Battery state of health estimation method and system based on knowledge enhanced large language model

Abstract

The application provides a battery state of health evaluation method and system based on knowledge enhanced large language model, aiming to solve the problems of cross-modal gap between battery time sequence signal and large language model, physical inconsistency illusion and lack of uncertainty safety redundancy, comprising: screening constant current charging segment, constructing time sequence type and statistical type health factor and standardizing; generating field knowledge enhancement prompts by fusing battery chemical system, use condition and safety boundary rules; converting continuous charging signal into continuous semantic vector through time sequence semantic mapper; inputting prompt vector and continuous semantic vector into frozen or partially frozen large language model to train the sum of physical consistency constraint loss and regression loss as objective function; calculating uncertainty index by multiple random forward propagation during reasoning, adaptively weighting and fusing large language model output and physical model output according to uncertainty safety threshold, realizing high-precision and strong-generalization battery state of health evaluation.

Description

Technical Field

[0001] This invention relates to the fields of new energy vehicle battery management and artificial intelligence technology, specifically to a battery health status assessment method and system based on a knowledge-enhanced large language model. Background Technology

[0002] With the rapid development of the electric vehicle industry, the state of health (SOH) of power batteries, as the core energy carrier, directly determines the vehicle's driving range, operational safety, and residual value. Accurately assessing battery SOH is crucial for early warning, lifespan prediction, and secondary utilization of battery systems. However, power batteries exhibit highly nonlinear, time-varying, and multi-factor coupled degradation characteristics under actual operating conditions, making high-precision, highly generalized SOH estimation in real dynamic environments extremely challenging.

[0003] Currently, mainstream SOH estimation methods are mainly divided into two categories: model-based methods and data-driven methods.

[0004] Model-based methods (such as equivalent circuit models and electrochemical models) evaluate batteries by establishing mathematical relationships between internal degradation mechanisms and external measurable parameters. These methods heavily rely on precise prior knowledge of the battery's electrochemical system (such as open-circuit voltage curves and internal resistance characteristics) and complex parameter identification processes. In practical applications, parameter drift throughout the battery's lifespan, manufacturing inconsistencies, and dynamic operating conditions can significantly reduce model accuracy, making it difficult to adapt to complex scenarios across different time periods and environments.

[0005] Data-driven methods (such as support vector machines, neural networks, and Gaussian process regression) do not rely on precise physical models; they directly learn the mapping relationship between input features (such as voltage, current, and temperature) and state of equilibrium (SOH) from historical operating data. However, these methods are essentially statistical fitting paradigms and have the following inherent drawbacks: 1. Single modality: Most methods rely on single modal information (such as using only voltage curves or capacity increment curves), failing to fully utilize the complementary information in multi-source heterogeneous data (such as charging behavior, driving conditions, temperature series, etc.); 2. Insufficient generalization ability: The model tends to memorize the distribution characteristics of the training data. When faced with cross-battery types, abnormal charging conditions, or out-of-distribution samples, its estimation accuracy drops sharply. 3. Poor interpretability: It is difficult to explain the rationality of its prediction results from the perspective of physical mechanisms.

[0006] In recent years, large language models (LMs) have demonstrated their application potential in numerous scientific and engineering fields due to their powerful cross-domain knowledge transfer, complex semantic understanding, and reasoning capabilities. However, directly applying general-purpose large language models to battery SOH estimation faces three major technical obstacles: First, there is the modal gap problem. Large language models are pre-trained for discrete text data, while battery monitoring signals (such as current and voltage time series) are continuous signals constrained by physical laws. There is a fundamental difference in representation between the two, and there is a lack of effective semantic alignment mechanisms.

[0007] Secondly, there are issues of domain adaptation costs and knowledge forgetting. Full-fine-tuning of large language models requires enormous computational resources and labeled data in the battery domain, and it is easy for the model to forget its general reasoning abilities; while simple prompt engineering is difficult to deeply activate the battery degradation prior knowledge embedded in the model.

[0008] Third, there is the issue of physical inconsistency and hallucination. As a generative model, large language models are prone to producing "hallucination" outputs that deviate from the actual physical facts of battery degradation in regression tasks that strictly follow physical laws. For example, they may predict that the SOH (Solubility in Harmonic Drive) increases abnormally with usage time, which poses an unacceptable risk to the underlying safety of the vehicle.

[0009] In summary, how to fully utilize the powerful capabilities of large language models while effectively integrating physical knowledge from the battery field, solving cross-modal representation problems, and suppressing illusory outputs with physical inconsistencies are key technical challenges that urgently need to be addressed in this field. Summary of the Invention

[0010] In view of this, embodiments of the present invention provide a battery health status assessment method and system based on a knowledge-enhanced large language model, in order to solve the problems of modal gap between multi-source continuous time-series signals of batteries and discrete text representations of large language models in the prior art, high cost of full parameter fine-tuning which easily leads to forgetting of general reasoning ability, while simple hint engineering is difficult to fully inject prior knowledge of battery degradation, and safety risks of large language models easily generating illusory outputs that deviate from the actual degradation law of batteries in physical regression tasks and lacking uncertainty assessment and physical correction mechanisms.

[0011] A first aspect of this invention provides a battery health status assessment method based on a knowledge-enhanced large language model, comprising the following steps: Long-term monitoring data of the battery during actual operation is obtained from the vehicle terminal or cloud battery management platform. Based on current stability constraints and temperature and voltage range constraints, effective charging segments in the constant current charging stage are selected from the monitoring data. Based on the effective charging segments, time-series health factors and statistical health factors are constructed respectively. The time-series health factors are used to characterize the local evolution of current, voltage and temperature during the charging process. The time-series health factors and statistical health factors are standardized respectively. The time-series health factors adopt a channel-by-channel sequence normalization method, and the statistical health factors adopt a scalar normalization method based on the mean and standard deviation of the reference vehicle group. The standardized health factors are integrated with battery chemistry system rules, vehicle operating condition rules, and safety boundary rules to generate domain knowledge-enhanced prompts. The multi-dimensional temporal signals in the effective charging segment are input into a temporal semantic mapper composed of a trainable neural network, and the output is a continuous semantic vector consistent with the embedding dimension of the large language model. The domain knowledge-enhanced prompts and the continuous semantic vectors are input into the large language model. During the training phase, the pre-trained parameters of the large language model are frozen, while the trainable parameters of the adaptation layer, output mapping layer, or learnable prefix vectors attached to the large language model are retained. The trainable parameters are trained with the sum of physical consistency constraint loss and regression loss as the objective function. After processing the input, the large language model outputs a hidden layer embedding that represents the battery degradation state. The hidden layer embedding is mapped to a continuous battery health state estimate through a trainable numerical regression output layer. The physical consistency constraint loss includes a monotonicity constraint term that constrains the health state estimate from abnormally increasing with the aging process, and a consistency constraint term that constrains the deviation between the model prediction result and the reference health state obtained based on the ampere-hour integral. The effective charging segment is independently estimated based on at least one of the ampere-hour integration method, equivalent circuit model method, or electrochemical model method to obtain the physical model output. During the inference phase, multiple random forward propagations or multi-model ensemble predictions are performed on the same input sample to obtain a set of battery health state prediction results and calculate uncertainty indices. Based on the relationship between the uncertainty indices and the uncertainty safety threshold, the output weights of the large language model are calculated using a monotonically decreasing weight function with the uncertainty indices as input, and then adaptively weighted and fused with the physical model outputs to obtain the final battery health state estimate.

[0012] In one embodiment, the selection criteria for the effective charging segment are expressed as follows: in, This is the effective constant current charging sampling set; Let be the charging current at time t; This represents the change in current between adjacent sampling points; This is the preset constant current stability threshold; Let be the unit voltage at time t; , These are the lower and upper limits of the effective voltage window, determined based on the battery chemistry system. Different voltage windows can be set for lithium iron phosphate batteries and ternary lithium batteries, respectively. Let t be the battery temperature at time t; , These are the preset lower and upper limits of the temperature filtering range, respectively.

[0013] In one embodiment, the statistical health factors include cumulative mileage, rated capacity, charging duration, average charging current, charging temperature statistics, single-cell voltage difference statistics, and charging capacity statistics within an effective voltage window determined based on the battery chemistry system. The effective voltage window is the same as the voltage range constraint interval used to screen effective charging segments. The standardization process is performed according to the following formula: in, This is the kth original health factor; This is to provide a reference for the average value of this health factor in the vehicle population; The standard deviation of this health factor in the reference vehicle group; ε is a regularization constant to prevent the denominator from being zero; The normalized health factor is the kth health factor; the standardization process is used to eliminate dimensional differences caused by different vehicles, different capacity levels and different data collection conditions.

[0014] In one embodiment, the generation of the domain knowledge-enhanced prompt information is represented as follows: in, Enhance the hint vector for domain knowledge; This is the normalized set of health factors; Rules for battery chemistry systems; Rules for vehicle operating conditions; For safety boundary rules; Ψ(·) is the prompt generation function; The prompt generation function is implemented using text templates, continuous soft prompt vectors, key-value prefix vectors, or a combination of the above forms; the safety boundary rules include constraints that the health state is monotonically non-increasing with the aging process and that the estimated health state value is within a preset physical range; the battery chemistry system rules, vehicle operating condition rules, and safety boundary rules are encoded and injected into the domain knowledge-enhanced prompt information in the form of natural language text or structured fields to constrain the output of the large language model to conform to the physical boundary of battery degradation.

[0015] In one embodiment, the temporal semantic mapper employs any one of a one-dimensional convolutional network, a recurrent neural network, a Transformer encoder, an attention matching network, or a combination of the above networks to inject battery current changes, voltage plateau shifts, temperature fluctuations, and individual cell consistency degradation during the charging process into the large language model in the form of continuous semantic vectors, thereby achieving cross-modal alignment between physical temporal signals and the embedding space of the large language model.

[0016] In one embodiment, the physical consistency constraint loss is expressed as: Wherein, the t-th charging segment and the (t+1)-th charging segment are two constant current charging segments that are adjacent in time for the same vehicle; This represents the health state estimate of the model at the t-th charging segment; This represents the health state estimate of the model at the (t+1)th charging segment; This indicates a reference health status based on ampere-hour integration; Adjustable weighting coefficients for monotonicity constraints that prevent health status from abnormally increasing with the aging process; An adjustable weighting coefficient for the consistency constraint term between the model prediction results and the reference health status obtained based on the ampere-hour integral; The overall objective function of the training is the sum of the physical consistency constraint loss and the regression loss, that is: in, Total loss; For regression loss, mean absolute error, mean squared error, Huber loss, or a combination of the above losses are used.

[0017] In one embodiment, the final battery health state estimate of the adaptive weighted fusion is expressed as: The weighting function is expressed as follows: in, This is the final output estimate of the battery health status; The health status estimate output by the large language model; The health status estimate is output by the physical model; U is the uncertainty index; τ is the uncertainty safety threshold; γ is the weight switching coefficient. When the uncertainty index does not exceed the safety threshold, the system determines that the prediction result of the large language model is in the confidence range, and the weighted output is mainly based on the estimation of the large language model; when the uncertainty index exceeds the safety threshold, the system determines that the model is in a low confidence state, automatically reduces the output weight of the large language model and enhances the output weight of the physical model. The health status estimate output by the physical model is obtained by independently estimating the effective charging segment based on the ampere-hour integral method, the equivalent circuit model method, or the electrochemical model method. The physical model and the time-series semantic mapper share the same effective charging segment as input.

