A battery health state estimation method with unbiased uncertainty quantification capability

Abstract

This invention belongs to the field of battery management technology and discloses a battery health state estimation method with unbiased uncertainty quantification capability. It collects process quantity signals during the charging process of a lithium-ion battery, and vertically concatenates several continuous process quantity signals, flattened into row vectors, into a matrix. A hierarchical stochastic spatiotemporal Transformer is established, and a stochastic self-attention mechanism based on the reparameterization of the Günbel normalized exponential function is used to capture and express the uncertainty of the model during feature extraction and aggregation. Based on this stochastic self-attention mechanism, an uncertainty-aware step-by-step dilated convolution and a distance-aware hierarchical Transformer are implemented to capture the dynamic inter-cyclic dependencies in the battery degradation process. An uncertainty calibration auxiliary module is established, which performs post-processing correction on the original quantified uncertainty based on ordinal-preserving regression to reflect the true confidence level of the uncertainty corresponding to the estimated battery health state.

Description

Technical Field

[0001] This invention relates to the field of battery management technology, and in particular to a battery health state estimation method with unbiased uncertainty quantification capability. Background Technology

[0002] Lithium-ion batteries, due to their high energy density and long cycle life, are widely used in electric vehicles, portable electronic devices, and energy storage systems—fields with high requirements for safety and reliability. To ensure the safe operation of these systems and effectively manage their lifespan, accurate estimation of the state of health (SQS) of lithium-ion batteries is crucial. SQS is typically defined as the ratio of the battery's current maximum usable capacity to its nominal capacity, and is a core indicator for evaluating the degree of battery performance degradation.

[0003] In existing technologies, data-driven battery health state estimation methods, especially those based on deep learning, have attracted widespread attention because they do not require the construction of complex electrochemical models and can learn battery degradation patterns from historical data. These methods typically utilize time-series data such as voltage, current, and temperature during the charging and discharging process to predict battery health state by constructing models such as neural networks. However, existing data-driven health state estimation methods still face the following major technical problems in practical applications, limiting their reliability and further promotion. First, the deterministic predictions implemented by existing methods have limitations. Most existing models only provide point estimates of health state, lacking the quantification of uncertainty in the prediction results. This makes it difficult to assess the reliability of the prediction under complex or unknown operating conditions, thus affecting subsequent maintenance decisions and risk management. Even if some existing methods have uncertainty quantification capabilities, they still suffer from imperfections in the uncertainty quantification methods themselves, failing to fully address the uncertainty propagation problem in the feature extraction and information fusion processes within the model. Secondly, battery degradation involves complex spatiotemporal dynamics. How to effectively capture and utilize the inherent uncertainties in these spatiotemporal dependencies to extract more robust degradation features remains a critical challenge for data-driven battery health state estimation models. Finally, data-driven battery health state estimation methods suffer from a lack of calibration for prediction confidence. The uncertainty measure output by the model may not match the actual distribution of prediction errors, significantly impacting its practical guiding significance and potentially misleading decision-making. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a lithium battery health state estimation method with unbiased uncertainty quantification capabilities. This method describes the attention score as an uncertainty weight characterized by a Gumbel distribution, and based on this, decouples the uncertainty in the feature extraction process into spatial and temporal uncertainties. To make the uncertainty output by the model corresponding to the health state more reliable, this invention proposes an uncertainty calibration strategy based on ordinal-preserving regression, and minimizes the distance between the model-quantified distribution and the true data distribution.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A battery health state estimation method with unbiased uncertainty quantification capability includes a data preprocessing module, a hierarchical stochastic spatiotemporal Transformer module, and an uncertainty calibration auxiliary module. The data preprocessing module receives timing signals from the battery charging and discharging process obtained by the sensor module of the battery management system, and processes them to obtain a battery degradation sample set. , Indicates the first Characteristics of the sample Indicates the first The label for each sample represents the battery health state corresponding to the last charge-discharge cycle of that sample. It is the total number of degraded samples in the sample set; The hierarchical random spatiotemporal Transformer module This includes a spatial degradation pattern extractor for extracting spatial degradation dependencies and a temporal degradation pattern extractor for extracting temporal degradation dependencies. The spatial degradation pattern extractor takes a battery degradation sample set as input and calculates sensor channel weights based on weights obtained from a random self-attention layer. A dilated convolutional layer processes the degradation data with sensor channel weights to obtain a convolutional feature map that reveals the spatial degradation information. During the spatial degradation pattern extraction process, the pre-clustering submodule and the time-step distance-sensitive position encoding submodule remain closed. The convolutional feature map is flattened and vertically stacked before being fed into the temporal degradation pattern extractor. The time degradation time information is processed; the time degradation pattern extractor includes a pre-clustering submodule, a time-step distance-sensitive location encoding submodule, and a multi-head random self-attention submodule; the pre-clustering submodule and the time-step distance-sensitive location encoding submodule remain active during the time degradation pattern extraction process; the time degradation pattern extractor has a residual connection structure, and the inputs are respectively processed by the multi-head random self-attention submodule and direct mapping, then added and normalized to generate the final hierarchical spatiotemporal degradation pattern; the hierarchical spatiotemporal degradation pattern is processed by a multilayer perceptron regressor to generate battery health state point estimates; The uncertainty calibration auxiliary module is used to correct the deviation between the prediction confidence of the hierarchical stochastic spatiotemporal Transformer module and the actual error distribution, and to perform unbiased uncertainty quantification; the uncertainty calibration auxiliary module relies on a calibration dataset. The supervised correction process uses hierarchical random spatiotemporal Transformer modules to correct samples in the correction dataset. The output predicted cumulative distribution function is used as a feature, based on the battery degradation sample set. The true cumulative distribution function of the degraded samples is calculated as the label; based on an order-preserving regressor trained using a calibration dataset. The uncertainty calibration auxiliary module calibrates the hierarchical stochastic spacetime Transformer module, and the hierarchical stochastic spacetime Transformer module with unbiased uncertainty quantification capability is finally expressed as follows: ,in The function composition is represented, and the output is the estimated battery health state and its corresponding uncertainty.

