A multi-step probability prediction method for battery state of health and related devices

Abstract

The embodiment of the application provides a battery health state multi-step probability prediction method and related equipment, and belongs to the technical field of lithium ion battery management. The method comprises the following steps: obtaining short-time input sequence charging feature data of a target battery at a current time, and deducing charging feature probability distribution of at least 150 future cycles through a feature probability deduction model; obtaining charging features and a health state label of the target battery at the initial stage of the life cycle for several cycles as context data; inputting the deduced future cycle charging features into a health state probability estimation model in combination with the context data, and outputting health state probability distribution of multiple future cycles, so as to realize long-term multi-step probability prediction. Through the double-layer optimization structure of the meta-learning framework, the application learns cross-domain degradation prior knowledge from the multi-battery domain, dynamically adjusts the body difference by combining the meta-adaptive mechanism, and realizes high-precision, long-term and probabilistic health state prediction under the support of a small amount of context data, thereby providing a reliable basis for battery health management and risk assessment.

Description

Technical Field

[0001] This application relates to the field of lithium-ion battery health status estimation and prediction technology, and in particular to a multi-step probability prediction method and related equipment for battery health status. Background Technology

[0002] As the core power unit of electric vehicles, large-scale energy storage, and other systems, the accurate prediction of the State of Health (SOH) of lithium-ion batteries is a key technological guarantee for the safe and long-term operation of battery systems. SOH directly characterizes the battery's lifespan degradation state, and its reliable long-term prediction is crucial for preventing failures and optimizing operation and maintenance strategies, thus possessing significant engineering and economic value.

[0003] Currently, data-driven SOH prediction methods have become the mainstream research paradigm. However, when applied to real-world long-term prediction scenarios, existing technologies reveal several technical bottlenecks that urgently need to be addressed:

[0004] First, there is a fundamental contradiction between deterministic prediction models and the inherent uncertainty of State of Harm (SOH). Battery SOH is influenced by multiple uncertainties, including manufacturing tolerances, sensor measurement noise, fluctuations in operating conditions, and the randomness of internal electrochemical processes; it is essentially a probabilistic problem. However, most existing models still focus on deterministic point prediction, outputting only a single SOH estimate. This approach completely ignores the inherent uncertainty of the prediction results themselves, failing to provide any quantitative basis for risk assessment (such as failure probability and performance confidence intervals), and thus failing to meet the needs of high-reliability applications.

[0005] Secondly, long-term multi-step prediction faces the challenge of error accumulation. The widely used iterative multi-step prediction strategy has inherent flaws when predicting SOH over long periods. This method uses the prediction output of the previous step as the prediction input for the next step, causing the prediction error to propagate and amplify continuously as the prediction step size increases, resulting in a sharp decline in the reliability of long-term prediction results.

[0006] Furthermore, the model's reliance on large amounts of historical data leads to the cold start dilemma. Many high-performance models require a sufficiently long historical data set of the target battery itself to make accurate personalized predictions. However, in practical applications, for newly developed and marketed batteries or online monitoring systems, available data is extremely limited. This data-scarce cold start scenario makes it impossible for traditional models to achieve effective early warning.

[0007] To address uncertainty, some studies have attempted to introduce probabilistic models such as Bayesian methods or Gaussian processes. However, these methods are often computationally complex, have poor scalability, and fail to fundamentally solve the problems of error accumulation and data dependence in long-term forecasting. Summary of the Invention

[0008] The main objective of this application is to propose a multi-step probability prediction method, electronic device, storage medium, and program product for battery health status based on future operating trajectory projection, in order to solve the problems in the prior art such as the inability to quantify uncertainty in deterministic prediction, accumulation of long-term prediction errors, and strong data dependence under cold start conditions.

[0009] To achieve the above objectives, one aspect of this application proposes a multi-step probability prediction method for battery health status, the method comprising:

[0010] Feature probability inference steps: Obtain charging feature data of the target battery for at least two cycles at the current time, input them into the pre-trained feature probability inference model, and directly infer the probability distribution of the charging feature data of the target battery for at least 150 future cycles, wherein the probability distribution includes the feature prediction mean and feature prediction variance of each future cycle point;

[0011] Health status probability estimation step: Obtain real or virtual charging feature data of the target battery in the early stage of its life cycle for use as context information, and obtain the real health status label of each cycle. Combine it with the charging feature data of multiple future target cycles obtained in the feature probability inference step, and input them into the pre-trained health status probability estimation model to estimate the health status probability distribution of the multiple future target cycles. The health status probability distribution includes the health status prediction mean and the health status prediction variance.

[0012] The feature probability inference model is trained by meta-learning and can infer the feature probability distribution for multiple future steps based on a small amount of cyclic input data.

[0013] In some embodiments, the feature probability inference model is a stochastic variational autoencoder model, comprising:

[0014] Encoder: Employs a multi-layer bidirectional long short-term memory network structure to extract bidirectional time-dependent features from the input sequence. It obtains a fixed-length sequence representation through sequence average pooling and maps it to the mean vector and log-variance vector of the latent space via two independent fully connected networks to characterize the probability parameters of the latent distribution.