[0018] A second aspect of this invention provides a battery health status assessment system based on a knowledge-enhanced large language model, comprising: The data acquisition and segment selection module is used to acquire long-term monitoring data of the battery during actual operation from the vehicle terminal or cloud battery management platform, and select effective charging segments of the constant current charging stage from the monitoring data based on current stability constraints and temperature and voltage range constraints. The health factor construction and standardization module is used to construct a time-series health factor that characterizes the local evolution law of the charging process and a statistical health factor that characterizes the macroscopic degradation state of the battery based on the effective charging segment, and to standardize the time-series health factor and the statistical health factor. The domain knowledge enhancement prompt generation module is used to integrate standardized health factors with battery chemistry system rules, vehicle operating condition rules, and safety boundary rules to generate domain knowledge enhancement prompt information. The temporal semantic mapping module is used to convert the multi-dimensional temporal signals in the effective charging segment into continuous semantic vectors consistent with the embedding dimension of the large language model through a trainable neural network. The large language model inference module is used to receive the domain knowledge-enhanced prompt information and the continuous semantic vector, and train its trainable parameters with the sum of physical consistency constraint loss and regression loss as the objective function, and output the battery health status estimate. The uncertainty assessment and adaptive fusion module is used to perform multiple random forward propagation or multi-model ensemble predictions on the same input sample, calculate the uncertainty index, and adaptively weight and fuse the output of the large language model and the output of the physical model based on the uncertainty index to obtain the final battery health state estimate.

[0019] In one embodiment, the large language model in the large language model inference module adopts the method of freezing all parameters or freezing some parameters during the training phase to achieve efficient parameter adaptation; the physical consistency constraint loss includes a monotonicity constraint term that constrains the estimated health status value not to increase abnormally with the aging process, and a consistency constraint term that constrains the deviation between the model prediction result and the reference health status obtained based on the ampere-hour integral. The domain knowledge enhancement prompt generation module's prompt generation function is implemented using text templates, continuous soft prompt vectors, key-value prefix vectors, or a combination of the above forms; the temporal semantic mapper of the temporal semantic mapping module is any one of a one-dimensional convolutional network, a recurrent neural network, a Transformer encoder, an attention matching network, or a combination of the above networks.

[0020] In one embodiment, the long-term monitoring data acquired by the data acquisition and segment selection module includes individual cell voltage, total voltage, total current, battery temperature, state of charge, cumulative mileage, charging time, and vehicle identification information. The statistical health factors constructed by the health factor construction and standardization module include cumulative mileage, rated capacity, charging duration, average charging current, charging temperature statistics, individual cell voltage difference statistics, and charging capacity statistics within an effective voltage window determined based on the battery chemistry system. The effective voltage window is the same as the voltage range constraint interval used to select effective charging segments. The uncertainty assessment and adaptive fusion module uses any one of prediction variance, prediction entropy, quantile interval width, model integration divergence, or confidence score to measure the uncertainty index, and uses any one of hard switching or weighted fusion processing to achieve the fusion of the large language model output and the physical model output.

[0021] The battery health status assessment method and system based on a knowledge-enhanced large language model provided in this invention constructs a cross-modal alignment mechanism that includes constant current charging segment screening and temporal semantic mapping. This translates the battery's high-frequency continuous temporal signals into semantic representations that the large language model can understand, preserving electrochemical degradation characteristics and activating the complex reasoning capabilities of the large language model. This breaks down the modal barriers between physical signals and the large language model, solving the problem that real-vehicle multi-source monitoring data is difficult to directly utilize by general large models. By designing a domain knowledge-enhanced prompt generation strategy, statistical health factors, temporal health factors, battery chemical system rules, operating condition rules, and safety boundary rules are integrated. This achieves low-cost injection of battery degradation prior knowledge without relying on full parameter fine-tuning, avoiding the model forgetting the general reasoning capabilities of pre-training. This solves the problems of high domain adaptation costs and insufficient professional priors in large models for battery health status assessment tasks. By proposing a physical consistency constraint loss function, the invention introduces monotonic non-increasing constraints on health status and ampere-hour integral reference deviation constraints. This method forces the model output to conform to the actual physical laws of battery degradation, effectively suppressing the risk of large language models producing illusions in physical regression tasks. It also solves the problem that traditional generative models easily output unreliable results that deviate from physical facts, leading to underlying vehicle safety hazards. By constructing an uncertainty assessment and adaptive weighted fusion correction framework, multiple random forward propagation or multi-model ensemble predictions are performed on the same input. The uncertainty index is calculated and the outputs of the large language model and the physical model are adaptively fused according to the safety threshold. Under low confidence conditions, the weight of the large model is automatically reduced and the correction effect of the physical model is enhanced, improving the robustness and safety redundancy of battery health status assessment under open conditions, abnormal conditions, and out-of-distribution samples. At the same time, this method only needs to rely on the constant current charging segment collected by the vehicle and a lightweight temporal semantic mapper. The large language model is frozen during the inference stage to achieve efficient parameter adaptation, without the need for full model retraining or high-frequency trajectory reconstruction. This provides highly reliable and low-cost intelligent technical support for online battery health status assessment, retirement screening, and safety early warning in vehicle-cloud collaborative scenarios.

[0022] It is understandable that the beneficial effects of the second aspect mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1This is a flowchart of the battery health status assessment method based on a knowledge-enhanced large language model according to an embodiment of the present invention; Figure 2 This is an architecture diagram of the battery health status assessment system based on a knowledge-enhanced large language model according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the domain knowledge enhancement prompt generation and temporal semantic mapping mechanism in an embodiment of the present invention; Figure 4 This is a schematic diagram of the uncertainty-aware adaptive weighted fusion mechanism in an embodiment of the present invention. Detailed Implementation

[0025] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0026] This invention provides a battery health status assessment method and system based on a knowledge-enhanced large language model. This method is particularly suitable for vehicle-side data collection, cloud-based health status assessment, and vehicle-cloud collaborative reasoning scenarios for power batteries in new energy vehicles. It can also be extended to battery system management engineering scenarios with continuous charge-discharge cycles and health status monitoring requirements, such as energy storage power stations, electric ships, and drone batteries.

[0027] During actual battery operation, the vehicle battery management system collects multi-source monitoring data, including cell voltage, total voltage, total current, battery temperature, state of charge, cumulative mileage, charging time, and vehicle identification information. Among these, the charging phase, especially the constant current charging segment, contains rich electrochemical degradation characteristics, making it an ideal data source for assessing battery health. However, traditional methods often struggle to utilize large language models to process continuous temporal physical signals: full parameter fine-tuning is costly and easily leads to the model forgetting its general reasoning capabilities, while simple hint engineering cannot fully incorporate prior knowledge of battery degradation. Furthermore, as generative models, large language models are prone to producing illusory outputs that deviate from the actual degradation patterns of batteries in physical regression tasks, such as predicting an abnormal increase in health status with aging, posing a significant risk to the underlying vehicle safety. While some improved methods introduce temporal feature encoding or health factor screening, they focus more on feature engineering itself and lack systematic design around cross-modal alignment, domain knowledge fusion, and physical consistency constraints of large language models. They also struggle to provide uncertainty assessment and adaptive correction mechanisms under low-confidence conditions. The aforementioned shortcomings make it difficult to simultaneously ensure the accuracy, generalization ability, and safety of battery health status assessment under real-world operating conditions, and the reliability of assessments across battery types, abnormal charging conditions, and out-of-distribution samples is severely inadequate.

[0028] This invention constructs a cross-modal semantic alignment mechanism for constant current charging segments, utilizing a temporal semantic mapper to translate high-frequency continuous temporal signals from the battery into continuous semantic vectors understandable by a large language model. It designs a domain knowledge-enhanced prompt generation strategy, fusing statistical health factors, temporal health factors, battery chemical system rules, operating condition rules, and safety boundary rules to achieve low-cost injection of prior knowledge about battery degradation. It introduces a physical consistency constraint loss, forcing the model output to conform to the monotonically non-increasing physical law of health state and suppressing the risk of illusion. Simultaneously, it constructs an uncertainty assessment and adaptive weighted fusion correction framework, performing multiple random forward propagation or multi-model ensemble predictions on the same input, calculating uncertainty indices, and adaptively fusing the outputs of the large language model and the physical model based on a safety threshold. This automatically enhances the role of the physical model under low-confidence conditions, thus constructing a complete, lightweight, intelligent battery health state assessment method that can be embedded in a vehicle-cloud collaborative architecture. The technical solutions provided in the embodiments of this invention are described in detail below.

[0029] Example 1 This embodiment provides a battery health status assessment method based on a knowledge-enhanced large language model. This embodiment uses battery health status (SOH, unit: %) as the assessment target. Figure 1As shown, the method in this embodiment includes six stages: data acquisition and effective charging segment selection, health factor construction and standardization, domain knowledge enhancement prompt generation, temporal semantic cross-modal mapping, frozen large language model inference and physical consistency constraint training, uncertainty assessment and adaptive weighted fusion. Each stage will be described in detail below with reference to the accompanying drawings.

[0030] S101, Data Acquisition and Valid Charging Segment Selection: In applications, in-vehicle terminals or cloud-based battery management platforms continuously record long-term monitoring data generated by electric vehicles during actual driving and charging. This data covers multiple dimensions, including cell voltage, total voltage, total current, battery temperature, state of charge, cumulative mileage, charging time, and vehicle identification information, forming the original data foundation for subsequent health status assessments. Due to the highly random nature of driving behavior, data from the discharge phase fluctuates drastically due to road conditions, driving habits, and climate, making it difficult to stably reflect the internal electrochemical characteristics of the battery. In contrast, the current during the constant current charging phase remains relatively stable, and the temporal evolution of voltage and temperature is directly related to the internal battery reactions, providing a more reliable indication of battery degradation. The screening process uses current stability constraints to eliminate segments where the current change at adjacent sampling points exceeds a set threshold, temperature range constraints to exclude segments with distorted characteristics due to extreme thermal effects, and voltage range constraints to limit effective segments to a voltage window corresponding to the characteristic range of the battery chemical system. These three constraints jointly ensure the data quality and consistency required for subsequent health factor extraction.

[0031] S102. Construction and Standardization of Health Factors: In application, based on the selected effective charging segments, the system simultaneously constructs health factors from both local dynamic and macroscopic perspectives. Time-series health factors capture subtle manifestations of battery degradation in local charging behavior by recording the sequential changes of current, voltage, and temperature over time during charging. For example, as aging deepens, the voltage rise rate during constant current charging accelerates, and the charging plateau gradually shortens. These local characteristics carry real-time information about internal electrochemical changes in the battery. Statistical health factors quantify the degradation state from a macroscopic perspective. Cumulative mileage reflects the battery's usage history and the degree of cycle fatigue accumulation; rated capacity represents the highest level of currently available energy; charging duration and average charging current reflect changes in charging acceptance; charging temperature statistics reflect the effectiveness of thermal management; individual cell voltage difference statistics reveal the consistent degradation trend of each cell; and charging capacity statistics within the effective voltage window determined by the battery chemistry system (which is the same as the voltage range constraint range used to select effective charging segments) highly correspond to capacity measurements under standard test conditions. Since different vehicles have different rated capacity levels, the absolute value range of each factor varies under different data collection conditions. Directly using the original values ​​will make the model sensitive to changes in dimensions and difficult to generalize. Therefore, the mean and standard deviation of the reference vehicle group are used to normalize each health factor so that each factor falls into a similar value range, providing a unified feature representation basis for cross-vehicle generalization.

