Battery state of health assessment method and equipment considering causal relationship enhancement

Abstract

The invention provides a battery health state evaluation method and equipment considering causal relationship enhancement. The method comprises the following steps: acquiring battery measurement data, acquiring a multi-time-sequence nonlinear causal relationship between a battery health state and a health factor based on a causal inference method, and constructing a battery multi-time-sequence nonlinear causal relationship directed graph with generality; according to the battery multi-time-sequence nonlinear causal relationship directed graph and a battery health factor and health state deduction method based on causal accumulation response, establishing a causal relationship enhancement model based on causal-observation fusion deduction, and obtaining a battery multi-time-sequence nonlinear causal enhancement feature; the causal enhancement feature is used as input, and the causal enhancement-bidirectional gating cycle unit model is used for outputting the battery health state, so that the interpretability, accuracy and generalization of health state estimation are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of battery state of health assessment, and relates to a method and device for battery state of health assessment, and particularly to a method and device for battery state of health assessment with enhanced consideration of causal relationships. Background Art

[0002] With the rapid development of the new power system, in order to support the access of high-penetration renewable energy to the power system, various new energy storage technologies represented by batteries are more and more widely used. During the recycling process, the loss of ions and active materials inside the battery will cause the battery to age continuously and its performance to decline, leading to safety accidents such as battery leakage, insulation damage, and local short circuits. Accurately estimating the state of health of the battery and real-time monitoring of the battery capacity attenuation can provide parameter support for functions such as battery safety warning, fault diagnosis, and equalization grouping, and ensure the safe and reliable operation of the battery.

[0003] Existing battery state of health estimation methods mainly include model-based estimation methods and data-driven estimation methods. The model-based estimation method describes the battery aging behavior by building a state-space model. However, in actual situations, the electrochemical mechanism is complex and variable, and it is difficult to apply in practice. To avoid complex electrochemical mechanism analysis and solution, most current research uses data-driven methods to estimate the battery state of health. The data-driven method is to measure parameters such as the current, voltage, and temperature of the battery, extract health factors highly correlated with the change of the state of health, and use them as training data to build an estimation model, so as to achieve the estimation of the state of health.

[0004] Existing research shows that there are significant differences in the capacity degradation curves of batteries under different operating conditions, which puts higher requirements on the generalization ability of the data-driven model. Under the condition of sufficient data samples, the data-driven model based on correlation analysis constructed for a specific operating condition can guarantee the generalization ability and accuracy within the distribution to a certain extent. However, for scenarios with unknown operating conditions such as newly built energy storage power stations, the traditional data-driven model based on correlation analysis is prone to pseudo-correlation problems, resulting in insufficient interpretability of the model and affecting its generalization ability outside the distribution. To address the above problems, some research calculates the contribution degree of each health factor in the battery state of health estimation under different operating conditions based on the attention mechanism to improve the interpretability and generalization ability of the state of health estimation. However, the state of health estimation method based on the attention mechanism still relies on correlation analysis to weigh the contribution degree of health factors, and fails to fundamentally solve the pseudo-correlation problem, restricting the cross-condition generalization ability of battery state of health estimation.

[0005] In recent years, the theory of causal inference has provided new ideas for solving the above problems. Compared with the traditional data-driven method that only relies on correlation analysis, causal inference can reveal the internal action mechanism between variables, effectively avoid the problem of spurious correlation, and thus improve the interpretability and cross-operating condition generalization ability of battery state of health estimation. Causal inference has been applied to fields such as power load and new energy prediction, and power system state estimation. However, in the field of battery state of health estimation, existing research mainly applies causal inference to health factor screening, and fails to fully incorporate causal relationships containing deep causal information such as causal graphs and causal effects into the state of health estimation model, resulting in room for improvement in the accuracy and generalization ability of the state of health estimation method considering causal relationships. Therefore, it is necessary to develop a battery state of health estimation method considering enhanced causal relationships. Summary of the Invention

[0006] To solve the above problems, the present invention proposes a battery state of health assessment method and device considering enhanced causal relationships, uses the causal inference method to obtain the multi-temporal non-linear causal relationship between the battery state of health and health factors, generates battery multi-temporal non-linear causal enhanced features through causal-observation fusion deduction, and based on the battery multi-temporal non-linear causal enhanced features, estimates the battery state of health based on a bidirectional gated recurrent unit network to improve the interpretability, accuracy, and generalization of the state of health estimation.

[0007] The technical solution adopted by the present invention is as follows:

[0008] A battery state of health assessment method considering enhanced causal relationships, comprising the following steps: obtaining battery measurement data, obtaining the multi-temporal non-linear causal relationship between the battery state of health and health factors based on the causal inference method, and constructing a directed graph of the battery multi-temporal non-linear causal relationship with commonality; according to the directed graph of the battery multi-temporal non-linear causal relationship and the battery health factor and state of health deduction method based on causal cumulative response, establishing a causal relationship enhancement model based on causal-observation fusion deduction to obtain battery multi-temporal non-linear causal enhanced features; using the causal enhanced features as input and outputting the battery state of health using the causal enhancement - bidirectional gated recurrent unit model.

[0009] Further, the battery measurement data includes the mean, minimum value, maximum value, standard deviation, and integral of voltage, current, and temperature measurement data in each charge and discharge cycle, as well as the time used for each charge and discharge cycle.

[0010] Further, the multi-temporal non-linear causal relationship between the battery health state and the health factors obtained by the causal inference method specifically includes: using the latent Peter-Clark instantaneous conditional independence algorithm to discover the causal graph between the battery health state and the health factors; using the optimal adjustment set theory to estimate the non-linearity causal effect between the battery health state and the health factors according to the causal graph, and combining the causal graph discovery result and the causal effect estimation result to obtain the multi-temporal non-linear causal relationship between the battery health state and the health factors.

[0011] Further, the directed graph of the multi-temporal non-linear causal relationship of the battery with commonality includes a node set and an edge set. The node set consists of a set of health factors with common causality and a battery health state index, and the edge set describes the multi-temporal non-linear common causal relationship between the battery health state and the health factors.

[0012] Further, according to the directed graph of the multi-temporal non-linear causal relationship of the battery and the battery health factor and health state deduction method based on causal cumulative response, a causal relationship enhancement model based on causal-observational fusion deduction is established to obtain the multi-temporal non-linear causal enhancement feature of the battery. The specific steps include: establishing a battery health state and health factor deduction model based on causal cumulative response according to the directed graph of the multi-temporal non-linear causal relationship of the battery, and using the deduction model to obtain the deduction values of the health state and health factors in each charge and discharge cycle according to the historical health state and health factor observation values; aiming at minimizing the difference between the health factor deduction value and the health factor observation value, introducing the optimal health state proxy value to replace the unknown health state observation value, and constructing a causal relationship enhancement model based on causal-observational fusion deduction; solving the causal relationship enhancement model to obtain the deduction values of the battery health state and health factors, which together constitute the multi-temporal non-linear causal enhancement feature of the battery.