[0015] Reparameterization module: used to perform differentiable sampling from the probability distribution of the latent space using reparameterization techniques to generate latent vectors;

[0016] Decoder: Employs a multi-layer fully connected neural network structure to decode the latent vector into feature prediction parameters for the complete future sequence, with outputs including the feature prediction mean and the feature prediction log variance;

[0017] The decoder further includes a residual connection module, which is used to map the feature mean of the input sequence through a projection network and then perform weighted fusion with the decoder output to enhance prediction stability.

[0018] In some embodiments, the meta-learning framework for training the feature probability inference model employs a two-layer optimization structure, including:

[0019] Meta-task construction steps: Randomly select one domain and several samples in the domain from multiple battery domains to form a meta-task. Several meta-tasks form a meta-batch. The samples of each meta-task are divided into a support set and a query set. Each sample consists of a short-term input sequence (several current cyclic input features) and a long-term target sequence (several future cyclic features to be deduced).

[0020] Inner loop optimization layer: In each metatask in the metatask batch, the current model parameters are cloned as their respective task-specific (initial) parameters. All metatasks are traversed and multi-step gradient descent is performed independently on their respective support sets to update their respective task-specific parameters. During the update process, the global model parameters remain unchanged.

[0021] Outer loop optimization layer: After traversing all meta-tasks in the inner loop optimization layer, the meta-loss is calculated on the query set of each meta-task using its own updated task-specific parameters. The meta-losses of all meta-tasks are aggregated and the global model parameters are updated through backpropagation of the second gradient.

[0022] In some embodiments, the health status probability estimation model is a neural process model based on a meta-adaptation mechanism, which includes:

[0023] Battery type embedding layer: used to map battery type to embedding vector;

[0024] Encoder: Used to concatenate and encode the context charging feature data, its health status label and the battery type embedding vector to generate potential distribution parameters for each context point;

[0025] Aggregator: Used to average and aggregate the potential distribution parameters of multiple context points to obtain a global potential representation characterizing the battery degradation mode;

[0026] Reparameterization module: used to perform differentiable sampling from the distribution of the global latent representation to generate latent vectors;

[0027] Meta-adaptive network: used to generate an adaptive adjustment vector based on the distribution parameters of the global latent representation, and to fuse it with the latent vector through residual connections;

[0028] Decoder: Used to combine charging feature data of future long-term target cycles, latent vectors with meta-adaptive correction, and the battery type embedding vector to output the predicted mean and standard deviation of the health status for each target point.

[0029] In some embodiments, obtaining real or virtual charging characteristic data of the target battery during a number of cycles in the early stages of its lifespan as context information includes:

[0030] Prioritize obtaining real context charging feature data of the target battery in the early stages of its life cycle; only when real context charging feature data is difficult to obtain, obtain virtual context charging feature data of the target battery in the early stages of its life cycle; the virtual context charging feature data is generated by selecting all batteries of the same type and operating conditions as the target battery in the training set of the health state probability estimation stage, and calculating the average value of each feature in the battery dimension in the early stages of its life cycle.

[0031] In some embodiments, the virtual data is specifically calculated as follows:

[0032] Assume the target battery is a 3500mAh NCA battery, charged at 25℃ and 0.5C. First, select batteries of the same type and under the same conditions from the training set, assuming a total of 3 batteries, denoted as B1, B2, and B3, and denote feature 1 of B1 in the first cycle. F 1) For F B1_cycle_1 B2 is F B2_cycle_1 B3 is F B3_cycle_1 The virtual data in the first period F The feature calculation method is as follows: ( F B1_cycle_1 +F B1_cycle_1 +F B1_cycle_1 ) / 3; the calculation method for each feature in each of the remaining cycles is the same.

[0033] In some embodiments, the charging characteristic data includes the charging capacity during the constant current charging phase, the charging time, the difference between the termination voltage and the starting voltage, and the voltage integral over time.

[0034] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0035] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described above.

[0036] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer program product, including a computer program that, when executed by a processor, implements the method described above.

[0037] The embodiments of this application include at least the following beneficial effects: This application provides a multi-step probabilistic prediction method for battery health status, an electronic device, a storage medium, and a program product. This solution endows the model with powerful cross-battery domain generalization and adaptive capabilities through a meta-learning mechanism, and realizes high-precision, long-term, probabilistic SOH prediction with the support of a small amount of data, providing a reliable basis for battery health management and risk assessment. Attached Figure Description

[0038] Figure 1 This is a flowchart of the meta-learning driven training process of the feature probability inference model in this embodiment of the invention.

[0039] Figure 2 This is a diagram of the core architecture of the probabilistic variational autoencoder based on Bi-LSTM in the embodiments of this application.

[0040] Figure 3 This is a schematic diagram illustrating the predicted characteristics of a 45°C NCM battery over the next 150 cycles in an embodiment of this application.

[0041] Figure 4 This is the core architecture diagram of the health status probability estimation model (MNPs) in the embodiments of this application.

[0042] Figure 5 This is a schematic diagram showing the results of multi-step probability prediction of the SOH of a 45℃ NCM battery in an embodiment of this application.

[0043] Figure 6 This is a flowchart illustrating the implementation of the multi-step probability prediction method for battery health status in this application.