[0032] For time-series health factors, standardization employs a channel-by-channel sequence normalization approach. The normalization benchmark is derived from the global statistical characteristics of the reference vehicle group within the same channel. Specifically, the current, voltage, and temperature sequences within a charging segment are normalized independently for each channel: the global maximum and minimum values ​​(MinMax normalization) or the global mean and standard deviation (ZScore normalization) of the channel within the reference vehicle group are used as benchmark parameters to linearly transform the channel sequence values ​​of the current segment to a uniform numerical range, rather than using the mean and standard deviation of the current segment itself for normalization. This global benchmark normalization method ensures consistent comparability of time-series health factors across different charging segments and vehicles, eliminating dimensional biases caused by segment length, sampling time, and instantaneous fluctuations. For statistical health factors, the mean of the reference vehicle group is used as the benchmark. and standard deviation The scalar normalization method is based on the factorial scalar normalization formula.

[0033] Since the effective charging segment time lengths differ across charging sessions, when calculating reference group channel statistical parameters across segments, one approach is to unify the segment length by truncating each segment to its shortest segment length and then aligning them. The channel mean and channel standard deviation of the aligned segments are then calculated step-by-step within the reference group as a normalization benchmark. Alternatively, one approach is to linearly interpolate each segment along the time axis to a preset standard length before statistical analysis to achieve cross-segment comparability of the normalization benchmark.

[0034] S103, Domain Knowledge Enhancement Hint Generation: In application, the generation of domain knowledge-enhanced prompts organically integrates numerical health factors with three types of structured domain rules. Battery chemistry system rules describe the differences in degradation mechanisms among different cathode materials. Lithium iron phosphate batteries, due to their unique olivine structure, have a nearly flat charge / discharge voltage plateau. The capacity decay of ternary lithium batteries is more tightly coupled with temperature stress and depth of charge. This type of chemical system knowledge enables the large language model to adopt matching evaluation logic for different material systems. Vehicle usage condition rules describe usage behavior patterns such as cumulative mileage rate, charge / discharge frequency, and depth of charge as prior background information affecting the degradation rate, enabling the model to distinguish between normal aging and accelerated degradation caused by abnormal usage behavior. Safety boundary rules define a physically reasonable output range for the health state estimation results, preventing the large language model from producing output results exceeding the battery's physical limits when performing numerical regression. After encoding these three types of rules and normalized health factors into a prompt vector, the large language model possesses complete battery domain background knowledge upon receiving input, thus achieving effective activation of domain knowledge without relying on full parameter fine-tuning.

[0035] S104, Temporal Semantic Mapping: In applications, the multi-dimensional continuous time-series signals such as current, voltage, and temperature sequences collected during battery charging are essentially physical measurement signals constrained by electrochemical laws. These signals exhibit fundamental modal differences in signal morphology and semantic expression compared to the discrete text symbols used in the pre-training of the large language model. The temporal semantic mapper plays a crucial role in bridging this modal difference. By learning the mapping relationship from the physical signal space to the language model embedding space, it transforms the degradation information contained in the charging signals into continuous vectors aligned with the internal representation of the large language model. After mapping, phenomena such as increased internal resistance revealed by the rate of change of current, loss of active material revealed by the shift in the voltage plateau position, decreased thermal management capability revealed by abnormal temperature distribution, and cell consistency degradation revealed by increased individual cell voltage dispersion are all injected into the reasoning process in the form of continuous vectors perceptible to the large language model, fully activating the large language model's ability to understand and infer physical degradation laws.

[0036] S105, Large Language Model Training and Inference: During the training phase, the pre-trained backbone parameters of the large language model remain frozen, and gradients are not propagated to them. Only the adaptation layers attached outside the backbone (including but not limited to low-rank decomposition adaptation matrices, learnable prefix vectors, or output linear mapping layers) remain trainable and receive gradient updates from the objective function through backpropagation. During the inference phase, all parameters are fixed, and there is no parameter update process. This efficient parameter adaptation strategy enables the large language model to guide output through the adaptation layers while freezing the pre-trained parameters, rather than updating the backbone parameters. This is the specific meaning of "frozen or partially frozen training" in this invention.

[0037] In applications, the training strategy of freezing or partially freezing the parameters of a large language model means fixing most of the model's weights in the pre-training state during training, allowing only the output layer, adapter layer, or lightweight add-ons to update parameters. This strategy allows the model to retain the general reasoning and semantic understanding capabilities accumulated during the pre-training phase, even with limited battery data, while significantly reducing the computational and storage resources required for training. The physical consistency constraint loss consists of two parts: a monotonicity constraint term requires that the health state estimates between adjacent charging segments should not show an abnormally increasing trend with the aging process, thus preventing the model from generating false recovery predictions that violate the continuous degradation law of the battery; and a consistency constraint term requires that the deviation between the model output and the reference health state calculated based on the ampere-hour integration method be kept within a reasonable range, using the reliability of physical methods to constrain the output boundary of the large language model. The overall training objective function is the sum of the physical consistency constraint loss and the regression loss, both of which guide the model to conform to the basic physical laws of battery degradation while fitting the data.

[0038] Domain knowledge augmentation hint vector The combination of continuous semantic vectors output by the temporal semantic mapper can be achieved by concatenating them along the token sequence dimension, whereby the continuous semantic vectors are appended as an additional preceding token sequence before the text prompt token sequence to form the complete input sequence of the large language model; alternatively, the continuous semantic vectors can be used as conditional embeddings and injected into the intermediate layers of the large language model through a cross-attention mechanism. When using the former approach, the output dimension of the temporal semantic mapper must be consistent with the word embedding dimension of the large language model; when using the latter approach, the mapper's output dimension must be consistent with the cross-attention key-value dimension of the corresponding intermediate layer of the large language model.

[0039] S106. Uncertainty Assessment and Adaptive Fusion: In applications, when performing multiple random forward propagations on the same input sample during the inference phase, each propagation generates differentiated prediction results by randomly deactivating some neurons or applying slight perturbations to the input. The variance or entropy of the multiple prediction results reflects the model's confidence in the current input. When using random deactivation, the random deactivation layer in the large language model must be kept active during the inference phase. The deactivation rate during the inference phase can be consistent with that during the training phase, typically set in the range of 0.1 to 0.3. A higher deactivation rate results in greater dispersion of the prediction distribution and stronger sensitivity to the uncertainty index. If the large language model used does not contain a random deactivation layer or the deactivation layer is disabled by default during the inference phase, randomization can be achieved by applying Gaussian noise perturbation to the input features. The perturbation amplitude should preferably not exceed 5% of the signal standard deviation of the corresponding physical quantity in the effective charging segment to ensure that the perturbed sample remains within a physically reasonable range. The number of samples in multiple random forward propagations is usually no less than 20. In practical applications, this number can be adaptively determined based on the convergence of uncertainty estimation. When the relative change in the estimated prediction variance or prediction entropy is less than 1% when the number of samples is continuously increased, the number of samples can be considered to have met the stability requirements. For multi-model ensemble prediction methods, the number of ensemble models is usually no less than three. Each model should have different random initialization weights or use different training subsets to ensure the diversity of the ensemble. When different uncertainty measurement methods are used, the resulting uncertainty index U has different dimensions. The corresponding safety threshold τ needs to be calibrated on the validation set for each selected measurement method. The τ values ​​of different measurement methods cannot be used interchangeably. When the prediction results are highly concentrated, it indicates that the model has a stable response to the current input sample, and the reliability of the health status estimation is high. At this time, the system mainly gives the results based on the output of the large language model. When the prediction results have a large dispersion, it indicates that the current input conditions may deviate from the distribution of the training data and the model is in a low confidence state. The system automatically reduces the weight of the output of the large language model, while increasing the output weight based on physical methods such as ampere-hour integral or equivalent circuit model. The weighted fusion of the two is achieved through a smooth switching function with uncertainty index as input, ensuring a smooth transition between different confidence states and avoiding output jitter caused by sudden weight changes.

[0040] Specifically, in a complete battery health status assessment, the system first selects effective segments of the constant current charging stage from real-time or historically stored vehicle monitoring data based on three constraints: current stability, temperature range, and voltage range. Based on these segments, it extracts temporal health factors reflecting local charging dynamics and statistical health factors describing macroscopic degradation. After normalization based on the reference population mean and standard deviation, the normalized health factors, along with battery chemistry system rules, operating condition rules, and safety boundary rules, are input into a prompt generation function and encoded as a domain knowledge-enhanced prompt vector. Simultaneously, the multidimensional physical time-series signals in the charging segments are mapped to continuous semantic vectors consistent with the embedding space dimension of the large language model via a temporal semantic mapper. The dimension of the continuous semantic vectors is consistent with the hidden layer dimension of the large language model's embedding layer. The domain knowledge-enhanced prompt vector and the temporal semantic vector are input into a large language model with frozen or partially frozen parameters. The large language model then integrates pre-trained general knowledge and prior information from the battery domain to output a health status estimate. The training process optimizes the sum of physical consistency constraint loss and regression loss. During the inference phase, multiple random predictions are performed on the same input and uncertainty indices are calculated. The weighting ratio of the output of the large language model and the output of the physical model is dynamically adjusted by a weight switching function with an uncertainty safety threshold as the dividing point. Finally, a health status estimate that balances accuracy and safety is output.

[0041] The selection of effective charging segments can be achieved through a multi-faceted joint judgment method combining current stability constraints with temperature range constraints and voltage range constraints. Alternatively, voltage slope constraints, state-of-charge interval constraints, or charging power stability constraints can be used to replace current stability constraints to adapt to the differences in data acquisition formats of different battery management systems. The generation of domain knowledge-enhanced prompts can be achieved by using natural language text templates to concatenate health factors and domain rules into readable text input. Alternatively, a parameterized form of continuous soft prompt vectors can be used, with the soft prompt vectors and health factor values ​​jointly optimized to improve expressiveness while maintaining the compactness of the prompts.

[0042] This solution integrates six key steps—multi-source monitoring data screening, dual-modal health factor construction and standardization, proactive injection of domain rules, cross-modal semantic alignment, physical constraint-guided training, and uncertainty-aware adaptive fusion—to form a complete end-to-end evaluation chain. Compared to single data-driven or purely physical modeling solutions, this solution simultaneously possesses the complex reasoning and cross-domain knowledge transfer capabilities of large language models, as well as the interpretability and reliability of physical methods. It maintains stable evaluation performance across vehicle types, battery chemistry systems, and operating conditions. The physical consistency constraint loss effectively mitigates the risk of large language models producing predictions that violate physical laws when performing numerical regression on battery degradation. The uncertainty-aware fusion mechanism enables the system to autonomously identify low-confidence operating conditions and proactively seek physical model support, providing necessary safety redundancy for automotive-grade power battery health status assessment applications.

[0043] In an optional embodiment, the selection criteria for valid charging segments are expressed as follows: in, This is the effective constant current charging sampling set; Let be the charging current at time t; This represents the change in current between adjacent sampling points; This is the preset constant current stability threshold; Let be the unit voltage at time t; , These are the lower and upper limits of the effective voltage window, determined based on the battery chemistry system. Different voltage windows can be set for lithium iron phosphate batteries and ternary lithium batteries, respectively. Let t be the battery temperature at time t; , These are the preset lower and upper limits of the temperature filtering range, respectively.