[0013] Further, the deduction model starts from the health state and health factor observation values of the first charge and discharge cycle, and sequentially calculates the average causal response of each health state and health factor in each charge and discharge cycle affected by all other health states or health factors at historical moments according to the time sequence.

[0014] Further, a loss function based on the invariant risk minimization theory is adopted during the training process of the causal enhancement - bidirectional gated recurrent unit model.

[0015] Further, before the causal enhancement feature is input, it needs to go through domain calibration batch normalization processing, and each different operating condition has independent batch normalization parameters.

[0016] Further, for battery samples with known operating conditions, directly apply the batch normalization parameters corresponding to the operating conditions for processing; for battery samples with unknown operating conditions, calculate the similarity distances between them and all known operating conditions, and select the batch normalization parameters corresponding to the operating condition with the smallest distance for processing.

[0017] A computer device, comprising:

[0018] One or more processors;

[0019] A memory for storing one or more programs;

[0020] When the one or more programs are executed by the one or more processors, the one or more processors implement the above battery health state assessment method considering causal relationship enhancement.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Traditional data-driven methods based on correlation analysis have problems of insufficient interpretability and are difficult to ensure the accuracy and generalization of battery health state estimation outside the training samples. The present invention establishes a battery causal relationship enhancement model based on causal-observational fusion deduction, constructs multi-temporal non-linear causal enhancement features from battery measurement data, and improves the interpretability and in-distribution generalization of health state estimation.

[0022] 2. Based on the multi-temporal non-linear causal enhancement features of the battery, the present invention proposes a lithium-ion battery health state estimation method based on a bidirectional gated recurrent unit network and the theory of invariant risk minimization, further improving the accuracy and out-of-distribution generalization of health state estimation. Description of the Drawings

[0023] Figures 1 to 4 It is a multi-temporal non-linear causal relationship diagram between the inferred health states of different batteries and the battery health factors in the embodiments of the present invention.

[0024] Figure 5 It is a comparison diagram of the actual observed values and the causal-observational fusion deduction results of the battery health state and the integral of the discharge current of Battery B0005 in the embodiments of the present invention.

[0025] Figures 6 to 7 It is a comparison diagram of the t-distributed stochastic neighbor embedding dimensionality reduction visualization results of the original health factor feature distributions and the causal enhancement feature distributions of different batteries in the embodiments of the present invention (t-SNE is the t-distributed stochastic neighbor embedding).

[0026] Figures 8 to 15Comparison diagram of health state estimation curves of the causal enhancement - bidirectional gated recurrent unit and comparative methods in 29 test batteries in the embodiments of the present invention (in the figure, SVR, LSTM, BRU, BiLSTM, BiGRU, GC - LSTM, CE - BiGRU are support vector regression, long short - term memory network, gated recurrent unit, bidirectional long short - term memory network, bidirectional gated recurrent unit, Granger causality - long short - term memory network, and causal enhancement - bidirectional gated recurrent unit respectively).

[0027] Figure 16 Comparison diagram of health state estimation results of battery B0042 by different methods in the embodiments of the present invention.

[0028] Figures 17 to 19 Comparison diagram of health state estimation results of three typical working conditions of batteries by different methods in the embodiments of the present invention. Detailed implementation manners

[0029] Taking a lithium - ion battery as an example below, the technical solution of the present invention will be further clearly and detailedly described with reference to the accompanying drawings.

[0030] A battery health state assessment method considering causal relationship enhancement includes the following steps: obtaining battery measurement data, obtaining the multi - time - series non - linear causal relationship between the battery health state and health factors based on a causal inference method, and constructing a directed graph of the battery's multi - time - series non - linear causal relationship with commonality; according to the directed graph of the battery's multi - time - series non - linear causal relationship and the battery health factor and health state deduction method based on causal cumulative response, establishing a causal relationship enhancement model based on causal - observation fusion deduction to obtain the battery's multi - time - series non - linear causal enhancement features; using the causal enhancement features as input and outputting the battery health state by using the causal enhancement - bidirectional gated recurrent unit model.

[0031] First, to improve the interpretability and in - distribution generalization of battery health state estimation, the present invention studies the multi - time - series non - linear causal relationship between battery measurement data and battery health state, establishes a battery causal relationship enhancement model based on causal - observation fusion deduction, and thus constructs battery multi - time - series non - linear causal enhancement features. The specific method is as follows:

[0032] Related research on the process of lithium-ion capacity decay shows that operating conditions such as voltage, current, temperature, and time during the lithium-ion charge and discharge cycles have important effects on the battery health state over the entire life cycle. Specifically, too high a battery operating temperature will accelerate the side reactions inside the battery, while too low a temperature will lead to lithium deposition and loss of active substances; too high a battery charging voltage will cause overcharging and trigger problems such as electrolyte decomposition and lithium deposition, and too low a discharge cut-off voltage will result in over-discharge and loss of active substances; too high a battery charge and discharge rate will also lead to loss of active substances and an increase in temperature, thus accelerating the internal side reactions. Based on the above analysis, to avoid insufficient interpretability of causal relationships caused by complex processing of battery measurement data, starting from the basic statistical characteristics of voltage, current, temperature, and time measurement data, the mean, minimum value, maximum value, standard deviation, and integral of the voltage, current, and temperature measurement data in each charge and discharge cycle, as well as the time taken for each charge and discharge cycle, are extracted to form the time series data of the battery health factors to be screened for subsequent causal inference and battery health state estimation. The time series data of the battery health factors to be screened for battery sample i can be obtained as follows :

[0033]

[0034] where is the nth health factor data in the tth charge and discharge cycle of battery sample i; is the set of charge and discharge cycles of battery sample i; is the set of battery health factors to be screened; is the set of battery samples.

[0035] To study the multi-temporal non-linear causal relationship between the battery health state and battery health factors, an appropriate causal inference method needs to be adopted to analyze their causal relationship. The capacity degradation process of lithium-ion batteries is usually accompanied by continuous shedding and insertion on the solid electrodes and the simultaneous occurrence of positive electrode oxidation reactions and negative electrode reduction reactions, which may result in a multi-temporal non-linear causal relationship between the battery health state and battery health factors that includes unobserved health factors. Causal inference generally includes two steps: causal graph discovery and causal effect estimation. For the causal graph discovery between the battery health state and battery health factors, time series variable causal discovery methods such as the method based on Granger causality and the method based on constraints can be used. Considering that the simple causal graph obtained by the method based on Granger causality usually focuses on the causal relationship between the battery health state and battery health factors from an overall perspective, while the window causal graph obtained by the method based on constraints can clearly show the dynamic causal relationship between the battery health state and battery health factors over multiple time series, the method based on constraints is more suitable for the discovery of the multi-temporal causal graph between the health state and battery health factors during battery degradation. The potential Peter-Clark instantaneous conditional independence algorithm based on constraints is an extension of the Peter-Clark algorithm and the instantaneous conditional independence test in the presence of potential unobserved variables, and can effectively discover the multi-temporal causal graph of the battery under the influence of potential unobserved health factors. For the causal effect estimation between the battery health state and battery health factors, causal effect estimation methods such as the method based on propensity scores and the method based on regression can be used. Considering that the estimation results of the method based on propensity scores usually show the expected probability of the values of the battery health state or battery health factors, while the method based on regression can clearly estimate the non-linear influence of battery health factors on the values of the battery health state, the method based on regression is more interpretable in the non-linear causal effect estimation between the battery health state and battery health factors. The optimal adjustment set theory based on regression can, based on a non-linear non-parametric estimator, achieve accurate estimation of the non-linear causal effect between the battery health state and battery health factors by minimizing the asymptotic variance of the causal effect estimation.

