Battery health assessment system based on AI large model

Abstract

The invention relates to the technical field of battery health assessment, and provides a battery health assessment system based on an AI large model, and the system comprises a first module which obtains the ultrasonic characteristic data of a battery in different energy flow states through non-invasive ultrasonic detection; the second module is used for constructing a large battery health assessment model; and the third module is used for inputting the battery electrical parameter data and the ultrasonic characteristic data into a large battery health assessment model and outputting a battery health assessment result. According to the invention, the health degree of the battery can be accurately and rapidly detected in multiple environments.

Description

Battery Health Assessment System Based on AI Large Model Technical Field

[0001] This invention relates to the field of battery health assessment technology, and more specifically to a battery health assessment system based on an AI large model. Background Technology

[0002] With the explosive growth of the global new energy industry, batteries, as core energy storage and supply units, have been widely used in key areas such as new energy vehicles, large-scale energy storage power stations, and portable electronic devices. Battery State of Health (SOH), as a core indicator for measuring battery aging, remaining lifespan, and safety performance, directly affects the operational safety, reliability, and economy of the equipment. For example, in the new energy vehicle sector, excessive SOH assessment errors can lead to misjudgments of driving range or overcharging and over-discharging of the battery, causing safety accidents such as thermal runaway. In energy storage power station scenarios, inaccurate health assessments can cause deterioration in battery cluster consistency, significantly shortening the overall lifespan and increasing operation and maintenance costs.

[0003] However, existing battery health assessment technologies have several significant drawbacks: while existing electrochemical detection methods can reflect the internal state of the battery to some extent, the detection process is time-consuming, easily damages the battery, and cannot achieve real-time online monitoring. Indirect assessment methods based on electrical parameters such as voltage and current are greatly affected by ambient temperature, charge / discharge rate, and other operating conditions, lack sensitivity, and are difficult to capture early changes in the battery's internal microstructure. Furthermore, existing assessment models often use single-feature modeling, failing to effectively integrate complementary information from multi-dimensional data, and some black-box models have poor interpretability, making it difficult to quantify the intrinsic correlation between features and battery health status, resulting in high assessment errors. Summary of the Invention

[0004] The present invention aims to provide a battery health assessment system based on an AI large model, which can accurately and quickly detect the health of batteries in various environments.

[0005] The battery health assessment system based on an AI-powered large-scale model includes: a first module that acquires ultrasonic characteristic data of the battery under different energy flow states through non-invasive ultrasonic detection; a second module that constructs a large-scale battery health assessment model; and a third module that inputs both the battery electrical parameter data and the ultrasonic characteristic data into the large-scale battery health assessment model and outputs the battery health assessment results.

[0006] Beneficial effects: By acquiring ultrasonic feature data of different energy flow states through non-invasive ultrasonic detection, and combining it with battery electrical parameter data, a dedicated AI model is input to output evaluation results. This avoids damage to the battery caused by invasive detection, and achieves multi-dimensional data fusion, solving the limitations of single data evaluation in existing methods, and achieving accurate and rapid detection of battery health under various environments.

[0007] Preferably, the different energy flow states include battery SOC values ​​of 5%-35%, 35%-65%, and 65%-95%.

[0008] Beneficial effects: It clearly defines the three SOC ranges corresponding to different energy flow states: 5%-35%, 35%-65%, and 65%-95%, accurately delineating the key stages of battery energy change, making subsequent data collection and model training more targeted, avoiding feature confusion caused by indiscriminate analysis of the entire SOC range, and laying the foundation for improving evaluation accuracy.

[0009] Preferably, the battery SOC values ​​of 5%-35%, 35%-65%, and 65%-95% are set as independent stages. In each stage, ultrasonic feature data collected every few seconds is extracted. The ultrasonic feature dataset of each stage is used as an independent dataset, and the dataset is used to train a large battery health assessment model.

[0010] Beneficial effects: Training the model with the three SOC intervals as independent datasets can fully explore the unique patterns of ultrasonic feature data at each stage, reduce cross-interference between data from different SOC stages, enable the model to more accurately capture battery health-related features at each stage, and further improve the reliability and accuracy of the evaluation results.

[0011] Preferably, the battery electrical parameter data includes battery voltage and current; the ultrasonic feature data includes time-domain features and frequency-domain features; the time-domain feature data includes flight time, maximum value, minimum value, and peak-to-peak value; the frequency-domain features include the maximum amplitude of the characteristic frequency band.

[0012] Beneficial effects: The specific types of battery electrical parameter data (voltage, current) and ultrasonic characteristic data are clearly defined. Among them, time-domain characteristics (time of flight, peak-to-peak value, etc.) can reflect changes in the internal structure of the battery, and frequency-domain characteristics (maximum amplitude of characteristic frequency band) can capture the evolution of electrochemical characteristics. It comprehensively covers the key dimensions required for battery health assessment, provides rich and physically meaningful input information for the model, and completely solves the problem of assessment bias caused by incomplete feature coverage in existing methods.

[0013] Preferably, the large-scale battery health assessment model includes a generalized additive model, and the expression of the generalized additive model includes...

[0014] Where μ is the learnable intercept term, Let represent the smoothing function for the i-th feature, where i and N are positive integers; E(y) represents the expected value of the model's output random variable y; the architecture of the generalized additive model includes a multi-layer differentiable forgetting decision tree, which achieves data segmentation through feature selection functions; the feature selection functions include:

[0015] in, Characterized as a feature selection function, R (d×D) Choose a matrix for learnable features. To achieve sparse feature selection, j and D are both positive integers, where D represents the total dimension of the input features, and x i It is represented by the value of the i-th input feature variable.

[0016] Beneficial effects: The architecture combining a generalized additive model with a multi-layer differentiable forgetting decision tree allows for the individual quantification of the contribution of each feature to the SOH, overcoming the core pain point of the traditional black-box model's lack of interpretability. At the same time, the sparse feature selection achieved through a learnable feature selection matrix effectively eliminates redundant information, reduces the model's computational complexity, and balances the model's strong expressive power and interpretability, resolving the contradiction between "high accuracy and interpretability" in existing models.