[0044] In applications, the effective constant current charging sampling set The effective data is defined by sampling time t that meets four conditions: charging direction, current stability, voltage range, and temperature range. A charging current I(t) less than zero is the identification condition for charging behavior. In general engineering conventions, the current in the charging direction is defined as negative. This condition distinguishes between the charging and discharging stages, ensuring that only sampling points in the charging state are included in subsequent analysis. The absolute value of the current change ΔI(t) between adjacent sampling points is less than the constant current stability threshold. This is the core constraint for determining whether the current is in a stable constant current state. When the current fluctuation exceeds the threshold, the sampling point is considered to be in the transition period of charging mode switching or has been affected by external interference, and is not included in the valid set. Individual cell voltage. It needs to be within the effective voltage window determined according to the battery chemistry system. and Electrode materials in different chemical systems have different characteristic charging plateau ranges. Limiting data to these characteristic ranges effectively avoids cross-chemical system characteristic distortion caused by differences in the plateau ranges of different materials. For lithium iron phosphate batteries and ternary lithium batteries, separate voltage windows can be set. The battery temperature T(t) must be within the temperature screening range. and Between extreme high or low temperatures, battery characteristics can deviate significantly due to thermal effects. Excluding data from extreme temperature conditions can avoid the impact of temperature interference on the accuracy of health factor extraction.

[0045] Specifically, during the actual data screening process, the system scans the collected raw charging time series point by point, and verifies the above four constraints at each sampling time. Only when the current direction is charging and the current fluctuation meets the constant current stability requirement, the single-cell voltage... A time is included in the valid constant current charging sampling set only when it falls within the characteristic voltage window corresponding to the chemical system and the temperature T(t) is within the effective range. This process ultimately generates a valid data segment consisting of multiple eligible time points, which serves as the input source for subsequent health factor construction and analysis. For different battery chemistry systems, the system filters data based on pre-set voltage window parameters or those derived from historical data analysis, thereby ensuring the validity and comparability of data across different chemistry system scenarios.

[0046] The charging direction can be determined using the current sign, or alternatively using the monotonically increasing trend of the state of charge, to address situations where insufficient accuracy of some current sensors leads to misjudgments of the sign. The effective voltage window... and The determination of the charging platform can be achieved through manual experience setting based on the characteristics of electrode materials in different battery chemical systems, or by automatically identifying the charging platform range from historical charging data through capacity increment curve analysis or clustering algorithms. The latter can adapt to the situation where the charging platform drifts with the degree of degradation during long-term battery use.

[0047] Through joint screening under four constraints, an effective constant current charging sampling set was determined. The system simultaneously manages data quality across four levels: charging behavior recognition, constant current stability assurance, chemical system matching, and temperature rationality. This significantly reduces the impact of random operating condition interference and outliers under extreme conditions on subsequent health factor extraction. The design of linking the voltage window to the battery chemical system allows the screening scheme to adapt to batteries with different chemical systems, such as lithium iron phosphate and ternary lithium, without modifying the main algorithm structure, thereby improving the scheme's versatility and scalability.

[0048] In an optional embodiment, statistical health factors include cumulative mileage, rated capacity, charging duration, average charging current, charging temperature statistics, single-cell voltage difference statistics, and charging capacity statistics within an effective voltage window determined based on the battery chemistry system. The effective voltage window is the same as the voltage range constraint interval used to screen effective charging segments, such as... Figure 3 As shown, the upper part displays a set of health factors. With three types of domain rules ( , , The hint vector is generated by fusing the Ψ function. The complete path is shown in the lower half, which shows the path of the effective charging segment being transformed into an embedded aligned semantic vector by the temporal semantic mapper. The two paths enter the vertical main body box of the large language model respectively, intuitively presenting the cross-modal input structure.

[0049] In applications, statistical health factors quantify the macroscopic degradation state of batteries from seven dimensions. Cumulative mileage records the total distance the battery has traveled since it was put into use, serving as a comprehensive indicator of historical usage intensity and the degree of cyclic fatigue accumulation. Rated capacity represents the maximum amount of electricity the battery can currently release under standard test conditions, providing the most intuitive reference benchmark for assessing battery health; its degradation directly corresponds to the loss of active materials and the increase in internal resistance. Charging duration reflects the time required to complete charging within the same voltage window; as battery capacity declines, the time required to charge to the cutoff voltage during the constant current phase shortens accordingly. Average charging current reflects the acceptable level of charging power; batteries with accelerated aging often trigger current derating protection earlier in the constant current phase, thus showing a downward trend in average charging current. Charging temperature statistics include statistical characteristics such as the temperature mean and standard deviation during the charging process; increased internal resistance leads to enhanced thermal effects during charging, and rising charging temperature and decreased heat distribution uniformity are typical manifestations of battery aging. The individual cell voltage difference statistic reflects the statistical characteristics of the voltage dispersion between different cells within the same battery pack. As aging progresses, the differences in capacity and internal resistance among cells gradually increase, and the increase in the individual cell voltage difference statistic is a direct manifestation of the degradation of battery pack consistency. The charging capacity statistic within the effective voltage window is obtained by analyzing the internal […]. , The voltage range is obtained by integrating the current flowing within the voltage range over time. This voltage range is consistent with the voltage range constraint used to screen effective charging segments in step S101. The resulting charging capacity is highly correlated with the current capacity of the battery and is less affected by the operating conditions, thus serving as an effective supplementary characterization of the rated capacity.

[0050] In application, standardization is performed according to the following formula: in, This is the kth original health factor; This is to provide a reference for the average value of this health factor in the vehicle population; The standard deviation of this health factor in the reference vehicle group; ε is a regularization constant to prevent the denominator from being zero; This represents the k-th health factor after normalization; the standardization process is used to eliminate dimensional differences caused by different vehicles, capacity levels, and data collection conditions. The reference vehicle group typically selects statistical data from a large number of vehicles of the same platform or model during normal operation, with the mean value... This reflects the typical level of the group at its current stage of degradation, with a standard deviation of [value missing]. This reflects the range of natural differences within the group. After standardization, health factors of different absolute levels are compressed to a similar numerical range, eliminating the inconsistency of dimensions caused by differences in rated capacity levels of different vehicles and different data collection conditions. This allows the model to treat each health factor in a balanced way when training with multi-vehicle data fusion, while also improving the model's cross-vehicle generalization ability when facing new vehicles.

[0051] Specifically, when extracting and standardizing health factors for a specific effective charging segment of a vehicle, the system first calculates the raw values ​​(hk) of statistical health factors from the time-series data of that charging segment, including charging duration, average charging current, average and standard deviation of charging temperature, average and maximum voltage differences between individual cells, and the integral value of charging capacity within a fixed voltage window. Simultaneously, it reads the cumulative mileage and rated capacity from the vehicle's historical records, and then uses pre-stored statistical parameters from a group of reference vehicles. and Based on the mean, a normalization operation is performed on each health factor sequentially by subtracting the mean and dividing by the standard deviation to obtain the normalized health factors. The seven normalized statistical health factors and the time-series health factors are combined to form a complete health factor vector, which serves as the input for subsequent prompt generation and model inference.

[0052] The calculation of the charging capacity statistics within the effective voltage window determined by the battery chemistry system, where the effective voltage window is the same as the voltage range constraint interval used to select effective charging segments, can be achieved by using an ampere-hour metering method that integrates the current of the constant current charging segment within the preset voltage range over time. Alternatively, it can be achieved by using a voltage increment analysis method to statistically analyze the incremental capacity distribution characteristics within the characteristic voltage window. The latter method is more sensitive to capacity degradation and less dependent on the magnitude of the charging current. The standardization process references population statistical parameters. and The update can be achieved using fixed parameters through offline batch statistics, or it can be achieved using dynamic parameters through rolling window statistics. The dynamic update method can adapt to situations where the reference distribution drifts due to the expansion of the fleet size or the addition of new vehicle models.

[0053] The seven-dimensional statistical health factor comprehensively characterizes battery degradation from multiple independent yet complementary dimensions, including usage history, state of capacity, charging behavior, thermal management effectiveness, cell consistency, and capacity decay characteristics. This avoids the risk of reduced assessment accuracy due to the failure of a single indicator under specific operating conditions, making the health status characterization more comprehensive and robust. The standardized processing strategy based on the statistical characteristics of the reference population effectively eliminates the inconsistency in dimensions introduced by differences in vehicle capacity levels and data collection conditions. This enables the solution to maintain stable assessment performance in scenarios involving the fusion of data from multiple vehicle models and operating conditions, significantly broadening the solution's applicability.

[0054] In an optional embodiment, the generation of domain knowledge-enhanced prompts is represented as follows: in, Enhance the hint vector for domain knowledge; This is the normalized set of health factors; Rules for battery chemistry systems; Rules for vehicle operating conditions; Ψ(·) is the safety boundary rule; Ψ(·) is the prompt generation function.

[0055] In application, domain knowledge-enhanced hint vectors The four types of inputs are obtained by comprehensively encoding the four types of inputs through a prompt generation function. The four types of inputs are normalized sets of health factors. Battery chemical system rules Vehicle operating condition rules and security boundary rules Normalized set of health factors Numerical vectors containing standardized time-series and statistical health factors provide quantified information about the current state of the battery. Battery chemistry system rules. The degradation mechanisms of different electrode materials, such as lithium iron phosphate and ternary lithium, are described in the form of structured knowledge, including the charging voltage plateau of each chemical system and the coupling relationship between capacity decay and temperature and depth of charge. This type of knowledge may only exist in the pre-training corpus of the large language model in the form of general text. Explicit injection ensures that the large language model calls upon the correct electrochemical knowledge in battery evaluation tasks. Vehicle operating condition rules. The current vehicle's mileage accumulation rate, charging habits, and depth of charge / discharge are described as background information affecting the degradation rate, enabling the large language model to combine usage behavior to make a reasonable assessment of the current health status and degree of degradation. Safety boundary rules. Explicitly define a reasonable range and physical constraints for battery health status, injecting these as output constraints to reduce the probability of the large language model generating unreasonable numerical outputs. Hint generation function. It undertakes the task of encoding the above four types of information into a unified vector representation, and its specific implementation is not limited to a single structure.

[0056] In the application, the function to generate suggestions is provided. The implementation can be chosen based on the interpretability requirements and computational resource conditions of the application scenario. The text template method fills health factor values ​​and domain rules into human-readable text strings according to predefined natural language templates. The generated prompts are directly fed into the large language model as text input. Its advantage lies in the transparency of the generation process and the ease of human review of the results. The continuous soft cue vector method... Represented as learnable parameter vectors matching the dimensions of the embedding layer of the large language model, these parameter vectors are optimized through gradient descent during training. This achieves highly compact and expressive cue representations without increasing the number of main parameters in the large language model, making it particularly suitable for scenarios with limited computational resources and large training data. The key-value prefix vector approach introduces additional key-value pairs into the attention layer of the large language model, injecting domain knowledge into the model's attention computation process in a key-value structure, allowing domain knowledge to influence the model's feature extraction behavior at a deeper level. These various forms can also be combined; for example, a text template can be used to generate a textual description of chemical system rules, while continuous soft cue vectors represent the numerical features of health factors. The two types of cue are then concatenated and input into the large language model.