[0036] In summary, the present invention adopts a battery causal inference method based on the latent Peter-Clark instantaneous conditional independence algorithm and the optimal adjustment set theory to obtain the multi-temporal non-linear causal relationship between the battery health state and the battery health factors without observing the influence of health factors, providing a basis for subsequent construction of multi-temporal non-linear causal enhanced features of the battery. Among them, the discovery of the battery causal graph based on the latent Peter-Clark instantaneous conditional independence algorithm is mainly achieved through three main steps: First, the possible causal connections between the battery health state and the battery health factors to be screened are initially identified through the conditional independence test based on the non-linear Gaussian process distance correlation. Secondly, the causal direction is determined based on the instantaneous conditional independence test and the time series relationship. Finally, combining the results of the first two steps, a complete battery multi-temporal causal graph considering the influence of unobserved health factors is constructed. The estimation of the battery causal effect based on the optimal adjustment set theory is mainly achieved through three steps: First, all battery health factors with a deterministic causal path to the battery health state are screened according to the battery multi-temporal causal graph obtained by the latent Peter-Clark instantaneous conditional independence algorithm. Secondly, the optimal adjustment set that can simultaneously block the non-causal paths of the battery and minimize the estimation variance is identified by applying the optimal adjustment set theory, and the non-linear causal effect between the screened battery health factors and the battery health state is estimated using a random forest non-linear non-parametric regression estimator. Finally, combining the results of the battery multi-temporal causal graph discovery and the non-linear causal effect estimation results, a multi-temporal non-linear causal relationship between the battery health state and the battery health factors is formed.

[0037] For the i-th battery sample, according to the battery causal inference method based on the latent Peter-Clark instantaneous conditional independence algorithm and the optimal adjustment set theory, the battery health factors with a causal path to the battery health state can be screened, and the multi-temporal non-linear causal relationship between the battery health state and the screened battery health factors can be obtained. The above process expression is as follows:

[0038] In the formula, is used to represent the battery causal inference method based on the latent Peter-Clark instantaneous conditional independence algorithm and the optimal adjustment set theory; is the time series data of the health factors with a deterministic causal path to the battery health state after screening by the causal inference method for the i-th battery sample, is the set of battery health factors with a deterministic causal path to the battery health state for the i-th battery sample; is the actual health state time series data of the i-th battery sample; is the directed graph of the multi-temporal non-linear causal relationship between the battery health state and the health factors obtained after identification by the causal inference method for the i-th battery sample.

[0039] Considering that in actual situations, new batteries cannot provide time-series data of the state of health for inferring the causal relationship between the state of health and health factors, relevant research shows that there are partial common causal relationships between the state of health and health factors of batteries under different operating conditions. Therefore, a common causal relationship directed graph can be extracted from the causal relationship directed graphs of different battery samples with known time-series data of the state of health to obtain a common causal relationship directed graph . Estimating the state of health of the battery based on the common causal relationship helps to improve the in-distribution generalization ability. Based on this hypothesis, a common battery multi-time-series non-linear causal relationship directed graph can be obtained as follows:

[0040]

[0041] In the formula, the node set of the common battery multi-time-series non-linear causal relationship directed graph is composed of the set of screened health factors with common causality and the battery state of health index . Among them is the set of screened health factors derived from each battery sample i; the edge set describes the multi-time-series non-linear common causal relationship between the battery state of health and health factors represents the non-linear common causal effect with a time lag of τ from the m-th index to the n-th index of the battery. The estimator characterizes the non-linear common causal effect by calculating the average value of the non-linear non-parametric estimators based on random forests in different battery samples. It quantifies the expected value of the n-th index corresponding to the value of the m-th index of the in-distribution battery after a time lag of τ; is the common causal threshold, which characterizes the minimum threshold of the proportion of the number of occurrences of a certain causal relationship in the battery sample .

[0042] Considering that existing research mainly applies causal inference theory to health factor screening and fails to fully incorporate causal relationships containing deep causal information such as causal graphs and causal effects into the health state estimation model, the present invention adopts a method for deducing the battery state of health and health factors based on causal cumulative response to fully capture the battery multi-time-series non-linear causal relationship directed graph Coupling effect of multi-time sequence causal diagram and non-linear causal effect. This method takes iterative cumulative causal response as the core idea. Starting from the health state and health factor observation values of the first charge-discharge cycle, the average causal response of each health state and health factor within each charge-discharge cycle affected by all other health states or health factors at historical moments is calculated in chronological order. For the t-th charge-discharge cycle of battery sample i, the expression of the deduction results of battery health state and health factors based on causal cumulative response is as follows:

[0043]

[0044] In the formula, and are the deduction results of the battery health state and the n-th health factor in the t-th charge-discharge cycle of battery sample i respectively, and their initial values and are the initial observation values of the battery health state and health factors respectively and ; and are used to represent the deduction methods of battery health state and health factors based on causal cumulative response respectively. Their inputs are the deduction results of all battery health states and health factors before the (t - 1)-th charge-discharge cycle of battery sample i, and the outputs are the deduction results of the battery health state and the n-th health factor in the t-th charge-discharge cycle; and represent the total number of battery health states or health factors with time-lag causal effects on the battery health state and the n-th health factor in the t-th charge-discharge cycle of battery sample i respectively; is the maximum time-lag step considered in the causal deduction process.

