A method for analyzing micro-failure of lithium ion battery and related products

Abstract

The present application relates to the technical field of battery management, and provides a lithium ion battery micro-failure analysis method and related products, comprising: obtaining a discrete swelling displacement sequence of a lithium ion battery in a constant current charging stage, reconstructing a swelling amount probability density curve by using a parameter-free kernel density estimation, extracting multi-dimensional high-order mechanical degradation features, combining the cycle number of the battery, estimating the state of health of the battery through a physical information neural network; wherein, in the forward propagation of the physical information neural network, an attention mechanism is introduced to dynamically decouple and weight the input features; in the backward propagation, a first-order monotone decreasing and second-order aging acceleration double partial differential physical law hard constraint is introduced; the attention weight of the physical information neural network is extracted, the estimated state of health of the battery is combined as a prior boundary, and the micro-aging mode of the battery is decoupled through a multivariate nonlinear mapping equation. Efficient deep diagnosis from macro-mechanical evolution to micro-electrochemical failure is realized.

Description

Technical Field

[0001] This invention belongs to the field of battery management technology, and in particular relates to a method for analyzing microscopic failures in lithium-ion batteries and related products. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] The degradation of the state of health (SOH) of lithium-ion batteries is closely related to significant external mechanical expansion and internal stress evolution. Accurate and rapid online estimation of the SOH directly determines the battery's service life, secondary utilization strategy, and control of thermal runaway risk. However, existing SOH estimation techniques face the following technical bottlenecks in practical engineering applications: Traditional SOH estimation methods rely heavily on external characteristic electrical signals (voltage, current). However, in certain specific aging stages, electrical parameters exhibit a characteristic insensitivity, making it difficult to comprehensively and intuitively reflect the mechanical damage caused by side reactions such as solid electrolyte interphase (SEI) film thickening, microcrack initiation in active particles, and local lithium plating. Since the nonlinear mechanical deformation of the battery cell is accompanied by a complex viscoelastic strain relaxation effect, directly performing time-domain differential calculation on discrete sensing data can easily amplify high-frequency measurement noise and local sudden spikes in the sensor. This places extremely high demands on the smoothing preprocessing of the data, resulting in poor engineering robustness of feature extraction. Existing SOH estimation methods based on deep learning models (such as Long Short-Term Memory Network LSTM and Gated Recurrent Unit GRU) are essentially black-box models that rely on data. They lack boundary constraints from underlying electrochemical failure mechanisms or thermodynamic evolution laws. When faced with complex and variable actual working conditions or early small sample training conditions, the models are prone to overfitting, and even produce fitting oscillations that violate electrochemical common sense, such as an abnormally large rebound in output capacity when testing long sequence prediction. Most existing health status assessment methods can only provide macroscopic capacity decay values ​​(i.e., SOH percentage), and cannot further trace the microscopic physical causes of battery capacity degradation (such as loss of available lithium ions (LLI) and loss of active materials (LAM) in the positive and negative electrodes). This makes it impossible to provide differentiated early warnings and dynamic charging rate control for high-risk failure modes such as lithium plating and electrode structure collapse, and it is also difficult to provide refined safety ratings when retiring batteries are reused. Summary of the Invention

[0004] To address the technical problems mentioned above, this invention provides a method for analyzing microscopic failures in lithium-ion batteries and related products. It abandons the traditional method of differential differentiation of time-domain expansion signals, thus avoiding high noise defects, and achieves robust decoupling of mechanical degradation characteristics from a global statistical perspective. Simultaneously, it innovatively introduces a dual partial differential physical law hard constraint of first-order monotonically decreasing and second-order aging acceleration into the backpropagation computation graph, significantly suppressing non-physical oscillations in the deep learning model and achieving highly robust SOH precision diagnosis under multiple operating conditions and small sample sizes. Using this high-precision macroscopic SOH as a thermodynamic prior boundary, it precisely decouples the microscopic aging modes inside the battery through multi-objective optimization, achieving efficient and deep diagnosis from macroscopic mechanical evolution to microscopic electrochemical failure, thus constructing a high-precision closed-loop battery safety management system.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of this invention provides a method for analyzing the microscopic failures of lithium-ion batteries, comprising: The discrete expansion displacement sequence of the lithium-ion battery during the constant current charging stage is obtained, and the expansion probability density curve is reconstructed using parameterless kernel density estimation. Based on probability density curves, multidimensional higher-order mechanical degradation features are extracted; Based on multidimensional high-order mechanical degradation characteristics and combined with the number of battery cycles, the battery health status is estimated through a physical information neural network. In the forward propagation of the physical information neural network, an attention mechanism is introduced to dynamically decouple and weight the input features, while in the back propagation, a dual partial differential physical law hard constraint of first-order monotonically decreasing and second-order aging acceleration is introduced. The attention weights of the physical information neural network are extracted and combined with the estimated battery health state as a prior boundary. The micro-aging mode of the battery is decoupled through a multivariate nonlinear mapping equation.

[0006] Furthermore, the multidimensional higher-order mechanical degradation features include: Peak position, representing the amount of expansion displacement corresponding to the main peak of the probability density curve; Mathematical expectation, representing the centroid of the morphology of the inflationary probability distribution; Peak height represents the maximum probability density value of the probability density curve. Standard deviation characterizes the range of fluctuation of an expansion displacement sequence around its mathematical expectation; Skewness characterizes the direction and degree of asymmetry of the inflation probability density distribution; Information entropy characterizes the state complexity of an inflated probability density distribution.

[0007] Furthermore, the global optimization objective function of the physical information neural network is: ; in, , , For the control weights of physical partial differential equations; The regularization coefficient is used. To optimize space globally; The loss is the mean square error data fitting loss; the first-order monotonically decreasing partial differential penalty functional. Second-order aging acceleration physical loss penalty functional Differential penalty N is the total number of loop samples. The SOH value of the battery under the i-th cycle estimated by the model. For the model estimate of the i-th Battery SOH value after 1 cycle This represents the number of iterations corresponding to the i-th iteration.