[0057] Specifically, in the health status assessment scenario for lithium iron phosphate batteries, the prompt generation function... First, read the numerical vector composed of the normalized seven-dimensional statistical health factors and the time-series health factors. Subsequently, the degradation mechanism description and charging platform characteristics of lithium iron phosphate batteries were retrieved from a pre-defined chemical system rule base as... The vehicle's usage patterns, reflecting its intensity of use and charging habits, are extracted from the vehicle's records as... And load the health state value range and physical meaning constraints set for lithium iron phosphate batteries as... Finally, the above four types of information are encoded into domain knowledge-enhanced hint vectors. Together with the continuous semantic vectors output by the temporal semantic mapper, they constitute the complete input of the large language model.

[0058] Among them, the rules of battery chemical system The injection can be achieved using text templates that describe the degradation mechanisms and charging characteristics of different electrode materials in natural language, or using soft hints that represent the prior knowledge of the chemical system using learnable continuous vector parameterization. The latter has stronger adaptive expressive power when there is sufficient training data. Among these, safety boundary rules... This can be achieved by embedding a numerical range description of the physical reasonable range of the health status in the prompt vector, or by post-processing and truncating the output during the decoding stage of the large language model. The two methods can be combined to form a dual security constraint.

[0059] Domain knowledge enhancement hint generation mechanism generates rules for battery chemistry systems. Use working condition rules and security boundary rules By deeply integrating with quantified health factors at the prompting level, this approach effectively injects prior knowledge of battery degradation without relying on full-parameter fine-tuning of a large language model. Compared to full-parameter fine-tuning, this mechanism significantly reduces the computational and storage resources required for training and effectively mitigates the risk of loss of general reasoning ability due to overfitting in small-sample scenarios. Support for multiple prompting implementations allows the solution to flexibly select the most suitable prompting strategy based on the interpretability requirements of the application scenario, computational resource conditions, and training data scale, thus improving the solution's engineering adaptability.

[0060] In an optional embodiment, the temporal semantic mapper employs any one of a one-dimensional convolutional network, a recurrent neural network, a Transformer encoder, an attention matching network, or a combination of the above networks to inject battery current changes, voltage plateau shifts, temperature fluctuations, and individual cell consistency degradation during the charging process into the large language model in the form of continuous semantic vectors, so as to achieve cross-modal alignment between physical temporal signals and the embedding space of the large language model.

[0061] In applications, the temporal semantic mapper receives multi-dimensional temporal signals from effective charging segments as input. These signals include a current sequence reflecting the evolution of charging current, a voltage sequence reflecting voltage response characteristics, and a temperature sequence reflecting thermal management status. These continuous physical signals record the dynamic response characteristics of the battery during charging at a fixed sampling rate. The core task of the temporal semantic mapper is to transform these physical measurement signals into continuous vector representations consistent with the embedding layer dimension of the large language model, enabling the large language model to perceive and infer the degradation information contained in these physical signals. One-dimensional convolutional networks extract local pattern features of signals along the time axis using sliding convolution kernels, making them suitable for capturing short-range temporal patterns and local morphological changes of voltage plateaus within stable current segments. Recurrent neural networks establish temporal dependencies through the mechanism of information transmission between time steps using hidden states, exhibiting strong modeling capabilities for features reflecting long-term charging behavior evolution trends. Transformer encoders calculate the correlation weights between any two positions in a temporal sequence using a self-attention mechanism, enabling simultaneous capture of short-range local features and long-range dependencies globally, making them suitable for processing high-dimensional multivariate temporal signals. Attention matching networks calculate attention weights between temporal signals and reference templates, comparing and matching the input signal with known degradation patterns. This approach is suitable for scenarios involving degradation stage assessment and abnormal condition identification. The aforementioned network structures can also be combined; for example, a one-dimensional convolutional network can be used to extract local features, which can then be passed to a Transformer encoder for global modeling to obtain richer temporal semantic representations.

[0062] Specifically, in the process of converting the charging physical signal into a continuous semantic vector that can be processed by the large language model, the system concatenates the current sequence, voltage sequence, and temperature sequence from the selected effective charging segments into a multivariate time-series matrix after time alignment. This matrix is ​​then input into a time-series semantic mapper. The mapper uses its internal network structure to extract features and transform the dimensions of the multivariate time-series signal, outputting a fixed-length continuous vector that matches the dimensions of the large language model's embedding layer. This vector is then used as a continuous prompt along with the domain knowledge-enhanced prompt vector. By inputting a common large language model, the large language model can simultaneously perceive numerical statistical features, domain knowledge background, and degradation information in the original physical signal, thereby achieving deep fusion of multimodal information.

[0063] The temporal semantic mapper can employ a one-dimensional convolutional network to extract local pattern features of the charging signal along the time axis, or it can use gated recurrent unit structures such as long short-term memory networks to capture long-range dependencies across multiple time steps. These two structures can also be used in series to balance local feature extraction and long-range dependency modeling. The input organization of multivariate time-series signals can be achieved by concatenating current, voltage, and temperature sequences into a multi-channel input along the channel dimension, or by fusing each signal through an independent mapper along the feature dimension in a multi-stream parallel approach. The latter approach preserves the independent features of each physical quantity and reduces mutual interference between different physical quantities.

[0064] The temporal semantic mapper establishes a cross-modal semantic alignment channel between continuous physical signals and discrete semantic representations by learning the mapping relationship from the battery physical signal space to the large language model embedding space. This breaks down the modal barrier that prevents large language models from directly processing physical measurement signals, allowing electrochemically significant physical phenomena such as battery current changes, voltage plateau shifts, temperature fluctuations, and cell consistency degradation to be injected into the inference process as continuous vectors perceptible to the large language model. This fully activates the large language model's ability to infer physical degradation patterns. The design of multiple optional network structures allows the temporal semantic mapper to flexibly select the most suitable feature extraction architecture based on actual conditions such as the sampling frequency, segment length, and signal quality of the charging data, ensuring the adaptability of the solution under different data conditions.

[0065] In an optional embodiment, the physical consistency constraint loss is expressed as: Among them, the t-th charging segment and the t+1-th charging segment are two constant current charging segments that are adjacent in time for the same vehicle; This represents the health state estimate of the model at the t-th charging segment; This represents the health state estimate of the model at the (t+1)th charging segment; This indicates a reference health status based on ampere-hour integration; Adjustable weighting coefficients for monotonicity constraints that prevent health status from abnormally increasing with the aging process; An adjustable weighting coefficient is used to constrain the consistency between the model's prediction results and the reference health status obtained based on the ampere-hour integral.

[0066] In applications, physical consistency constraint loss It consists of two parts: a monotonicity constraint term and a consistency constraint term. These two terms penalize two types of physical violations that may occur when the large language model performs health state regression. The monotonicity constraint term is calculated based on the health state estimate corresponding to the (t+1)th charging segment. Health state estimate corresponding to the t-th charging segment The difference is taken as a large value relative to zero to identify instances where the estimated value increases between adjacent charging segments and to penalize such instances. During normal aging, the battery's health state should monotonically decrease with increasing cycle count or at least remain unchanged. If the model predicts an increase in health state between adjacent charging segments, it indicates a false recovery prediction that violates the continuous degradation pattern of the battery. The monotonicity constraint term eliminates such illegal predictions by incorporating the positive portion of this difference into the loss function during training. To control the adjustable coefficient of this constraint term's weight, a larger... The value makes the monotonicity constraint more stringent. The consistency constraint term is calculated by using the health status estimate predicted by the model. Reference health status based on ampere-hour integration The absolute difference between the current and rated capacity is calculated and penalized. The ampere-hour integral method estimates the ratio of the current to the rated capacity in a physically reliable manner by integrating the current over time during the charging and discharging process. The resulting reference health state has high physical reliability in the short term and can be used to constrain the reasonable range of the output of large language models.

[0067] Adjacent charging segment pairs ( , The charging sessions should originate from the same vehicle and be arranged in chronological order of the charging session occurrence. For the t-th effective constant current charging segment of the vehicle, For the (t+1)th effective constant current charging segment of the same vehicle, there is a clear temporal order between the two segments to ensure the physical meaning of the monotonicity constraint. During the construction of the training batch, adjacent segments are sampled from the time-series data of the same vehicle, and charging segments from different vehicles do not constitute an adjacent relationship.

[0068] The overall objective function for training is the sum of the physical consistency constraint loss and the regression loss, i.e.: in, Total loss; For regression loss, mean absolute error, mean squared error, Huber loss, or a combination of the above losses are used.

[0069] In application, the overall objective function of the training process From regression loss And physical consistency constraint loss The regression loss is composed of two parts added together. Used to measure the gap between model predictions and the actual labeled health status, driving the model to improve prediction accuracy. Conventional regression losses can take the form of mean absolute error, mean squared error, or Huber loss. Mean absolute error is more robust to outliers, mean squared error imposes a stronger penalty on large error samples, and Huber loss combines the advantages of both, using mean squared error when the error is small and switching to mean absolute error when the error is large. The sum of the two types of losses allows the model to minimize prediction error while being constrained by physical laws, preventing large language models from producing numerically close but physically unreasonable predictions in order to fit the data.

[0070] Specifically, in each batch of model training, the system accumulates the positive values ​​of the difference between the predicted health status of adjacent charging segments of each sample pair within the batch, and multiplies them by a weighting coefficient. The monotonicity constraint term is obtained, and the predicted value and corresponding ampere-hour integral reference value of each sample within the batch are calculated. The absolute error between them, multiplied by the weighting coefficient The consistency constraint term is obtained, and the two constraint terms are added together to obtain the physical consistency constraint loss. In addition, the regression loss is calculated from the error between the predicted value and the labeled true value. This results in a total loss. The trainable parameters are then updated via backpropagation. and The value of can be determined in the early stages of training by performing ablation experiments on the validation set to balance the trade-off between prediction accuracy and physical consistency.

[0071] Among them, regression loss The calculation can use the mean absolute error (MAE) to measure the average deviation between the predicted and labeled values, or it can use Huber loss to reduce the interference of outliers on the training process while maintaining accurate fitting to small error samples. Among these, the reference health status... The calculation can be achieved by using the basic ampere-hour metering method of integrating the constant current charging section current over time and comparing it with the rated capacity. Alternatively, an improved ampere-hour integration method that incorporates coulomb efficiency correction and temperature compensation can be used. The latter can provide a more accurate reference health status estimate under operating conditions with large temperature variations.

[0072] Physical consistency constraint loss Through a dual mechanism of monotonicity constraints and consistency constraints with physical reference values, the fundamental physical laws of battery degradation are embedded into the optimization objective of the large language model during the training phase. This effectively reduces the risk of the large language model generating illusory predictions that violate physical facts when performing numerical regression tasks of health status. For automotive-grade power battery health status assessment applications, this constraint mechanism provides a crucial safety guarantee, ensuring that the prediction results of the large language model possess basic physical reliability. Adjustable weight coefficients and The introduction of this feature enables the scheme to flexibly adjust the strength of physical constraints based on actual data quality and the security requirements of the application scenario, achieving a reasonable balance between prediction accuracy and physical rationality.

