voltage values) of the three channels of each nearest-neighbor clean sample fragment at the key alignment point are extracted. Current value Temperature value ). This The group alignment point features are constructed as follows Guide vectors Each guiding vector records the specific operational state observed at the corresponding time in the pure cycle with the most similar morphology.

[0047] Step S32: Constructing joint input features and conditional information during model training. In this step, the model learns posterior features containing the true lifetime distribution. Therefore, the noisy preliminary prediction sequence obtained in step S2 is concatenated and fused with the true remaining lifetime labels in the training set to construct the joint input features for the training phase. Simultaneously, the guiding feature matrix constructed in step S31 is concatenated with the true remaining lifetime labels in the training set, and together they serve as the conditional information input into the model for this training phase.

[0048] Step S33: Construct joint input features and conditional information during the actual prediction stage of the model. During actual prediction, since true lifetime labels are unavailable, the model must rely on known physical quantities for inference. Therefore, the noisy preliminary prediction sequence is concatenated and fused with the historical operating data of the battery under test (i.e., test set features including terminal voltage, charge / discharge current, and temperature) to construct the joint input features for the prediction stage. Simultaneously, the guiding feature matrix is ​​concatenated with the historical operating data, and together they serve as the conditional information input into the model for this prediction stage.

[0049] In one feasible embodiment of the present invention, see [link to relevant documentation]. Figure 3 The multiconditional variational autoencoder model with fused attention network includes multiple composite encoders, a final encoder, an intermediate connection module, multiple composite decoders, and a final decoder connected in series; each composite encoder includes a basic encoder and an attention network module (e.g., ...). Figure 7 Each of the composite decoders includes a basic decoder and an attention network module; See Figure 4 Step S4 specifically includes the following sub-steps: Step S41: Nonlinear mapping and latent variable sampling of the basic encoder. Taking the first composite encoder as an example, its basic encoder input receives the joint input features constructed in the corresponding stage of step S3, and receives the corresponding conditional information. This basic encoder uses a first fully connected layer with a ReLU activation function to perform a nonlinear transformation on the concatenated vector to obtain fused features; then, two fully connected layers are used in parallel to process the fused features, calculating the mean vector and variance vector of the latent variable probability distribution, respectively. Based on these mean and variance vectors, a reparameterization technique is used to sample from the distribution to obtain the initial latent variables. This process maps a deterministic input to a probabilistic latent space representation, where the latent variable... The input feature matrix will be used as the input feature matrix. It is fed into the attention network of the same level.

[0050] Step S42: Dimensional alignment and feature modulation of the cross-attention network. In one embodiment of the invention, the cross-attention network employs a multi-dimensional temporal unfolding and feature modulation mechanism to efficiently guide noisy features with clean sample features. The attention network receives the aforementioned input feature matrix. and the guiding feature matrix constructed in step S3 (Depend on (Constituted by several guiding vectors).

[0051] First, using the first trainable weight matrix right Perform a linear transformation to map its feature dimensions to the target feature dimensions. And a single-row query vector is generated using the Sigmoid activation function. (The aforementioned) And the following mentioned , These are all parameters of the fully connected layers within the attention network, which are randomly initialized during model building and iteratively optimized to fixed values ​​through backpropagation during subsequent offline joint training phases. To achieve temporal alignment from single-step latent features to sequence features, this query vector... Copy in the row direction (of which) To set the sequence length, construct the initial query matrix. , to use a preset position encoding matrix With the initial query matrix Element-by-element addition, with the temporal sequence information injected, generates the final query matrix. .

[0052] Simultaneously, utilizing the second and third trainable weight matrices and For the guiding feature matrix respectively Perform linear transformations and Sigmoid activation to generate the bond matrix. Sum matrix ,in Map the feature dimension to the target feature dimension .

[0053] calculate and The inner product is used to obtain the intermediate score matrix. The calculation formula is as follows:

[0054] right The attention weight matrix is ​​obtained by normalization using the Softmax function. Finally, the attention weight matrix... With the value matrix Perform matrix multiplication to obtain the weighted modulated feature matrix. .