[0008] Furthermore, the multivariate nonlinear mapping equation is: ; in, The lithium-ion inventory loss index, the positive electrode active material loss index, and the negative electrode active material loss index are respectively characterized. To maintain the degeneracy characteristics of higher-order mechanics; Dynamic weights for attention; and The optimal mapping matrix and bias are trained using multi-task physical boundary constraints. It is the product of Hadama.

[0009] Furthermore, the micro-aging mode includes loss of active material and loss of available lithium ions.

[0010] Furthermore, it also includes implementing differentiated closed-loop battery safety management strategies based on the battery's micro-aging patterns: When the loss index of negative electrode active material exceeds the preset safety threshold, a safety warning is triggered and the maximum charging rate is reduced or the battery cell is isolated. When capacity degradation is dominated by the loss of available lithium ions and the loss of positive and negative electrode active materials remains relatively intact, adjust the charge and discharge cutoff voltage range. When the battery health deteriorates to the retirement threshold, the failure utilization rating of the retired battery is determined based on the degree of loss of active materials.

[0011] Furthermore, in the forward propagation of the physical information neural network, a single-expert multilayer perceptron is used as a feature extractor, and the dimensionality reduction and mapping of the feature domain are completed through nested hyperbolic tangent nonlinear activation functions.

[0012] Furthermore, the attention mechanism maps the input to a high-dimensional latent space through a two-layer cascaded fully connected network, autonomously learns and calculates the weight coefficients of dynamic features, and applies the weight coefficients to the input through the Hadamard product.

[0013] A second aspect of the present invention provides a lithium-ion battery microscopic failure analysis system, comprising: The probability density space reconstruction module is configured to: acquire the discrete expansion displacement sequence of the lithium-ion battery during the constant current charging stage, and reconstruct the expansion probability density curve using parameterless kernel density estimation. The high-order mechanical feature extraction module is configured to extract multi-dimensional high-order mechanical degradation features based on probability density curves. The neural network estimation module is configured to estimate the battery health status based on multidimensional high-order mechanical degradation features and the number of battery cycles through a physical information neural network. In the forward propagation of the physical information neural network, an attention mechanism is introduced to dynamically decouple and weight the input features, and in the backward propagation, a dual partial differential physical law hard constraint of first-order monotonically decreasing and second-order aging acceleration is introduced. The micro-mode decoupling module is configured to: extract the attention weights of the physical information neural network, combine them with the estimated battery health state as a priori boundary, and decouple the battery's micro-aging mode through a multivariate nonlinear mapping equation.

[0014] A third aspect of the present invention provides a lithium-ion battery microscopic failure analyzer, characterized in that it comprises: The sensing and acquisition layer, deployed on a physical test bench or battery pack side, is used to acquire discrete expansion displacement sequences of lithium-ion batteries during the constant current charging phase. The core layer, as the edge computing master control platform, executes the steps in the lithium-ion battery micro-failure analysis method described in the first aspect to obtain the micro-aging mode. The interaction and execution layer is used to receive micro-aging modes and is controlled by the battery management system to perform safety management.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention abandons the traditional method of differential differentiation of time-domain dilated signals to avoid high noise defects, and achieves robust decoupling of mechanical degradation characteristics from a global statistical perspective. At the same time, it innovatively introduces the dual partial differential physical law hard constraint of first-order monotonically decreasing and second-order aging acceleration into the backpropagation calculation graph, which significantly suppresses non-physical oscillations of deep learning models and achieves highly robust SOH precision diagnosis under multiple operating conditions and small samples. Using this high-precision macroscopic SOH as a thermodynamic a priori boundary, it precisely decouples the microscopic aging mode inside the battery through multi-objective optimization, realizing efficient deep diagnosis from macroscopic mechanical evolution to microscopic electrochemical failure, and constructing a high-precision battery safety management closed loop. Attached Figure Description

[0016] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0017] Figure 1 This is a flowchart of a method for analyzing the microscopic failures of a lithium-ion battery according to Embodiment 1 of the present invention; Figure 2 This is a full life-cycle expansion curve of Embodiment 1 of the present invention; Figure 3 This is a PDF feature curve diagram of Embodiment 1 of the present invention; Figure 4 This is a correlation diagram between the six-dimensional features of Embodiment 1 of the present invention and the battery capacity; Figure 5 This is a graph showing the battery capacity evaluation results of Embodiment 1 of the present invention; Figure 6 This is a hardware architecture diagram of a lithium-ion battery micro-failure analyzer according to Embodiment 3 of the present invention; Figure 7 This is a schematic diagram of a benchtop physical deployment case according to Embodiment 3 of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0019] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0020] Example 1 This embodiment provides a method for analyzing the microscopic failures of lithium-ion batteries.

[0021] This embodiment provides a method for analyzing microscopic failures in lithium-ion batteries. It integrates macroscopic mechanical expansion probability density characteristics with a deep learning network constrained by physical boundaries to achieve accurate assessment of the health status of lithium-ion batteries and classification of microscopic failure mechanisms (LLI / LAM).

[0022] This embodiment provides a method for analyzing the microscopic failures of lithium-ion batteries. It achieves battery health status assessment and microscopic failure analysis through dual driving forces of multidimensional expansion probability density morphological features and enhanced physical information neural network (Attention-PINN).

[0023] This embodiment provides a method for analyzing the microscopic failures of lithium-ion batteries. It abandons the traditional method of differential differentiation of time-domain expansion signals, thus avoiding high noise defects, and achieves robust decoupling of mechanical degradation characteristics from a global statistical perspective. At the same time, it innovatively introduces the dual partial differential physical law hard constraint of first-order monotonically decreasing and second-order aging acceleration into the backpropagation computation graph, which significantly suppresses non-physical oscillations of deep learning models and achieves highly robust SOH precision diagnosis under multiple operating conditions and small samples. Using this high-precision macroscopic SOH as a thermodynamic prior boundary, the microscopic aging modes inside the battery are precisely decoupled through multi-objective optimization.