[0073] In an optional embodiment, the final battery health state estimate of the adaptive weighted fusion is expressed as: The weighting function is expressed as: in, This is the final output estimate of the battery health status; The health status estimate output by the large language model; τ is the health status estimate output by the physical model; U is the uncertainty index; τ is the uncertainty safety threshold; and γ is the weight switching coefficient.

[0074] The above weighting function It has the following characteristics: when uncertainty indicators When not exceeding the safety threshold τ, When the value is large, the system determines that the prediction result of the large language model is within the confidence range, and the weighted output is mainly based on the estimation of the large language model; when the uncertainty index When the safety threshold τ is exceeded, The system rapidly reduces the confidence level, indicating the model is in a low-confidence state. It then automatically decreases the output weights of the large language model and increases the output weights of the physical model. This smooth weight switching design, based on the Sigmoid function, allows the system to smoothly transition between different confidence levels, avoiding sudden output changes caused by hard switching and ensuring the reliability of evaluation results under abnormal conditions.

[0075] like Figure 4 As shown, the three parallel inputs are clearly distinguished: the upper input calculates the uncertainty index U from the prediction set generated by multiple random forward propagations, and then drives it downwards via the weight function ω(U); the middle input is the direct output of the large language model; and the lower input is the output of the physical model. The three inputs converge in a fusion box, with a threshold judgment logic annotation for the uncertainty safety threshold τ appended at the bottom. In application, the final battery health state estimate... By using uncertainty indicators To adjust the adaptive weighting of the signal, the health status estimate output by the large language model is... Health status estimate output by the physical model Obtained through linear fusion. Weights The output of the large language model is assigned complementary weights 1 minus The outputs assigned to the physical model are always weighted the same as the sum of the weights of the two estimates, ensuring a meaningful and smooth transition between the two estimates in the fusion result. This is especially important when the large language model is in a high-confidence state. tending towards one, the final output This is mainly determined by the high-precision estimation of the large language model, fully leveraging its advantages in understanding complex features. When the large language model is in a low-confidence state... Approaching zero, the final output It is mainly determined by the reliable estimation of the physical model, and the safety and rationality of the output are guaranteed by physical methods.

[0076] In application, when the uncertainty index U equals the safety threshold τ, the value of ω(U) is 0.5, indicating that the output of the large language model and the output of the physical model each account for half of the weight; this is the balance point between the two. When U < τ, The value >0.5 and approaches 1 as U decreases, indicating that the large language model is in a high-confidence state and its output weights gradually increase. When U > τ, The value is <0.5 and approaches zero as U increases, indicating that the large language model is in a low-confidence state. The output weights of the physical model gradually increase. The weight switching coefficient γ controls the steepness of the weight transition as the uncertainty index changes. A larger γ value makes the weight switching more rapid, while a smaller γ value makes the transition smoother. It can be adjusted according to the different requirements of system response speed and smoothness in the application scenario. The uncertainty safety threshold τ represents the critical point for the system to judge the credibility of the large language model output. Its specific value can be determined by statistically analyzing the correlation between the prediction error of the large language model and the uncertainty index on the validation set.

[0077] In application, the uncertainty index U quantifies the model's confidence in the current input, and its measurement can be implemented from multiple perspectives. Prediction variance is calculated by performing multiple random forward propagations on the same input to obtain a set of predictions, and then calculating the statistical variance of this set of results. A larger variance indicates a more unstable prediction and lower model confidence. Prediction entropy is calculated by analyzing the distribution of multiple predictions. A higher entropy value indicates a more uniform distribution of predictions, a less clear bias in the model towards a specific prediction, and lower confidence. Quantile interval width measures uncertainty by calculating the range between the upper and lower quantiles of multiple predictions. A wider interval indicates a higher degree of dispersion in the predictions and lower confidence. Model ensemble divergence measures uncertainty by calculating the degree of difference between the predictions of multiple independently trained models for the same input. Consistent predictions from multiple models indicate high confidence, while significant divergence indicates low confidence. Confidence score is achieved by introducing a dedicated confidence estimation branch into the model architecture to directly output a score on the reliability of the prediction results.

[0078] Specifically, when processing a health status assessment request for a certain charging segment during the inference phase, the system performs several random forward propagations with random deactivation enabled on the input sample, collects multiple health status prediction values, and calculates their variance or entropy as an uncertainty index U. Subsequently, it substitutes a pre-determined safety threshold τ and a weight switching coefficient γ into the weight function to calculate the current output weights of the large language model. Simultaneously, it calls the physical model to calculate reference health status estimates based on methods such as ampere-hour integration. Finally, ω(U) is multiplied by the health status estimate of the large language model. Add 1 minus ω(U) multiplied by the physical model estimate The final output is the health status estimate. The results are then returned to the upper-level system. The entire process rapidly outputs high-precision results from large language models when uncertainty is low, and automatically introduces physical models to support the output when uncertainty is high, ensuring the security of the output.

[0079] The calibration method for the uncertainty safety threshold τ is as follows: Samples in the validation set that cause the absolute prediction error of the large language model to exceed a preset accuracy target value (e.g., 3% SOH) are considered high-error samples. A relatively high quantile (e.g., 85th to 95th quantile) of the uncertainty index U corresponding to this high-error sample set is calculated, and this quantile value is taken as the initial setting value of τ. Further adjustments to the quantile are made to balance the detection rate (recall rate) of the high-error samples with the false alarm rate of the large language model output being excessively restricted under normal operating conditions, achieving an engineering-acceptable trade-off between the two. The above calibration process should be performed independently for each selected uncertainty metric; the τ value obtained from different metric methods is not universally applicable. The weight switching coefficient γ is typically set in the range of 1 to 10. A larger γ value results in a steeper weight transition of U near τ, while a smaller γ value results in a smoother transition. In engineering applications, a grid search can be used on the validation set to determine the final fusion output. The value of γ that minimizes the mean square error.

[0080] The uncertainty index U can be measured using a Monte Carlo deactivation method, which calculates the variance of the prediction results after multiple forward propagations with random deactivation enabled. Alternatively, it can be achieved using an ensemble learning approach, which trains multiple independent models with different initializations or subsets and calculates the divergence of their prediction results. Ensemble learning typically provides more accurate uncertainty estimation but incurs higher computational costs. and The fusion can be adopted using The continuous fusion method with smooth weighting can be used, or the hard switching method with the threshold τ as the dividing point for switching the output source can be used. Hard switching is simple to implement, and the smooth transition of continuous weighted fusion is more suitable for scenarios where the online evaluation of vehicle-cloud collaboration has high requirements for output stability.

[0081] The uncertainty-aware adaptive fusion mechanism introduces the ability to proactively perceive and adaptively adjust the reliability of the large language model's output into the health status assessment system. Under normal operating conditions with high confidence, the system fully utilizes the high-precision estimation capability of the large language model to provide accurate health status assessment results. Under abnormal or out-of-distribution operating conditions with low confidence, the system proactively reduces the output weight of the large language model. And introduce reliable support for the physical model. This effectively controls the risk of unreliable predictions from large language models under abnormal operating conditions within an acceptable range. The smooth weight switching design based on the Sigmoid function effectively avoids sudden output changes that may be caused by hard switching, ensuring the smoothness of the system's transition between different confidence states. This gives the solution the necessary safety redundancy and engineering reliability for automotive-grade power battery health status assessment applications.

[0082] During the inference phase, the physical model and the large language model are executed in parallel for the same charging segment. The physical model estimates the reference health state by the ratio of the charging current time integral to the rated capacity, without relying on the output of the large language model. The two generate estimates independently and then merge them in the fusion module.

[0083] Example 2 This embodiment provides a battery health status assessment system based on a knowledge-enhanced large language model. This embodiment uses battery health status as the assessment target. Figure 2 As shown, the system in this embodiment includes functional modules corresponding to the method steps in Embodiment 1. The following is a detailed description of each module in conjunction with the accompanying drawings: The data acquisition and segment selection module serves as the system's data entry point. It is responsible for acquiring long-term monitoring data continuously generated by the battery during actual operation from the onboard terminal or cloud-based battery management platform, and for quality screening of the raw data to obtain effective data segments suitable for health status analysis. The onboard terminal collects multi-dimensional data in real time via the vehicle's internal bus, including cell voltage, total voltage, total current, battery temperature, state of charge, cumulative mileage, charging time, and vehicle identification information, and uploads the data to the cloud-based battery management platform for persistent storage. The selection process reviews the raw time-series data point by point. Based on current stability constraints, it determines whether the current sampling point is in a stable constant-current charging state. Based on temperature range constraints, it eliminates abnormal data segments under extreme temperature conditions. Based on voltage range constraints and the characteristic charging platform ranges of each battery chemical system, it selects effective segments with degradation analysis value. These three types of constraints work together to ultimately output a set of valid charging segments that meet the conditions for use by subsequent modules.

[0084] The health factor construction and standardization module receives valid charging segments output by the data acquisition and segment selection module, and extracts characteristic information representing the battery degradation state from two levels. The construction process of time-series health factors records and organizes the time series of current, voltage, and temperature in the valid charging segments, preserving the dynamic evolution of each physical quantity over time during charging, for deep feature extraction by the time-series semantic mapping module. The construction process of statistical health factors calculates statistics for the valid charging segments, extracting cumulative mileage, rated capacity, charging duration, average charging current, charging temperature statistics, single-cell voltage difference statistics, and charging capacity statistics within an effective voltage window determined based on the battery chemistry system. This effective voltage window is the same as the voltage range constraint interval used to select valid charging segments, resulting in seven dimensions of numerical features. Each feature quantifies the degradation state from the perspectives of usage history, capacity status, charging behavior, thermal management status, cell consistency, and capacity decay. The standardization process uses the statistical characteristics of the reference vehicle group as a benchmark to normalize the mean and standard deviation of each statistical health factor, eliminating the dimensional differences introduced by different vehicles and operating conditions, and providing a consistent feature representation basis for cross-vehicle generalization.

[0085] The domain knowledge enhancement prompt generation module receives standardized health factors and retrieves domain rules from the knowledge base that match the battery being evaluated. It then encodes numerical features and structured knowledge into a domain knowledge enhancement prompt vector using a prompt generation function. This module maintains three types of domain rule bases: a chemical system rule base storing descriptions of degradation mechanisms and prior information on charging characteristics for different electrode materials; a usage condition rule base storing descriptions of the impact of different usage modes on degradation rates; and a safety boundary rule base storing physical constraints for various battery health status assessment results. Upon receiving input containing vehicle identification information and battery model information, the module automatically retrieves matching rule entries from the corresponding rule bases. These entries, along with the normalized health factors, are processed by the prompt generation function to generate a complete prompt vector containing numerical information and domain background knowledge. This vector fully activates the pre-trained knowledge related to battery degradation analysis in the large language model, injecting domain expertise under conditions of efficient parameter adaptation.

[0086] The temporal semantic mapping module receives multi-dimensional temporal signals from the effective charging segment. Through its internal temporal semantic mapper, it transforms the physical measurement signals into continuous vector representations that match the embedding space dimension of the large language model, achieving cross-modal mapping from the physical signal domain to the semantic representation domain. The module performs time alignment and multi-channel organization on the input current, voltage, and temperature sequences, constructing a multivariate temporal matrix before inputting it to the temporal semantic mapper. The mapper analyzes the temporal signals through its internal feature extraction network, capturing current change rate, voltage plateau shape, temperature distribution characteristics, and the collaborative variation patterns between multivariate signals. Finally, it outputs a fixed-length continuous semantic vector. This vector, concatenated with a domain knowledge-enhanced cue vector in a continuous cue format, serves as input to the large language model, enabling the model to simultaneously perceive numerical statistical features, domain background knowledge, and degradation information inherent in the original physical signals.

