A deep learning-based electric vehicle battery life prediction method

Abstract

The application discloses a kind of electric vehicle battery life prediction methods based on deep learning, it is related to electric vehicle power battery management and health state evaluation technical field, the application is by constructing attenuation knowledge graph and prior matching to target condition characteristics, make the adaptation starting point of target battery change from global model search to the directional retrieval associated with degradation mode, reduce the drift risk caused by cross-system mismatch;By sparse gating, only a small number of expert subnetworks are activated and an initial model is formed, the substructures involved in modeling are consistent with the target conditions, and the interference of irrelevant subnetworks on the output is reduced;In addition, by freezing the feature encoder and sharing the output head, and limiting the update range to the activated expert subnetworks, the input representation and output caliber remain stable, the parameters are changed in the local degradation mode branch, and the controllability of the adaptation process is enhanced.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle power battery management and health status assessment technology, and in particular to a method for predicting the lifespan of electric vehicle batteries based on deep learning. Background Technology

[0002] Currently, the health management of electric vehicle power batteries requires the estimation of battery state of health (SOH) and remaining service life (RUL) to support safety warnings and maintenance decisions. In recent years, with the improvement of the ability of vehicle sensors and battery management systems (BMS) to collect operational data (such as voltage, current, temperature, etc. during charging and discharging), data-driven methods based on deep learning have been gradually adopted for SOH / RUL prediction. Research and engineering explorations have emerged that focus on online prediction and use sparse fragment data for life inference.

[0003] At the same time, new energy vehicle products and battery technology routes continue to evolve, and some new battery systems (such as sodium-ion batteries and semi-solid / quasi-solid routes) have begun to enter the demonstration or mass production application stage. For newly launched models or new battery pack systems, there is often a lack of historical data covering the entire life cycle, which makes it difficult to form reliable predictions in a timely manner by relying solely on the target battery's own data to train deep models.

[0004] To address the aforementioned data scarcity issue, existing technologies have adopted transfer learning or domain adaptation approaches: after training the model on a source battery / source operating condition with sufficient historical data, the parameters are fine-tuned using a small amount of early data from the target battery, or feature alignment is used to reduce the data distribution difference between the source and target domains, in order to improve the prediction accuracy in the target scenario.

[0005] However, when the target is a battery with a chemical system, structural design, or operating conditions that differ significantly from the source domain, the degradation mechanism and data patterns may change significantly. The applicability of source domain knowledge to the target domain is difficult to guarantee. Existing schemes based on fine-tuning or distribution alignment are prone to problems such as decreased prediction accuracy or insufficient result stability, thereby affecting the timely assessment and management of the battery life of the new system. Summary of the Invention

[0006] In view of the aforementioned existing problems, the present invention is proposed.

[0007] This invention provides a deep learning-based method for predicting the lifespan of electric vehicle batteries, addressing the problems of scarce battery data for new systems, instability in migration fine-tuning and alignment, and poor generalization.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0009] In a first aspect, embodiments of the present invention provide a method for predicting the battery life of an electric vehicle based on deep learning, comprising:

[0010] Step S1: Construct a decay knowledge graph and a sparse modular neural network. The decay knowledge graph stores the association between conditional feature vectors and degradation representation vectors extracted from the full lifespan operation data of the source battery. The sparse modular neural network includes a feature encoder, a prior index layer, a sparse gating layer, multiple expert subnetworks, a shared output head, and a fine-tuning controller.

[0011] Step S2: Obtain the static parameters of the target battery and the data of the first N complete charge-discharge cycles; form a target condition feature vector from the static parameters and the statistics of the cycle data, and obtain a sparse gating vector by calculating the similarity between the prior index layer and the attenuation knowledge graph; the statistics of the cycle data are the operating stress statistics calculated from the data of the first N complete charge-discharge cycles.

[0012] Step S3: Activate at least one expert subnetwork according to the sparse gating vector to form an initial model;

[0013] Step S4: Under the control of the fine-tuning controller, only the parameters of the activated expert sub-network are updated to obtain the final model; the remaining lifespan of the target battery is output using the final model.