[0024] This embodiment provides a method for analyzing microscopic failures in lithium-ion batteries. First, discrete expansion signals during constant-current charging are acquired. A smooth probability density (PDF) curve is reconstructed using kernel density estimation, and six-dimensional high-order mechanical degradation features, such as peak position and information entropy, are extracted, overcoming the high noise deficiency of traditional time-domain differential equations. Then, an enhanced physical information neural network (Attention-PINN) with an attention decoupling mechanism is constructed. In backpropagation, a dual partial differential physical law hard constraint of monotonically decreasing and second-order aging acceleration is introduced, significantly suppressing non-physical oscillations in the data-driven model and achieving highly robust and accurate macroscopic state of equilibrium (SOH) estimation. Finally, without relying on traditional complex open-circuit voltage reconstruction, the dynamic attention weights of the Attention-PINN network are extracted. Combined with the macroscopic SOH prior boundary, a multivariate nonlinear mapping equation is constructed to accurately decouple microscopic aging modes such as loss of active material (LAM) and loss of lithium inventory (LLI). This achieves efficient and deep diagnosis from macroscopic mechanical evolution to microscopic electrochemical failure, constructing a high-precision closed-loop battery safety management system.

[0025] This embodiment provides a method for analyzing the microscopic failures of lithium-ion batteries, such as... Figure 1 As shown, it includes the following steps: Step 1: Acquisition of expansion signal and spatial reconstruction of probability density (PDF) of constant current charging segment.

[0026] Step 101, Expansion Data Acquisition: Obtain the discrete expansion displacement sequence of the battery during the constant current charging (CC) stage.

[0027] Because batteries suffer irreversible mechanical damage during long-cycle cycles due to side reactions such as thickening of the solid electrolyte interphase (SEI) film and accumulation of dead lithium, their macroscopic expansion signals contain underlying mechanical degradation characteristics.

[0028] Under constant current charging conditions, the charged amount Q is directly proportional to the charging time t. When the expansion of the cell remains within a certain micro-element range, the accumulated charged amount within that range directly corresponds to the number of lithium ions embedded in the negative electrode at that time. Therefore, from an electrochemical thermodynamic perspective, the probability density distribution of the expansion essentially maps the relationship between the residence probability of the phase change reaction occurring within a unit mechanical deformation range and the amount of lithium ion embedding. By extracting the probability density (PDF) morphology of the expansion, the decay of the microscopic phase transition dynamics inside the battery can be intuitively quantified.

[0029] In actual engineering implementation, in order to avoid the problem of local high-frequency noise amplification caused by directly differentiating the time domain signal, the discrete expansion displacement time series acquired online by the hardware sensor is directly used as the core underlying input.

[0030] Step 102: Hampel pretreatment.

[0031] Let the effective expansion displacement sampling sequence after Hampel filtering preprocessing to remove abnormal abrupt changes in the sensor be: .

[0032] Step 103: Construct the dilated PDF curve using the Gaussian kernel function.

[0033] The parameterless kernel density estimation (KDE) algorithm is used to discretize sampling points for each true inflation amount. Using the anchor point, a smooth and continuous probability density curve can be directly reconstructed in the spatial domain.

[0034] The mathematical formula for KDE is as follows: ; in, The continuous probability density curve is reconstructed from discrete expansion data; n is the total number of samples of effective expansion within the current constant current charging segment; h is the adaptive bandwidth parameter that controls the smoothness. This represents the value of the i-th actual expansion measurement point in the discrete sampling sequence; This is the kernel function.

[0035] Among them, kernel function Using a standard Gaussian kernel function, high-frequency glitches are naturally filtered out through spatial probability superposition, and its expression is as follows: ; in, The independent variable of the Gaussian kernel function is substituted into the actual implementation as... .

[0036] Through the above calculation steps, without relying on any complex time or charge integration, the discrete dilated signal accompanied by noise in the time domain can be directly calculated. The data is robustly transformed into a continuous probability density statistical curve that can characterize battery aging, thus establishing a high signal-to-noise ratio data foundation for the subsequent extraction of higher-order mechanical features.

[0037] Step 2: Extraction and analysis of multidimensional higher-order mechanical degradation features.

[0038] The continuous probability density curve reconstructed in step 1 From a probabilistic and statistical perspective, six independent high-order mechanical features are extracted to deeply characterize the microscopic morphological distortions inside the battery caused by phase transition kinetic decay and residual stress accumulation. (1) Peak Position F1: The main peak of the PDF curve directly corresponds to the mechanical response range (i.e., the main phase transition plateau) where the battery has the highest probability of remaining in the constant current charging process and the expansion rate is the most gradual. The main peak position characterizes the absolute expansion displacement of the main phase transition plateau, and its mathematical definition is the horizontal axis corresponding to the maximum probability density. Throughout the battery's entire life cycle, the monotonically increasing peak position directly quantifies the macroscopic residual irreversible expansion accumulation caused by side reactions such as continuous thickening of the SEI film and accumulation of dead lithium. It is the most intuitive absolute scale for assessing the overall mechanical aging of the battery. The calculation formula is: ; Here, argmax is a mathematical function used to find the independent variable that makes a certain function reach its maximum value; denoted as the probability density function of the expansion displacement reconstructed through kernel density estimation; x represents the independent variable of the battery's constant current charging expansion displacement.

[0039] (2) Expected Mean F2: Represents the centroid of the overall expansion probability distribution, reflecting the global concentration trend of the expansion signal throughout the entire charging cycle. Compared to the peak position at a single point, F2 is more robust to local stress concentration and distribution distortion, comprehensively reflecting the macroscopic thickness change caused by the consumption of recyclable lithium ions. The calculation formula is: ; in, It represents the mathematical expectation.

[0040] (3) Peak Height F3: Characterizes the density or proportion of expansion data points under a specific deformation state. Furthermore, according to the basic properties of probability theory, the total area of ​​the PDF curve is always 1. At the end of the battery life, especially in the severe aging stage, due to the sharp increase in polarization resistance and the large loss of active materials, the battery reaches the cutoff voltage very quickly in the constant current (CC) charging stage, which leads to a significant compression of the constant current charging window. This greatly weakens the reversible lattice expansion of the battery in this stage, and the expansion deformation almost stops. Since a large amount of expansion sampling data is highly concentrated in a very narrow numerical range, in order to maintain the mathematical constraint that the total area is 1, the peak height of the PDF curve will rise significantly at the end of the life, and the curve will become extremely sharp. This height increase phenomenon accurately reflects the physical essence of the shortening of the effective charging range and the loss of reversible deformation capability at the end of the aging battery.