[0044] Figure 7 This is a schematic diagram of the hardware structure of the electronic device in the embodiments of this application. Detailed Implementation

[0045] 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 of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0046] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0047] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0048] 1) Battery State of Health (SOH) is the core indicator for measuring battery health. It reflects the ratio of the battery's current capacity to its factory capacity and is used to assess battery performance and remaining lifespan.

[0049] Current technologies lack an integrated framework capable of directly probabilistically extrapolating the characteristics of a target battery over multiple future cycles with limited current cycle data, and thereby predicting the state of health (SOH) probability distribution in advance through multiple steps. Therefore, developing an innovative prediction method to overcome these bottlenecks and achieve early, long-term, and quantifiable SOH prediction has become an urgent need for advancing battery management technology. Based on this, this application provides a multi-step probabilistic prediction method for battery health status, an electronic device, a storage medium, and a program product.

[0050] Due to manufacturing tolerances and sensor noise and environmental interference during data acquisition, the State of Harmony (SOH) of a battery is not fixed; it is essentially a probability distribution problem. To improve estimation accuracy and quantify risk assessment, the model needs to be capable of learning probability distributions and making probabilistic inferences. Furthermore, multi-step prediction of battery SOH is a key prerequisite for optimizing battery lifecycle maintenance strategies. This application proposes a method that uses the charging capacity during the constant current charging phase. Charging time The difference between the termination voltage and the starting voltage Integral of voltage over time Four features are used as input features of the model. A complete charge-discharge process is regarded as a cycle. First, based on the feature data of the current two cycles, the probability distribution of the four features in the next 150 cycles is inferred. Then, based on the inference results of these features, the SOH probability distribution at each cycle point is estimated. In the inference of the feature probability distribution of the next 150 cycles, a meta-learning mechanism is incorporated. Through a two-layer optimization structure, the model can learn cross-domain prior knowledge from the degradation patterns of multiple battery domains. Thus, when facing a new battery, even with only a very small amount of initial cycle data as context input, it can overcome individual differences and perform personalized and accurate inferences, giving the model generalization and adaptive capabilities. In the health state estimation task, a meta-adaptation mechanism is introduced. This mechanism maps the context cycle data of the target battery point by point to latent distribution parameters through an encoder. After obtaining the global latent distribution parameters through an aggregator, the latent vector is sampled from this distribution. The meta-adaptation network generates an adjustment vector based on the global latent distribution parameters and dynamically corrects the latent vector through residual connections. By incorporating the individualized degradation patterns of different batteries as prior knowledge into the model, personalized long-term health status predictions for batteries with different chemical systems were achieved.

[0051] This application utilizes data from only the current two cycles to accurately estimate the SOH probability distribution for the next 150 cycles. This significantly shortens the testing cycle required for the development and verification of new battery products, saving time and economic costs. Furthermore, the long-term probabilistic predictions provided by this method can offer forward-looking guidance for the rational use and health management of batteries.

[0052] This application provides a multi-step probability prediction method for battery health status, relating to the field of lithium-ion battery state of health (SOH) estimation technology. This method can be applied to terminals, servers, or software running on either a terminal or server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or in-vehicle terminal, but is not limited to these. The server can be configured as an independent 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 functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing the multi-step probability prediction method for battery health status, but is not limited to the above forms.

[0053] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0054] like Figure 6 As shown, this embodiment provides a multi-step probabilistic prediction method for battery health status based on future operating trajectory extrapolation, including the following steps:

[0055] S1: Feature probability deduction steps.

[0056] The charging characteristic data of the target battery for at least two cycles at the current moment are acquired and input into a pre-trained feature probability inference model to directly deduce the probability distribution of the target battery's charging stage characteristic data for at least 150 future cycles. The probability distribution includes the feature prediction mean and feature prediction variance for each future cycle point.

[0057] In some embodiments, the feature probability inference model is a probabilistic variational autoencoder (PVAE) trained by meta-learning. The encoder of this model employs a bidirectional long short-term memory (Bi-LSTM) network to extract deep feature representations from the input short-time series and maps them to a probability distribution in the latent space through variational inference, generating a mean vector. μ Sum of logarithmic variance vector Subsequently, latent variables were sampled from this distribution using the reparameterization technique. z The input is fed into a decoder composed of a fully connected network to directly predict the distribution parameters (i.e., the feature prediction mean and feature prediction variance) of the generated long-term target sequence. To enhance prediction stability, the model also introduces a residual connection mechanism based on the mean of the input sequence.

[0058] In some embodiments, the feature probability inference model is trained using meta-learning, enabling it to learn prior knowledge from multiple battery domains (i.e., battery data from different chemical systems and operating conditions). This allows it to overcome individual differences and perform personalized, accurate inferences when faced with new batteries, even with only a very small amount of initial cycling data. The meta-learning framework includes meta-task construction, inner loop, and outer loop:

[0059] The meta-task construction steps shown are used to: randomly select a domain and several samples in the domain from multiple battery domains to form a meta-task, and several meta-tasks form a meta-batch. The samples of each meta-task are divided into a support set and a query set. Each sample corresponds to a long-term target sequence (e.g., 150 loops) with a short-term input sequence (e.g., 2 loops).