[0055] Step S43: Dimensional alignment and composite encoding output of residual connections. The weighted modulated feature matrix described above... A non-linear transformation is performed through a fully connected layer, mapping its feature dimensions to the original input feature matrix. With the same feature dimensions, the processed feature matrix is ​​obtained. During the residual join operation, a similar dimension augmentation operation is performed, transforming the original single-row input feature matrix... Copy in the row direction Next, matrix dimension alignment is performed, and then element-wise addition is performed with the processed feature matrix. The result of the addition is the output of the first composite encoder, which serves as the latent variable feature after deep correction of the clean sample features.

[0056] Step S44: Multi-level cascading and generation of deep latent variables in the final stage. When the composite encoder is not the first composite encoder, the input feature matrix of its basic encoder is replaced with the corrected latent variable features output by the previous composite encoder. After passing through a preset number of composite encoders for progressive abstraction and attention correction, the features enter the final encoder. In this embodiment, to achieve final feature convergence, the final encoder consists only of the basic encoder and no longer includes an attention network. It receives the output of the last composite encoder and performs latent variable distribution parameter calculation and reparameterization sampling again, finally outputting deep latent variables, thereby completing the dimensionality reduction transformation from noisy sequences to a clean and robust deep latent space.

[0057] In one feasible embodiment of the present invention, step S5 inputs the deep latent variables to the locally weighted linear regression module, which serves as an intermediate connection module, to perform regression smoothing correction based on local weighting, and outputs the smoothed latent variable features; see also Figure 6 Specifically, it includes the following sub-steps: Step S51: Calculate the distance between the query point and the reference point set. Use the deep latent variable output by the final encoder in step S4 as the query point. Within the locally weighted linear regression module, a set of reference points learned and fixed during the model training phase is maintained. ,in The number of reference points. Calculate the query points. With each reference point in this reference point set Euclidean distance between The calculation formula is:

[0058] Step S52: Calculate local weights based on the Gaussian kernel function. Based on the distance calculated above... Local weights are assigned to each reference point using a kernel function. In this embodiment, a Gaussian kernel function is used to calculate the local weights. This allows reference points in the feature space that are closer to the query point to receive a larger weight, and the calculation formula is as follows:

[0059] in, This represents a hyperparameter that controls the bandwidth of the kernel function (i.e., the range of local smoothing).

[0060] Step S53: Perform weighted linear regression and smoothing correction on the output. Utilize the calculated set of local weights... For query points A weighted linear regression is performed within the neighborhood of its local feature space. Specifically, the local regression coefficients are solved by minimizing the weighted mean square error, and the mapped value after regression fitting is calculated accordingly. This mapped value is the transformed feature after smoothing correction. After this step, high-frequency noise and abrupt changes in the deep latent variables are effectively filtered out. This smoothed latent variable feature is then used as the first input to the decoder for subsequent high-precision reconstruction.

[0061] In one feasible embodiment of the present invention, see [link to relevant documentation]. Figure 5 The decoder structure of the multiconditional variational autoencoder model has a symmetrical mapping relationship with the encoder structure. Step S6 specifically includes the following sub-steps: Step S61: Forward reconstruction calculation and final prediction at the decoder. The smoothed latent variable features output from step S5 are used as the first-level input to the first composite decoder at the decoder end, and the conditional information constructed in the corresponding stage in step S3 is concatenated (concatenation of guiding information and real labels in the training stage, and concatenation of guiding information and historical running data in the prediction stage). This input feature is sequentially upsampled and reconstructed through a composite decoder with the same number of layers as the encoder end. In each composite decoder, a guiding feature matrix of clean samples is also introduced using a cross-attention network. Feature modulation and enhancement are performed. After processing by a multi-stage composite decoder, the features enter a final-stage decoder containing only the basic decoder, ultimately mapping and outputting the predicted remaining lifetime sequence. In the actual prediction phase, this sequence This is the final high-precision remaining lifetime prediction result after depth correction of the clean sample.

[0062] Step S62: Construct the variational lower bound joint loss function. During the model training phase, to ensure the model can accurately reconstruct the lifetime sequence while maintaining the continuity and robustness of the latent variable space, this embodiment constructs a joint loss function based on the evidence lower bound theory of variational autoencoders. The joint loss function is composed of a weighted sum of the reconstruction loss and the KL divergence.