[0041] The mathematical definition of peak height is the global maximum probability density value of the PDF curve: ; Where max represents the maximum value.

[0042] (4) Standard Deviation (F4): Based on the theory of continuous probability distribution, this feature reflects the fluctuation range and dispersion of the expansion signal around its mathematical expectation. At the end of the aging lifespan, with the sharp reduction in effective constant current charging time, the range of mechanical strain variation shrinks significantly. The extreme concentration of data distribution leads to a sharp decrease in this feature. This gradual reduction in overall dispersion is an important precursory feature for capturing the end of a battery's lifespan. The formula for calculating the standard deviation is: .

[0043] (5) Skewness (F5): Characterizes the skewness and asymmetry of the expansion probability density distribution. In the later stages of battery aging, uneven degradation within the electrodes (such as localized lithium plating and uneven porosity distribution) often occurs. These phenomena lead to localized stress concentration and asymmetric deformation within the electrodes. This characteristic variation can extremely sensitively capture the asymmetry of this micromechanical evolution. The skewness calculation formula is: .

[0044] (6) Shannon Entropy (F6): An index introduced based on information theory, used to quantify the system state complexity or uncertainty of the expansion probability density distribution. Information entropy reflects the diversity of mechanical deformation states within the battery from the perspective of the overall thermodynamic system. In a new battery state, the reversible lithium insertion / extraction process is uniform and sufficient, and the cell experiences a wide and smooth deformation range. At this time, the battery state is relatively active, and the information entropy is high. However, at the end of battery aging, the constant current charging range is severely compressed, and the battery loses its slow, uniform lattice reversible expansion capability, with deformation extremely concentrated in a single narrow range. This loss of diversity in microscopic mechanical responses and the accumulation of extreme states directly manifest as a significant decrease in information entropy on a macroscopic level. Therefore, the sharp reduction in information entropy is an extremely advanced physical evaluation index characterizing the degradation of the battery's internal phase transition capability and the intensification of system irreversibility. Its main calculation formula is: .

[0045] Step 3: Construct a physical information neural network and a multi-scale joint loss function to perform SOH estimation.

[0046] Step 301: To overcome the pain points of physical amnesia and feature redundancy in traditional data-driven models, a lightweight architecture of Information Neural Network (PINN) integrating a feature attention decoupling mechanism and a Single-Expert (SE) is constructed. This architecture can not only adaptively mine the temporal sensitivity of mechanical features in forward propagation, but also reshape the parameter space through partial differential physical laws in backpropagation. (1) A forward propagation network that integrates attention mechanisms.

[0047] The forward propagation of the physical information neural network aims to establish a nonlinear mapping from higher-order mechanical characteristics to battery health status.

[0048] The input to the forward propagation network includes the 6-dimensional PDF high-order features constructed in step 2 and the number of battery cycles C.

[0049] The forward propagation network process can be decomposed into the following three core modules.

[0050] (A) Input layer tensor construction and preprocessing.

[0051] To eliminate gradient oscillations caused by features of different dimensions (such as micrometer-level expansion displacement and dimensionless probability density) during the reverse update of the physical information neural network, the preferred feature group sequence is first subjected to Min-Max (minimum-maximum) extreme value normalization.

[0052] Let the original multidimensional feature vector of the Cth cycle be... The corresponding input tensor is constructed as follows: ; Finally, the normalized feature vector With scalar number of loops The components are concatenated to form the underlying input tensor: ; in, This represents the total dimension of the input after concatenation. This is the original multidimensional higher-order mechanical degradation feature matrix in the model training set.

[0053] (B) Feature attention decoupling mechanism.

[0054] In the complex operating conditions throughout the battery's entire lifecycle, the degradation sensitivity of various PDF physical features is not constant. In particular, when facing extreme conditions such as low temperatures, indicators representing asymmetry, such as skewness features, will exhibit strong distortion and a decrease in signal-to-noise ratio. If traditional fully connected networks are used for direct mapping, the model is prone to global failure due to the collapse of local features. To address this, a self-attention module (SAM) was designed to achieve dynamic decoupling of multidimensional physical field information.

[0055] First, the self-attention module maps the input to a high-dimensional latent space through a two-layer cascaded fully connected network. It can autonomously learn and calculate the weight coefficients of dynamic features, which can effectively reduce the impact of features with low correlation on the accuracy of SOH estimation. ; ; In the above formula, This represents the intermediate feature vector mapped from the attention module to the high-dimensional latent space; Indicates the activation function; and The projection matrix for attention-based dimensionality reduction and enhancement; and For bias terms; Sigmoid activation function The attention weight matrix is ​​rigorously constrained to the probability distribution space of (0,1).

[0056] Subsequently, the Hadamard product is used. Adaptive weight coefficients Applied to the original input, the output is an enhanced feature stream: ; The above mechanism can automatically assign high weights to key mechanical degradation information during training and shield redundant background noise features.

[0057] (C) Nonlinear mapping of physical information.

[0058] Unlike complex multi-gated hybrid expert networks, which are prone to overfitting on small sample data, this embodiment targets decoupled, high-purity enhancement features. We employ a lightweight and robust Single-Expert MLP as the core extractor, and use nested hyperbolic tangent nonlinear activation functions to achieve deep dimensionality reduction and mapping of the feature domain. ; ; in, and This represents the weight matrix of the first and second hidden layers in a single-expert multilayer perceptron. and This represents the bias vectors of the first and second hidden layers. This represents the output feature vector of the first hidden layer. This represents the output feature vector of the second hidden layer.

[0059] Finally, by linearly collapsing the scalar space of the output layer, the estimated SOH value for the Cth cycle is obtained: ; in, Represents the weight matrix; This represents a bias constant.

[0060] The set of all learnable parameters in the above forward propagation path, including the weights and biases of the attention and expert layers, is denoted as the network's global optimization space. .

[0061] (2) Analysis and backpropagation of Jacobian matrix under physical constraints.

[0062] Purely data-driven model optimization relies solely on minimizing the mean squared error data fitting loss between the estimated value and the true label. In the formula, The mean squared error is the data fitting loss. This represents the total number of samples in the training batch. The current model is in the Battery SOH estimates for each sample; For the first The actual SOH (State of Health) value of the battery corresponding to each sample. Because it does not introduce the underlying failure mechanism, its parameter assumption space is extremely broad. When faced with unknown long-sequence test estimations, the model is highly prone to exhibiting random or abnormal oscillations on the test set.