[0006] The sensor module includes a timer, a voltage sensor, and a current sensor, which record the relative time, battery voltage, and battery current at the start of the battery's charging state, respectively. The data preprocessing module sequentially performs timestamp alignment, outlier removal, sequence downsampling, sensor channel stitching, and charge / discharge cycle stitching on the received sensor signals. The sensor signals are processed by the preprocessing module to form data samples, denoted as... , This represents the number of consecutive charge-discharge cycles contained in a sample, for any one of those cycles. Its corresponding vector All data samples are obtained by downsampling and horizontal stitching of sensor signals from the current charging cycle; multiple data samples are merged to form a battery degradation sample set. .

[0007] The hierarchical stochastic spatiotemporal Transformer module is based on a stochastic self-attention layer with uncertainty quantification capabilities, and sequentially performs spatial degradation pattern extraction and temporal degradation pattern extraction; the stochastic self-attention layer is a dot product self-attention calculation module based on the Gumbel normalized exponential function resampling method; Based on the traditional deterministic self-attention calculation mechanism, randomization is achieved by introducing a resampling method based on the Gumbel normalized exponential function, and further uncertainty quantification is realized; for data samples The query is obtained through a linear transformation. ,key Sum ,in, This represents the head index in a multi-head self-attention mechanism, and the random self-attention layer. The calculation method is as follows: in, This represents the Gumbel normalized exponential function. The temperature coefficient is used to replace the scaling factor in traditional sub-attention mechanisms. It is set to a larger value at the beginning of training and gradually decreases as training progresses. The Gumbel normalization exponential function relies on Gumbel noise in its implementation, and is defined as follows: in It is the column index of the column vector obtained by dividing the input data sample matrix by column. It is a random variable that follows a uniform distribution; the calculation process of stochastic self-attention based on the Gumbel normalized exponential function is as follows: in, It is the first in the attention score vector Each component.

[0008] The spatial degradation pattern extractor is as follows: The spatial degradation pattern extractor uses convolution operations to extract spatial dependency information of the sequence at each time step and models the inter-sensor channel dependency relationship for a single charge-discharge cycle. The spatial degradation pattern extractor is based on a dilated convolutional layer with a kernel size of 2 and a dilation rate of 2. It processes the charge-discharge cycle sequence after being weighted by a random self-attention layer, and the convolutional feature map is flattened to form a spatial feature vector representing the time step.

[0009] In the temporal degradation pattern extractor, the pre-clustering submodule clusters the spatial feature vectors column-wise to obtain a set of trainable cluster centers. Based on a stochastic self-attention mechanism, the input key sequence is dynamically assigned to the cluster centers, forming a compact representation of the key space. When calculating the attention between the query and the clustered keys, the temporal step distance-sensitive position encoding submodule introduces a learnable weighting function related to the actual distance of the cycle period, dynamically adjusting the attention score. Finally, the multi-head stochastic self-attention submodule includes multiple stochastic self-attention layers, generating the final attention weights that simultaneously consider feature clustering and temporal distance, and aggregating the value sequence accordingly to obtain the final spatiotemporal feature representation.