[0063] The predicted remaining lifetime sequence output by the computational model Compared with the actual remaining life label The mean square error between them is used as the reconstruction loss. :

[0064] Calculate the posterior probability distribution of each stage output at the encoder. With prior probability distribution KL divergence between Used to regularize the latent space and prevent the model from overfitting:

[0065] in, This represents the latent variables extracted by the model. This represents the noisy preliminary prediction sequence input in step S2. This represents the actual remaining lifetime label that is passed in as conditional information during the training phase. Let be the posterior distribution of the latent variables given the input sequence and the true labels. Let be the prior distribution of the latent variables given the input sequence.

[0066] Combining the two parts above, we construct the final end-to-end joint training loss function. :

[0067] in, is a weighted hyperparameter used to balance reconstruction accuracy and latent space regularization.

[0068] See Figure 2This invention discloses a lithium battery remaining life prediction system based on a multi-conditional variational autoencoder, comprising a data acquisition module, a preliminary prediction module, a feature construction module, a latent variable generation module, a smoothing module, and a prediction output module. The data acquisition module acquires historical operating data of the lithium battery under test, dividing the historical operating data into contaminated samples and clean samples. The preliminary prediction module inputs the contaminated samples into a pre-trained Transformer model for preliminary prediction, obtaining a preliminary prediction sequence. The feature construction module constructs K guiding vectors based on the clean samples and fuses the historical operating data and the preliminary prediction sequence to construct joint input features. The latent variable generation module uses the joint input features as the primary input, combined with the historical operating data and the guiding vectors as conditional information, and inputs them into a multi-conditional variational autoencoder model with an attention network for feature extraction and correction, mapping to generate deep latent variables. The smoothing module is used to smooth the deep latent variables and obtain the smoothed latent variable features; the prediction output module is used to take the smoothed latent variable features as the first input, combine them with the historical running data as conditional information and the guiding vector, and input them into the multi-conditional variational autoencoder model of the fused attention network for reconstruction and decoding, and output the final remaining lifetime prediction result.

[0069] The core innovation of this invention lies in constructing a cascaded prediction framework based on preliminary prediction and deep correction, specifically designed for industrial data scenarios characterized by "mostly contaminated, a few pure samples." First, a pre-trained Transformer model extracts degradation trends from noisy data to obtain a preliminary prediction sequence. Then, a novel multi-conditional variational autoencoder model incorporating attention networks is proposed. Crucially, the model deeply injects a dynamically constructed guiding vector from a small number of pure samples into each level of the cross-attention network, ensuring that noisy features are continuously guided and corrected by pure prior information throughout the process of dimensionality reduction and reconstruction. Furthermore, a local weighted regression module smooths deep latent variables, effectively suppressing high-frequency noise fluctuations. Finally, under the constraint of conditional information in the differentiation stage, the decoder progressively reconstructs high-precision remaining life prediction results. This framework, through end-to-end joint training, achieves high robustness and accuracy in prediction even under strong noise environments, providing effective technical support for reliable management of lithium battery health status and remaining life prediction.

[0070] In one embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions from the computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used in the operation of a lithium battery remaining life prediction method based on a multi-condition variational autoencoder.

[0071] This invention also provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the lithium battery remaining life prediction method with a multi-condition variational autoencoder in the above embodiments.

[0072] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0073] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0074] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0075] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting the remaining life of a lithium battery using a multi-condition variational autoencoder, characterized in that, Includes the following steps: Acquire historical operating data of the lithium battery to be tested, and divide the historical operating data into contaminated samples and clean samples; The contaminated sample is input into a pre-trained Transformer model for preliminary prediction to obtain a preliminary prediction sequence; K guiding vectors are constructed based on the pure samples, and the historical running data is fused with the preliminary prediction sequence to construct joint input features; The joint input features are used as the first input, combined with the historical running data and the guiding vector as conditional information, and input into the multi-conditional variational autoencoder model with fused attention network for feature extraction and correction, and mapping to generate deep latent variables. The deep latent variables are smoothed to obtain the smoothed latent variable features; The smoothed latent variable features are used as the primary input, combined with the historical running data and the guiding vector as conditional information, and then input into the multi-conditional variational autoencoder model of the fused attention network for reconstruction and decoding, outputting the final remaining lifetime prediction result.