[0063] Therefore, this embodiment deeply integrates the inherent physical laws of battery aging and constructs a physical loss function that includes explicit partial differential equations. This is to punish parameter iterations that violate prior physical mechanisms.

[0064] The microcracks induced by alternating charge-discharge stress and the continuous growth of the SEI film will irreversibly consume the limited cyclic lithium ions within the cell. Although under complex actual operating conditions, the battery may experience localized capacity regeneration due to physical factors such as rest relaxation, from a long-term aging perspective, its true health status exhibits a macroscopic decreasing trend. Therefore, to extract intrinsic degradation characteristics and filter out local abnormal fluctuations, the macroscopic SOH trend estimated by the model is constrained to a function that decreases with the number of cycles, with the necessary mathematical condition being: .

[0065] To accurately capture this derivative in the backpropagation computation graph, an automatic differentiation (AD) mechanism from deep learning frameworks is introduced. This is due to the output of the Attention layer. The input is a nonlinear function. The partial derivative of the physical information neural network output with respect to the cycle number must be analytically expanded using the chain rule into a series of Jacobian matrices of each hidden layer: ; in, This represents the battery SOH estimate output by the physical information neural network.

[0066] If we let the basis vector partial derivative of the input vector with respect to the number of iterations be... By differentiating the forward equation, the derivative backpropagation expression for the single-expert hidden layer can be obtained as follows: ; in, Represents the gradient operator. Represents the weight matrix. This indicates an operation operator.

[0067] Based on the Jacobi expansion described above, and utilizing the non-negative penalty property of the corrected linear unit (ReLU), an absolutely monotonically decreasing partial differential penalty functional is constructed: ; That is, any positive gradient that violates the physical law of macroscopic capacity decay will be captured by the ReLU activation function, thereby introducing a corresponding numerical penalty into the global objective function to effectively suppress the fitting output that violates thermodynamic common sense.

[0068] Furthermore, in the later stages of battery life, due to the initiation of microcracks within the active particles caused by alternating shear stress, and the surge in polarization resistance under high-rate and heterogeneous operating conditions, battery aging exhibits a significant "accelerated degradation (capacity plunge)" macroscopic thermodynamic characteristic. To prevent unreasonable linear passivation or extrapolation distortion in long-sequence predictions, an automatic differentiation technique is used to extract the second-order Hessian matrix by further differentiating the cycle number, thus constructing a second-order aging acceleration physical loss penalty functional: ; Furthermore, internal side reactions within the battery, such as electrolyte decomposition and stress accumulation, evolve extremely continuously and slowly on a macroscopic timescale. Real-world state of equilibrium (SOH) evolution should not exhibit high-frequency spikes; localized capacity regeneration in experimental time-series test curves typically stems from measurement errors caused by abrupt changes in equipment operating conditions or strain relaxation effects. This embodiment utilizes a backward differential operator. Characterizing the accelerated transitions between adjacent discrete states, a difference penalty is imposed on the network to constrain the estimation smoothness: .

[0069] Where N is the number of samples currently being processed. For the model in the first Estimated SOH values ​​for each sample of batteries For the model estimate of the i-th Battery SOH value after 1 cycle For the first The actual number of cycles corresponding to each discrete state.

[0070] Step 302: Overall Optimization Objectives and Model Training Process of PINN

[0071] Combining data-driven approaches with mechanistic models, the global optimization objective function of Attention-PINN is rigorously defined as a multivariate joint loss mechanism: ; in, , , For the control weights of physical partial differential equations; The regularization coefficient is used to suppress the global norm of network weights. This further prevents overfitting.

[0072] As can be seen from the above, when the Attention-PINN model is performing test estimation, even under extreme conditions where the original features fluctuate abnormally, its powerful internal Jacobian gradient matrix can still force the inverse intervention of the joint objective function, ensuring that the final SOH decay trajectory is effectively locked within the degradation trend allowed by physical laws.

[0073] Step 303: Use the trained Attention-PINN model with an integrated attention decoupling mechanism to predict SOH.

[0074] Step 4: Classification of micro aging patterns (LLI / LAM) based on macroscopic SOH prior boundary constraints of Attention-PINN.

[0075] Based on obtaining high-precision macroscopic state of health (SOH) of the battery, in order to overcome the drawbacks of traditional electrochemical open-circuit voltage (OCV) reconstruction methods being computationally cumbersome and prone to getting trapped in local optima, this step fully explores the physical connotation of the internal attention weights and 6-dimensional higher-order mechanical features of the Attention-PINN network in the previous step, and constructs a microscopic aging mode (which can be correlated with lithium ion loss (LLI) and active material loss (LAM)) based on feature sensitivity decoupling and macroscopic SOH prior boundary constraints.

[0076] During long-cycle battery operation, the self-attention module (SAM) embedded in the forward propagation of the Attention-PINN network in step 3 dynamically generates adaptive weight coefficients for 6-dimensional mechanical features (peak position, expected value, peak height, standard deviation, skewness, and information entropy) for different cycle stages. Since each mechanical feature has a strong physical orientation towards different microscopic responses (e.g., the monotonically increasing peak position and expected value are directly and strongly correlated with SEI film thickening and residual irreversible expansion accumulation; while the increase in peak height and the sharp decrease in standard deviation and information entropy characterize the compression of the effective charge / discharge window and the loss of lattice reversible deformation capability), this embodiment extracts the dynamic distribution vector of the attention weights after training convergence to quantify the contribution of each mechanical degradation channel to macroscopic SOH decay in the current cycle.

[0077] To establish a deterministic mapping between macroscopic mechanical degradation characteristics and microscopic electrochemical causes, a highly robust macroscopic state of equilibrium (SOH) output by Attention-PINN is introduced as a global thermodynamic prior boundary. Based on the "bottleneck effect" and complementary constraints of battery internal capacity decay, a multivariate nonlinear regression mapping equation from multidimensional mechanical characteristics to microscopic failure indices is constructed: ; in, The lithium-ion inventory loss index, the positive electrode active material loss index, and the negative electrode active material loss index are respectively characterized. It is a 6-dimensional higher-order mechanical eigenvector; Dynamic weights for attention; and The optimal mapping matrix and bias are trained using multi-task physical boundary constraints. This is the Hadamard product, also known as the term-by-term product.