[0010] Specifically, the pre-clustering submodule involves the following process: from the input sample being projected onto the query, key, and value, to the calculation process of the pre-clustering submodule: in, , and Let be a mapping matrix, where the superscripts are the head indices of the multi-head self-attention mechanism. These are predefined cluster centers.

[0011] The attention calculation method obtained after encoding by the time-step distance-sensitive position encoding submodule is as follows: in, It represents the Hadamah accumulation. Represents the ReLU activation function. For the mapping matrix, For temperature coefficient, This is the parameter matrix used for distance reweighting, defined as follows: in, , , Used to characterize the index in the sequence and The absolute distance between the data.

[0012] The uncertainty calibration auxiliary module is described as follows: First, construct a calibration dataset. The calibration dataset consists of two parts: one part is the sequence of cumulative distribution function values ​​predicted by the initial hierarchical stochastic spatiotemporal Transformer for each sample in the calibration dataset before calibration. The other part is the sequence of empirical cumulative distribution function values ​​corresponding to each sample in the calibration dataset. The empirical cumulative distribution function value is obtained based on the actual distribution statistics of the true health status in the calibration dataset; the predicted cumulative distribution function... The definition is as follows: in, Indicate the indicator function; train an order-preserving regressor based on the above calibration dataset. A monotonically non-decreasing mapping function is learned; this mapping function takes the original cumulative distribution function value predicted by the uncalibrated hierarchical stochastic spatiotemporal Transformer as input and outputs a calibrated cumulative distribution function value; thus, the calibrated cumulative distribution function value is made as close as possible to the empirical cumulative distribution function value; finally, when estimating the health status of new test samples, the uncalibrated hierarchical stochastic spatiotemporal Transformer first outputs the original point estimate and uncertainty, and then transforms this original predicted cumulative distribution function through a trained ordinal-preserving regressor to obtain the calibrated cumulative distribution function. Based on the calibrated cumulative distribution function, a battery health status estimation model with unbiased uncertainty quantification capability is obtained. .

[0013] The calculation process for the battery health state estimation results and the corresponding uncertainty quantification results is as follows: Based on Monte Carlo sampling, through The battery health status estimation result is obtained from the next forward propagation. and its corresponding uncertainty measure They are defined as follows: in It is a set of all sampled model parameters. Based on the obtained health state estimation results and the corresponding uncertainty measure, the hierarchical stochastic spatiotemporal Transformer gives the interval estimation results of battery health state. .

[0014] The beneficial effects of this invention are as follows: This invention designs a battery health state estimation method with unbiased uncertainty quantification capability, and innovatively proposes a data-driven battery health state estimation based on a stochastic self-attention mechanism using the Gumbell normalized exponential function to achieve uncertainty perception, while considering both temporal and spatial uncertainties in the battery degradation process; the obtained uncertainties are calibrated based on ordinal regression to ensure that the quantified uncertainties are unbiased; and the overall credibility of the battery health state estimation results is significantly improved, providing more solid and reliable data support for battery safe use, life management, and maintenance decisions. Attached Figure Description

[0015] Figure 1 This is an overall framework diagram of a battery health state estimation method with unbiased uncertainty quantification capability provided by an embodiment of the present invention.

[0016] Figure 2 This is a schematic diagram of a time-step dilated convolution operation provided in an embodiment of the present invention.

[0017] Figure 3 This is a schematic diagram of a pre-clustering process provided in one embodiment of the present invention.

[0018] Figure 4 This is a schematic diagram of distance-sensitive battery cycle signal position encoding provided in an embodiment of the present invention; (a) is a schematic diagram of battery cycle signal, and (b) is a schematic diagram of position encoding between cycles after distance-sensitive position encoding. Detailed Implementation

[0019] The following will refer to the appendix. Figures 1 to 4Specific embodiments of this disclosure are described in detail. While specific embodiments of this disclosure are shown in the accompanying drawings, it should be understood that this disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of this disclosure to those skilled in the art.

[0020] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions are preferred embodiments for carrying out the invention; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of the invention. The scope of protection of this disclosure is determined by the appended claims.

[0021] To facilitate understanding of the embodiments of this disclosure, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. The accompanying drawings do not constitute a limitation on the embodiments of this disclosure.