[0014] As a preferred embodiment of the deep learning-based electric vehicle battery life prediction method described in this invention, the construction of the degradation knowledge graph includes: dividing the full life operation data of the source battery into segments according to the degradation stage, extracting the conditional feature vector and degradation representation vector of each segment, and establishing a graph structure for storage with the conditional feature vector as the index and the degradation representation vector as the content.

[0015] As a preferred embodiment of the deep learning-based electric vehicle battery life prediction method of the present invention, the conditional feature vector includes at least one or more of the following: average charge / discharge rate, state of charge range, temperature statistics, and resting time statistics.

[0016] As a preferred embodiment of the deep learning-based electric vehicle battery life prediction method of the present invention, the similarity is obtained by measuring the correlation between the target conditional feature vector and each index vector in the graph, and the K index vectors with the highest correlation measurement scores are selected according to a predetermined number K, or the correlation measurement scores are filtered according to a predetermined threshold to generate sparse gating vectors; where K is a predetermined positive integer.

[0017] As a preferred embodiment of the deep learning-based electric vehicle battery life prediction method of the present invention, when the threshold screening fails to select an expert subnetwork, the expert subnetwork corresponding to the index vector with the highest similarity is selected as the activated expert subnetwork.

[0018] As a preferred embodiment of the deep learning-based electric vehicle battery life prediction method of the present invention, wherein: the multiple expert sub-networks share the feature encoder and the shared output head, and each has an independent intermediate hidden layer; the sparse gating vector is used to control the participation weight or on / off state of the output of each expert sub-network.

[0019] As a preferred embodiment of the deep learning-based electric vehicle battery life prediction method described in this invention, the fine-tuning controller freezes the parameters of the feature encoder and the shared output head, only iterates and updates the parameters of the activated expert sub-network for no more than T rounds, and stops early based on the performance of the validation set divided by the first N loop data.

[0020] As a preferred embodiment of the deep learning-based electric vehicle battery life prediction method described in this invention, the first N complete charge-discharge cycle data are the N complete cycles obtained first in time sequence after the target battery is put into use, and N is an integer from 3 to 50.

[0021] The output of the remaining service life is either the remaining number of cycles or the remaining mileage, and the corresponding uncertainty metric can be output simultaneously.

[0022] In a second aspect, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements any step of the deep learning-based electric vehicle battery life prediction method described in the first aspect of the present invention.

[0023] Thirdly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the deep learning-based electric vehicle battery life prediction method described in the first aspect of the present invention.

[0024] Through the above technical solution, the present invention can achieve at least the following beneficial effects:

[0025] To address the issues of weak source domain knowledge and lack of guidance in blind migration for new vehicle models or new chemical systems, this paper constructs a decay knowledge graph and performs prior matching on target condition features. This transforms the starting point for adapting the target battery from a global model search to a targeted retrieval associated with the degradation mode, reducing the drift risk caused by cross-system mismatch.

[0026] To address the problem of unconstrained overall model parameter updates and large prediction fluctuations caused by large differences in the distribution between the source and target domains, sparse gating is used to activate only a small number of expert subnetworks to form an initial model. This ensures that the substructures involved in modeling are consistent with the target conditions and reduces the interference of irrelevant subnetworks on the output.

[0027] To address the issues of catastrophic forgetting and unstable generalization caused by fine-tuning early data with small samples, this paper proposes freezing the feature encoder and the shared output head and limiting the update range to within the activated expert subnetwork. This keeps the input representation and output caliber stable, with parameter changes concentrated in the local degradation mode branch, thus enhancing the controllability of the adaptation process.

[0028] To address the issue of potential null activations leading to process interruptions during threshold screening, the expert subnetwork with the highest similarity to the candidate sequence is activated back, ensuring that gating selection remains closed-loop even under extreme conditions and avoiding unpredictable interruptions.

[0029] To address the issues of inconsistent quality, missing data, and outliers in early-stage cyclic data that introduce statistical bias, this paper addresses these problems by implementing complete cycle determination, invalid cycle removal, missing data completion, and outlier handling. This ensures consistent input statistics and improves the stability of map retrieval and gating selection.