2. The method for predicting the remaining life of a lithium battery using a multi-condition variational autoencoder according to claim 1, characterized in that, The multiconditional variational autoencoder model of the fusion attention network includes multiple composite encoders, a final encoder, an intermediate connection module, multiple composite decoders, and a final decoder connected in series. Each of the composite encoders includes a base encoder and an attention network module; The multiple composite encoders are connected in series with the final encoder to perform step-by-step feature extraction and correction using the joint input features as the first-level input, combined with the historical running data as conditional information and the guidance information, and the deep latent variables are output by the final encoder. The intermediate connection module is a locally weighted linear regression module, which is used to receive the deep latent variables output by the final encoder, perform regression smoothing based on local weighting, and output the smoothed latent variable features. Each of the composite decoders includes a basic decoder and an attention network module; The multiple composite decoders are connected in series with the final-stage decoder. They are used to perform step-by-step reconstruction decoding by taking the smoothed latent variable features as the first-stage input and combining the historical running data as conditional information and the guiding information. Finally, the final-stage decoder outputs the final remaining lifetime prediction result.

3. The method for predicting the remaining life of a lithium battery using a multi-condition variational autoencoder according to claim 2, characterized in that, The working process of the composite encoder is as follows: The input of the basic encoder receives a vector formed by concatenating the input feature matrix of the composite encoder with the historical running data as conditional information; The basic encoder uses a first fully connected layer with a ReLU activation function to perform a nonlinear transformation on the concatenated vector to obtain fused features; The fused features are processed in parallel using two fully connected layers to calculate the mean vector and variance vector of the latent variable distribution, respectively; based on the mean vector and variance vector, a reparameterization technique is used to sample and generate latent variables. The latent variables output by the basic encoder and the K guiding vectors are input to the attention network module. After feature correction by the attention network module, the corrected latent variable features are output as the output of the composite encoder. Wherein, when the composite encoder is the first composite encoder in the series, its input feature matrix is ​​the joint input feature; when the composite encoder is not the first composite encoder, its input feature matrix is ​​the corrected latent variable feature output by the previous composite encoder.

4. The method for predicting the remaining life of a lithium battery using a multi-condition variational autoencoder according to claim 2, characterized in that, The workflow of the composite decoder is as follows: The input of the basic decoder receives a vector formed by concatenating the input feature matrix of the composite decoder with the historical running data as conditional information. The basic decoder uses a second fully connected layer to perform a linear mapping on the concatenated vector, mapping the feature dimension to the data space; The mapped features are nonlinearly transformed using a third fully connected layer configured with a Sigmoid activation function to generate decoded features; The decoded features output by the basic decoder and the K guiding vectors are input into the attention network module. After feature correction by the attention network module, the corrected decoded features are output as the output of the composite decoder. Wherein, when the composite decoder is the first composite decoder in the series, its input feature matrix is ​​the smoothed latent variable features output by the intermediate connection module; When the composite decoder is not the first composite decoder, its input feature matrix is ​​the corrected decoded feature output by the previous composite decoder.