[0078] Through this feature-mechanism joint mapping, efficient and unambiguous classification and tracing of macroscopic mechanical expansion characteristics to microscopic failure causes can be achieved without the need for complex electrochemical half-cell reconstruction.

[0079] Step 5: Intelligent battery safety early warning and dynamic control strategy based on closed-loop microscopic diagnostic results.

[0080] After accurately decoupling the current macroscopic state of health (SOH) and microscopic causes of aging in step 4, the battery management system (BMS) will execute differentiated closed-loop safety control and operation and maintenance strategies based on multidimensional diagnostic boundaries: Scenario 1, Electrode Structure Failure and Lithium Plating Safety Warning: When microscopic diagnostic results show an index of negative electrode active material loss... When the set security threshold is exceeded (e.g.) If the charge rate is >15%, it is determined that the negative electrode lattice inside the cell has severely collapsed. Continuing to charge at a high rate poses an extremely high risk of local lithium plating and dendrite piercing the separator. The BMS will immediately trigger a high-level safety warning, actively reduce the maximum allowable charging rate (C-rate) of the single cell, and isolate it through the equalization circuit if necessary, fundamentally blocking the thermal runaway path. Scenario 2, Dynamic charging optimization based on available lithium consumption: If diagnostic results indicate that capacity degradation is primarily dominated by lithium-ion loss (LLI) (e.g., The threshold can be adjusted according to the actual situation, while the positive and negative electrode active materials LAM remain relatively intact (e.g., (The threshold can be adjusted according to the actual situation) will be judged as normal stable aging; the BMS will adaptively adjust the charge and discharge cutoff voltage range to delay the consumption of remaining active lithium by relieving polarization pressure, thereby maximizing the remaining lifetime (RUL).

[0081] Scenario 3, Multi-dimensional Decommissioned Battery Failure Utilization Rating: When the macroscopic SOH decays to the decommissioning threshold (e.g., 80%), a comprehensive microscopic report including LLI and LAM will be generated; for batteries with lower LAM (e.g., Cells with severe structural damage (e.g., [missing information]) are directly assigned a high safety rating, allowing them to be used in large-scale energy storage systems even if they fail due to degradation; for cells with severe structural damage (e.g., [missing information]), or ), and are subject to mandatory scrapping and material recycling.

[0082] This step integrates deep physicochemical characteristics with macroscopic mechanical models, ultimately transforming them into executable device-side control commands, thus achieving high-precision, robust, and intelligent health and safety management throughout the entire lifecycle of lithium-ion batteries.

[0083] To fully verify the effectiveness of the method proposed in this embodiment, an aging cycle experiment covering the entire life cycle of a single battery was carried out. This embodiment provides a complete operation process for assessing the health status of lithium-ion batteries by integrating mechanical expansion characteristics and physical information neural networks.

[0084] The experimental subject was a 5Ah lithium-ion pouch battery with a discharge depth of 100%, a charge rate of 0.5C, a CCCV charging strategy, and a discharge rate of 0.5C. The experimental environment was room temperature (25℃). Figure 2 The figure shows the expansion curves of the battery throughout its entire lifecycle. From the global trend, it can be observed that the mechanical expansion signal throughout the entire lifecycle exhibits a superposition of two typical mechanical responses: first, high-frequency periodic elastic expansion and contraction accompanying the reversible lithium-ion insertion / extraction behavior during charge-discharge cycles; second, the curve baseline shows a significant and continuous monotonically increasing trend as cycle time progresses. This irreversible drift of the macroscopic baseline directly characterizes the residual strain (irreversible expansion) accumulated by side reactions such as the continuous growth of the solid electrolyte interphase (SEI) film, the accumulation of dead lithium, and damage to the structure of electrode active particles. Furthermore, at the end of the battery's aging life, the slope of the irreversible expansion increases sharply, accurately mapping the physical characteristics of the nonlinear accelerated degradation within the battery at the end of its lifespan.

[0085] Subsequently, following step 1, the expansion curve is reconstructed into a PDF characteristic curve. For example... Figure 3 As shown, the reconstruction results of the expansion displacement during the constant current charging stage of the battery are presented over 300 aging cycles. With the increase in the number of cycles (the graph color gradually transitions from dark blue representing the early stage of life to dark red representing the late stage), the global morphology and distribution range of the PDF feature curves undergo extremely significant mechanical evolution. For example, the most intuitive mechanical response feature is that the main peak position of the PDF curve exhibits a continuous and strong monotonically shifting trend to the right. Furthermore, as the probability density distribution shifts overall towards the high expansion range, the envelope shape, discrete span, and peak density of the curves all undergo microscopic distortions. Based on these reconstructed PDF curves, as described in step 2, six high-order mechanical degradation features—peak position, expected value, peak height, standard deviation, skewness, and information entropy—were rigorously extracted, laying a data foundation containing rich degradation mechanisms for the subsequent establishment of deep learning networks.

[0086] To further verify the characterization effectiveness of the proposed feature set for the evolution of battery health status, this embodiment calculates the Pearson Correlation Coefficient (PCC) between each feature and the actual battery capacity decay sequence. For example... Figure 4 As shown, the extracted high-order mechanical features and capacity degradation exhibit a highly linear and nonlinear mapping relationship. The absolute values ​​of the correlation coefficients (|PCC|) between the peak position and the mathematical expectation are as high as 0.9984 and 0.9984, respectively, strongly demonstrating that the centroid displacement of the macroscopic expansion morphology is the most crucial indicator characterizing the accumulation of residual stress and loss of active lithium within the electrode. Simultaneously, the correlation coefficients of skewness (0.9855), information entropy (0.9678), and standard deviation (0.9653) are also consistently at extremely high levels above 0.9, indicating that the asymmetry of the expansion probability distribution and the decay of system state complexity are highly synchronized with battery aging.