[0017] Preferably, the characteristic function is made controllable by introducing annealing parameters, including: Wherein, the annealing parameter is T, which is a constant; D represents the total number of dimensions of the input features and is a positive integer; and Cij represents the selection weight of the i-th decision node for the j-th input feature.

[0018] Beneficial effects: By introducing annealing parameters, the feature selection process is dynamically controllable. In the early stage of training, a wide range of features are explored, and in the later stage, core features are focused. This effectively solves the problem of uncontrollable feature selection in the NODE model, avoids feature entanglement in the latent space, ensures that each decision tree focuses on only one core healthy feature, improves the stability of model training and the targeting of feature recognition, and provides a reliable feature selection basis for subsequent accurate evaluation.

[0019] Preferably, in order to achieve that each layer of the differentiable forgetting decision tree ultimately selects only one feature, an annealing parameter is introduced, and a learnable feature selection matrix is ​​used, including: setting a learnable feature selection matrix S∈R(F×K); where F represents the number of features and K represents the number of tree nodes, and the learnable feature selection matrix is ​​used to achieve sparse feature selection.

[0020] Beneficial effects: By combining the annealing parameters with a learnable feature selection matrix S, the matrix dimension relationship between the number of features and the number of tree nodes is clarified, achieving efficient sparse feature selection. This forces each decision tree layer to focus on only one core feature, completely avoiding evaluation bias caused by the cross-interference of multiple features. This not only improves the accuracy of feature recognition but also simplifies the model operation process, effectively balancing evaluation accuracy and computational efficiency.

[0021] Preferably, a gating connection mechanism and a low-rank attention mechanism are configured; the gating connection mechanism includes: setting a gating function: A connection is established only when adjacent tree nodes select the same feature; otherwise, it is broken. Where C... j The feature index of the j-th tree node is represented as the final selected feature index; the low-rank attention mechanism includes: A = A1A2, where A1∈R (I×E) A2∈R (E×(I-1)N) .

[0022] Beneficial effects: The gating connection mechanism ensures the purity of feature paths and avoids confusion of different feature contributions; the low-rank attention mechanism reduces the computational complexity from O(N²) to O(NE), meeting the needs of real-time monitoring; the design of dynamically adjusting the information flow enables the model to automatically focus on the feature path with the greatest impact on SOH, and strengthens the core contribution through weighted linear combination, solving the problems of chaotic feature paths, large computational load and insufficient accuracy in traditional models, and achieving dual optimization of efficiency and accuracy.

[0023] Preferably, by setting a dynamically adjusted information flow, the large-scale battery health assessment model focuses on the feature paths that have the greatest impact on SOH; the dynamically adjusted information flow includes: Where l', j', j, and l are all positive integers; This is used to characterize the fact that connections are established only when adjacent tree nodes select the same feature. =1, otherwise disconnect. G j Represented as the output result of the j-th tree node; g l’j’ (x) represents the response function of the j'-th tree node in the previous layer to the input x; x represents the input feature vector of the model; a lj’j This is represented by the attention coefficients between adjacent layer nodes; based on the calculation formula for attention weights, the weights w are trained using a weighted linear combination. lj This enables the large-scale battery health assessment model to emphasize the most salient feature paths; the formula for calculating attention weights includes:

[0024] Where L is a positive integer, g ij (x l The representation is the pair of input feature x between the i-th tree node in the l-th layer.l The nonlinear response output, w lj The learnable weighting coefficients corresponding to the j-th tree node in the l-th layer are represented by the attention weights; the contribution of each feature path to the SOH output is dynamically adjusted through the calculation formula of the attention weights.

[0025] Beneficial effects: For scenarios with a tree depth greater than 2, the strategy of using two independent feature selection matrices alternately effectively captures the synergistic interaction effect of multiple features in deep decision trees. This solves the problem that a single feature selection matrix cannot take into account the coupling of multiple failure mechanisms in complex aging scenarios, enabling the model to adapt to the mode switching from single feature dominance to multi-feature coupling dominance during battery aging, and further improving the comprehensiveness and accuracy of evaluation under complex working conditions.

[0026] Preferably, two independent feature selection matrices C¹ and C² are introduced to replace the single learnable feature selection matrix S; when the tree depth d is greater than 2, according to C... d The rule of =C(d-1)mod2+1 alternates between two matrices C¹ and C²; when the tree depth is no greater than 2, a single learnable feature selection matrix S is used.

[0027] Beneficial effects: Expanding the assessment results into SOH estimates, battery health trend predictions, and quantitative interpretation reports breaks the limitations of traditional assessments that only provide single numerical results. It not only allows users to understand the current health status of the battery but also predicts future degradation trends. Furthermore, by clarifying the contribution of features and decision-making logic through quantitative interpretation, it provides full-chain decision support for battery charging and discharging strategy optimization, maintenance cycle formulation, and replacement early warning, significantly improving the engineering implementation value and operational practicality of the technology.

[0028] Preferably, the battery health assessment results include SOH (State of Health) estimates, battery health trend predictions, and a battery quantitative explanation report.

[0029] The first approach offers several advantages. First, the non-invasive testing method provides comprehensive data dimensions, ensuring battery safety and the integrity of the assessment. Utilizing non-invasive ultrasonic testing technology, characteristic data of the battery under different energy flow states is acquired without disassembling or damaging the battery. This fundamentally avoids the damage to the battery's structural integrity and sealing caused by invasive testing, reducing safety risks such as short circuits and leakage during testing. It is particularly suitable for batteries in service in scenarios such as new energy vehicles and energy storage power stations. Second, this approach does not rely on a single data type but combines ultrasonic characteristic data with electrical parameters such as battery voltage, current, temperature, and SOC value. This achieves multi-dimensional data coverage of "external operating status + internal structural changes"—time-domain data (time of flight, peak-to-peak value, etc.) in the ultrasonic characteristics can reflect microstructural changes such as electrode expansion and lithium dendrite accumulation inside the battery, while frequency-domain data (maximum amplitude of characteristic frequency bands) can capture electrochemical characteristics such as electrolyte drying and interfacial impedance changes. Combined with basic electrical parameters, this provides a more comprehensive and relevant data source for health assessment, solving the problem of one-sided assessment caused by the single data dimension of traditional methods.