[0087] The large language model inference module is the core computing unit of the system. It embeds a large language model trained with efficient parameter adaptation. This model receives a combination of domain knowledge-enhanced cue vectors and temporal semantic vectors as input, outputs a hidden layer embedding representing the battery degradation state, and then maps this hidden layer embedding to a numerical estimate of the battery health state by a trainable numerical regression output layer. Most of the weights of the large language model remain frozen during training, allowing only a small number of output adaptation layer or prefix vector parameters to be updated. This preserves the general inference capability of the large language model and avoids overfitting to small samples. During training, the sum of the physical consistency constraint loss and the regression loss is optimized. The physical consistency constraint loss prevents unreasonable increases in the health state estimate through monotonicity constraints, and the consistency constraint ensures that the model's prediction results maintain a reasonable deviation range from the ampere-hour integral reference value, ensuring that the battery health state estimate obtained through the numerical regression output layer conforms to the basic physical laws of battery degradation.

[0088] The uncertainty assessment and adaptive fusion module is a core component of the system's safety redundancy mechanism. It is responsible for assessing the confidence level of the large language model's output and dynamically adjusting the source structure of the final output based on the assessment results. The module triggers multiple random forward propagations of the large language model's inference module for the same input sample. Each propagation generates differentiated predictions by randomly deactivating some network nodes. After collecting multiple prediction results, it calculates uncertainty indicators such as variance or entropy. Subsequently, the module compares the calculated uncertainty indicator U with a preset safety threshold τ, calculates the weight coefficient ω(U) of the large language model's output using a sigmoid function with the difference between the two as the independent variable, and simultaneously calls physical methods to calculate a reference health status estimate. The final module estimates the large language model using weighted coefficients. and physical estimates A weighted summation is performed to obtain a final health status estimate that integrates data-driven capabilities and physical reliability. And export it to other countries.

[0089] Specifically, in the complete system workflow, the data acquisition and segment filtering module reads the charging records of a vehicle within a specified time period from the cloud-based battery management platform. After triple constraint filtering, the valid charging segments are passed to the health factor construction and standardization module. The latter extracts seven-dimensional statistical health factors and multi-dimensional temporal health factors and performs normalization. The normalized statistical health factors are then passed to the domain knowledge enhancement hint generation module, and the temporal signal is passed to the temporal semantic mapping module. The domain knowledge enhancement hint generation module combines the rule base to generate domain knowledge enhancement hint vectors. The temporal semantic mapping module outputs continuous semantic vectors. The combination of these two types of vectors is then input into the large language model inference module, which outputs the initial health state estimate. The uncertainty assessment and adaptive fusion module triggers multiple random inferences on the input and calculates the uncertainty index U, which is then weighted with adaptive weights ω(U). Compared with physical model estimates Weighted fusion is performed to output the final battery health state estimate. .

[0090] The data acquisition and segment selection module can be sourced from online streaming data uploaded in real-time via the vehicle terminal, or from offline batch processing by reading historical stored data in batches from the cloud-based battery management platform. The choice between these two methods depends on the real-time requirements of the application scenario. For the large language model inference module, efficient parameter adaptation can be achieved using a low-rank adaptation method that updates only the output and adapter layer parameters, or by adding learnable prefix vectors to the input. The former offers a good balance between parameter efficiency and performance, while the latter is more conservative in modifying the original parameters of the large language model.

[0091] This system integrates multi-source monitoring data acquisition, health factor construction, domain knowledge injection, cross-modal semantic alignment, physical constraint training, and uncertainty perception into a complete automotive-grade battery health status assessment solution through the orderly collaboration of six functional modules. Each module has a clear responsibility and standardized interfaces, facilitating flexible system configuration and expansion in different scenarios such as vehicle deployment, cloud training, and vehicle-cloud collaborative inference. The uncertainty assessment and adaptive fusion module acts as a safety guardian for the system. When the large language model encounters out-of-distribution samples, it proactively identifies low-confidence states and activates the physical model as a fallback, endowing the system with proactive safety protection capabilities under abnormal operating conditions. This ensures that the system meets the stringent requirements of automotive-grade applications for safety redundancy and robustness.

[0092] In application, the large language model in the inference module employs parameter efficiency adaptation by freezing all parameters (corresponding to the inference scenario) or freezing most parameters (corresponding to the training scenario). Freezing parameters means that the gradients of these parameters are set to zero and not updated during backpropagation, ensuring that the general semantic understanding and complex reasoning capabilities formed by the large language model in the pre-training stage are preserved. Only a small number of adaptation layer parameters, prefix vector parameters, or output mapping layer parameters are allowed to be adjusted with the training data, significantly reducing the risk of overfitting due to the limited scale of training data. At the same time, the number of trainable parameters is compressed to a very small proportion of the total number of parameters in the large language model, significantly reducing the computing power and GPU memory resources required for training. Physical consistency constraint loss. The monotonicity constraint term in the estimate of the health state of adjacent charging segments and The positive portion of the difference is penalized, constraining the large language model from predicting unreasonable results such as a rebound in battery health during aging. A consistency constraint term applies to the predicted values. Reference health status calculated based on the ampere-hour integral method The absolute deviation between the two is penalized, and the output of the large language model is constrained not to deviate too far from the physical reliable reference value. The two types of constraints work together to play the role of guarding the physical laws during the training phase.

[0093] In application, the suggestion generation function of the domain knowledge-enhanced suggestion generation module offers flexibility in its implementation. The text template implementation uses predefined natural language sentence structures as a framework, filling health factor values, chemical system characteristics, operating condition descriptions, and safety boundary information into a text string according to a fixed format. The generated content is human-readable, facilitating manual review and debugging, and is suitable for application scenarios with high interpretability requirements. The continuous soft suggestion vector implementation provides suggestions... Represented as a learnable continuous parameter vector, it is jointly optimized with the large language model during training through gradient backpropagation. It can automatically learn the implicit cue representation that is most conducive to activating the large language model's battery evaluation capability. It has few parameters but strong expressive power and is suitable for scenarios with large training data scales. The key-value prefix vector implementation introduces additional key-value pair parameters into the multi-head attention layer of the large language model, allowing domain knowledge to be deeply involved in the model's feature extraction process at the attention computation level. The temporal semantic mapping module's temporal semantic mapper also provides a variety of network structures to choose from. One-dimensional convolutional networks are good at extracting local patterns of temporal signals, recurrent neural networks are good at establishing dependencies across time steps, Transformer encoders achieve global temporal modeling through self-attention mechanisms, and attention matching networks are suitable for matching and recognition in the degradation stage. The most suitable structure can be selected according to the characteristics of the charging data.

[0094] Specifically, in the scenario of training a model on a ternary lithium battery dataset, the system freezes all weights of the transformer encoding layer of the large language model, allowing only the final linear regression output layer and a set of low-rank adaptation matrices to participate in gradient updates. During training, a physical consistency constraint loss, including monotonicity and consistency constraints, is calculated for each training batch. and mean squared error regression loss The sum of the two gives the total loss. The trainable parameters are then updated via gradient backpropagation. After training convergence, the large language model retains the general capabilities of pre-training while possessing the ability to accurately assess the health status of ternary lithium batteries. The cue generation function adopts a combination of text templates and continuous soft cue vectors, with the text templates incorporating chemical system rules. The soft cue vector represents the health factor value. The temporal semantic mapper uses a cascaded structure of a one-dimensional convolutional network and a Transformer encoder to first extract local temporal features and then perform global modeling.

[0095] Efficient parameter adaptation for the large language model can be achieved by decomposing the weight matrix using low-rank adaptation and training only the low-rank portion. Alternatively, it can be achieved by adding a learnable prefix vector before the input sequence for fine-tuning. Low-rank adaptation strikes a good balance between parameter efficiency and expressive power, while prefix fine-tuning requires minimal modification to the original model structure and is easier to version manage. The temporal semantic mapper can be trained using an end-to-end joint training approach with the large language model, or it can be trained in stages by first pre-training the mapper independently and then fine-tuning it jointly with the large language model. Staged training reduces the computational overhead of joint training while avoiding the gradient signal attenuation problem in deep networks.

[0096] Efficient parameter adaptation strategy for frozen parameters and loss due to physical consistency constraints The combination of these technologies simultaneously enhances the applicability of large language models in battery health status assessment tasks from both computational efficiency and output reliability perspectives. Efficient parameter adaptation effectively injects domain knowledge with only a minimal number of parameter updates, enabling large language models to achieve high-quality domain adaptation on small-sample battery datasets without losing pre-trained general reasoning capabilities. The training cost is significantly lower than full-parameter fine-tuning, making it particularly suitable for real-world scenarios where data annotation costs are high in automotive-grade applications. The flexible combination of various prompting implementations and mapper structures further broadens the solution's adaptability to different data conditions, hardware resources, and application requirements, improving the overall engineering feasibility of the solution.

[0097] In application, the data acquisition and segment filtering module obtains long-term monitoring data covering eight categories: individual cell voltage, total voltage, total current, battery temperature, state of charge, cumulative mileage, charging time, and vehicle identification information. Individual cell voltage... Recording the terminal voltage of each individual cell in the battery pack is a direct source for assessing cell consistency and individual cell health. The dispersion of multiple cell voltages directly reflects the degradation trend of battery pack consistency. The total voltage is the end-to-end voltage of the entire battery pack, reflecting the overall energy state of the battery pack. The total current is the charge transfer rate flowing into or out of the battery pack; it is negative in the charging direction and positive in the discharging direction, and is the basic data for identifying charge / discharge modes and calculating ampere-hour integrals. Battery temperature is collected by temperature sensors placed at different locations in the battery pack, reflecting the operating status of the thermal management system and the internal heat generation of the battery. The state of charge (SOC) is estimated in real time by the battery management system, reflecting the ratio of the current remaining charge to the fully charged state. Cumulative mileage is recorded by the vehicle odometer and is a comprehensive indicator of battery usage history. Charging time records the start and end times of each charging session, used to calculate charging duration and identify charging behavior patterns. Vehicle identification information includes vehicle model and battery batch information, used to retrieve matching chemical system rules and safety boundary rules from the rule base.

[0098] In application, the statistical health factors constructed by the health factor construction and standardization module total seven items. Among them, the cumulative mileage is directly read from historical records, reflecting the battery's accumulated driving history throughout its entire service life. Rated capacity is obtained through standardized capacity testing or capacity estimation algorithms, representing the upper limit of the battery's usable capacity under complete charge-discharge cycle conditions, and is the most core reference indicator for health status assessment. Charging duration is calculated from the time span of the effective charging segment; within the same voltage window, the charging duration shortens as capacity decreases. The average charging current is the arithmetic mean of the sampled current I(t) values ​​within the effective charging segment; this value is relatively stable within the constant current segment, but decreases when the charging current is derating due to protection limitations. The charging temperature statistics include the mean and standard deviation of the temperature within the effective charging segment; the mean reflects the average thermal level during charging, and the standard deviation reflects the spatial non-uniformity of temperature within the battery pack. The individual cell voltage difference statistics include the mean and maximum values ​​of the difference between the maximum and minimum individual cell voltages at each moment within the effective charging segment; this statistic increases with cell consistency degradation. Based on the charging capacity statistics within the effective voltage window determined by the battery chemistry system, the effective voltage window is the same as the voltage range constraint range used to screen effective charging segments. It is obtained by integrating the current over time within a preset standardized voltage range. It is highly correlated with the capacity measurement under standard test conditions and has low sensitivity to changes in charging strategy.