[0045] In an ideal situation, if the health state and health factor of the battery under different operating conditions strictly abide by the same and complete multi-temporal non-linear causal relationship of the battery, the deduced results of the battery health factor and health state based on the causal cumulative response should be exactly the same as their observed values. However, there are three key limitations in the actual situation: First, there are still some differences in the causal relationship between the battery health state and health factor under different operating conditions, which is due to the heterogeneity caused by factors such as battery material composition, manufacturing process, and environmental conditions; Second, some potential health factors cannot be observed, resulting in an incomplete deduced causal relationship. These unobserved variables may include factors that are difficult to directly measure, such as changes in the internal microstructure of the battery and decomposition of electrolyte components; Third, there may be cumulative biases in the causal cumulative deduction. Small early biases will continuously amplify during the deduction process, eventually leading to a significant deviation of the later deduced results from the actual values. These three limitations make the deduction method of the battery health factor and health state based only on the causal cumulative response insufficient in accuracy and difficult to meet the actual application requirements. At the same time, if only relying on the observed values of the battery health factor to estimate the health state, the inherent mechanism information provided by the causal relationship cannot be fully utilized, lacking interpretability and in-distribution generalization ability, especially performing poorly when the battery operating conditions change greatly. Therefore, to solve the above problems, the present invention combines causal deduction with observed data to establish a causal relationship enhancement model based on causal-observational fusion deduction to form multi-temporal non-linear causal enhancement features. This model combines the multi-temporal non-linear causal relationship between the obtained battery health state and health factor with the actual observed data. By introducing adjustable weight parameters, the fusion ratio of the causal deduction result and the actual observed value is actively weighted according to the data quality and the strength of the causal relationship, so as to correct the causal difference of individual batteries to a certain extent, make up for the influence of potential unrecognized causal relationships, and suppress the diffusion of cumulative errors. The constructed multi-temporal non-linear causal enhancement features simultaneously retain the inherent mechanism interpretation ability of the causal relationship and the real-time information of the observed data, thus taking into account accuracy, interpretability, and in-distribution generalization ability in battery health state estimation. In addition, considering that the observed value of the battery health state is usually unknown in actual applications, this poses a challenge to the causal relationship enhancement model based on causal-observational fusion deduction. To solve this problem, the present invention finds the optimal health state proxy value to replace the unknown health state observed value, so as to minimize the gap between the deduced result of the health factor based on causal-observational fusion deduction and the observed value of the health factor. Since both the deduced result of the health state and the deduced result of the health factor are obtained from causal-observational fusion deduction, this deduced result of the health state theoretically has a small gap with the real health state observed value. Therefore, the deduced results of the health state and the health factor obtained from causal-observational fusion deduction can be used as enhancement features to estimate the health state observed value, thereby improving the interpretability and in-distribution generalization ability of the health state estimation.

[0046] In summary, based on the multi-temporal non-linear causal relationship directed graph between the obtained battery health state and health factors and the battery health factor and health state deduction method based on causal cumulative response, a causal relationship enhancement model based on causal-observational fusion deduction is established to obtain the battery multi-temporal non-linear causal enhancement features composed of the battery health state deduction result and the health factor deduction result. The expression of the causal relationship enhancement model based on causal-observational fusion deduction is as follows:

[0047]

[0048] In the formula, the objective function is to minimize the sum of the squares of the difference between the battery health factor deduction result and the observed value, and the constraint condition is the causal-observational fusion deduction equation. A heuristic algorithm can be used to solve this non-linear model to obtain the multi-temporal non-linear causal enhancement features of battery sample i ; and are respectively the battery health state deduction result and the nth health factor deduction result of the tth charge-discharge cycle of battery sample i based on causal-observational fusion deduction, which together constitute its multi-temporal non-linear causal enhancement features ; is all the battery health state proxy values before the (t - 1)th charge-discharge cycle of battery sample i. is the weight used to balance the causal relationship and the proportion of observational data in the causal-observational fusion deduction, and its value range is (0, 1). When is close to 0, the deduction result tends to rely on the causal relationship. When is close to 1, the deduction result tends to adopt the observed value. When is between 0 and 1, the deduction result is a weighted combination of causal deduction and observed value. A smaller value is suitable for the case where the causal relationship is stable and the observational noise is large, and a larger value is suitable for the case where the causal relationship is incomplete but the observational data is accurate. In practical applications, the value can be optimized and selected through cross-validation and other methods for specific batteries and application scenarios.

[0049] Furthermore, considering that the battery health state has obvious temporal characteristics and the battery capacity attenuation is closely related to the historical degradation curve, the present invention, based on the constructed battery multi-temporal non-linear causal enhancement features estimates the battery health state by using a causal enhancement - bidirectional gated recurrent unit model to improve the accuracy and out-of-distribution generalization ability of battery health state estimation. The specific method is as follows:

[0050] As a variant of the recurrent neural network, the standard gated recurrent unit network alleviates the vanishing gradient problem by introducing reset gates and update gates, and can effectively extract temporal features. The bidirectional gated recurrent unit fuses the outputs of the forward and backward gated recurrent units, can simultaneously obtain historical and future information, and enhances the prediction ability for time series data. In the single-direction gated recurrent unit layer, the multi-temporal non-linear causal enhanced features corresponding to the \(t\)-th charge and discharge cycle of battery sample \(i\) are used as the input, corresponding to the candidate hidden state of the gated recurrent unit as follows:

[0051]

[0052] where tanh is the hyperbolic tangent activation function; , and are the input weight matrix, recurrent weight matrix, and bias respectively; is the hidden state of the previous gated recurrent unit; is the reset gate of the current gated recurrent unit. The hidden state of the gated recurrent unit is updated as follows:

[0053]

[0054] where is the update gate of the current gated recurrent unit. The bidirectional gated recurrent unit obtains the final output through the concatenation of the outputs of the forward and backward gated recurrent units, and the expression is as follows:

[0055]

[0056] where and are the forward and backward hidden states of the \(t\)-th charge and discharge cycle respectively; and are the forward and backward gated recurrent units respectively. Mapping the hidden state of the last layer of the bidirectional gated recurrent unit through a fully connected neural network, the estimated value of the lithium-ion battery health state of the \(t\)-th charge and discharge cycle of battery sample \(i\) can be obtained.

[0057] Considering that the battery exhibits different degradation modes under different operating conditions, this operating condition heterogeneity is one of the main challenges in battery state of health (SOH) estimation. Existing research has shown that different operating conditions can be regarded as different data distributions, but there are partial common causal relationships between the battery SOH and health factors under different operating conditions. Therefore, the present invention introduces the invariant risk minimization theory to improve the out-of-distribution generalization ability of the SOH estimation model. The traditional model learning paradigm based on the empirical risk minimization theory focuses on minimizing the average loss of the training data and often performs poorly when there are systematic differences between the test distribution and the training distribution. In contrast, invariant risk minimization can capture the causal structure of the data rather than the condition-specific correlations by learning an invariant predictor that performs well in all conditions, thus maintaining good performance in unseen conditions. For the battery SOH estimation problem, the present invention adopts a causal enhancement-bi-directional gated recurrent unit (CE-BiGRU) battery SOH estimation model learning method based on the invariant risk minimization theory to learn the SOH estimation rules that are effective in different conditions. The objective function of the CE-BiGRU model based on the invariant risk minimization theory is as follows:

[0058]

[0059] In the formula, the objective function is to minimize the causal enhancement-bi-directional gated recurrent unit training loss with an invariant risk minimization regularization term. The first term is the standard empirical risk minimization loss, which ensures that the CE-BiGRU has good average performance in different battery operating conditions. The second term is the invariant risk minimization regularization term, which promotes the model to learn causal feature representations that are invariant across conditions by calculating the sum of the squared norms of the gradients of the loss function with respect to the parameters of the virtual classifier in each condition; denotes the parameters of the battery SOH estimation model based on the CE-BiGRU; is the training error of the battery SOH estimation model, usually using the mean squared error function; is the weight parameter that balances the empirical risk minimization objective and the invariant risk minimization regularization term; are the parameters of the virtual linear classifier, which are used to verify the stability of the model representation in each condition. When the invariant risk minimization regularization term approaches zero, it indicates that the model has learned an approximately optimal prediction rule in all conditions, which is beneficial to improving the out-of-distribution generalization of the battery SOH estimation.