[0030] Secondly, the model design balances accuracy and interpretability. The constructed large-scale battery health assessment model integrates the interpretability of the Generalized Additive Model (GAM) with the strong expressive power of multi-layer differentiable forgetting decision trees, specifically addressing the core pain points of traditional assessment models. On the one hand, through innovative designs such as introducing annealing parameters, learnable feature selection matrices, and gating connection mechanisms, the model can achieve sparse feature selection, ensuring that each layer of the decision tree focuses on core features, avoiding feature entanglement in the latent space, and dynamically adjusting the information flow to accurately capture the dynamic evolution of feature importance during battery aging (such as the sudden change in the importance of the Max feature in fast charging scenarios), significantly reducing prediction errors. On the other hand, the additive structure of the GAM allows the contribution of each feature to SOH to be quantified and visualized individually. Combined with the feature interaction matrix for quantitative analysis of the synergistic effect of dual features, it breaks the dilemma of traditional black-box models that "only provide results, not reasons," making the generation logic of assessment results transparent and traceable, providing a clear basis for subsequent battery maintenance and fault diagnosis.

[0031] Finally, it adapts to multiple environments and responds quickly. The first approach divides the battery energy flow state into three SOC stages: 5%-35%, 35%-65%, and 65%-95%. The ultrasonic feature data for each stage is used as an independent dataset for model training. This staged data processing method can fully explore the unique patterns of battery health characteristics within different SOC ranges, avoiding feature confusion caused by indiscriminate training across the entire SOC range, and ensuring that the model maintains high adaptability across all energy states. Simultaneously, optimization techniques such as cross-validation, dynamic feature weighting, and outlier removal are employed during model training to further enhance the model's robustness. Attached Figure Description

[0032] Figure 1 is a flowchart of the battery health assessment system based on the AI ​​large model in Embodiment 1; Figure 2 is a signal amplitude-time diagram of the pneumatic signal and sensor signal in Embodiment 1; Figure 3 is a Fourier transform diagram of the sensor signal in Embodiment 1; Figure 4 is a schematic diagram of the battery health assessment large model in Embodiment 1. Detailed embodiments of the present invention will be described below with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solutions of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0033] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0034] Before describing specific implementations, it is necessary to understand that: SOC: State of Charge, which represents the percentage of the battery's current remaining charge relative to its total capacity, ranging from 0% (fully discharged) to 100% (fully charged).

[0035] SOH: Indicates the degree of degradation of the battery's current performance relative to its initial factory state, and is a core indicator for measuring the degree of battery aging.

[0036] Energy flow states: including battery charging and discharging states. Charging: Energy flows into the battery; discharging: Energy flows out of the battery.

[0037] Time of Flight (ToF): The total time difference between when an ultrasonic wave is emitted from the probe, penetrates the battery casing, is reflected by key internal interfaces (electrode-electrolyte, electrolyte-separator, etc.), and then returns to the probe. The unit is microseconds (μs). It primarily reflects the thickness changes of the internal structure of the battery (such as electrode expansion and separator deformation).

[0038] Maximum value (Max): The maximum amplitude (voltage value) of the ultrasonic reflected signal in the time domain waveform, in volts (V), reflecting the degree of energy attenuation of ultrasonic waves when they propagate inside the battery (related to electrolyte concentration and electrode porosity).

[0039] Minimum value (Min): The minimum amplitude (voltage value) of the ultrasonic reflected signal in the time domain waveform, with the unit being volts (V). Together with Max, it reflects the fluctuation range of the signal and indirectly reflects the uniformity of the microstructure inside the battery.

[0040] Peak-to-peak value (PP): The difference between Max and Min in the same time domain waveform (PP=Max-Min), with the unit being volts (V). It directly reflects the intensity fluctuation of the ultrasonic signal and can amplify the influence of abnormal areas inside the battery (such as lithium dendrite accumulation) on the signal.

[0041] Maximum amplitude of characteristic frequency band (AF): First, perform FFT transformation on the ultrasonic reflection signal in the time domain to convert it from the time-amplitude dimension to the frequency-amplitude dimension. Then, extract the maximum amplitude of the signal in the characteristic frequency band, which is an effective frequency band that matches the internal electrochemical-mechanical characteristics of the battery. This is the AF feature, which reflects the frequency domain energy distribution law of ultrasonic waves propagating inside the battery (related to electrolyte drying and changes in electrode interface impedance).

[0042] In existing technologies, the interpretability dilemma of black-box models and the lack of feature interaction terms are the current shortcomings of AI-based large-scale model evaluation. First, the lack of feature interaction terms is one of the core shortcomings of existing evaluation methods. The battery aging process is essentially a complex process involving the coupling of multiple physical fields and the synergistic effect of multiple failure mechanisms. The feature parameters corresponding to different failure modes do not exist in isolation, but rather exhibit significant electrochemical-mechanical coupling effects. For example, when electrode expansion (affecting ToF) and electrolyte drying (affecting AF) occur simultaneously, existing models can only analyze the effects of the two in isolation, while experimental data show that this coupling effect leads to a 40% increase in prediction error.

[0043] Secondly, the interpretability dilemma of black-box models has become a key obstacle restricting the implementation of technologies and the study of mechanisms. Current mainstream evaluation models are mostly based on deep learning architectures such as deep neural networks (DNNs) and recurrent neural networks (RNNs). These models fit the complex nonlinear relationship between features and the state of matter (SOH) through multiple fully connected layers and convolutional layers. While they can achieve a certain level of prediction accuracy, they are essentially "black-box" decision-making models. The root cause of their lack of interpretability lies in the irreversibility of their mathematical construction: the fully connected layers of deep neural networks map the input original features, often single original features, to a high-dimensional latent space. This is because the fully connected layers of deep neural networks cause entanglement of features in the latent space. When multiple original features are analyzed, the entanglement becomes excessive. During multiple matrix operations and nonlinear activation processes, the original physical meaning of the features is gradually diluted and entangled, ultimately making it impossible to decode the feature vectors in the latent space back into specific physical parameters. For example, after propagating through a 10-layer fully connected network, the correlation between high-frequency AF features (corresponding to changes in electrolyte concentration) and low-frequency ToF features (corresponding to changes in electrode thickness) and their original physical meaning can be lost by up to 35%. Engineers cannot know which feature combinations or decision-making logic the model uses to arrive at the SOH assessment results. This lack of interpretability leads to two serious problems: First, in fields with extremely high safety requirements, such as new energy vehicles and energy storage power stations, models that only provide prediction results without explaining the decision-making basis are difficult to gain industry trust and regulatory approval, greatly limiting the industrialization of the technology.