[0022] In one embodiment, such as Figure 1 As shown, the present invention provides a battery health state estimation method with unbiased uncertainty quantification capability, comprising three parts: a data preprocessing module, a hierarchical stochastic spatiotemporal Transformer module, and an uncertainty calibration auxiliary module.

[0023] The data preprocessing module receives timing signals from the sensor modules of the battery management system during the battery charging and discharging process. These sensors include timers, voltage sensors, and current sensors, which record the relative time, battery voltage, and battery current at the start of the charging state, respectively. The data preprocessing module sequentially performs timestamp alignment, outlier removal, sequence downsampling, sensor channel stitching, and charge / discharge cycle stitching on the received sensor signals. The original sensor signals are processed by the preprocessing module to form data samples, denoted as... ,in This represents the number of consecutive charge-discharge cycles contained in a sample, for any one of those cycles. Its corresponding vector All data samples are obtained by downsampling and horizontal stitching of sensor signals from the current charging cycle; multiple data samples are merged to form a battery aging sample set. ,in Indicates the first Characteristics of the sample Indicates the first The label of each sample indicates the battery health state corresponding to the last charge-discharge cycle of that sample. It is the total number of degraded samples in the sample set; the battery degradation samples sampled from this sample set are used for training the hierarchical stochastic spatiotemporal Transformer module and the uncertainty calibration auxiliary module.

[0024] The hierarchical stochastic spatiotemporal Transformer module is based on a stochastic self-attention mechanism with uncertainty quantification capability. This stochastic self-attention mechanism with uncertainty quantification capability is a variant of the self-attention mechanism. By introducing the Gumbel-Softmax Reparameterization (GSR) method, the deterministic self-attention calculation process is transformed into a probabilistic self-attention calculation process with uncertainty quantification capability.

[0025] The hierarchical random spatiotemporal Transformer module This includes a spatial degradation pattern extractor for extracting spatial degradation dependencies and a temporal degradation pattern extractor for extracting temporal degradation dependencies. The spatial degradation pattern extractor takes the battery degradation sample set output by the data preprocessing module as input, performs random self-attention calculation on the samples to obtain sensor channel weights, and uses dilated convolutional layers to process the degradation data with sensor channel weights to obtain convolutional feature maps that reveal the spatial degradation information. During this process, the pre-clustering submodule and the temporal step distance-sensitive location encoding submodule remain closed. The convolutional feature maps are flattened and vertically stacked to form new feature maps. The degradation time information is fed into the temporal degradation pattern extractor. During the temporal degradation pattern extraction process, the pre-clustering submodule and the time-step distance-sensitive position encoding submodule remain on to enhance degradation-related information and target cycle-related information in the feature map. The temporal degradation pattern extractor has a residual connection structure. The input of this residual connection structure is added and normalized after passing through a multi-head random self-attention submodule and a direct mapping to generate the final hierarchical spatiotemporal degradation pattern. This hierarchical spatiotemporal degradation pattern is then passed through a multilayer perceptron regressor containing three linear layers to generate battery health state point estimates.

[0026] The uncertainty calibration auxiliary module is used to correct possible deviations between the model prediction confidence level and the actual error distribution, achieving unbiased uncertainty quantification; the uncertainty calibration auxiliary module relies on a calibration dataset. The supervised correction process uses hierarchical random spatiotemporal Transformer modules to correct samples in the correction dataset. The output predicted cumulative distribution function is used as a feature, based on the aging sample set. The true cumulative distribution function of the degraded samples is calculated as the label; based on an order-preserving regressor trained using a calibration dataset. The uncertainty calibration auxiliary module calibrates the hierarchical stochastic spatiotemporal Transformer module, and the battery health state estimator with unbiased uncertainty quantification capability is finally expressed as follows: .

[0027] The stochastic self-attention mechanism with uncertainty quantification capability is described as follows: Stochastic self-attention with uncertainty quantification capability achieves randomization and further uncertainty quantification by introducing a GSR process based on the traditional deterministic self-attention calculation mechanism; for those with Degradation features of each component after processing by the data preprocessing module The query is obtained through a linear transformation. Key Sum (Value) ,in This represents the head index in a multi-head self-attention mechanism, or random self-attention. The calculation is as follows: in This represents the Gumbel normalized exponential function. The temperature coefficient is used to replace the scaling factor in traditional sub-attention mechanisms. This temperature coefficient is set to a large value at the beginning of training and gradually decreased as training progresses to facilitate model convergence to the global minimum. The Gumbel normalization exponential function relies on Gumbel noise in its implementation, and is defined as follows: in It is the column index of the column vector obtained by dividing the input degradation feature matrix into columns. It is a random variable that follows a uniform distribution; based on this, the calculation process of stochastic self-attention based on the Gumbel normalized exponential function is as follows: in It is the first in the attention score vector Each component.