5. The method for predicting the remaining life of a lithium battery using a multi-condition variational autoencoder according to claim 2, characterized in that, The attention network module is configured to process the input feature matrix. and stated by The guiding feature matrix consists of several guiding vectors. Perform the following operations: Using the first trainable weight matrix obtained through pre-training For the input feature matrix Perform a linear transformation to map its feature dimensions to the target feature dimensions. And a query vector is generated using the Sigmoid activation function. ; Using the second trainable weight matrix obtained through pre-training and the third trainable weight matrix obtained through pre-training For the guiding feature matrix respectively Perform a linear transformation, where Map the feature dimension to the target feature dimension The key matrices are generated by applying the Sigmoid activation function. Sum matrix ; The query vector Copy in the row direction Next, construct an initial query matrix. ,in The length of the input sequence; Then a preset position encoding matrix is ​​used. With the initial query matrix Add them together to get the final query matrix. ,Right now ; According to the formula Calculate the middle score matrix ; then The attention weight matrix is ​​obtained by normalizing using the Softmax function. ; The attention weight matrix With the value matrix Perform matrix multiplication to obtain the weighted modulated feature matrix. ; The weighted modulated feature matrix The feature dimensions are processed through a fully connected layer and mapped to the input feature matrix. With the same feature dimensions, the processed feature matrix is ​​obtained; a residual connection operation is performed to connect the input feature matrix. Copy in the same direction. After performing dimension alignment, the result is added to the processed feature matrix to output the final processed feature matrix. ; Wherein, when the attention network module is applied to the composite encoder, the input feature matrix The latent variables output by the basic encoder When the attention network module is applied to the composite decoder, the input feature matrix Decoding features output by the basic decoder .

6. The method for predicting the remaining life of a lithium battery using a multi-condition variational autoencoder according to claim 1, characterized in that, The workflow of the intermediate connection module is as follows: Using the deep latent variables output by the final encoder as query points, the distance between them and a set of reference points learned during model training is calculated, where the distance is calculated using the Euclidean distance formula: in, This represents the deep hidden variable. Represents the first of the set of reference points A reference point, Represents the deep hidden variables With the Reference points The Euclidean distance between them; Based on the distance, a set of local weights is calculated using a kernel function, wherein the kernel function is a Gaussian kernel, and the calculation formula is as follows: in, Indicates that for the first Local weights calculated from reference points Indicates the distance, Indicates bandwidth parameter; The deep latent variables are subjected to weighted linear regression using the local weights. The weighted least squares problem is solved to obtain the regression coefficients, and then the transformed features after regression fitting are output as the smoothed latent variable features. The smoothed latent variable features serve as the input feature matrix of the first composite decoder in the composite decoder cascade structure.

7. The method for predicting the remaining life of a lithium battery using a multi-condition variational autoencoder according to claim 1, characterized in that, The step of constructing K guiding vectors based on the pure samples specifically includes: Complete charge-discharge cycles are extracted from historical operating data as query timing segments, which include synchronization sequences of three channels: terminal voltage, charge-discharge current, and battery casing temperature. Extract the corresponding segments of each cycle from the complete historical running data of the clean sample to form a time-series segment library of the clean sample; The multivariate dynamic time warping (DTW) algorithm is used to calculate the similarity between the query time series segment and each segment in the time series segment library, and the K clean sample cyclic segments that are most similar to the query time series segment are selected as the K nearest neighbors. Based on the optimal curved path obtained by the DTW algorithm, each nearest pure sample loop segment is aligned with the query time sequence segment point by point on the time axis. The terminal voltage, charging and discharging current and temperature values ​​corresponding to each alignment point are extracted to construct K guiding vectors.

8. A lithium battery remaining life prediction system based on a multi-condition variational autoencoder, characterized in that, include: The data acquisition module is used to acquire historical operating data of the lithium battery under test and divide the historical operating data into contaminated samples and pure samples. The preliminary prediction module is used to input the contaminated sample into a pre-trained Transformer model for preliminary prediction to obtain a preliminary prediction sequence; The feature construction module is used to construct K guiding vectors based on the clean samples and fuse the historical running data with the preliminary prediction sequence to construct joint input features; The latent variable generation module is used to take the joint input features as the first input, combine them with the historical running data as conditional information and the guiding vector, and input them into the multi-conditional variational autoencoder model with fused attention network for feature extraction and correction, and map to generate deep latent variables. The smoothing module is used to smooth the deep latent variables and obtain the smoothed latent variable features; The prediction output module is used to take the smoothed latent variable features as the first input, combine them with the historical running data as conditional information and the guiding vector, and input them into the multi-conditional variational autoencoder model of the fused attention network for reconstruction and decoding, and output the final remaining lifetime prediction result.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the lithium battery remaining life prediction method according to any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the lithium battery remaining life prediction method of a multi-condition variational autoencoder as described in any one of claims 1-7.

Images