[0060] Considering that there are significant differences in the degradation characteristics of the battery under different operating conditions, the constructed multi-time series non-linear causal enhanced features of the battery There may be a distribution shift between different operating conditions. Traditional batch normalization methods usually assume that all training data come from the same distribution and cannot effectively handle the feature distribution differences of data under different battery operating conditions. To solve this problem, based on the causal augmentation-bi-directional gated recurrent unit battery state of health estimation model learning method based on the theory of invariant risk minimization, the present invention introduces a domain calibration batch normalization method, maintains independent normalization parameters for different operating conditions, and processes unknown operating conditions through a feature similarity matching mechanism. For the battery sample i under known operating conditions, the normalization process of its multi-temporal non-linear causal augmentation features is as follows:

[0061]

[0062] In the formula, is the result of normalizing the multi-temporal non-linear causal augmentation features of the battery sample i through the domain calibration batch normalization method; and are learnable affine parameters for the specific operating condition of the battery sample i, which are used to restore the expression ability of the normalized features; and are the mean and variance of the multi-temporal non-linear causal augmentation features under the operating condition to which the battery sample i belongs; is a small constant to prevent division by zero. For battery samples under out-of-distribution unknown operating conditions, the causal augmentation-bi-directional gated recurrent unit model needs to dynamically select the most suitable batch normalization parameters. Therefore, the present invention adopts an operating condition matching method based on the distribution similarity of battery causal augmentation features, calculates the feature distribution distance between the unknown condition sample and the known condition samples, and selects the batch normalization parameters of the most similar condition for application. The expression of the causal augmentation feature similarity distance between battery samples is as follows:

[0063]

[0064] In the formula, is the causal augmentation feature similarity distance between the battery sample i and the battery sample j; is the variance difference weight, which is used to balance the importance of the mean difference and the variance difference in the calculation of the condition similarity. When facing a battery sample under an unknown condition, by calculating its similarity distance with all known conditions, the batch normalization parameters corresponding to the condition with the smallest distance are selected as follows:

[0065] In the formula, k is the index of the battery sample of the known operating condition that is most similar to the causal enhanced feature distribution of the battery sample j under the unknown operating condition. In the training stage, domain calibration batch normalization maintains independent normalization parameters for each known condition. In the testing stage, for the battery samples of the known conditions, the domain calibration batch normalization parameters corresponding to the conditions are directly applied; for the battery samples of the unknown conditions, first calculate the similarity between their causal enhanced feature distributions and each known condition, and then apply the domain calibration batch normalization parameters of the most similar condition. This operating condition matching method based on the similarity of the causal enhanced feature distribution of the battery enables the causal enhanced-bi-directional gated recurrent unit network model to adaptively process the feature distribution differences inside and outside the distribution, further improving the generalization ability of the state of health estimation under unknown conditions.

[0066] In summary, in the present invention, the causal enhanced-bi-directional gated recurrent unit model can make full use of the internal mechanism information provided by the multi-temporal non-linear causal enhanced features, capture the temporal features and long-term dependencies through the bi-directional gated recurrent unit, and improve the generalization ability of the model under unknown conditions through the invariant risk minimization theory and the domain calibration batch normalization method, realizing accurate, interpretable and well-generalized state of health estimation.

[0067] The technical effects of the present invention are further illustrated below by using the NASA lithium-ion battery dataset. This dataset contains cyclic charge and discharge data of multiple groups of lithium-ion batteries under different operating conditions, and each group of batteries has experienced a complete aging process from brand new to the capacity decaying to 80% or 70% of the rated capacity. In this embodiment, four groups of battery data, namely B0005, B0006, B0007, and B0018, are selected to construct the training set and the validation set (where the training set and the validation set are divided according to a ratio of 4:1, and the training set and the validation set both obtain data by equally spaced sampling on the complete degradation curve), and 29 groups of battery data including B0025 to B0056 are selected as the test set. Each group of battery data contains measurement data such as voltage, current, temperature, time, and capacity during the charge and discharge process.

[0068] Table 1 lists the operating condition parameters of each group of batteries. Each group of batteries uses the same charging method, first charging at a constant current of 1.5 A to 4.2 V, and then charging at a constant voltage until the current drops to 20 mA. The differences in the operating conditions of different groups of batteries are mainly reflected in the ambient temperature, discharge cut-off voltage, and discharge current. The ambient temperature includes room temperature (24 °C), low temperature (4 °C), and high temperature (43 / 44 °C), the discharge current includes 1 A, 2 A, 4 A, and square wave (amplitude 4 A, duty cycle 50%, frequency 0.05 Hz), and the cut-off voltage includes 2.0 V, 2.2 V, 2.5 V, and 2.7 V.

[0069] Table 2 lists the parameter settings, search space, and determination methods of the constructed battery state of health estimation model based on causal enhancement - bidirectional gated recurrent unit. In this embodiment, the constructed causal relationship enhancement model based on causal - observation fusion deduction uses the dual annealing algorithm to solve. The dual annealing algorithm is a stochastic global optimization algorithm that combines simulated annealing and local search strategies, with the advantages of strong global optimization ability and fast convergence speed. The dual annealing algorithm sets the initial temperature to 5230.0, the annealing restart temperature ratio to 0.00002, and the convergence criterion is to stop when the maximum number of iterations is reached or the no - improvement condition is triggered during the local search process. The causal enhancement - bidirectional gated recurrent unit model is modeled and tested using the Pytorch framework of Python 3.8 version.

[0070] Multiple health factors to be screened are extracted from the battery measurement data, including the mean, minimum, maximum, standard deviation, and integral of voltage, current, and temperature measurement data, as well as the time used for each charge - discharge cycle. Through the potential Peter - Clark instantaneous conditional independence algorithm and the optimal adjustment set theory, a causal inference considering the influence of unobserved health factors on the multi - time - series non - linear causal relationship between the state of health of four groups of batteries B0005, B0006, B0007, and B0018 and the battery health factors in the training set is carried out. Figures 1 to 4 The multi - time - series non - linear causal relationships between the inferred state of health and the battery health factors of the four groups of batteries B0005, B0006, B0007, and B0018 are respectively shown (SOH represents the state of health in the figure), and the health factors that do not have a deterministic causal path with the state of health have been screened out. It should be noted that Figures 1 to 4 the positive / negative causal effects marked on the causal direction arrows, the values of which are obtained by calculating the change in the output of the non - linear causal effect estimator when the input ranges from 0 to 1 (normalized value), which reflects the causal impact on the target variable under the unit intervention of the independent variable.