[0028] The spatial degradation pattern extractor of the hierarchical stochastic spatiotemporal Transformer module is described as follows: The proposed spatial degradation pattern extractor uses convolution operations to extract spatial dependency information for the sequence at each time step, that is, to model the inter-channel dependencies of a single charge-discharge cycle. This spatial degradation pattern extractor is based on a dilated convolutional network with a kernel size of 2 and a dilation rate of 2. After random self-attention weighting, the explicit channel importance is represented as a multivariate time series charge-discharge cycle sequence. The processed convolutional feature map is flattened to form a vector representing the spatial features of that time step.

[0029] Figure 2 This is a schematic diagram of the proposed time-step dilated convolution operation. Figure 2 This diagram visually demonstrates a sparse sampling mechanism based on a convolutional kernel size of 2 and a dilation rate of 2. The connecting lines shown in the figure indicate that the convolutional kernel does not cover consecutive adjacent feature points, but rather performs skip sampling on the input multivariate time series according to a set dilation interval. This sampling method allows the network to effectively expand the receptive field while maintaining a constant parameter size, thereby capturing a wider range of inter-channel dependencies within a single charge-discharge cycle. After processing by this dilated convolutional structure, the original channel correlations are mapped to a high-level abstract feature map, and finally transformed into a one-dimensional feature vector containing global spatial dependency information through a flattening operation.

[0030] The temporal degradation pattern extractor of the hierarchical stochastic spatiotemporal Transformer module is described as follows: The temporal degradation pattern extractor first introduces a pre-clustering submodule to reduce the computational complexity of the degradation pattern extraction process. The pre-clustering submodule initializes a set of trainable pre-clustering centers and dynamically assigns the input key sequence to these pre-clustering centers, forming a compact representation of the key space. Second, a temporal step distance-sensitive location encoding submodule is constructed to more accurately model the impact of the actual time interval between different cycle periods on the degradation trend. Specifically, when calculating the attention between the query and the clustered keys, the temporal step distance-sensitive location encoding submodule introduces a learnable weighting function related to the actual distance of the cycle period, which dynamically adjusts the attention score. Finally, a stochastic self-attention mechanism is applied to generate the final attention weights that simultaneously consider feature clustering and temporal distance, and the aggregated value sequence is obtained based on these weights to obtain the final spatiotemporal feature representation.

[0031] The process from the input sample being projected onto the query, key, and value, to the computation of the pre-clustering submodule, is as follows: in, , and Let be a mapping matrix, where the superscripts are the head indices of the multi-head self-attention mechanism. These are predefined cluster centers.

[0032] Figure 3 This diagram illustrates the proposed pre-clustering process, visually demonstrating the transformation from discrete key sequences to compact feature representations within a two-dimensional feature space. The light-colored triangular scatter points represent the original input key sequences, exhibiting a relatively broad and sparse distribution before clustering. The dark squares represent a set of trainable pre-cluster centers initialized by the sub-module. The pre-clustering sub-module dynamically distributes and aggregates the originally scattered key sequences to the vicinity of the corresponding center points by measuring the spatial distance or similarity between the input sequence and each center point, thus forming the post-clustering feature distribution shown by the dark circles in the diagram. This mechanism successfully maps the high-dimensional and discrete key space into a compact feature representation, effectively reducing the data processing dimensionality and computational complexity of the subsequent degradation pattern extraction process while preserving the key feature structure.

[0033] The attention calculation method obtained after encoding by the time-step distance-sensitive position encoding submodule is as follows: in, It represents the Hadamah accumulation. Represents the ReLU activation function. For the mapping matrix, For temperature coefficient, This is the parameter matrix used for distance reweighting, defined as follows: in, , , Used to characterize the index in the sequence and The absolute distance between the data.

[0034] Figure 4 This is a schematic diagram of the proposed distance-sensitive battery cycle signal position encoding. (a) shows the input data containing multiple sensor channels, each sensor channel containing process quantities at several time steps, and the timestamps of each process quantity are aligned. (b) shows the process quantity sequence weights after being encoded by the distance-sensitive position encoding submodule. These weights decrease as the time step distance increases.