[0060] The inner loop optimization layer is used to: clone the current global model parameters as their respective task-specific parameters for each meta-task, traverse all meta-tasks and independently perform multi-step gradient descent on their respective support sets to update their respective task-specific parameters, keep the global model parameters unchanged during the update process, and only update the task-specific parameters.

[0061] The outer loop optimization layer is used to: after traversing all meta-tasks (completing a meta-batch) in the inner loop optimization layer, calculate the meta-loss using the updated task-specific parameters on the query set of each meta-task, aggregate the meta-losses of all meta-tasks, and update the global model parameters through backpropagation of the second-order gradient.

[0062] S2: Health status probability estimation step.

[0063] The charging characteristic data of the target battery in the early stages of its life cycle, both real and virtual, along with its corresponding health status labels, are obtained. These data are then combined with the charging characteristic data of the future long-term target cycle derived from step S1 and input into a pre-trained health status probability estimation model to estimate the health status probability distribution of the future long-term target cycle. The health status probability distribution includes the predicted mean and predicted variance of the health status.

[0064] In some embodiments, the health status probability estimation model is a meta-adaptive neural process (MNP) model. This model includes:

[0065] Battery type embedding layer, used to map battery type to embedding vector;

[0066] The encoder is used to concatenate and encode the context charging feature data, its health status label and the battery type embedding vector to generate potential distribution parameters for each context point.

[0067] The aggregator is used to average and aggregate the potential distribution parameters of multiple context points to obtain a global potential representation characterizing the battery degradation mode.

[0068] A reparameterization module is used to perform differentiable sampling from the distribution of the global latent representation to generate a latent vector;

[0069] A meta-adaptive network is used to generate an adaptive adjustment vector based on the distribution parameters of the global latent representation, and to fuse it with the latent vector through residual connections;

[0070] The decoder combines charging feature data from future long-term target cycles, meta-adaptive modified latent vectors, and the battery type embedding vector to output the predicted mean and standard deviation of the health status for each target point.

[0071] In a preferred embodiment, when actual data for the target battery at the beginning of its lifespan is difficult to obtain, the virtual context charging feature data of the target battery at the beginning of its lifespan, as described in step S2, is generated by selecting all batteries of the same type and operating conditions as the target battery in the training set of step S2, and calculating the average value of each feature in the battery dimension for several cycles at the beginning of their lifespan. This method effectively solves the cold start problem when a new battery is activated.

[0072] In some embodiments, the charging characteristic data includes, but is not limited to, the charging capacity during the constant current charging phase. Charging time The difference between the termination voltage and the starting voltage and the integral of voltage over time .

[0073] The solutions of the embodiments of this application will be described in detail below with reference to the accompanying drawings and specific application examples.

[0074] Firstly, in the future feature inference stage, a meta-learning-driven Stochastic Variational Autoencoder (SVAE) is provided. It uses a multi-layer bidirectional long short-term memory (Bi-LSTM) network to extract deep feature representations from historical time series. Fixed-length sequence representations are obtained through sequence average pooling, and after layer normalization, they are input into two independent multi-layer fully connected networks, which are then mapped to the mean vector of the latent space. μ Sum of logarithmic variance vector Then, the reparameterization technique is used to sample the latent vector from this latent distribution. zThe input multilayer fully connected decoder directly generates the distribution parameters (predicted mean and predicted log-variance) of the complete predicted sequence. To enhance prediction stability, the model introduces a residual connection mechanism based on the input sequence mean, mapping historical mean information through a projection network and then weighting and fusing it with the decoder output. During training, negative log-likelihood loss is used to quantify the goodness of fit between the predicted distribution and the true values ​​and to reflect prediction uncertainty. KL divergence regularization is combined to align the latent space distribution to a standard normal distribution. A two-layer optimization structure within the meta-learning framework enables rapid adaptation across battery domains, allowing the model to quickly adapt to new battery degradation patterns with limited sample size. This architecture not only outputs the predicted mean but also simultaneously outputs the predicted log-variance, providing reliable uncertainty quantification information for subsequent decision-making. Figure 1 A flowchart of the meta-learning-driven process is given. Figure 2 The core architecture diagram of Bi-LSTM-SVAE is given.

[0075] for Figure 1 It is important to note that in this embodiment, when constructing the task batch in step 1 of the inner loop, four domains (which can be repeated) are randomly sampled from the multi-domain battery data as four meta-tasks. Each task contains 16 random samples within that battery domain, with the first eight samples serving as a support set for task-specific adaptation and the last eight samples serving as a query set for meta-loss evaluation. Since the inner loop does not directly update the global model parameters, step 2 requires cloning the current model parameters. θ → θ' This serves as the basis for the inner loop computation. In step 3, K is set to 5, and five steps of gradient descent are performed on the support set of each task to update its clone parameters. θ' Each gradient calculation must preserve the computation graph so that the elementary gradient can be calculated via backpropagation using the second derivative in step 2 of the outer loop. θ L meta This enables cross-task meta-knowledge learning. Inner loop step 4 uses the quickly adapted parameters. θ' Calculate the meta-loss on the query set of this task. L task ( θ' In the inner loop, step 5 iterates through all four tasks, repeating the above process, and finally aggregates the meta-loss of all tasks. L meta =Σ L task ( θ' The results are passed to the outer loop. After aggregating the meta-loss of all tasks in step 1, the outer loop calculates the meta-gradient using the second-order gradient in step 2 and updates the global parameters. θ = θ - β θL meta This completes one iteration of the outer loop. It's important to emphasize that each meta-task checks the current global parameters before starting the inner loop. θ clone independent θ' Furthermore, the global parameter update of the outer loop is only performed once after all the inner loop tasks have been traversed. The entire process iterates... N outer This continues until convergence.