[0071] From Figures 1 to 4It can be seen that among the three batteries B0005, B0006, and B0007, the integrals of the discharge currents all show strong negative causal effects on the state of health, which are -0.56, -0.58, and -0.53 respectively (where -0.53 is the indirect effect). This indicates that the gradual decrease in the cumulative charge transfer during the discharge process is the common mechanism leading to battery capacity degradation. It should be noted that in the NASA lithium-ion battery dataset used in this embodiment, the discharge current data is presented as negative values. Therefore, the integral of the discharge current shows a negative causal effect on the state of health. At the same time, in all four groups of batteries, it is shown that the mutual influence between the health factor and the state of health has obvious time-delay characteristics (ranging from t-1 to t-11), indicating that battery aging is a sequential cumulative process, and the changes in the health factor or the state of health in the current cycle will gradually show their effects in multiple subsequent cycles. In addition, from Figures 1 to 4 It can also be seen that the inferred causal relationship of the B0005 battery is the most complex among the four groups of batteries, including 12 causal nodes and 11 causal paths, indicating that the degradation mechanism of this battery under the given working conditions is more complex than that under other working conditions. The B0018 battery only shows a single causal path of "state of health(t) → integral of discharge current(t+1)", which may indicate that the battery is in a special aging stage, and the change in the state of health has become the dominant factor guiding the change in the health factor, rather than being affected by the health factor.

[0072] Taking the B0005 and B0007 batteries as examples, for Figures 1 to 4Analyze the key causal paths. The long-time-delay path of Battery B0005, "Minimum charging current (t-11) → Discharge single-cycle duration (t-6) → Discharge current integral (t-1) → State of health of the battery (t)", shows the mechanism that abnormal charging current (possibly insufficient charging) will affect the discharge duration and ultimately lead to a decline in the state of health of the battery through this path. The effect intensity of this causal path is "+0.35 → -0.28 → -0.56", and the comprehensive effect is approximately +0.055, indicating that insufficient charging current in the early stage may ultimately slightly reduce the state of health through a complex path. This causal chain with a long time span explains why some battery anomalies may appear only after a long time. The direct positive effect of the path "Discharge single-cycle duration (t-5) → State of health of the battery (t)" of Battery B0005 is +0.26, indicating that appropriately extending the discharge time may be beneficial to the battery life, which is consistent with the battery usage recommendation of avoiding deep discharge. The direct negative effect of the path "Minimum charging voltage (t-3) → State of health of the battery (t)" of Battery B0007 is -0.80. This strong negative effect indicates that too high an initial charging voltage significantly damages the battery life. This is consistent with the mechanism that overcharging of lithium-ion batteries will cause problems such as electrolyte decomposition and lithium deposition. The effect intensity of the causal path "State of health (t) → Standard deviation of charging current (t+1)" of Battery B0007 is +0.34. This positive effect indicates that a decline in the state of health will lead to a decrease in the standard deviation of the charging current, which may mean that the battery loses its fast charging ability as it ages and the charging curve becomes flatter. When the battery ages, due to the thickening of the solid electrolyte interface film and the loss of active materials, it often shows an increase in internal resistance and polarization, resulting in a flattening of the charge-discharge curve and a reduction in the fluctuation of the charging current.

[0073] In summary, there is a significant causal cascade effect between the state of health of the battery and the health factors. This complex causal structure with multiple nodes and multiple paths reveals the potential physical and chemical mechanisms of lithium-ion battery degradation. The differences in the causal structures under different battery operating conditions also indicate that it may be difficult to effectively adapt to the battery degradation characteristics under different operating conditions by relying solely on data correlation for state-of-health estimation. This phenomenon indicates the necessity of combining mechanism and data-driven methods, using the causal structure to enhance the feature distribution, and then improving the accuracy, generalization, and interpretability of battery state-of-health estimation.

[0074] The present invention constructs a causal relationship enhancement model based on causal-observational fusion deduction, and uses multi-temporal non-linear causal relationships to solve for multi-temporal non-linear causal enhancement features of the battery. To verify the effectiveness of the causal enhancement features, in Figure 5 shows the comparison results of the actual observed values and the causal-observational fusion deduction results of the state of health of Battery B0005 and the discharge current integral. From Figure 5It can be seen that the actual health status observation value of the B0005 battery shows a clear downward trend, gradually decreasing from the initial close to 1.0 to about 0.1, and contains more local capacity regeneration phenomena, which is in line with the general law of lithium-ion battery capacity degradation. The causal-observation fusion deduction results of the health status better retain the overall degradation trend of the actual health status, and to a certain extent restore the capacity regeneration phenomenon that appears in the observation value in the early stage (number of cycles <80), confirming the accuracy of the identified causal relationship. However, it is difficult to track the change pattern of the actual observation value in the middle and late stages (number of cycles >80). This may be because the battery has a degradation mechanism in the deep aging stage that is difficult to capture through the identified causal relationship. From Figure 5 It can also be seen that the actual discharge current integral observation value of the B0005 battery shows a clear upward trend. Since the actual value of the discharge current is negative, it actually reflects that the battery's dischargeable charge gradually decreases with the increase in cycles, which is also in line with the general law of lithium-ion battery capacity degradation. The causal-observation fusion deduction results of the discharge current integral can also well track the variation pattern of the actual observation value. Overall, although the causal-observation fusion deduction results are significantly different from the observed values in some cycles, the differences between the health state and discharge current integral deduction results and the observed values are similar in each charge and discharge cycle. In particular, around cycle number 80, the intersection point of the four curves in the fusion deduction results and the actual observation value almost coincides. This verifies the rationality of using the health state and health factor deduction results obtained by causal-observation deduction as causal enhancement features to estimate the battery health state.

[0075] To further analyze the distribution characteristics of causal enhancement features, Figures 6 to 7 The distribution of the original health factor features and causal enhancement features of different batteries after projection in the two-dimensional feature space is shown. This embodiment uses the t-distributed random neighbor embedding dimensionality reduction method to project the original health factor features and causal enhancement features into the two-dimensional space for visualization, and distinguish them by color according to the battery samples. Figure 6 、 Figure 7 It can be seen that compared with the original health factor features, the causal enhancement features show a more consistent distribution pattern across different batteries. Figure 6 In , there are obvious boundaries between the feature distributions of different batteries, indicating that the original features are more sensitive to individual differences of batteries. Figure 7 In the , this boundary is significantly weakened, and the distributions of different battery characteristics tend to cluster, indicating that the causal enhancement feature better captures the common health degradation mechanisms among different batteries. This shows that the causal enhancement feature is of great significance for improving the in-distribution generalization of the health state estimation model, helping to reduce the model's dependence on specific individual battery characteristics and enhance its ability to learn common degradation mechanisms.