[0035] The uncertainty calibration auxiliary module is described as follows: First, construct a calibration dataset. The dataset consists of two parts: first, the sequence of cumulative distribution function values ​​predicted by the model for each sample in the calibration dataset when the model is not calibrated; and second, the sequence of cumulative distribution function values ​​predicted by the model for each sample in the calibration dataset. Secondly, the sequence of empirical cumulative distribution function values ​​corresponding to these samples. The cumulative distribution function predicted by the model reflects the probability that the model considers the true health state to be less than or equal to a certain specific value. The empirical cumulative distribution function value is obtained by statistically analyzing the actual distribution of the true health state in the calibration dataset. The definition is as follows: in The first step is to define the indicator function; the second step is to train an order-preserving regressor using the aforementioned calibration dataset. A monotonically non-decreasing mapping function is learned, which takes the original cumulative distribution function (CDF) value predicted by the model as input and outputs a calibrated CDF value. This ensures that the calibrated CDF value is as close as possible to the empirical CDF value. Finally, when estimating the health status of new test samples, the model first outputs the original point estimate and uncertainty. Then, this original predicted CDF value is transformed through a trained ordinal-preserving regressor to obtain the calibrated CDF value. Based on the calibrated CDF value, a battery health status estimation model with unbiased uncertainty quantification capability can be obtained. .

[0036] The calculation process for the battery health state estimation results and the corresponding uncertainty quantification results given by the model is as follows: The model is based on Monte Carlo sampling, through The battery health status estimation result is obtained from the next forward propagation. and its corresponding uncertainty measure They are defined as follows: in It is the set of all sampled model parameters. Based on the obtained health state estimation results and the corresponding uncertainty measure, the model gives the battery health state interval estimation results as follows: .

[0037] Table 1

[0038] Table 2 The technical solution of this invention is illustrated below through comprehensive verification experiments conducted on four publicly available lithium battery aging datasets from MIT, TRI, TJU, and XJTU. The datasets used are shown in Table 1. First, the method proposed in this invention is compared with seven existing state-of-the-art methods, including MRVR, GP-ICE, BMNN, SACN, DeNN, BDAE, and SNGP. To ensure the fairness of the experiment, all methods used the same hyperparameter settings, and the estimation accuracy and uncertainty quantification quality of all methods are shown in Table 2. Experimental results show that the method proposed in this invention exhibits significant advantages in all evaluation metrics on all datasets, including Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Negative Log-Likelihood (NLL), and Expected Calibration Error (ECE). For example, on the MIT dataset, compared to suboptimal models, the proposed method reduces MAE by 18.10%, RMSE by 9.88%, NLL by 23.48%, and ECE by 40.26%, fully demonstrating the effectiveness and superiority of the proposed method in terms of battery health estimation accuracy and uncertainty quantification quality. Furthermore, through comparison with different stochastic attention mechanisms such as MCDT, VIT, and GSRT, the proposed method maintains high performance while exhibiting competitive computational efficiency, making it suitable for industrial applications. Secondly, detailed ablation experiments and calibration effect visualization analysis were conducted. The ablation experiments verified the positive contributions of the GSR-based stochastic self-attention, hierarchical spatiotemporal feature representation structure, and various components (such as distance sensitivity and clustering mechanism) in the hierarchical self-attention mechanism. The calibration effect visualization results show that after processing with the proposed principled calibration strategy, the model's prediction confidence is more consistent with the actual observation frequency, and the calibration curve is closer to the ideal diagonal, significantly alleviating the underconfidence or overconfidence problem of the model; for example, the average ECE is reduced by 64.1%. Furthermore, experimental studies on in-distribution and out-of-distribution samples demonstrate that the proposed method can effectively distinguish between in-distribution and out-of-distribution samples based on their prediction uncertainty, further confirming its ability to provide reliable health status estimates. Finally, to explore the performance of the proposed method under different parameter settings, sensitivity analysis was conducted on its key hyperparameters, including the number of encoder layers, the number of attention heads, the attention embedding dimension, and the attention temperature factor. Experimental results show that the proposed method maintains good performance within a reasonable range of hyperparameters and can find the optimal parameter combination to balance model complexity and prediction accuracy.For example, when the number of encoder layers is 1, the number of attention heads is 2, the embedding dimension is 16, and the temperature factor is 1.0, the model can achieve good RMSE performance on multiple datasets, demonstrating the adaptability and robustness of the proposed model under different parameter configurations.

[0039] The specific embodiments described above are merely illustrative and instructive, not restrictive. Those skilled in the art, guided by this specification and without departing from the scope of the claims, can make many other variations, all of which are protected by this invention.