[0076] for Figure 2 The Bi-LSTM-PVAE model architecture shown in the figure has dashed boxes indicating that the portion used only during the training phase is for loss calculation and parameter updates. The input sequence during the encoding phase... x The length is 2, and the sliding window step size is 1; the decoding phase predicts the probability distribution of the next 150 loop points. An example of training data construction is as follows:

[0077] 1) Anchor point selection: For a given meta-task, a specific anchor point is randomly selected. t ,in t >10, and t The selectable range is t = t +1. That is, the anchor point time. t Starting from the 11th cycle, its value range is: t ∈{11, 12, 13, 14,..., L-152}, where L is the total number of cycles of the battery. L-152 is used to determine the end time of the anchor point to ensure that there is enough data to form a sample.

[0078] 2) Meta-batch construction: Randomly select 16 anchor timestamps from the current meta-task domain. For example, if the selected anchor timetamp is 13, then the input sequence... x The input sequence consists of the feature values ​​from cycles 13 to 14, while the target sequence consists of the feature values ​​from cycles 15 to 164 (150 cycles). This input-target sequence set constitutes a sample. If the selected anchor time is 100, then the input sequence... x The input sequence consists of two cycles of feature values ​​from 100 to 101, while the target sequence consists of 150 cycles of feature values ​​from 102 to 251. This input-target sequence set constitutes another sample. This process is repeated 16 times, selecting 16 samples in total.

[0079] 3) Meta-task construction: Steps 1 and 2 above complete the construction of one meta-task data. Repeat steps 1 and 2 4 times to complete the construction of 4 meta-task data.

[0080] The input sequence is processed using a Bi-LSTM encoder during the encoding stage. x And by averaging along the cyclic dimension, a fixed-length feature vector is obtained.h The probabilistic reasoning stage will use features h Mapped to a probability distribution, generated via a mean-based network. f μ ( h ) and log-variance generator network f σ ( h Output the mean of the latent distribution respectively. μ Sum of logarithms and variances σ ², and then differentiable sampling is performed using the reparameterization technique to obtain z ~ N( ) which follows a normal distribution. μ , σ ²) latent variables, thereby capturing the inherent uncertainty of the data; the decoding stage will identify latent variables z With input sequence x The mean features are input together into the fully connected decoder, and the prediction stability is enhanced through the residual connection mechanism. Finally, the predicted mean is output simultaneously. μ pre and prediction variance σ pre 2 Construct a complete probability prediction distribution Y ~ N( μ pre , σ pre 2 This enables quantitative prediction of uncertainty. The entire learning process consists of... Figure 1 Driven by the meta-learning framework, it achieves rapid cross-domain adaptation capabilities through the collaborative optimization of inner and outer loops.

[0081] To verify the effectiveness of the aforementioned future feature projection method, this embodiment selects three types of batteries—NCM, INR, and NCA+NCM—to train the model. The NCA+NCM battery is a hybrid battery whose cathode material consists of 42% NCA and 58% NCM. The three battery types are divided into three different domains. Table 1 provides the specific information about the batteries.

[0082] Table 1

[0083]

[0084] Table 1 shows a total of 9 NCA+NCM type batteries, with three test conditions: 0.5 / 1, 0.5 / 2, and 0.5 / 4. Each condition has 3 batteries with data, totaling 9 batteries. Two batteries were randomly selected from each condition, for a total of 6 batteries, as the training set. The remaining 3 batteries were used as the test set. INR batteries also have three test conditions, with WLCT being the globally unified light vehicle test cycle discharge condition. Each condition has 2 batteries with data, totaling 6 batteries. One battery was randomly selected from each condition, for a total of 3 batteries, as the training set. The remaining 3 batteries were used as the test set. Taking a 45℃ NCM battery as an example... Figure 3 The results show the extrapolation of features for the next 150 cycles using data from the current two cycles.