[0030] To address the challenges of expressing the reliability of new system predictions and their application in health management decisions, this paper addresses the issue of outputting uncertainty metrics during the inference phase. This allows lifespan prediction results to be characterized by stability, facilitating risk grading and threshold triggering in management strategies. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0032] Figure 1 This is a flowchart of a deep learning-based method for predicting the lifespan of electric vehicle batteries, as shown in the embodiment. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0034] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0035] Example 1:

[0036] like Figure 1 As shown, this application proposes a deep learning-based method for predicting the lifespan of electric vehicle batteries, including the following steps:

[0037] Step S1: Construct a decay knowledge graph and a sparse modular neural network. The decay knowledge graph stores the association between conditional feature vectors and degradation representation vectors extracted from the full lifespan operation data of the source battery. The sparse modular neural network includes a feature encoder, a prior index layer, a sparse gating layer, multiple expert subnetworks, a shared output head, and a fine-tuning controller.

[0038] A sparse modular neural network is trained under supervised conditions on source battery data. The training data consists of the full lifespan operation data of the source batteries. The input is the cyclic operation data and the corresponding conditional feature vectors. The output supervision signal is the remaining lifespan label value corresponding to each input sample. The remaining lifespan label value is generated based on the health status label and the lifespan termination criterion. The lifespan termination criterion is determined by the health status label reaching a predetermined termination state. During training, a regression supervision objective is used to update the network parameters, and regression error and stability indicators are monitored on validation data to determine the training termination time.

[0039] The degradation knowledge graph consists of index nodes, content nodes, and connecting edges. Each index node records a conditional feature vector and its source identifier, which includes the battery chemistry system identifier, vehicle model identifier, and operating condition identifier. Each content node records the degradation characterization vector and health status label corresponding to the index node. Connecting edges record the one-to-one association between index nodes and content nodes, and also record the similarity adjacency relationships between index nodes. Similarity adjacency relationships are established by calculating the similarity of the conditional feature vectors of the index nodes and are used for candidate expansion during retrieval.

[0040] Step S2: Obtain the static parameters of the target battery and the data of the first N complete charge-discharge cycles; form the target condition feature vector from the static parameters and the cycle data statistics, and obtain the sparse gating vector by calculating the similarity between the prior index layer and the attenuation knowledge graph; the cycle data statistics are the operating stress statistics calculated from the data of the first N complete charge-discharge cycles.

[0041] Static parameters are discretized, mapped, and normalized to form static parameter features. Chemical system identifiers are mapped to fixed-length embedding vectors via table lookup; vehicle model identifiers are mapped to fixed-length embedding vectors via table lookup; rated capacity and charging rate limits are mapped to continuous features using the same dimension normalization method. Cyclic data statistics are obtained from the aforementioned calculation process of operational stress statistics features and concatenated with static parameter features to form the target condition feature vector.

[0042] Step S3: Activate at least one expert subnetwork according to the sparse gating vector to form the initial model;

[0043] Step S4: Under the control of the fine-tuning controller, only the parameters of the activated expert sub-network are updated to obtain the final model; the target battery remaining lifespan is output using the final model.

[0044] In this embodiment, constructing the degradation knowledge graph includes: dividing the full lifespan operation data of the source battery into segments according to the degradation stage, extracting the conditional feature vector and degradation characterization vector of each segment, and establishing a graph structure for storage with the conditional feature vector as the index and the degradation characterization vector as the content;

[0045] Segmentation is based on the changing patterns of health status labels. Health status labels are jointly determined by capacity-related indicators and internal resistance-related indicators. When the capacity-related indicator shows a continuous downward trend with a sudden change in the rate of decline, the corresponding position is determined as the stage boundary. When the internal resistance-related indicator shows a continuous upward trend with a sudden change in the rate of increase, the corresponding position is determined as the stage boundary. When the stage boundaries of the capacity-related indicator and the internal resistance-related indicator are inconsistent, the earlier-appearing boundary is used as the stage boundary, and adjacent segments are merged and verified. The merging and verification is based on the monotonicity of the degradation representation vector within the segment.

[0046] The conditional feature vector is formed by concatenating static parameter features and operational stress statistical features. Static parameter features include chemical system identifiers, rated capacity identifiers, and upper limit of charge rate identifiers; operational stress statistical features are calculated from the charge / discharge current, terminal voltage, temperature, and state of charge sequence within each segment. The degradation characterization vector is composed of capacity change characterization, internal resistance change characterization, and voltage curve morphology characterization. Capacity change characterization includes the difference and rate of change of capacity-related indicators at the beginning and end of each segment; internal resistance change characterization includes the difference and rate of change of internal resistance-related indicators at the beginning and end of each segment; and voltage curve morphology characterization includes the voltage plateau position and slope characteristics of representative cycles within each segment.