[0044] As shown in Figure 1, this embodiment provides a battery health assessment system based on an AI large model, including: a first module, which obtains characteristic data of the battery in different energy flow states through non-invasive ultrasonic detection.

[0045] Specifically, non-invasive ultrasonic testing is performed using Olympus ultrasonic testing equipment. A 13mm diameter 2.5MHz ultrasonic probe is vertically fixed to the battery surface using existing clamps. A 0.1mm thick continuous coupling layer is formed using a high-viscosity ultrasonic coupling agent (UT-100A) to eliminate the influence of air gaps and ensure signal transmission efficiency. The Olympus ultrasonic testing equipment operates in pulse-echo mode with a 1kHz pulse repetition frequency, ensuring signal quality while avoiding interference from multiple reflections. A 50MHz high-speed acquisition card and a 200V high-voltage pulse generator constitute the signal acquisition system. A hardware synchronous triggering mechanism ensures time synchronization accuracy of <10μs with the battery charging and discharging process.

[0046] By connecting a high-voltage pulse generator and a high-speed data acquisition card with a frequency of at least 50MHz, the pulse generator produces an excitation signal at a pulse repetition frequency of at least 1kHz, driving the probe to emit ultrasonic waves. The acquisition card uses pulse echo mode to receive the reflected signal, ensuring accurate capture of microsecond-level time-domain signals.

[0047] In the preprocessing stage, a fifth-order Butterworth bandpass filter (0.5-5MHz) is used to retain the effective frequency band, combined with a minimum-phase FIR filter to eliminate phase distortion. This ensures the time accuracy of the time-domain signal while filtering low-frequency noise and high-frequency interference, retaining the effective frequency band signal related to the battery's internal structure. The envelope extracted by Hilbert transform highlights key feature points of the battery charging and discharging signal, such as peak and valley values. Dynamic baseline correction technology eliminates baseline drift caused by inherent noise of the device used to acquire the battery charging and discharging signal and environmental interference. This device includes a bandpass filter and connecting wires. The preprocessed battery charging and discharging signal is then processed. Specifically, by continuously acquiring the same signal 64 times and averaging it in the time domain, the signal-to-noise ratio (SNR) is improved to over 40dB. Random noise is effectively suppressed, thus significantly improving the signal-to-noise ratio.

[0048] As shown in Figure 2, during the ultrasonic feature data calculation stage, the Time of Flight (ToF) is accurately calculated using a cross-correlation algorithm, achieving a time-of-flight resolution of 0.01 μs. The automatically identified aerodynamic and sensor signal amplitude features (Max, Min, PP) are smoothed using a sliding window to eliminate instantaneous fluctuations. A 2048-point Fast Fourier Transform (FFT) is performed on the preprocessed time-domain signal to extract the maximum amplitude of the 1-3 MHz characteristic frequency band as the frequency amplitude (AF) feature, as shown in Figure 3. The Pearson correlation coefficient is used to analyze the correlation between each feature and the So-Hyperion of Oscillation (SOH), eliminating redundant features with an absolute correlation coefficient <0.6. Outliers are detected and removed using the 3σ principle to ensure the reliability of the feature data.

[0049] As shown in Figure 4, the second module is to construct a large-scale battery health assessment model.

[0050] Specifically, feature engineering employs a 3σ outlier detection mechanism, multi-feature joint standardization (Z-score) to eliminate the influence of dimensions, Pearson correlation coefficient analysis to select the optimal feature combination, and a dynamic feature weighting algorithm to automatically adjust weights based on battery type. Secondly, standardization is performed using Z-score standardization to unify the dimensions of the selected features, ensuring each feature follows a normal distribution with a mean of 0 and a variance of 1, thus eliminating the impact of dimensional differences on model training. Based on battery type (e.g., ternary lithium battery, lithium iron phosphate battery) and charge / discharge rate, a dynamic feature weighting algorithm automatically adjusts the initial weights of each feature to adapt to differences in feature importance under different scenarios.

[0051] Furthermore, after extracting relevant features from the ultrasonic data, the battery charging and discharging process is initially divided into three stages based on different SOC values ​​(5%-35%, 35%-65%, and 65%-95%). This means the battery is divided into three stages based on different energy flow states: the first stage is 5%-35% SOC, the second stage is 35%-65% SOC, and the third stage is 65%-95% SOC. For each stage, ultrasonic feature data collected every few seconds is extracted. In this embodiment, to improve extraction efficiency and accuracy and reduce complexity, ultrasonic feature values ​​are extracted every 10 seconds. Therefore, three independent datasets are obtained, each corresponding to a different SOC stage. The input data for the large-scale battery health assessment model includes ultrasonic feature data and battery electrical parameter data. Battery electrical parameter data includes battery voltage and current; ultrasonic feature data includes time-domain and frequency-domain features; time-domain features include time-of-flight, maximum value, minimum value, and peak-to-peak value; frequency-domain features include the maximum amplitude of the characteristic frequency band.