[0085] Figure 3 The extrapolation process is as follows: First, the data from the first 10 cycles is used as context data. Starting from the 11th cycle, the extrapolation of future features is performed (the specific reasons will be explained in the subsequent SOH probability prediction section). The data from cycles 11-12 is used to extrapolate the future features of cycles 13-162. Then, the data from cycles 163-164 is used to extrapolate the features of cycles 165-314, and this cycle continues until the end of the battery lifespan. In this embodiment, different colors represent different extrapolation stages. For 45℃ NCM, a total of 6 extrapolations were performed. Because two cycles are too short, the corresponding data points, even after being labeled, are still difficult to identify in the diagram. Figure 3 In the left-hand plot, the results of each cycle are represented by the mean (solid line) and standard deviation (light-colored dashed shadow). The mean represents the central tendency of the probability distribution, while the confidence interval of the standard deviation reflects the uncertainty of the model estimate. The narrower the confidence interval, the higher the certainty of the model at that point; the wider the interval, the greater the uncertainty. In this case, further diagnosis is needed to determine whether the problem originates from the input features or the model itself, so as to make targeted adjustments. Figure 3 The error distributions of the middle and right subgraphs verify that the feature inference has high accuracy, which lays a reliable foundation for the subsequent probability distribution prediction of SOH. The core objective of this invention is the final SOH estimation, and feature inference is an intermediate step serving this objective; therefore, detailed statistical analysis of the inference results of the intermediate process is not performed.

[0086] In the SOH probability estimation stage, this embodiment provides a Meta-Adaptive Neural Process (MNP) framework based on a meta-adaptive mechanism. This method achieves representation learning of battery degradation patterns through an encoder-decoder architecture. The encoder network integrates charging features from several context cycles, SOH labels, and battery type embeddings to model the probability distribution of each context point, outputting the mean and variance parameters of each point. The aggregator aggregates the latent distribution parameters of multiple context points through averaging operations, extracts global degradation statistical features, and generates global latent distribution parameters. The reparameterization module performs differentiable sampling from this distribution to obtain a latent vector. The meta-adaptive network generates an adaptive adjustment vector based on the global latent distribution parameters and dynamically corrects the latent vector through residual connections. The decoder combines the corrected latent vector, target point charging features, and battery type embeddings to output the mean and variance parameters of the SOH prediction for each target point.

[0087] The framework innovatively introduces a meta-adaptive network. First, it uses reparameterization techniques to sample latent vectors from the aggregated global latent distribution parameters. Then, it inputs these global latent distribution parameters into the meta-adaptive network to dynamically generate adaptive adjustment vectors. Residual connections are used to correct these latent vectors, injecting individual degradation prior knowledge into the prediction process, enabling rapid adaptation to different chemical systems and individual differences. The reparameterization mechanism not only ensures efficient gradient calculation during backpropagation but also enhances the model's generalization ability through random sampling of the latent space. The model loss uses negative log-likelihood to measure the goodness of fit between the predicted distribution and the true values. Simultaneously, a KL divergence regularization term is introduced to constrain the latent distribution to align with the standard normal distribution, preventing overfitting and enhancing generalization ability.

[0088] This framework not only outputs probability predictions for target points, but also provides 95% confidence intervals, offering reliable uncertainty quantification support for battery health management.

[0089] Figure 4 The core architecture diagram of MNPs during prediction is given. The data source used is consistent with Table 1, consisting of three different battery domains: NCA+NCM, NCM, and INR. The dashed box portion in the diagram is used only during the training phase for loss calculation and parameter updates. In each training batch, samples are randomly sampled from the three battery types. The construction of each sample follows these rules:

[0090] 1) Anchor point selection: Randomly select an anchor point at a given time. t Within a certain battery type, t ≥10, and t The selectable range is t = t +5. That is, the anchor point time. tStarting from the 10th loop, the sliding window step size is 5. Its possible values ​​are... t ∈{10, 15, 20, 25, ..., L-10}. L is the total number of cycles for this battery.

[0091] 2) Context window: Extracting time t -9 to t The features of the 10 loops are used as context features, and their corresponding SOH values ​​are used as context labels.

[0092] 3) Target window: Extraction time t +1 to t The features from the 10 + 10 loops are used as the target features, and their corresponding SOH values ​​are used as the target labels.

[0093] 4) Type encoding: Obtain the battery type (NCA+NCM / NCM / INR) of the current sample and perform embedding encoding.

[0094] Batch build example (batch size=16):

[0095] Sample 1 (NCM battery): Anchor point time t →Context[ t -9, t ],Target[ t +1, t +10].

[0096] Sample 2 (INR cell): Anchor point time i →Context[ i -9, i ],Target[ i +1, i +10].

[0097] …

[0098] Sample 16 (random type): Anchor time k →Context[ k -9, k ],Target[ k +1, k +10].

[0099] Each batch contains 16 such samples, which may come from different battery types and different degradation stages. This hybrid sampling strategy enables the model to learn common features across battery types and specific degradation patterns for each type.

[0100] For prediction, data from the first 10 cycles of the battery under test are used as the starting data for model conditionation. The prediction target is the SOH probability distribution at all cycle points from the 11th cycle until the end of the battery life. Taking a 45℃ NCM battery as an example, the method for generating data for the first 10 cycles of the battery under test is given below:

[0101] As shown in Table 1, the training phase includes 5 battery samples. First, the feature data of these 5 batteries are extracted in the first 10 loops, and then the arithmetic mean of the corresponding features in each loop is calculated among the battery samples.

[0102] Specifically, based on features For example: Battery 1's first 10 cycles The sequence is [ , ,..., The corresponding sequence for battery 2 is []. , , ..., ], and so on, the sequence for battery 5 is [ , , ..., By averaging the eigenvalues ​​at each time step across cells, we obtain:

[0103]

[0104]

[0105] …

[0106]

[0107] This yields the average feature sequence used for prediction. , , ..., ]. Features Δ V , Using the same processing method, we obtain [ , ..., ], [ , , ..., ], [ , , ..., ].