[0076] Evaluate the accuracy of the constructed causal enhancement - bidirectional gated recurrent unit model for battery state of health estimation. To verify the effectiveness of the battery state of health estimation method based on causal enhancement - bidirectional gated recurrent unit in the present invention, the following methods are selected for comparison: support vector regression, long short - term memory network, gated recurrent unit, bidirectional long short - term memory network, bidirectional gated recurrent unit, and Granger causality - long short - term memory network. Among them, the Granger causality - long short - term memory network method screens health factors that have a direct causal impact on the state of health through Granger causality test, and then estimates the battery state of health through the long short - term memory network. All models are evaluated using two indicators, root mean square error and mean absolute error, and the average value, minimum value, and maximum value of each evaluation indicator in 29 groups of test batteries are calculated. Table 3 shows the comparison of the overall performance of health state estimation of different methods on all test batteries. It can be seen from Table 3 that the causal enhancement - bidirectional gated recurrent unit method in the present invention is significantly superior to the comparison methods in all average value and maximum value evaluation indicators. Compared with the traditional machine learning method support vector regression, the average value and maximum value of the root mean square error of the causal enhancement - bidirectional gated recurrent unit are reduced by 23.5% and 21.5% respectively, and compared with the bidirectional gated recurrent unit, a deep learning method based on correlation analysis, they are reduced by 17.0% and 30.9% respectively; compared with the Granger causality - long short - term memory network based on causality analysis, they are reduced by 36.5% and 3.6% respectively. This shows that the causal enhancement - bidirectional gated recurrent unit effectively improves the accuracy of health state estimation by integrating multi - time - series non - linear causal enhancement features, bidirectional gated recurrent unit network, and invariant risk minimization theory. It is worth noting that the maximum value of the root mean square error of the causal enhancement - bidirectional gated recurrent unit, 0.7949, and the maximum value of the mean absolute error, 0.7085, are much lower than those of other methods, which is particularly important for the abnormal capacity detection and safety warning of the lithium - ion battery management system.

[0077] Figures 8 to 15 Shows the comparison results of the health state estimation curves of the causal enhancement - bidirectional gated recurrent unit and the comparison methods in 29 groups of test batteries. Here, taking the test battery B0042 as an example, the health state estimation curves of different methods are compared and analyzed, as Figure 16 shown. According to the operating condition parameters of each group of batteries listed in Table 1, the B0042 battery operates under low - temperature conditions, and the discharge current switches between 1A and 4A. From Figure 16 it can be seen that there are significant phenomena of sharp capacity decline and capacity regeneration in the actual degradation curve of the B0042 battery. From Figure 16It can be seen that the causal enhancement - bidirectional gated recurrent unit method can accurately track the change trend of the health state of battery B0042 throughout the battery life cycle. In particular, the estimation accuracy of the sharp capacity decline phenomenon in the middle stage (between 40 and 85 cycles) and the capacity regeneration phenomenon in the later stage (after 85 cycles) is significantly better than other methods. Health state estimation methods based on correlation analysis, such as bidirectional gated recurrent units, fail to identify the sharp capacity decline phenomenon in the middle stage. Although the Granger causality - long short - term memory network method based on causality analysis can identify the sharp capacity decline phenomenon in the middle stage, its estimation accuracy for the capacity regeneration phenomenon in the later stage is not as good as that of the causal enhancement - bidirectional gated recurrent unit. This indicates that the causal enhancement - bidirectional gated recurrent unit method not only outperforms the comparative methods in overall accuracy but also maintains stable estimation accuracy on a long - time scale, which is of great significance for long - term battery health state monitoring.

[0078] To verify the generalization of the proposed method under different working conditions, the test batteries are divided into three categories according to the similarity between the operating conditions and the training set: similar conditions (the ambient temperature is the same as that of the training set, and the discharge uses a square - wave current, but the average current value is the same as that of the training set, including B0025, B0026, B0027, B0028), medium - difference conditions (the ambient temperature is the same as that of the training set, but the discharge current size varies greatly, or the discharge current is the same as that of the training set, but the ambient temperature varies greatly, including a total of 10 groups of batteries B0033, B0034, B0036, B0049, B0050, B0051, B0053, B0054, B0055, B0056), and significant - difference conditions (the ambient temperature is high, low, or alternating high and low, and the discharge current is different from that of the training set, including all the remaining 15 groups of batteries in the test set). Table 4 shows the accuracy performance of different methods under the three types of working conditions. It can be seen from Table 4 that all methods perform best under similar conditions and worst under medium - difference or significant - difference conditions, which conforms to the general law of the model's generalization ability. However, the causal enhancement - bidirectional gated recurrent unit method has the smallest performance decay among different working conditions. The root - mean - square error increases by only 0.2074 from similar conditions to significant - difference conditions, while the Granger causality - long short - term memory network, bidirectional gated recurrent unit, and support vector regression increase by 0.3709, 0.2908, and 0.3329 respectively. It is worth noting that the root - mean - square errors of the causal enhancement - bidirectional gated recurrent unit under medium - difference and significant - difference conditions (0.3585 and 0.2476) are better than all comparative methods, fully demonstrating the advantage of the causal enhancement - bidirectional gated recurrent unit in improving cross - condition generalization.

[0079] Figures 17 to 19 Shows the comparison of the health state estimation results of the causal enhancement - bidirectional gated recurrent unit and the comparative methods on batteries under three typical working conditions, where Figure 17Among them is the battery B0028 under similar working conditions, Figure 18 Among them is the battery B0056 under medium-difference working conditions, Figure 19 Among them is the battery B0045 under significantly different working conditions. It can be seen from the figure that for the battery B0028 under similar working conditions, both the causal enhancement-bi-directional gated recurrent unit and other methods can estimate the change in the state of health well, but the causal enhancement-bi-directional gated recurrent unit has relatively better estimation accuracy. For the battery B0056 under medium-difference working conditions, the comparison method shows obvious deviation in the middle and late stages (cycle number > 50), while the causal enhancement-bi-directional gated recurrent unit maintains high accuracy. For the battery B0045 under significantly different working conditions, the comparison method shows large fluctuations and systematic deviations, while the causal enhancement-bi-directional gated recurrent unit can still accurately capture the change trend of the state of health. This indicates that under unknown or extreme working conditions, the causal enhancement-bi-directional gated recurrent unit learns more robust and generalizable battery degradation laws by combining multi-temporal non-linear causal enhancement features and the theory of invariant risk minimization.

[0080] To further verify the stability of the proposed method and guide the selection of actual application parameters, a sensitivity analysis is carried out on key parameters, including the common causal threshold and the weight used to balance the causal relationship and the proportion of observed data in the causal-observation fusion deduction .