[0047] In this embodiment, the conditional feature vector includes at least one or more of the following: average charge / discharge rate, state of charge range, temperature statistics, and resting time statistics.

[0048] The average charge / discharge rate is obtained by normalizing the current sequence within the cycle and the rated capacity identifier. Within the cycle, the charging and discharging phases are statistically analyzed separately and then combined into segmented statistical values. The state of charge (SOC) interval is determined by the maximum and minimum values ​​of the SOC sequence within the cycle, with separate statistics for the charging and discharging intervals. Temperature statistics are calculated from the temperature sequence within the cycle, including the mean, extreme values, and fluctuation range. Resting time statistics are obtained by accumulating consecutive time slices where the absolute value of the current sequence falls under the resting condition. The resting condition is determined by the noise level of the current sequence and remains consistent within the same battery data set.

[0049] In this embodiment, the similarity is obtained by measuring the correlation between the target conditional feature vector and each index vector in the graph, and the K index vectors with the highest correlation measurement scores are selected according to a predetermined number K, or the correlation measurement scores are filtered according to a predetermined threshold to generate sparse gating vectors; where K is a predetermined positive integer.

[0050] The target conditional feature vector and the conditional feature vectors of each index node are mapped to a unified representation vector by the same feature encoder and then normalized. The relevance measure is obtained by a weighted combination of one or both of dot product relevance and distance relevance, where dot product relevance is calculated after normalization, and distance relevance is calculated using a distance metric within the unified representation space. The relevance measure results of all index nodes are sorted to form a candidate sequence.

[0051] The sparse gating vector is generated from the candidate sequence. When selecting a predetermined number of nodes, the top-ranking index nodes in the candidate sequence are selected, and their corresponding gating positions in the expert subnetwork are activated; unselected gating positions are deactivated. When selecting a threshold, the set of index nodes whose relevance measurement results meet predetermined filtering criteria are selected, and their corresponding gating positions in the expert subnetwork are activated; gating positions that do not meet the filtering criteria are deactivated. The gating vector simultaneously generates a weight vector, which is obtained by normalizing the relevance measurement results of the activated index nodes and is used for weighting during expert fusion.

[0052] The output of the prior index layer is used to transform the correlation measurement results between the target conditional feature vector and each index vector in the decaying knowledge graph into sparse gating vectors, so as to control the activation of the expert subnetwork.

[0053] In one implementation, the process can be organized into links based on alignment, measurement, normalization, and sparsification; specifically as follows:

[0054] When the target conditional feature vector is denoted as And in the atlas the first The index vectors are denoted as To reduce the impact of differences in the units of different statistical measures on similarity, the prior index layer performs consistent linear alignment on both:

[0055] ,

[0056] in, This represents the aligned target conditional embedding vector; Indicates the aligned first 1 indexed embedding vector; This represents the projection matrix of the prior index layer; This represents the bias vector of the prior index layer; This represents the index vector number, which corresponds one-to-one with the expert subnetwork number.

[0057] Cosine similarity can be used to express directional consistency:

[0058] ,

[0059] in, Indicates the target and the first Similarity scores for each index; Indicates vector transpose; This represents the L2 norm.

[0060] Dot product attention and the introduction of a temperature coefficient can also be used to adjust the concentration of the score distribution.

[0061] ,

[0062] in, Indicates the dot product correlation score; This represents the normalized attention weights; Indicates the temperature coefficient; This indicates the total number of graph index vectors, which is equal to the total number of expert subnetworks. Indicates the summation number.

[0063] When there is a significant correlation between conditional statistics, Mahalanobis distance can be used to suppress artificially high similarity in the covariance direction and map the distance to similarity:

[0064] ,

[0065] in, Indicates the target and the first Squared Mahalanobis distance of each index; Represents the covariance matrix of the index embedding; This represents the inverse of the covariance matrix. It can be estimated by the embedding of the spectral index:

[0066] ,

[0067] in, Indicates all The mean vector; Indicates the diagonal perturbation coefficient; Represents the identity matrix.