[0108] The first 10 cycle characteristics of all tested cells at 45℃ NCM were replaced with the average characteristic sequence calculated above. This is because:

[0109] 1) In practical industrial applications, it may be difficult to obtain contextual data for 10 consecutive complete cycles of the test set battery. Furthermore, the model used for SOH probability prediction in this invention essentially uses contextual data to identify the degradation patterns of specific battery types, which is then used to calibrate the mapping relationship between features and SOH. The information contained in the feature values ​​of the target point truly plays a core role in SOH prediction.

[0110] 2) For any specific type of battery (referring to the broad category of chemical systems, such as NCM systems, INR systems, etc.), the differences between individual batteries in the first 10 cycles are very small and almost negligible. Even if the mean features of the training set batteries are used as replacements, it will not have an irreversible impact on the final prediction results.

[0111] This is also Figure 3 The reason why the initial 10 cyclic data points need to be retained without processing during future feature extrapolation.

[0112] During training, the MNPs model learns the battery aging degradation function from the training data. f (Degradation mode, target point features), where degradation mode = encode(context). This allows the model to accurately estimate the SOH probability distribution of the target point during the prediction process, based on the context data of the first 10 cycles and the feature data at any future time.

[0113] Figure 5 Using 45℃ NCM as an example again, the results of SOH probability estimation are presented. The calculation method for the context features of the first 10 cycles is as described above. The method for extrapolating future features is the same as... Figure 3 The methods used are consistent.

[0114] Figure 5 The left-middle figure shows the fitting results between the actual SOH value of the battery (solid red line) and the model's predicted mean (dashed green line). The green shaded area provides a 95% confidence interval, which is obtained through 100 Monte Carlo (MC) forward propagations. MC Dropout is used to activate different model subnetworks in each inference and sample from their respective predicted distributions, resulting in 100 possible distributions. Based on the statistical analysis of these samples, the 95% confidence interval is calculated to quantify the uncertainty of the prediction.

[0115] Figure 5This indicates that the model prediction has high certainty and a narrow confidence interval, reflecting good consistency of the prediction results. The middle and right figures further analyze the distribution of prediction errors. The predicted values ​​are in high agreement with the true values ​​(R² = 0.994), and the mean absolute error (MAE) is as low as 0.0045. These results fully verify that the model has extremely high prediction accuracy and successfully achieves the design goal of high-precision, multi-step SOH probability distribution estimation. Table 2 presents the overall results of multi-step SOH estimation for the test cells in each domain.

[0116] Table 2

[0117]

[0118] In summary, compared with the prior art, the method of this embodiment has at least the following advantages and beneficial effects:

[0119] 1) Achieved probabilistic prediction and uncertainty quantification: This application directly outputs the probability distribution (mean and variance) and confidence interval of SOH, providing a reliable basis for risk quantification for battery health management and overcoming the limitations of traditional deterministic point prediction.

[0120] 2) Solves the problem of error accumulation in long-term prediction: The feature distribution of future multiple steps (such as 150 cycles) is directly generated by the feature probability inference model, instead of iterative prediction, which fundamentally avoids the propagation and accumulation of errors and ensures the reliability of long-term prediction.

[0121] 3) Overcoming the cold start dilemma: By utilizing the meta-learning mechanism, the model learns general degradation prior knowledge from multiple battery domains, enabling it to start accurate long-term predictions with only a very small amount of data (such as 2 cycles) of data from the target battery, greatly shortening the product verification cycle and saving costs.

[0122] 4) Possesses strong generalization and adaptation capabilities: Through the meta-learning framework and meta-adaptation network, the model can quickly adapt to different chemical systems (such as NCA+NCM, NCM, INR) and different individual batteries, demonstrating excellent cross-domain generalization performance.

[0123] 5) High prediction accuracy: Experimental verification shows that the method of this invention can predict the SOH of various types of batteries with a mean absolute error (MAE) as low as 0.45%~0.92% and a coefficient of determination (R²) as high as 0.994 or above, which is significantly better than the existing technology.

[0124] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0125] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0126] Please see Figure 7 , Figure 7 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes:

[0127] The processor 701 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0128] The memory 702 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 702 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 702 and is called and executed by the processor 701 using the methods described in the embodiments of this application.

[0129] The input / output interface 703 is used to implement information input and output;

[0130] The communication interface 704 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0131] Bus 705 transmits information between various components of the device (e.g., processor 701, memory 702, input / output interface 703, and communication interface 704);

[0132] The processor 701, memory 702, input / output interface 703, and communication interface 704 are connected to each other within the device via bus 705.

[0133] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0134] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0135] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0136] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0137] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented in the embodiments of this program product are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments. The executable computer program code or "code" used to perform the various embodiments can be written in high-level programming languages ​​such as C, C++, Python, Smalltalk, Java, JavaScript, Visual Basic, Structured Query Language (e.g., Transact-SQL), Perl, or in various other programming languages.