[0081] Table 5 shows the influence of different and on the accuracy of the battery state of health estimation method based on the causal enhancement-bi-directional gated recurrent unit. It can be seen that when increases from 0.2 to 0.8, it represents that the causal enhancement-bi-directional gated recurrent unit requires a gradually more stringent proportion of the extracted common causal relationship in the 4 groups of batteries in the training set. At this time, the root mean square error under all working conditions shows a trend of first rising and then falling, and the lowest point of the root mean square error is located at equal to 0.2. Since the training set in this embodiment includes a total of 4 groups of batteries, this parameter value indicates that the extracted common causal relationship is the union of the causal relationships of each of the 4 groups of batteries, which shows that appropriately fully accommodating various causal relationships helps to extract cross-working-condition invariant features and enhance cross-working-condition generalization ability. When increases from 0.1 to 0.9, it represents that the causal enhancement-bi-directional gated recurrent unit gradually attaches more importance to the causal relationship during the causal-observation fusion deduction. At this time, the root mean square error under all working conditions shows a trend of first falling and then rising, and the lowest point of the root mean square error in the significantly different working conditions is located at equal to 0.5. This indicates that when the proportion of the causal relationship and the observed data in the causal-observation fusion deduction is 1:1, the model is optimal in both prediction accuracy and cross-working-condition generalization ability. To sum up, equal to 0.2 and Equal to 0.5 is a comprehensive balance point. Under this setting, the model can not only maintain the accuracy under similar working conditions, but also significantly improve the generalization under different working conditions.

[0082] To analyze the contributions of each module in the causal enhancement - bidirectional gated recurrent unit method, Table 6 shows the accuracy and generalization of model variants obtained from different module combinations. It can be seen from Table 6 that replacing the original features with causally enhanced features reduces the average root mean square error of the bidirectional gated recurrent unit from 0.2521 to 0.2381, a decrease of 5.6%. Especially under significantly different working conditions, it drops from 0.3281 to 0.2681, an improvement of 18.3%, indicating the importance of causal relationships in the estimation of battery health state. Introducing the domain calibration batch normalization mechanism reduces the root mean square error to 0.2173, a decrease of 8.7%, and an overall improvement of 13.8%, verifying the importance of feature distribution calibration under different working conditions. After adding invariant risk minimization regularization to obtain the complete causal enhancement - bidirectional gated recurrent unit model, the root mean square error drops to 0.2154, an improvement of 0.9%. Especially under moderately different working conditions, it drops from 0.3631 to 0.3585, an improvement of 1.3%, demonstrating the role of invariant risk minimization in enhancing the out - of - distribution generalization ability of the model. These results show that each module of the causal enhancement - bidirectional gated recurrent unit makes significant contributions to the improvement of model performance, and these improvements are more obvious when facing out - of - distribution data, proving the rationality and effectiveness of the constructed causal enhancement - bidirectional gated recurrent unit model.

[0083] Table 1 Operating condition parameters of lithium - ion batteries

[0084] Table 2 Parameter settings, search space and determination methods of the causal enhancement - bidirectional gated recurrent unit model

[0085] Table 3 Comparison of health state estimation accuracy of different methods

[0086] Table 4 Comparison of root mean square errors of health state estimation of different methods under different working conditions

[0087] Table 5 Comparison of root mean square errors of battery health state estimation of the causal enhancement - bidirectional gated recurrent unit with different parameters

[0088] Table 6 Comparison of root mean square errors of battery health state estimation of model variants with different module combinations

[0089] The above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A method for evaluating the state of health of a battery by enhancing causal relationship consideration, characterized in that It includes the following steps: Obtain battery measurement data, acquire the multi-temporal non-linear causal relationship between the battery health state and health factors based on the causal inference method, and construct a directed graph of the multi-temporal non-linear causal relationship of the battery with commonality; According to the directed graph of the multi-temporal non-linear causal relationship of the battery and the deduction method of the battery health factors and health state based on causal cumulative response, establish a causal relationship enhancement model based on causal-observational fusion deduction, and obtain the multi-temporal non-linear causal enhancement features of the battery; Use the causal enhancement features as input, and output the battery health state by using the causal enhancement-bi-directional gated recurrent unit model.

2. The method for evaluating the battery health state considering causal relationship enhancement according to claim 1, wherein The battery measurement data includes the mean, minimum value, maximum value, standard deviation and integral of the voltage, current and temperature measurement data in each charge and discharge cycle, as well as the time used for each charge and discharge cycle.

3. The method for evaluating the battery health state by considering the enhancement of causal relationship according to claim 1, characterized in that, The obtaining of the multi-temporal non-linear causal relationship between the battery health state and health factors based on the causal inference method specifically includes: using the latent Peter-Clark instantaneous conditional independence algorithm to discover the causal graph between the battery health state and health factors; according to the causal graph, using the optimal adjustment set theory to estimate the causal effect of the non-linearity between the battery health state and health factors, and combining the causal graph discovery result and the causal effect estimation result to obtain the multi-temporal non-linear causal relationship between the battery health state and health factors.

4. The method for evaluating the battery health state considering enhanced causal relationship according to claim 1, characterized in that The directed graph of the multi-temporal non-linear causal relationship of the battery with commonality includes a node set and an edge set. The node set consists of a set of health factors with common causality and a battery health state index, and the edge set describes the multi-temporal non-linear common causal relationship between the battery health state and health factors.

5. The method for evaluating the battery health state considering enhanced causality according to claim 1, characterized in that, According to the directed graph of the multi-temporal non-linear causal relationship of the battery and the deduction method of the battery health factors and health state based on causal cumulative response, establish a causal relationship enhancement model based on causal-observational fusion deduction, and obtain the multi-temporal non-linear causal enhancement features of the battery. The specific steps include: Establish a deduction model of the battery health state and health factors based on causal cumulative response according to the directed graph of the multi-temporal non-linear causal relationship of the battery, and use the deduction model to obtain the deduction values of the health state and health factors within each charge and discharge cycle according to the historical health state and health factor observation values; With the goal of minimizing the difference between the health factor deduction value and the health factor observation value, introduce the optimal health state proxy value to replace the unknown health state observation value, and construct a causal relationship enhancement model based on causal-observational fusion deduction; Solve the causal relationship enhancement model to obtain the deduction values of the battery health state and health factors, which together constitute the multi-temporal non-linear causal enhancement features of the battery.

6. The method for evaluating the battery health state considering the enhancement of causal relationship according to claim 5, wherein The deduction model starts from the health state and health factor observation values of the first charge and discharge cycle, and sequentially calculates the average causal response of each health state and health factor within each charge and discharge cycle affected by all other health states or health factors at historical moments in chronological order.

7. The method for evaluating the battery health state by considering the enhancement of causal relationship according to claim 1, characterized in that During the training process of the causal enhancement-bi-directional gated recurrent unit model, a loss function based on the invariant risk minimization theory is adopted.

8. The method for evaluating the battery health state considering the enhancement of causal relationship according to claim 1, characterized in that, Before the input of the causal enhancement feature, domain calibration batch normalization processing is required, and each different operating condition has independent batch normalization parameters.

9. The method for evaluating the battery health state considering the enhancement of causal relationship according to claim 8, characterized in that, For battery samples with known operating conditions, directly apply the batch normalization parameters corresponding to this operating condition for processing; For battery samples with unknown operating conditions, calculate the similarity distance between it and all known operating conditions, and select the batch normalization parameters corresponding to the operating condition with the smallest distance for processing.

10. A computer device, characterized in that, The computer device includes: One or more processors; A memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the battery health state evaluation method considering causal relationship enhancement as described in any one of claims 1 to 9.

Images