[0068] After obtaining a set of scores or weight After that, sparse gated vectors It can be generated using Top-K hard gating.

[0069] by For example, let's first take the set: ,

[0070] in, This represents the set of selected index numbers; This indicates selecting the one with the largest weight. One number; This indicates the number of Top-K values.

[0071] Reconstruct the sparse gating and normalize it:

[0072] ,

[0073] in, Represents unnormalized sparse gated components; This represents the normalized gated component; This represents a sparse gated vector.

[0074] When filtering by threshold, the selected set can be defined as:

[0075] ,

[0076] in, This represents the set of index numbers that satisfy the threshold. This indicates the gating threshold.

[0077] If it appears In this case, the expert subnetwork corresponding to the largest weight can be activated using the weighted backoff logic:

[0078] ,

[0079] in, This represents the index number with the highest weight. This represents the index corresponding to the maximum value. The above formula is then used subsequently. and The construction method yields sparse gated vectors.

[0080] When it is necessary to maintain differentiability of sparsity during the training phase in order to jointly optimize the prior index layer, one can... The generation is replaced by a sparse normalization operator, such as using Sparsemax to obtain naturally sparse weights:

[0081] ,

[0082] in, This represents the sparsely normalized weight vector; This represents the sparse maximization mapping operator; and Take the same vector as the sparse gating vector.

[0083] Specifically, in the above implementation, a computable correspondence is established between the conditional information of the target battery and the existing index information in the graph, so that the gating result is driven by data relevance rather than fixed allocation. A priori index layer is used to map different statistics to the same embedding space, reducing bias caused by differences in dimensions and scales. Similarity measures can be selected from directional consistency, dot product correlation, and covariance correlation, providing interpretable scoring criteria for matching under different operating conditions. After score normalization, a weight distribution is formed, and then sparse gating is generated through Top-K or threshold screening, ensuring that only a few expert subnetworks participate in subsequent modeling, thereby reducing interference from irrelevant experts. In the case of an empty set in threshold screening, maximum weight backoff is used to ensure that at least one expert is activated, allowing the process to remain closed-loop even under extreme conditions. Temperature adjustment is used to control the degree of weight concentration, facilitating an adjustable trade-off between the number of experts and the dispersion of gating.

[0084] In this embodiment, when the threshold screening fails to select an expert subnetwork, the expert subnetwork corresponding to the index vector with the highest similarity is selected as the activated expert subnetwork.

[0085] When the activation set generated by the threshold screening method is empty, the expert subnetwork corresponding to the first index node of the candidate sequence is used as the only activated expert subnetwork, and its gating weight is reset to the full weight state; when the candidate sequence contains multiple index nodes with the same correlation measurement results, the unique index node is selected according to the priority matching rule of chemical system identifier in the source identifier, and when the chemical system identifiers are consistent, the unique index node is selected according to the priority matching rule of vehicle model identifier.

[0086] In this embodiment, multiple expert subnetworks share a feature encoder and a shared output head, and each has an independent intermediate hidden layer; sparse gating vectors are used to control the participation weights or on / off states of the outputs of each expert subnetwork.

[0087] Each expert subnetwork receives a unified representation vector from the feature encoder as input and outputs an expert representation vector. For expert subnetworks with sparse gating vectors in an inactive state, their output expert representation vector is set to zero; for expert subnetworks in an active state, their output expert representation vector is multiplied by the corresponding gating weight to obtain a weighted expert representation vector. The shared output head receives the sum of all weighted expert representation vectors as input and outputs the remaining lifetime prediction.

[0088] In this embodiment, the fine-tuning controller freezes the parameters of the feature encoder and the shared output head, only iterates and updates the parameters of the activated expert subnetwork for no more than T rounds, and stops early based on the performance of the validation set divided by the first N loop data.

[0089] The frozen parameter set contains all learnable parameters of the feature encoder and all learnable parameters of the shared output head. The activated expert subnetwork parameter set contains all learnable parameters of its independent intermediate hidden layers. Fine-tuning iterations are performed using the training subset obtained from the initial cyclic data partition as input, and the regression error is calculated on the validation subset after each iteration. When the regression error of the validation subset no longer decreases in several consecutive iterations, the iteration update stops and the final prediction model is output. The fine-tuning objective includes both a lifetime regression error term and a parameter offset penalty term, which is determined by the change in the activated expert subnetwork parameters relative to the initial model parameters.