[0052] Dividing the battery into three stages based on different SOC values ​​serves several purposes. First, it aligns with the nonlinear changes in battery energy flow and internal state. In the 5%-35% (low SOC) stage, the electrolyte distribution inside the battery is uneven, and the electrode active materials do not react sufficiently. In the 35%-65% (medium SOC) stage, the reaction is stable, and characteristic signal fluctuations are small. In the 65%-95% (high SOC) stage, problems such as lithium dendrite growth and electrolyte polarization are prone to occur. The characteristic correlations and health sensitivity points of the three stages differ significantly, and staged analysis can accurately capture the unique patterns of each stage. Second, it avoids feature interference across SOC ranges. The correlation between ultrasonic features (such as ToF and AF) and electrical parameters varies abruptly across different SOC ranges. For example, in the low SOC stage, ToF is significantly affected by electrode shrinkage, while in the high SOC stage, it is dominated by lithium dendrites. Merging analyses can lead to feature confusion, while staged analysis allows the model to focus on the core influencing factors in each range. Simultaneously, it improves data utilization and model training efficiency. Treating each stage as an independent dataset allows for targeted mining of key features (e.g., features are more stable in the mid-SOC stage, improving estimation accuracy; low / high SOC stages focus on capturing anomalous features), avoiding the dilution of effective information across the entire range and reducing model training complexity. However, existing models (such as simple neural networks and linear regression) lack the ability to efficiently handle heterogeneous data across multiple ranges, easily leading to overfitting or a surge in computational complexity. Therefore, simplification to single-variable SOC modeling is chosen. However, the heterogeneity of feature distributions across different SOC ranges and the fundamental differences in health-sensitive features across different SOC ranges mean that stage-based segmentation allows the model to specifically capture the feature interaction effects of each range (e.g., the coupling effect of Max and AF in the high SOC stage), and also provides data support for dynamically adjusting information flow (e.g., focusing on stable feature paths in the mid-SOC stage), indirectly improving the accuracy and interpretability of the evaluation.

[0053] Three datasets were used to explore the optimal SOC range for ultrasound-based SOH estimation. The datasets were split into training and testing sets in an 8:2 ratio, and five cross-validations were performed to ensure model robustness. The GAN-DE and GAN-DEI models had maximum depths of 4 and 3 layers, respectively, with 200 IDT trees per layer, used to estimate SOH within different SOC ranges. A QHAdam optimizer with a learning rate of 0.01 and an SSE loss function were used. After training, the optimal SOC range was selected for interpretability analysis, and visualizations were generated to explore the relationships between ultrasound features, interaction terms, and SOH, as detailed in Example 2.

[0054] Specifically, the large-scale battery health assessment model integrates a generalized additive model and the expressive power of the Neural Decision Ensemble (NODE). The model structure consists of L layers (3 layers by default), each containing N differentiable forgetting decision trees (ODTs, 200 by default), with a tree depth of d=4.

[0055] The generalized additive model (GAM) is expressed as: (1) where μ is the learnable intercept term. Let represent the smoothing function of the i-th feature. i and N are both positive integers. This additive structure allows the contribution of each feature to the output to be visualized individually. Furthermore, through the learnable intercept term and the structure of accumulating the smoothing functions of each feature, the independent contribution path of each feature to the SOH output is clarified, mathematically ensuring the quantifiability and visualization of single feature contributions. This lays the foundation for overcoming the predicament of the lack of physical meaning in existing black-box models, enabling engineers to intuitively understand the impact of individual features (such as ToF) on battery health assessment. Based on the existing NODE architecture, improvements are made, consisting of L-layer differentiable forgetting decision trees (ODTs), each layer containing N trees of depth d. The ODT uses a feature selection function... Implement data partitioning: (2) where Choose a matrix for learnable features. Sparse feature selection is achieved, where j is a positive integer and D represents the total dimension of the input features. GAN-DE is a GAM model improved based on the NODE architecture and is part of a large battery health assessment model. Formula (2) defines the feature selection function of the differentiable forgetting decision tree (ODT). Sparse feature selection is achieved through a learnable feature selection matrix, which effectively filters out core features (ToF, Max, etc.) that are strongly correlated with SOH, eliminates redundant information, reduces the computational complexity of the model, and ensures the targeting of feature selection. It avoids irrelevant features from interfering with the assessment accuracy and provides an efficient and accurate feature input basis for subsequent model optimization. To solve the problem of uncontrollable feature selection in the NODE model, GAN-DE introduces the temperature annealing parameter T: (3) In the early stage of training, a high T value is set to allow for wide feature exploration. The T value is constant and defined according to the number of layers L. Then it is linearly reduced to 0 to ensure that each tree ultimately selects only one feature. In response to the problem of unsolvable feature entanglement in existing black box models, a feature path purity mechanism is designed in reverse. The gating function of formula (3) only allows the connection of the same feature path, forcibly separating the contributions of different features, so that the originally entangled latent space features are transformed into quantifiable independent contributions, breaking through the interpretability dilemma. Specifically, formula (3) introduces the temperature annealing parameter T. By using a strategy of high T value for wide-range feature exploration in the early stage of training and linearly reducing to 0 in the later stage to lock a single feature, the problem of uncontrollable feature selection in the NODE model is solved. This ensures that each decision tree focuses on only one core feature, avoids feature entanglement in the latent space, maintains the expressive power of the model, and strengthens the purity of the feature path. In this embodiment, the T value is constant and a learnable feature selection matrix S∈R is introduced according to NODE. (F×K) (F is the number of features, K is the number of tree nodes), achieving sparse feature selection. Simultaneously, a temperature annealing parameter T is used, initially T=10 during training, linearly decreasing to T=0 to ensure each tree ultimately selects only one feature, avoiding feature entanglement. A gating connection mechanism is introduced, and a gating function is designed. A connection is established only when adjacent tree nodes select the same feature. Otherwise disconnect ( This hard selection strategy preserves the purity of the feature paths. A low-rank attention mechanism A = A1A2 is introduced, where A1 ∈ R. (I×E) A2∈R (E×(I-1)N) This reduces the computational complexity from O(N²) to O(NE). Attention weight α ljj 'Dynamically adjusting the information flow makes the model focus more on feature paths that significantly affect SOH.' The attention weights αljj dynamically adjust the information flow, including: (4) Where l', j', j, and l are all positive integers; This is used to characterize the fact that connections are established only when adjacent tree nodes select the same feature. =1, otherwise disconnect. G j Represented as the output result of the j-th tree node; g l’j’ (x) represents the response function of the j'-th tree node in the previous layer to the input x; x represents the input feature vector of the model; a lj’j It is represented as the attention coefficient between adjacent layer nodes.