Claims

1. A battery health state estimation method with unbiased uncertainty quantification capability, characterized in that, It includes a data preprocessing module, a hierarchical stochastic spatiotemporal Transformer module, and an uncertainty calibration auxiliary module; The data preprocessing module receives timing signals from the battery charging and discharging process obtained by the sensor module of the battery management system, and processes them to obtain a battery degradation sample set. , Indicates the first Characteristics of the sample Indicates the first The label for each sample represents the battery health state corresponding to the last charge-discharge cycle of that sample. It is the total number of degraded samples in the sample set; The hierarchical random spatiotemporal Transformer module It includes a spatial degradation pattern extractor for extracting degradation spatial dependencies and a temporal degradation pattern extractor for extracting degradation temporal dependencies; the spatial degradation pattern extractor takes the battery degradation sample set as input and calculates the sensor channel weights based on the weights obtained from the random self-attention layer; A dilated convolutional layer processes degraded data with sensor channel weights to obtain a convolutional feature map that reveals the spatial information of degradation. During the spatial degradation pattern extraction process, the pre-clustering submodule and the time-step distance-sensitive position encoding submodule remain off. The convolutional feature map is flattened and vertically stacked before being fed into the temporal degradation pattern extractor to reveal the temporal information of degradation. The temporal degradation pattern extractor includes a pre-clustering submodule, a time-step distance-sensitive position encoding submodule, and a multi-head random self-attention submodule. During the temporal degradation pattern extraction process, the pre-clustering submodule and the time-step distance-sensitive position encoding submodule remain on. The temporal degradation pattern extractor has a residual connection structure. The inputs are respectively processed by the multi-head random self-attention submodule and direct mapping, then added and normalized to generate the final hierarchical spatiotemporal degradation pattern. The hierarchical spatiotemporal degradation pattern is processed by a multilayer perceptron regressor to generate battery health state point estimates. The uncertainty calibration auxiliary module is used to correct the deviation between the prediction confidence of the hierarchical stochastic spatiotemporal Transformer module and the actual error distribution, and to perform unbiased uncertainty quantification; the uncertainty calibration auxiliary module relies on a calibration dataset. The supervised correction process uses hierarchical random spatiotemporal Transformer modules to correct samples in the correction dataset. The output predicted cumulative distribution function is used as a feature, based on the battery degradation sample set. The true cumulative distribution function of the degraded samples is calculated as the label; Based on an order-preserving regressor trained using a calibration dataset The uncertainty calibration auxiliary module calibrates the hierarchical stochastic spacetime Transformer module, and the hierarchical stochastic spacetime Transformer module with unbiased uncertainty quantification capability is finally expressed as follows: ,in The function composition is represented, and the output is the estimated battery health state and its corresponding uncertainty.

2. The battery health state estimation method with unbiased uncertainty quantification capability according to claim 1, characterized in that, The sensor module includes a timer, a voltage sensor, and a current sensor, which record the relative time, battery voltage, and battery current at the start of the battery's charging state, respectively. The data preprocessing module sequentially performs timestamp alignment, outlier removal, sequence downsampling, sensor channel stitching, and charge / discharge cycle stitching on the received sensor signals. The sensor signals are processed by the preprocessing module to form data samples, denoted as... , This represents the number of consecutive charge-discharge cycles contained in a sample, for any one of those cycles. Its corresponding vector All data samples are obtained by downsampling and horizontal stitching of sensor signals from the current charging cycle; multiple data samples are merged to form a battery degradation sample set. .

3. The battery health state estimation method with unbiased uncertainty quantification capability according to claim 1, characterized in that, The hierarchical stochastic spatiotemporal Transformer module is based on a stochastic self-attention layer with uncertainty quantification capabilities, and sequentially performs spatial degradation pattern extraction and temporal degradation pattern extraction; the stochastic self-attention layer is a dot product self-attention calculation module based on the Gumbel normalized exponential function resampling method; Based on the traditional deterministic self-attention calculation mechanism, randomization is achieved by introducing a resampling method based on the Gumbel normalized exponential function, and further uncertainty quantification is realized; for data samples The query is obtained through a linear transformation. ,key Sum ,in, This represents the head index in a multi-head self-attention mechanism, and the random self-attention layer. The calculation method is as follows: in, This represents the Gumbel normalized exponential function. The temperature coefficient is used to replace the scaling factor in traditional sub-attention mechanisms. It is set to a larger value at the beginning of training and gradually decreases as training progresses. The Gumbel normalization exponential function relies on Gumbel noise in its implementation, and is defined as follows: in It is the column index of the column vector obtained by dividing the input data sample matrix by column. It is a random variable that follows a uniform distribution; the calculation process of stochastic self-attention based on the Gumbel normalized exponential function is as follows: in, It is the first in the attention score vector Each component.