[0138] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0139] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0140] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0141] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0142] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification 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 data can be interchanged 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.

[0143] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0144] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0145] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0146] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0147] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0148] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A multi-step probabilistic prediction method for battery health status, characterized in that, The method includes the following steps: Feature probability inference steps: Obtain charging feature data of the target battery for at least two cycles at the current time, input them into the pre-trained feature probability inference model, and directly infer the probability distribution of the charging feature data of the target battery for at least 150 future cycles, wherein the probability distribution includes the feature prediction mean and feature prediction variance of each future cycle point; Health status probability estimation step: Obtain real or virtual charging feature data of the target battery in the early stage of its life cycle for use as context information, and obtain the real health status label of each cycle. Combine it with the charging feature data of multiple future target cycles obtained in the feature probability inference step, and input them into the pre-trained health status probability estimation model to estimate the health status probability distribution of the multiple future target cycles. The health status probability distribution includes the health status prediction mean and the health status prediction variance. The feature probability inference model is trained by meta-learning and can infer the feature probability distribution of future multiple steps based on a small amount of cyclic input data. The feature probability inference model is a stochastic variational autoencoder model, including: Encoder: Employs a multi-layer bidirectional long short-term memory network structure to extract bidirectional time-dependent features from the input sequence. It obtains a fixed-length sequence representation through sequence average pooling and maps it to the mean vector and log-variance vector of the latent space via two independent fully connected networks to characterize the probability parameters of the latent distribution. Reparameterization module: used to perform differentiable sampling from the probability distribution of the latent space using reparameterization techniques to generate latent vectors; Decoder: Employs a multi-layer fully connected neural network structure to decode the latent vector into feature prediction parameters for the complete future sequence, with outputs including the feature prediction mean and the feature prediction log variance; The decoder further includes a residual connection module, which is used to map the feature mean of the input sequence through a projection network and then perform weighted fusion with the decoder output to enhance prediction stability. The health status probability estimation model is a neural process model based on a meta-adaptation mechanism, which includes: Battery type embedding layer: used to map battery type to embedding vector; Encoder: Used to concatenate and encode the context charging feature data, its health status label and the battery type embedding vector to generate potential distribution parameters for each context point; Aggregator: Used to average and aggregate the potential distribution parameters of multiple context points to obtain a global potential representation characterizing the battery degradation mode; Reparameterization module: used to perform differentiable sampling from the distribution of the global latent representation to generate latent vectors; Meta-adaptive network: used to generate an adaptive adjustment vector based on the distribution parameters of the global latent representation, and to fuse it with the latent vector through residual connections; Decoder: Used to combine charging feature data of future long-term target cycles, latent vectors with meta-adaptive correction, and the battery type embedding vector to output the predicted mean and standard deviation of the health status for each target point.

2. The method according to claim 1, characterized in that, The meta-learning framework for training the feature probability inference model adopts a two-layer optimization structure, including: Meta-task construction steps: Randomly select one domain and several samples in the domain from multiple battery domains to form a meta-task. Several meta-tasks form a meta-batch. The samples of each meta-task are divided into a support set and a query set. Each sample consists of a short-time input sequence and a long-time target sequence. Inner loop optimization layer: In each metatask in the metatask batch, the current model parameters are cloned as their respective task-specific parameters. All metatasks are traversed and multi-step gradient descent is performed independently on their respective support sets to update their respective task-specific parameters. During the update process, the global model parameters remain unchanged. Outer loop optimization layer: After traversing all meta-tasks in the inner loop optimization layer, the meta-loss is calculated on the query set of each meta-task using its own updated task-specific parameters. The meta-losses of all meta-tasks are aggregated and the global model parameters are updated through backpropagation of the second gradient.

3. The method according to claim 1, characterized in that, The acquisition of real or virtual charging characteristic data of the target battery during several cycles in the early stages of its life cycle, used as context information, includes: Prioritize obtaining real context charging feature data of the target battery in the early stages of its life cycle; only when real context charging feature data is difficult to obtain, obtain virtual context charging feature data of the target battery in the early stages of its life cycle; the virtual context charging feature data is generated by selecting all batteries of the same type and operating conditions as the target battery in the training set of the health state probability estimation stage, and calculating the average value of each feature in the battery dimension in the early stages of its life cycle.

4. The method according to claim 3, characterized in that, The specific calculation method for virtual data is as follows: Assume the target battery is a 3500mAh NCA battery, charged at 25℃ and 0.5C. First, select batteries of the same type and under the same conditions from the training set, assuming a total of 3 batteries, denoted as B1, B2, and B3, and denote feature 1 of B1 in the first cycle. F 1) For F B1_cycle_1 B2 is F B2_cycle_1 B3 is F B3_cycle_1 The virtual data in the first period F The feature calculation method is as follows: ( F B1_cycle_1 +F B1_cycle_1 +F B1_cycle_1 ) / 3; the calculation method for each feature in each of the remaining cycles is the same.

5. The method according to claim 1, characterized in that, The charging characteristic data includes the charging amount during the constant current charging phase, the charging time, the difference between the termination voltage and the starting voltage, and the voltage integral over time.

6. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 5.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 5.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 5.

Images