[0056] Formula (4) constructs the matrix factorization form of the low-rank attention mechanism (A=A1A2), reducing the computational complexity from O(N²) to O(NE), significantly improving the model's computational efficiency and meeting the microsecond-level real-time evaluation requirements. At the same time, by dynamically adjusting the information flow through attention weights, the model automatically focuses on feature paths that have a significant impact on SOH (such as the decision path corresponding to ToF), further improving the evaluation accuracy and achieving a balance between efficiency and accuracy.

[0057] Unlike NODE, which simply averages the output, GAN-DE uses a weighted linear combination to train the weights w. lj This enables the model to emphasize important feature paths: (5) Where l', j', j, l, and L are all positive integers. g ij (x l The representation is the pair of input feature x between the i-th tree node in the l-th layer. l The nonlinear response output, w lj The attention weights are represented by the learnable weighted coefficients corresponding to the j-th tree node in the l-th layer. The contribution of each feature path to the SOH output is dynamically adjusted using the attention weight calculation formula. Based on GAN-DE, and based on the principle of dual feature selection matrices, two independent feature selection matrices C¹ and C² are introduced to replace the single matrix S. When the tree depth d > 2, according to rule C... d =C （d-1） The two matrices, mod2+1, are used alternately to ensure that each decision tree considers the synergistic effect of at least two features and captures the feature interaction effect. Formula (5) fuses the outputs of each layer of decision trees through a weighted linear combination. The training weights emphasize the contribution of important feature paths, replacing the simple averaging strategy of the traditional NODE model. This allows the model to specifically amplify the influence of core features (such as the 70.3% contribution of ToF) and key interaction terms, reduce the interference of secondary paths, significantly improve the accuracy of SOH prediction, and ensure that the model can still achieve the high accuracy requirement of R²>0.99 under complex working conditions. Whether it is the GAN model or the GAN-DE model, its input features are one of the input data of the large battery health assessment model.

[0058] It can be seen that formula (1) lays the foundation for the additive interpretability of the model, formula (2) achieves accurate feature selection, formula (3) solves the problem of feature entanglement, formula (4) optimizes computational efficiency and feature focusing, and formula (5) strengthens the core contribution weight. This multi-formula combined effect not only resolves the contradiction between high accuracy and interpretability in existing models, but also enables the model to capture the contribution of a single feature and the interaction effect of multiple features through mathematical collaborative design, reducing prediction errors in complex aging scenarios, while meeting the microsecond-level real-time requirements.

[0059] Furthermore, the gating connection rules are adjusted so that a connection is established only when the C¹ and C² matrices of adjacent tree nodes are exactly the same (ω). ljj' =1), further enhancing the targeting of feature interaction paths. This quantifies interaction terms through the feature interaction matrix M∈R. F×F The quantification of the collaborative contribution of two features, M(i,j) represents the interaction contribution of the i-th feature and the j-th feature, to achieve accurate modeling of the coupling effect.

[0060] Through the above steps, the GAN model is improved to form the GAN-DE model, and then multiple mechanisms are introduced to form the final GAN-DEI model, thus constructing the final large-scale battery health assessment model.

[0061] In the output layer of the large-scale battery health assessment model, a weighted linear combination method is used to fuse the outputs of decision trees from each layer, and the weights ω are trained. ljj' Emphasizing the contribution of important feature paths, the model ultimately outputs a SOH estimate.

[0062] First, by setting up a GAN-DE / GAN-DEI model architecture, the interpretability of generalized additive models and the expressive power of neural decision integration are combined. Through triple improvements in temperature annealing parameters, gating mechanisms, and low-rank attention mechanisms, the feature entanglement problem of existing black-box models is solved, and the computational efficiency and expressive power of the model are improved. Second, the dual feature selection matrix design of the GAN-DEI model achieves accurate modeling of feature interaction effects in battery health assessment for the first time, effectively solving the deficiency of existing technologies in capturing electrochemical-mechanical coupling effects, and reducing the prediction error of the model in complex aging scenarios. Finally, the large-scale battery health assessment model adapts to the dynamic evolution of feature importance during battery aging through dynamic feature weighting and SOC segmentation modeling strategies, such as the sudden change in the importance of the Max feature in fast charging scenarios, ensuring the accuracy and stability of the model in the entire life cycle assessment.

[0063] The third module inputs both the battery electrical parameter data and the ultrasonic feature data into the large battery health assessment model and outputs the battery health assessment results.

[0064] Specifically, by decomposing the model output Σfi(xi), the contribution of each feature to SOH is obtained (e.g., the contribution of the ToF feature reaches 70.3%). Based on the interaction term matrix M of GAN-DEI, a dual-feature interaction heatmap is generated to quantitatively display the collaborative contribution of features (e.g., the collaborative contribution of Max and AF reaches 7.89%). Local weighted regression (LOWESS) is used to fit the nonlinear relationship curves between each feature and SOH, dynamically demonstrating the influence of features at different aging stages on SOH. The complete decision path of the model from input features to output SOH is recorded, including the feature selection results of each layer of decision trees and the distribution of attention weights, realizing the transparency of the decision-making process.

[0065] Battery health assessment results include State of Health (SOH) estimates, health trend predictions, and quantitative interpretation reports. These results provide a complete "detection-analysis-decision" solution for battery management, offering three major advantages over traditional methods: high speed (microseconds), high accuracy (R²>0.99), and strong interpretability.

[0066] Furthermore, the battery health assessment results include: First, a real-time monitoring panel is provided, displaying the raw ultrasonic signal waveform, preprocessed characteristic curves, and real-time SOH estimates, updated at a frequency of 10Hz. Second, a feature analysis view is provided, offering time-series curves and distribution histograms for each feature (ToF, Max, Min, PP, AF), supporting ranked display of feature contribution. Third, an interaction effect view is provided, showcasing the interaction effects of dual features through heatmaps and 3D surface plots, supporting user-defined feature combinations for analysis. Fourth, a trend prediction chart is provided, predicting future SOH degradation trends based on historical data, providing lifespan prediction curves under different operating conditions (such as regular charge / discharge, fast charging, and low-temperature environments).