[0087] Then, following step 3, as follows Figure 5 As shown, the absolute capacity decay trajectory of the experimental battery over its entire lifespan of nearly 400 cycles is illustrated. Figure 5 The solid black line represents the actual capacity, and the dashed red line represents the estimated output capacity (the overall curve divided by the initial capacity value indicates the healthy state). From the full-cycle tracking trajectory, it can be observed that under hundreds of test cycles, the estimated curve consistently and closely follows the actual physical degradation baseline. The above data and results analysis fully demonstrate that the method proposed in this embodiment, which integrates macroscopic mechanical characteristics and physical boundary constraints, can achieve high-precision and highly stable online assessment of the health state of lithium-ion batteries in practical engineering applications.

[0088] Example 2 This embodiment provides a lithium-ion battery micro-failure analysis system, comprising: The probability density space reconstruction module is configured to: acquire the discrete expansion displacement sequence of the lithium-ion battery during the constant current charging stage, and reconstruct the expansion probability density curve using parameterless kernel density estimation. The high-order mechanical feature extraction module is configured to extract multi-dimensional high-order mechanical degradation features based on probability density curves. The neural network estimation module is configured to estimate the battery health status based on multidimensional high-order mechanical degradation features and the number of battery cycles through a physical information neural network. In the forward propagation of the physical information neural network, an attention mechanism is introduced to dynamically decouple and weight the input features, and in the backward propagation, a dual partial differential physical law hard constraint of first-order monotonically decreasing and second-order aging acceleration is introduced. The micro-mode decoupling module is configured to: extract the attention weights of the physical information neural network, combine them with the estimated battery health state as a priori boundary, and decouple the battery's micro-aging mode through a multivariate nonlinear mapping equation.

[0089] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation processes are the same, so they will not be repeated here.

[0090] Example 3 Existing advanced SOH diagnostic and micro-failure analysis technologies are mostly limited to offline laboratory calculations. They lack hardware analysis devices that modularize and materialize high-order algorithms, making it difficult to implement complex deep learning networks in real-world edge computing scenarios, resulting in a serious disconnect between software and hardware.

[0091] This embodiment provides a lithium-ion battery micro-failure analyzer, which runs steps 1 to 5 of the lithium-ion battery micro-failure analysis method described in Embodiment 1. For example... Figure 6 As shown, the instrument's hardware architecture is divided into three functional layers from left to right: the sensing and acquisition layer, the microscopic failure analyzer core layer, and the interaction and execution layer.

[0092] The sensing and acquisition layer is deployed on a physical test bench or battery pack side, integrating a high-precision displacement sensor and a high-precision voltage / current tester to collect the discrete expansion displacement sequence and voltage and current signals of the target cell in real time during the constant current charging stage.

[0093] The core layer of the microscopic failure analyzer serves as the edge computing control platform, and includes a multi-channel signal conditioning and high-speed A / D conversion circuit, a KDE probability density space reconstruction engine, a 6-dimensional high-order mechanical feature extraction and operation unit, an Attention-PINN neural network estimation model, and a multivariate nonlinear mapping microscopic pattern decoupler, which completes the online calculation of macroscopic SOH and microscopic LLI / LAM ratio from the original sensing signal.

[0094] The interaction and execution layer receives diagnostic results via the onboard CAN / RS485 bus. The battery management system (BMS) then executes electrode damage and lithium plating alarms, dynamic charging optimization management, and retired battery cascade utilization ratings, achieving a closed-loop safety management system from perception and analysis to execution.

[0095] This embodiment integrates the lithium-ion battery micro-failure analysis method described in Embodiment 1 into a micro-failure analyzer that includes a sensing and acquisition layer, an edge computing layer, and an interactive execution layer, thereby driving the BMS to perform safety warnings and dynamic control.

[0096] This embodiment constructs a micro-failure analyzer hardware entity that encompasses signal sensing, edge computing calculation, and BMS closed-loop management, thereby driving the BMS to perform early warning and tiered utilization rating, realizing a closed-loop battery safety management system that combines software and hardware.

[0097] To verify the online diagnostic performance and hardware feasibility of the microscopic failure analyzer in a real physical system, a complete hardware-in-the-loop test bench was built, which includes a single-unit charge-discharge test device, a high-precision pressure distribution measurement system, and an edge computing core.

[0098] like Figure 7 As shown, the physical deployment of the analytical instrument hardware platform: Data acquisition layer: Electrochemical boundary conditions are provided by single-cell charge-discharge testing equipment, which supports high-precision electrical signal acquisition with a wide range of up to 24kW maximum power, voltage 0~5V, and current -300A~+300A. It is responsible for accurately recording the voltage, current, and time series of the battery during the constant current charging stage. Macroscopic mechanical sensing adopts a pressure distribution measurement system, which tightly attaches an ultra-thin and flexible film pressure sensor to the surface of the target cell and places it in a rigid test fixture to obtain the contact pressure distribution of the entire contact surface in real time with high fidelity.

[0099] The core layer of the analyzer utilizes a high-performance edge computing industrial computer as its core. The pressure data acquisition unit transmits the acquired data to the computer, where the software records pressure distribution video files at a set frequency and exports a .csv data table containing full matrix information. The edge computing core aligns the electrochemical data and expansion pressure data from the battery testing equipment using high-precision timestamps and feeds them into the internally integrated KDE spatial reconstruction engine and Attention-PINN deep learning model.

[0100] Interaction and Execution Layer: The edge computing core and the main control host computer of the battery charging and discharging test equipment establish closed-loop communication through TCP / IP or high-speed CAN bus to realize real-time feedback of diagnostic results and dynamic takeover of charging and discharging protocols.

[0101] The instrument's operating procedure is as follows: (1) Signal Synchronization and Status Assessment: After the test bench is started, the battery testing equipment executes a charging program on the pouch battery according to the 0.5C constant current constant voltage (CCCV) strategy. At the same time, the distributed pressure device enters real-time recording mode to record the stress process. After the constant current segment ends, the edge analyzer core automatically retrieves the exported time-series expansion pressure matrix, completes the smooth reconstruction of the continuous PDF curve, and extracts the 6-dimensional high-order mechanical degradation features. Subsequently, the Attention-PINN network calls the fixed weights to calculate the current macroscopic health status estimate of the battery.