[0090] In one implementation, updating only the parameters of the activated expert subnetwork can be achieved by a fine-tuning controller that limits the set of trainable parameters according to the gating result, and the update target is bound to the early stopping condition. The validation set is obtained by partitioning a complete cyclic dataset; as follows:

[0091] By sparse gated vectors The set of activated expert subnetwork IDs corresponding to the zero components in China and Africa can be denoted as:

[0092] ,

[0093] in, This represents the set of IDs for the activated expert subnetworks; Indicates the expert subnetwork number; Represents sparse gated vectors The Each component.

[0094] The fine-tuning controller moves the target battery forward. A complete cyclic data set is constructed into a sample set. The training set and validation set are divided into two sets in chronological order:

[0095] ,

[0096] in, Indicates from the previous The target sample set formed by a complete cyclic data set; Represents the training set; Represents the validation set; This represents the empty set.

[0097] In terms of parameter update range, the fine-tuning controller will adjust the feature encoder parameters. With shared output header parameters Set to frozen, and freeze the parameters of inactive expert subnetworks as well, allowing updates only to the set. Internal expert subnetwork parameters During forward computation, for each sample... The gated weighted expert output can be written as:

[0098] ,

[0099] in, Indicates the target battery sample Input features; This represents the hidden representation obtained by the feature encoder; This represents the forward mapping of the feature encoder; This represents the gated weighted expert fusion representation; Indicates the first Forward mapping of an expert subnetwork; Indicates sample The predicted remaining useful life; Indicates a forward mapping of the shared output header; Indicates the first Expert subnetwork parameters.

[0100] The update objective can employ a remaining useful life regression loss with added parameter variation constraints to suppress bias towards source domain knowledge. Taking mean squared error as an example:

[0101] ,

[0102] in, This indicates the regression loss of remaining useful life; Indicates the number of samples in the training set; Indicates sample The corresponding reference remaining service life value.

[0103] For regularization terms introduced by updating only the parameters of the activated expert subnetwork, the L2 norm constraint of the expert parameters relative to the initial values ​​can be used:

[0104] ,

[0105] in, This represents the constraint term for changes in expert parameters; Indicates the start of fine-tuning. Initial values ​​for the parameters of each expert subnetwork.

[0106] When further suppression of catastrophic forgetting is needed, source domain importance weights can be incorporated into the retention terms:

[0107] ,

[0108] in, Represents the parameter retention terms with importance weights; Indicates the first The parameter importance weight matrix obtained by the expert subnetwork during the source domain training phase.

[0109] The overall goal for fine-tuning can be written as:

[0110] ,

[0111] in, This indicates a fine-tuning and optimization objective; This represents the weighting coefficient of the constraint term for parameter variation; This represents the weight coefficient of the parameter retention term.

[0112] During iterative updates, the fine-tuning controller only adjusts... The expert parameters are used for gradient descent, while the remaining parameters are kept frozen:

[0113] ,

[0114] in, Indicates the first During the first iteration Parameters of an expert subnetwork; This indicates fine-tuning the learning rate; Indicates about The gradient operator; Indicates the first The fine-tuning target value corresponding to each iteration; This indicates the maximum number of iterations.

[0115] Early stopping based on validation set performance can be achieved using a sliding window to smooth the validation loss, and a waiting round mechanism can be introduced. The validation loss can be calculated using the same approach as the training loss.

[0116] ,

[0117] in, Indicates the first The verification loss of each iteration; Indicates the number of samples in the validation set; Indicates the first Round of iterations to validate samples The predicted value; This represents the validation loss after smoothing the sliding window. Indicates the length of the sliding window; This indicates the summation number within the window.

[0118] Under the waiting rounds mechanism, an effective descent threshold and a waiting count can be defined, and iteration stops when the count reaches the upper limit:

[0119] ,

[0120] in, Indicates the first The wait count corresponding to each round of iteration; Indicates as of the date The minimum smooth verification loss for round iterations; Indicates the effective descent threshold; This indicates the operation of taking the smaller value. The fine-tuning controller, when satisfying... or The update will end at that time. This indicates the maximum number of allowed waiting rounds.