[0067] The quantitative interpretation report includes basic battery information, test results, feature analysis, interaction effect analysis, and decision recommendations. Basic information includes battery model, test time, and test conditions (charge / discharge rate, temperature, and SOC range). Test results include the current SOH value, measurement accuracy (MAE, R²), and the deviation between the battery's current capacity and nominal capacity. Feature analysis includes the ranking of each feature's contribution, trend analysis of key features (such as ToF), and abnormal feature identification results. Interaction effect analysis includes quantifying the impact of major feature interactions on SOH and providing risk warnings for coupling effects (such as early warning of the synergistic effect of electrode expansion and electrolyte drying). Decision recommendations include suggestions for optimizing charge / discharge strategies based on SOH decay trends (such as avoiding prolonged fast charging and controlling the SOC operating range), maintenance cycle recommendations, and replacement warning thresholds.

[0068] In summary, firstly, by dividing the dataset into three independent datasets based on SOC values ​​(5%-35%, 35%-65%, and 65%-95%), the feature distribution in each scenario is made equipotential (eliminating feature shifts caused by SOC differences). This allows the model to maintain high accuracy in all scenarios, resolving the contradiction between generality and specificity.

[0069] Second, a dynamic feature weighting algorithm is adopted to pre-adjust feature weights based on battery type and charge / discharge rate, addressing differences in feature importance under different scenarios (such as increasing the Max feature weight in fast charging scenarios), avoiding fluctuations in the accuracy of static models, and ensuring the stability of evaluation during dynamic aging.

[0070] Third, by quantifying the coupling effect through the feature interaction matrix M, such as the synergistic contribution of Max and AF, the weight of the interaction terms is dynamically adjusted. Just like the hydraulic system's adaptive pressure change, this adapts to the mode switching from single feature dominance to multi-feature coupling dominance during the battery aging process, thereby improving the adaptability to complex working conditions.

[0071] The beneficial effects of this embodiment are as follows: First, the constructed GAN-DE / GAN-DEI dual-model system forms a large-scale battery health assessment model, achieving a dual breakthrough in accuracy and interpretability. The GAN-DE model integrates the interpretability of generalized additive models with the expressive power of neural decision integration. Through temperature annealing parameters, gating mechanisms, and low-rank attention mechanisms, it avoids the feature entanglement problem of traditional deep learning models and reduces the computational complexity from O(N²) to O(NE), ensuring that each decision tree focuses on only a single core feature, allowing the contribution of each feature to SOH to be quantified individually (e.g., the ToF contribution reaches 70.3%). Meanwhile, the GAN-DEI model, through a dual feature selection matrix and strict gating conditions, achieves accurate modeling of feature interaction effects for the first time, effectively addressing the complex scenario of the synergistic effect of electrode expansion and electrolyte drying, reducing prediction errors by more than 40%. With the construction of a three-segment dataset based on SOC value, an 8:2 data split ratio, and five-fold cross-validation, the model achieves ultra-high accuracy of MAE < 0.0034 and R² > 0.991 in the mid-SOC stage (35%-65%), far exceeding the 1% threshold required for industrial applications. This completely solves the defects of traditional models, such as rigid static feature selection and inability to adapt to the dynamic aging process of batteries.

[0072] Secondly, in terms of practicality assessment, Example 1 forms a complete "detection-analysis-decision" solution, significantly enhancing the industrial application value of the technology. The data processing stage employs multiple steps, including fifth-order Butterworth filtering, Hilbert transform, and Z-score standardization, to improve the signal-to-noise ratio to over 40dB, eliminating redundant features and outliers to ensure the purity and standardization of the input model's feature set, laying a solid foundation for high-precision evaluation. The evaluation output not only includes SOH estimates but also achieves full-process interpretability through feature contribution quantification, interaction effect heatmaps, and decision path tracing, allowing engineers to clearly understand the model's decision-making basis and quickly pinpoint the root cause of battery health issues. Simultaneously, the real-time monitoring panel, feature analysis view, trend prediction charts, and other visual interfaces, along with quantitative explanation reports containing basic information, detection results, and decision recommendations, transform abstract technical outputs into concrete engineering guidance. This can be directly applied to practical work such as charging / discharging strategy optimization, maintenance cycle formulation, and replacement warnings, significantly reducing operation and maintenance costs and providing scientific support for battery lifecycle management.

[0073] Finally, it has strong compatibility and scalability. The modular hardware design supports quick probe replacement to adapt to different types and sizes of batteries. The dynamic feature weighting algorithm can automatically adjust the weights according to the battery type (ternary lithium, lithium iron phosphate, etc.) and charge / discharge rate. The dataset construction and model training process is clear and standardized, which facilitates engineering reproduction and promotion.

[0074] The difference between Example 2 and Example 1 is that the battery health assessment system based on the AI ​​large model includes model training and optimization.

[0075] By employing the QHAdam optimizer with a learning rate of 0.01, the parameter update step size is adaptively adjusted to improve training stability. The sum of squared errors (SSE) is used as the loss function, i.e., Loss = Σ(SOH) true -SOH pred )², minimizing the deviation between the predicted and actual values. Where SOHtrue is used to characterize the actual value of SOH, and SOH... pred Predicted values ​​used to characterize SOH.

[0076] During training, GAN-DE and GAN-DEI models were trained on datasets from three SOC stages. Optimal hyperparameters (number of tree layers, number of trees, attention dimension, etc.) were selected using five-fold cross-validation. The MAE and R² of different models at each SOC stage were compared, and the optimal evaluation interval was determined to be the SOC stage (35%-65%), within which the model had MAE < 0.0034 and R² > 0.991.

[0077] The beneficial effects of this embodiment are achieved by adding a model training and optimization stage, employing a precise combination of the QHAdam optimizer and the SSE loss function, and adaptively adjusting the parameter update step size with a learning rate of 0.01, which effectively improves the stability and convergence speed of model training. At the same time, through the squared error and minimization strategy, the deviation between the SOH predicted value and the true value is significantly reduced. Combined with five-fold cross-validation, hyperparameters such as the number of tree layers and attention dimension are optimized, and the MAE and R² performance of GAN-DE and GAN-DEI models are compared in the low, medium and high SOC stages. Finally, 35%-65% is identified as the optimal evaluation interval, so that the model MAE < 0.0034 and R² > 0.991 in this interval. This not only far exceeds the 1% threshold required for industrial applications, but also specifically solves the problem of large accuracy fluctuations in traditional models in the whole life cycle evaluation. It allows the model to maintain stable high-precision output under complex working conditions, providing more reliable and industrially valuable technical support for battery health evaluation, and further strengthening the core advantages of the overall solution in terms of accuracy and robustness.