4. The battery health state estimation method with unbiased uncertainty quantification capability according to claim 1, characterized in that, The spatial degradation pattern extractor is as follows: The spatial degradation pattern extractor uses convolution operations to extract spatial dependency information of the sequence at each time step and models the inter-sensor channel dependency relationship for a single charge-discharge cycle. The spatial degradation pattern extractor is based on a dilated convolutional layer with a kernel size of 2 and a dilation rate of 2. It processes the charge-discharge cycle sequence after being weighted by a random self-attention layer, and the convolutional feature map is flattened to form a spatial feature vector representing the time step.

5. The battery health state estimation method with unbiased uncertainty quantification capability according to claim 1, characterized in that, In the temporal degradation pattern extractor, the pre-clustering submodule clusters the spatial feature vectors column-wise to obtain a set of trainable cluster centers; according to the random self-attention mechanism, the input key sequence is dynamically assigned to the cluster centers to form a compact representation of the key space; when calculating the attention between the query and the clustered keys, the time-step distance-sensitive position encoding submodule introduces a learnable weighting function related to the actual distance of the cycle period to dynamically adjust the attention score; Finally, the multi-head random self-attention submodule includes multiple random self-attention layers, which generate the final attention weights that simultaneously consider feature clustering and temporal distance, and aggregate the value sequence accordingly to obtain the final spatiotemporal feature representation.

6. The battery health state estimation method with unbiased uncertainty quantification capability according to claim 5, characterized in that, Specifically, the pre-clustering submodule involves the following process: from the input sample being projected onto the query, key, and value, to the calculation process of the pre-clustering submodule: in, , and Let be a mapping matrix, where the superscripts are the head indices of the multi-head self-attention mechanism. These are predefined cluster centers.

7. The battery health state estimation method with unbiased uncertainty quantification capability according to claim 5, characterized in that, The attention calculation method obtained after encoding by the time-step distance-sensitive position encoding submodule is as follows: in, It represents the Hadamah accumulation. Represents the ReLU activation function. For the mapping matrix, For temperature coefficient, This is the parameter matrix used for distance reweighting, defined as follows: in, , , Used to characterize the index in the sequence and The absolute distance between the data.

8. The battery health state estimation method with unbiased uncertainty quantification capability according to claim 1, characterized in that, The uncertainty calibration auxiliary module is described as follows: First, construct a calibration dataset. The calibration dataset consists of two parts: one part is the sequence of cumulative distribution function values ​​predicted by the initial hierarchical stochastic spatiotemporal Transformer for each sample in the calibration dataset before calibration. The other part is the sequence of empirical cumulative distribution function values ​​corresponding to each sample in the calibration dataset. The empirical cumulative distribution function value is obtained based on the actual distribution statistics of the true health status in the calibration dataset; the predicted cumulative distribution function... The definition is as follows: in, Indicate the indicator function; train an order-preserving regressor based on the above calibration dataset. A monotonically non-decreasing mapping function is learned; this mapping function takes the original cumulative distribution function value predicted by the uncalibrated hierarchical stochastic spatiotemporal Transformer as input and outputs a calibrated cumulative distribution function value; thus, the calibrated cumulative distribution function value is made as close as possible to the empirical cumulative distribution function value; finally, when estimating the health status of new test samples, the uncalibrated hierarchical stochastic spatiotemporal Transformer first outputs the original point estimate and uncertainty, and then transforms this original predicted cumulative distribution function through a trained ordinal-preserving regressor to obtain the calibrated cumulative distribution function. Based on the calibrated cumulative distribution function, a battery health status estimation model with unbiased uncertainty quantification capability is obtained. .

9. The battery health state estimation method with unbiased uncertainty quantification capability according to claim 1, characterized in that, The calculation process for the battery health state estimation results and the corresponding uncertainty quantification results is as follows: Based on Monte Carlo sampling, through The battery health status estimation result is obtained from the next forward propagation. and its corresponding uncertainty measure They are defined as follows: in It is a set of all sampled model parameters. Based on the obtained health state estimation results and the corresponding uncertainty measure, the hierarchical stochastic spatiotemporal Transformer gives the interval estimation results of battery health state. .

Images