[0121] When the remaining lifetime output also provides an uncertainty metric, the convergence of the validation set uncertainty can be used as an additional stopping condition:

[0122] ,

[0123] in, Indicates the first The average uncertainty on the round-iteration validation set; Indicates the first Round of iterations to validate samples The uncertainty measure of the output. The fine-tuning controller can be used... Continuous satisfaction The iteration stops at round-robin time, where, Indicates the threshold for uncertainty changes; This indicates the number of rounds required for convergence under uncertainty.

[0124] In practice, parameter settings can be organized around small sample sizes, fewer rounds, and low learning rates: Take 5 to 20; Pick to ; Take 2 to 4; Take 2 to 6; Pick to When adopting season Pick to The frozen layer list includes the feature encoder, the shared output head, and the inactive expert subnetworks; the prior index layer and the sparse gating layer remain unchanged during this stage, ensuring that the gating selection and fine-tuning update action boundaries are consistent.

[0125] Specifically, the fine-tuning process in the above implementation restricts the update range to a small number of gated expert subnetworks, ensuring that the small sample data of the target battery only performs local corrections on degradation modes similar to its operating conditions. After freezing the feature encoder and shared output head, the input representation and output caliber remain stable, and the update amount is concentrated in the intermediate hidden layers of the experts, reducing fluctuations caused by parameter linkage. The regression loss provides a fitting direction towards the remaining lifetime output, and parameter change constraints pull the update amplitude back to near the fine-tuning starting point, thus incorporating both adapting to the target battery and preserving source domain knowledge into the optimization objective. The validation set is obtained by dividing the first N complete cycle data by time, and the validation loss is used to describe the error changes on the non-training subset during the iteration process. The sliding window smooths the single-round fluctuations into a trend quantity, and the waiting mechanism transforms the state of no improvement into a stopping signal, enabling the number of iteration rounds to adaptively converge within the upper limit of T. If the output also gives uncertainty, the change in the average uncertainty can be used as an additional stopping criterion to identify the stable interval of the model on the target data.

[0126] Example 2:

[0127] Based on Example 1, the first N complete charge-discharge cycle data are the N first complete cycles obtained in chronological order after the target battery is put into use, where N is an integer from 3 to 50. The determination of a complete cycle is based on the coverage range of the state of charge sequence and the change in current direction. Cycles that meet the predetermined minimum state of charge span and include both charging and discharging phases are determined to be complete cycles; cycles that do not meet the predetermined minimum state of charge span are determined to be incomplete cycles and are removed from the initial cycle data. Cycles with missing sensing segments and a missing duration exceeding a predetermined upper limit are determined to be invalid cycles and removed; cycles with a missing duration not exceeding the predetermined upper limit are interpolated and completed according to the time axis. Sampling points with abnormal peaks and peak amplitudes exceeding the predetermined abnormality determination conditions within a cycle are determined to be abnormal points and removed, and then smoothly completed according to the time axis.

[0128] The remaining service life output can be either the remaining number of cycles or the remaining driving range, and can also output the corresponding uncertainty measure at the same time;

[0129] Uncertainty metrics are generated during the inference phase. During inference, the random deactivation mechanism in the network is enabled, and multiple forward inferences are performed on the same input sample to obtain a set of remaining lifetime predictions. The uncertainty metric is calculated from the dispersion of these predictions and is used as the output along with the prediction mean. In another implementation, multiple sparse modular neural networks with different initializations but identical structures are used to form an ensemble model. This model outputs multiple sets of predictions for the same input sample, and the uncertainty metric is generated from the dispersion of these predictions.

[0130] This embodiment also provides a computer device, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the deep learning-based electric vehicle battery life prediction method proposed in the above embodiment.