[0078] In this experimental example, the detection time at different SOC stages was tested, and the results are shown in Table 1 below: Table 1 Real-time Verification Results

[0079] It can be seen that the average total time is ≤26.1μs, which is much lower than the several hours required by existing electrochemical detection methods, thus meeting the requirements for real-time online monitoring.

[0080] In this test, the same battery (SOH=80%) was continuously tested 100 times under different environments (including normal temperature environment and extreme environment), and the fluctuation range of the estimated SOH value was recorded. The results are shown in Table 2 below: Table 2 Stability Verification Results

[0081] It can be seen that the fluctuation range is ≤ ±0.2% under extreme or normal temperature conditions, indicating that the battery health evaluation system based on the AI ​​large model of the aforementioned embodiment has good stability over a wide temperature range.

[0082] Numerous specific details are set forth in this specification. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, systems, and techniques have not been shown in detail so as not to obscure the understanding of this specification. In the description of this specification, references to the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., mean that a specific feature, method, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this specification.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A battery health assessment system based on an AI large-scale model, characterized in that, include: The first module acquires ultrasonic characteristic data of the battery under different energy flow states through non-invasive ultrasonic detection. The second module involves building a large-scale battery health assessment model. The third module inputs both the battery electrical parameter data and the ultrasonic feature data into the large battery health assessment model and outputs the battery health assessment results.

2. The battery health assessment system based on an AI large model according to claim 1, characterized in that, The different energy flow states include battery SOC values ​​of 5%-35%, 35%-65%, and 65%-95%.

3. The battery health assessment system based on an AI large model according to claim 2, characterized in that, The battery's 5%-35% SOC, 35%-65% SOC, and 65%-95% SOC values ​​were set as independent stages. In each stage, ultrasonic feature data was extracted every few seconds. The ultrasonic feature dataset of each stage was used as an independent dataset to train a large battery health assessment model.

4. The battery health assessment system based on an AI large model according to claim 1, characterized in that, The battery electrical parameter data includes battery voltage and current; the ultrasonic feature data includes time domain features and frequency domain features; the time domain feature data includes flight time, maximum value, minimum value, and peak-to-peak value; the frequency domain feature includes the maximum amplitude of the feature frequency band.

5. The battery health assessment system based on an AI large model according to claim 1, characterized in that, The large-scale battery health assessment model includes a generalized additive model, the expression of which includes: Where μ is the learnable intercept term, Let represent the smoothing function for the i-th feature, where i and N are positive integers; E(y) represents the expected value of the model's output random variable y; the architecture of the generalized additive model includes a multi-layer differentiable forgetting decision tree, which achieves data segmentation through feature selection functions; the feature selection functions include: in, Characterized as a feature selection function, R (d×D) Choose a matrix for learnable features. To achieve sparse feature selection, j and D are both positive integers, where D represents the total dimension of the input features, and x i It is represented by the value of the i-th input feature variable.

6. The battery health assessment system based on an AI large model according to claim 5, characterized in that, By introducing annealing parameters, the characteristic function becomes controllable, including: Where the annealing parameter is T, which is a constant, D represents the total dimension of the input features and is a positive integer, and C... ij The weight of the i-th decision node for the j-th input feature is represented as the selection weight of the j-th input feature. To ensure that each layer of the differentiable forgetting decision tree ultimately selects only one feature, an annealing parameter is introduced, along with a learnable feature selection matrix, including: setting the learnable feature selection matrix S∈ R(F×K) Where F represents the number of features and K represents the number of tree nodes, the learnable feature selection matrix is ​​used to achieve sparse feature selection.

7. The battery health assessment system based on an AI large model according to claim 6, characterized in that, Configure gating connection mechanism and low-rank attention mechanism; gating connection mechanism includes: setting gating function: A connection is established only when adjacent tree nodes select the same feature; otherwise, it is broken. Where C... j The feature index of the j-th tree node is represented as the final selected feature index; the low-rank attention mechanism includes: A = A1A2, where A1∈R (I×E) A2∈R (E×(I-1)N) .

8. The battery health assessment system based on an AI large model according to claim 7, characterized in that, By setting up a dynamically adjusted information flow, the large-scale battery health assessment model focuses on the feature paths that have the greatest impact on state of health (SOH). The dynamically adjusted information flow includes: Where l', j', j, and l are all positive integers; This is used to characterize the fact that connections are established only when adjacent tree nodes select the same feature. =1, otherwise disconnect. G j Represented as the output result of the j-th tree node; g l’j’ (x) represents the response function of the j'-th tree node in the previous layer to the input x; x represents the input feature vector of the model; a lj’j This is represented by the attention coefficients between adjacent layer nodes; based on the calculation formula for attention weights, the weights w are trained using a weighted linear combination. lj This enables the large-scale battery health assessment model to emphasize the most salient feature paths; the formula for calculating attention weights includes: Where L is a positive integer, g ij (x l The representation is the pair of input feature x between the i-th tree node in the l-th layer. l The nonlinear response output, w lj The learnable weighting coefficients corresponding to the j-th tree node in the l-th layer are represented by the attention weights; the contribution of each feature path to the SOH output is dynamically adjusted through the calculation formula of the attention weights.

9. The battery health assessment system based on an AI large model according to claim 7, characterized in that, By introducing two independent feature selection matrices C¹ and C², a single learnable feature selection matrix S is replaced; when the tree depth d is greater than 2, according to C... d The rule of =C(d-1)mod2+1 alternates between two matrices C¹ and C²; when the tree depth is no greater than 2, a single learnable feature selection matrix S is used.

10. The battery health assessment system based on an AI large model according to claim 1, characterized in that, The battery health assessment results include SOH (State of Health) estimates, battery health trend predictions, and a battery quantitative interpretation report.

Images