[0102] (2) Micro-mode decoupling and command issuance: Under the strong constraint of macro-level prior boundaries, the analyzer separates the micro-indices of the current aging state. , , It generates a microscopic physical examination and diagnosis report and sends it to the bus network.

[0103] (3) Dynamic control closed-loop execution: If the message output by the analyzer determines the loss index of the negative electrode active material Exceeding the set 15% threshold will immediately trigger an electrode structure failure warning. The bus then sends a dynamic derating command to the Nebula test equipment, utilizing its wide-band control capability of -300A to +300A to forcibly reduce the maximum allowable charging current limit for the next cycle to below 0.2, directly blocking potential lithium plating and thermal runaway paths from the physical test equipment end, and verifying the closed-loop safety control management.

[0104] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for analyzing microscopic failures in lithium-ion batteries, characterized in that, include: The discrete expansion displacement sequence of the lithium-ion battery during the constant current charging stage is obtained, and the expansion probability density curve is reconstructed using parameterless kernel density estimation. Based on probability density curves, multidimensional higher-order mechanical degradation features are extracted; Based on multidimensional high-order mechanical degradation characteristics and combined with the number of battery cycles, the battery health status is estimated through a physical information neural network. In the forward propagation of the physical information neural network, an attention mechanism is introduced to dynamically decouple and weight the input features, while in the back propagation, a dual partial differential physical law hard constraint of first-order monotonically decreasing and second-order aging acceleration is introduced. The attention weights of the physical information neural network are extracted and combined with the estimated battery health state as a prior boundary. The micro-aging mode of the battery is decoupled through a multivariate nonlinear mapping equation.

2. The method for analyzing microscopic failures in lithium-ion batteries as described in claim 1, characterized in that, The multidimensional higher-order mechanical degradation features include: Peak position, representing the amount of expansion displacement corresponding to the main peak of the probability density curve; Mathematical expectation, representing the centroid of the morphology of the inflationary probability distribution; Peak height represents the maximum probability density value of the probability density curve. Standard deviation characterizes the range of fluctuation of an expansion displacement sequence around its mathematical expectation; Skewness characterizes the direction and degree of asymmetry of the inflation probability density distribution; Information entropy characterizes the state complexity of an inflated probability density distribution.

3. The method for analyzing microscopic failures in lithium-ion batteries as described in claim 1, characterized in that, The global optimization objective function of the physical information neural network is: ； in, , , For the control weights of physical partial differential equations; The regularization coefficient is used. To optimize space globally; The loss is the mean square error data fitting loss; the first-order monotonically decreasing partial differential penalty functional. Second-order aging acceleration physical loss penalty functional Differential penalty N is the total number of loop samples. The SOH value of the battery under the i-th cycle estimated by the model. For the model estimate of the i-th Battery SOH value after 1 cycle This represents the number of iterations corresponding to the i-th iteration.

4. The method for analyzing microscopic failures in lithium-ion batteries as described in claim 1, characterized in that, The multivariate nonlinear mapping equation is: ； in, The lithium-ion inventory loss index, the positive electrode active material loss index, and the negative electrode active material loss index are respectively characterized. To maintain the degeneracy characteristics of higher-order mechanics; Dynamic weights for attention; and The optimal mapping matrix and bias are trained using multi-task physical boundary constraints. It is the product of Hadama.

5. The method for analyzing microscopic failures in lithium-ion batteries as described in claim 1, characterized in that, The micro-aging mode includes the loss of active material and the loss of available lithium ions.

6. The method for analyzing microscopic failures in lithium-ion batteries as described in claim 1, characterized in that, This also includes implementing differentiated closed-loop battery safety management strategies based on the battery's micro-aging patterns: When the loss index of negative electrode active material exceeds the preset safety threshold, a safety warning is triggered and the maximum charging rate is reduced or the battery cell is isolated. When capacity degradation is dominated by the loss of available lithium ions and the loss of positive and negative electrode active materials remains relatively intact, adjust the charge and discharge cutoff voltage range. When the battery health deteriorates to the retirement threshold, the failure utilization rating of the retired battery is determined based on the degree of loss of active materials.

7. The method for analyzing microscopic failures of lithium-ion batteries as described in claim 1, characterized in that, In the forward propagation of the physical information neural network, a single-expert multilayer perceptron is used as the feature extractor. The feature domain is reduced and mapped by a nested hyperbolic tangent nonlinear activation function.

8. The method for analyzing microscopic failures in lithium-ion batteries as described in claim 1, characterized in that, The attention mechanism maps the input to a high-dimensional latent space through a two-layer cascaded fully connected network, autonomously learns and calculates the weight coefficients of dynamic features, and applies the weight coefficients to the input through the Hadamard product.

9. A microscopic failure analysis system for lithium-ion batteries, characterized in that, include: The probability density space reconstruction module is configured to: acquire the discrete expansion displacement sequence of the lithium-ion battery during the constant current charging stage, and reconstruct the expansion probability density curve using parameterless kernel density estimation. The high-order mechanical feature extraction module is configured to extract multi-dimensional high-order mechanical degradation features based on probability density curves. The neural network estimation module is configured to estimate the battery health status based on multidimensional high-order mechanical degradation features and the number of battery cycles through a physical information neural network. In the forward propagation of the physical information neural network, an attention mechanism is introduced to dynamically decouple and weight the input features, and in the backward propagation, a dual partial differential physical law hard constraint of first-order monotonically decreasing and second-order aging acceleration is introduced. The micro-mode decoupling module is configured to: extract the attention weights of the physical information neural network, combine them with the estimated battery health state as a priori boundary, and decouple the battery's micro-aging mode through a multivariate nonlinear mapping equation.

10. A microscopic failure analysis instrument for lithium-ion batteries, characterized in that, include: The sensing and acquisition layer, deployed on a physical test bench or battery pack side, is used to acquire discrete expansion displacement sequences of lithium-ion batteries during the constant current charging phase. The core layer, as the edge computing master control platform, executes the steps in the lithium-ion battery micro-failure analysis method as described in any one of claims 1-7 to obtain the micro-aging mode. The interaction and execution layer is used to receive micro-aging modes and is controlled by the battery management system to perform safety management.

Images