[0131] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0132] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a deep learning-based method for predicting the battery life of an electric vehicle as proposed in the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0134] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for predicting the lifespan of electric vehicle batteries based on deep learning, characterized in that, Includes the following steps: Step S1: Construct a decay knowledge graph and a sparse modular neural network. The decay knowledge graph stores the association between conditional feature vectors and degradation representation vectors extracted from the full lifespan operation data of the source battery. The sparse modular neural network includes a feature encoder, a prior index layer, a sparse gating layer, multiple expert subnetworks, a shared output head, and a fine-tuning controller. Step S2: Obtain the static parameters of the target battery and the data of the first N complete charge-discharge cycles; form a target condition feature vector from the static parameters and the statistics of the cycle data, and obtain a sparse gating vector by calculating the similarity between the prior index layer and the attenuation knowledge graph; the statistics of the cycle data are the operating stress statistics calculated from the data of the first N complete charge-discharge cycles. Step S3: Activate at least one expert subnetwork according to the sparse gating vector to form an initial model; Step S4: Under the control of the fine-tuning controller, only the parameters of the activated expert sub-network are updated to obtain the final model; The final model is used to output the remaining lifespan of the target battery. The construction of the degradation knowledge graph includes: dividing the full life operation data of the source battery into segments according to the degradation stage, extracting the conditional feature vector and degradation characterization vector of each segment, and establishing a graph structure for storage with the conditional feature vector as the index and the degradation characterization vector as the content; The segmentation is based on the changing patterns of health status labels; the health status labels are jointly determined by capacity-related indicators and internal resistance-related indicators; when the capacity-related indicators show a continuous downward trend and the rate of decline changes abruptly, the corresponding position is determined as the stage boundary; when the internal resistance-related indicators show a continuous upward trend and the rate of increase changes abruptly, the corresponding position is determined as the stage boundary; when the stage boundaries of the capacity-related indicators and the internal resistance-related indicators are inconsistent, the boundary that appears first is taken as the stage boundary and adjacent segments are merged and verified, and the merging and verification is based on the monotonicity of the degradation representation vector within the segment; The conditional feature vector is formed by concatenating static parameter features and operational stress statistical features. The static parameter features include chemical system identifier, rated capacity identifier, and upper limit of charging rate identifier. The operational stress statistical features are calculated from the charging and discharging current, terminal voltage, temperature, and state of charge sequence within the segment. The degradation characterization vector is composed of capacity change characterization, internal resistance change characterization, and voltage curve morphology characterization. Among them, the capacity change characterization includes the difference and rate of change of capacity-related indicators at the beginning and end of the segment, the internal resistance change characterization includes the difference and rate of change of internal resistance-related indicators at the beginning and end of the segment, and the voltage curve morphology characterization includes the voltage plateau position and slope characteristics of representative cycles within the segment.

2. The method for predicting the lifespan of electric vehicle batteries based on deep learning according to claim 1, characterized in that, The conditional feature vector includes one or more of the following: average charge / discharge rate, state of charge range, temperature statistics, and resting time statistics.

3. The method for predicting the lifespan of electric vehicle batteries based on deep learning according to claim 1, characterized in that, The similarity is obtained by measuring the correlation between the target conditional feature vector and each index vector in the graph, and the K index vectors with the highest correlation measurement scores are selected according to a predetermined number K, or the correlation measurement scores are filtered according to a predetermined threshold to generate sparse gating vectors; where K is a predetermined positive integer.

4. The method for predicting the lifespan of electric vehicle batteries based on deep learning according to claim 3, characterized in that, When the threshold fails to filter out expert subnetworks, the expert subnetwork corresponding to the index vector with the highest similarity is selected as the activated expert subnetwork.

5. The method for predicting the lifespan of electric vehicle batteries based on deep learning according to claim 1, characterized in that, The multiple expert subnetworks share the feature encoder and the shared output head, and each has an independent intermediate hidden layer; the sparse gating vector is used to control the participation weight or on / off state of the output of each expert subnetwork.

6. The method for predicting the lifespan of electric vehicle batteries based on deep learning according to claim 1, characterized in that, The fine-tuning controller freezes the parameters of the feature encoder and the shared output head, only iteratively updates the parameters of the activated expert subnetwork for no more than T rounds, and stops early based on the performance of the validation set divided by the first N cycles of data.

7. The method for predicting the lifespan of electric vehicle batteries based on deep learning according to claim 1, characterized in that, The first N complete charge-discharge cycle data are the N complete cycles obtained first in chronological order after the target battery is put into use, and N is an integer from 3 to 50; The output of the remaining service life is either the remaining number of cycles or the remaining mileage, and the corresponding uncertainty metric can be output simultaneously.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the deep learning-based electric vehicle battery life prediction method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the deep learning-based electric vehicle battery life prediction method according to any one of claims 1 to 7.

Images