[0099] In applications, the uncertainty assessment and adaptive fusion module offers several optional implementations for measuring the uncertainty index U. Prediction variance is calculated by performing multiple random forward propagations with random deactivation enabled on the same input, using the statistical variance of the resulting predictions as the uncertainty index U. This method is simple to implement and has moderate computational cost, making it suitable for real-time inference scenarios. Prediction entropy is calculated by using the information entropy of the discrete distribution of multiple random prediction results as the uncertainty index U. It is more sensitive to the overall distribution characteristics of the prediction results and is suitable for detecting out-of-distribution samples. Quantile interval width is calculated by using the width between the upper and lower quartiles of multiple prediction results as the uncertainty index U, exhibiting some robustness to abnormal predictions. Model ensemble divergence is calculated by training multiple independent models with different initializations or different training subsets, using the standard deviation or mean absolute deviation between the prediction results of each model as the uncertainty index U. This method provides high-quality uncertainty estimation but requires maintaining multiple model replicas, making it suitable for cloud-based inference scenarios. Confidence scoring is calculated by attaching a confidence estimation branch to the output layer of the large language model, directly outputting an evaluation score of the reliability of the prediction results as the uncertainty index U. This method has the lowest inference cost but requires additional confidence-labeled data for supervised training. The fusion method can be either a continuously weighted smooth fusion method or a hard switching method with a threshold τ as the judgment point. The former has a smooth output and is suitable for online evaluation, while the latter has simple logic and is suitable for embedded deployment scenarios.

[0100] Specifically, in a typical deployment scenario of vehicle-cloud collaborative reasoning, the data acquisition and fragment filtering module is deployed on a cloud-based battery management platform, responsible for the centralized filtering and management of a large amount of charging data uploaded by vehicles. The health factor construction and standardization module completes the calculation and normalization of seven-dimensional statistical health factors in the cloud. The domain knowledge enhancement prompt generation module automatically matches chemical system rules based on vehicle identification information. and operating condition rules Generate hint vectors The temporal semantic mapping module transforms the temporal signals of effective charging segments into semantic vectors, while the large language model inference module performs inference on high-performance computing nodes in the cloud and outputs a health status estimate. The uncertainty assessment and adaptive fusion module performs multiple random samplings on the inference results and calculates the prediction variance or prediction entropy as the uncertainty index U. Based on a preset safety threshold τ, it uses the Sigmoid weighting function ω(U) to... Reference estimate of ampere-hour integral The weighted fusion is performed, and the final fused health status assessment results are then... Push the battery management system to the input terminal for use in upper-layer applications such as range prediction, charging strategy optimization, and safety warnings.

[0101] The uncertainty index U can be measured using either Monte Carlo deactivation (calculating the prediction variance after multiple random deactivation forward propagations) or ensemble learning (calculating the ensemble divergence after independently training multiple models). Monte Carlo deactivation has lower overhead on a single model, while ensemble learning offers higher reliability in uncertainty estimation but requires deploying multiple model instances. The large language model estimate... Compared with physical model estimates The fusion can be achieved by a continuous weighted fusion method with the Sigmoid function ω(U) as the weight switching function, or by a hard switching method with the uncertainty safety threshold τ as the judgment boundary to switch the output source. The continuous weighted fusion method outputs more smoothly within the confidence transition range, while the hard switching method has clear logic and is suitable for resource-constrained embedded edge deployment environments.

[0102] By meticulously defining eight input sources for monitoring data, the complete composition of seven-dimensional statistical health factors, and combining five uncertainty measurement methods with two fusion implementation options, the system forms a rich technical layer at the key implementation details of each functional module. This enables the overall solution to maintain effective operation under different deployment modes, including high-computing-power training in the cloud, lightweight inference on the vehicle side, and vehicle-cloud collaborative online evaluation. The optional design of multiple uncertainty measurement methods fully considers the differences in computing power resources and real-time requirements, allowing the system to flexibly select the optimal uncertainty estimation strategy according to the actual deployment environment while ensuring output safety. This provides comprehensive technical support for the full life cycle health status management of automotive-grade power batteries, balancing accuracy, safety, and engineering feasibility.

[0103] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A battery health status assessment method based on a knowledge-enhanced large language model, characterized in that, Includes the following steps: Effective charging segments during the constant current charging phase were selected from battery monitoring data, and health factors were constructed. The health factors are integrated with domain rules to create domain knowledge-enhanced prompts. The domain rules include battery chemistry system rules, vehicle operating condition rules, and safety boundary rules. The safety boundary rules include constraints that the health status is monotonically non-increasing with the aging process and that the estimated health status value is within a preset physical range. The battery chemistry system rules, vehicle operating condition rules, and safety boundary rules are encoded and injected into the domain knowledge enhancement prompt information in the form of natural language text or structured fields to constrain the output of the large language model to conform to the physical boundary of battery degradation. The effective charging segments are mapped into continuous semantic vectors in the large language model embedding space by a temporal semantic mapper; The domain knowledge-enhanced prompts and the continuous semantic vectors are input into the large language model, and the model is trained with the sum of physical consistency constraint loss and regression loss as the objective function to output the battery health status estimate. The physical consistency constraint loss includes a monotonicity constraint term that constrains the health status estimate from abnormally increasing with the aging process, and a consistency constraint term that constrains the deviation between the model prediction result and the reference health status obtained based on the ampere-hour integral. The physical model output is obtained by independently estimating the effective charging segment using the physical model; the physical model is at least one of the ampere-hour integration method, the equivalent circuit model method, or the electrochemical model method. The uncertainty index is calculated and predicted multiple times for the same input sample. Based on the uncertainty index, the battery health state estimate is weighted and fused with the output of the physical model to obtain the final battery health state estimate.

2. The battery health status assessment method based on a knowledge-enhanced large language model according to claim 1, characterized in that, The battery monitoring data is long-term monitoring data of the battery during actual operation, obtained from the vehicle terminal or cloud-based battery management platform. The screening is based on current stability constraints and temperature and voltage range constraints. The screening criteria for the effective charging segment are expressed as follows: in, This is the effective constant current charging sampling set; Let be the charging current at time t; This represents the change in current between adjacent sampling points; This is the preset constant current stability threshold; Let be the unit voltage at time t; , The effective voltage window is determined based on the lower and upper limits of the battery chemistry system, and different voltage windows are set for lithium iron phosphate batteries and ternary lithium batteries, respectively. Let t be the battery temperature at time t; , These are the preset lower and upper limits of the temperature filtering range, respectively.

3. The battery health status assessment method based on a knowledge-enhanced large language model according to claim 1, characterized in that, The health factors include time-series health factors that characterize the local evolution of current, voltage and temperature during charging and statistical health factors that characterize the macroscopic degradation state of the battery. The time-series health factors are standardized using a channel-by-channel sequence normalization method, and the statistical health factors are standardized using a scalar normalization method based on the mean and standard deviation of the reference vehicle group. The statistical health factors include cumulative mileage, rated capacity, charging duration, average charging current, charging temperature statistics, single-cell voltage difference statistics, and charging capacity statistics within an effective voltage window determined based on the battery chemistry system. The effective voltage window is the same as the voltage range constraint interval used to screen effective charging segments. The standardization process is performed according to the following formula: in, This is the kth original health factor; This is to provide a reference for the average value of this health factor in the vehicle population; The standard deviation of this health factor is used as a reference for the vehicle population; ε is a regularization constant to prevent the denominator from being zero; This represents the k-th health factor after normalization.

4. The battery health status assessment method based on a knowledge-enhanced large language model according to claim 1, characterized in that, The generation of the domain knowledge-enhanced prompt information is represented as follows: in, Enhance the hint vector for domain knowledge; This is the normalized set of health factors; Rules for battery chemistry systems; Rules for vehicle operating conditions; For safety boundary rules; Ψ(·) is the prompt generation function; The prompt generation function is implemented using text templates, continuous soft prompt vectors, key-value prefix vectors, or a combination of the above.

5. The battery health status assessment method based on a knowledge-enhanced large language model according to claim 1, characterized in that, The temporal semantic mapper is composed of a trainable neural network, employing any one of a one-dimensional convolutional network, a recurrent neural network, a Transformer encoder, an attention matching network, or a combination of the above networks. It injects battery current changes, voltage plateau shifts, temperature fluctuations, and individual cell consistency degradation during the charging process into the large language model in the form of continuous semantic vectors, thereby achieving cross-modal alignment between physical temporal signals and the embedding space of the large language model.

6. The battery health status assessment method based on a knowledge-enhanced large language model according to claim 1, characterized in that, The training includes freezing the pre-trained parameters of the large language model, retaining the trainable parameters of the adaptation layer, output mapping layer, or learnable prefix vectors attached to the large language model, and training the trainable parameters with the objective function. The large language model processes the input and outputs a hidden layer embedding representing the battery degradation state. This hidden layer embedding is mapped to an estimate of the battery health state via a trainable numerical regression output layer. The physical consistency constraint loss is expressed as: in, Loss due to physical consistency constraints; This represents the health state estimate of the model at the t-th charging segment; This represents the health status estimate of the model at the (t+1)th charging segment; the tth charging segment and the (t+1)th charging segment are two constant current charging segments that are adjacent in time for the same vehicle. This indicates a reference health status based on ampere-hour integration; Adjustable weighting coefficients for monotonicity constraints that prevent health status from abnormally increasing with the aging process; An adjustable weighting coefficient for the consistency constraint term between the model prediction results and the reference health status obtained based on the ampere-hour integral; The overall objective function of the training is the sum of the physical consistency constraint loss and the regression loss, that is: in, Total loss; For regression loss.

7. The battery health status assessment method based on a knowledge-enhanced large language model according to claim 1, characterized in that, The multiple predictions refer to multiple random forward propagations or multi-model ensemble predictions performed on the same input sample. The weighted fusion is performed based on the relationship between the uncertainty index and the uncertainty safety threshold, and after calculating the output weights of the large language model using a monotonically decreasing weight function with the uncertainty index as input. The final battery health state estimate is expressed as follows: The weighting function is expressed as follows: in, This is the final output estimate of the battery health status; The health status estimate output by the large language model; The health status estimate is output by the physical model; U is the uncertainty index; τ is the uncertainty safety threshold; γ is the weight switching coefficient. The physical model and the temporal semantic mapper share the same valid charging segments as input.

8. A battery health status assessment system based on a knowledge-enhanced large language model, characterized in that, include: The data acquisition and segment filtering module is used to filter effective charging segments during the constant current charging phase from battery monitoring data. A health factor construction and standardization module is used to construct health factors based on the effective charging segment; A domain knowledge enhancement prompt generation module is used to integrate the health factors with domain rules into domain knowledge enhancement prompt information; The domain rules include battery chemistry system rules, vehicle operating condition rules, and safety boundary rules. The safety boundary rules include constraints that the health status is monotonically non-increasing with the aging process and that the estimated health status value is within a preset physical range. The battery chemistry system rules, vehicle operating condition rules, and safety boundary rules are encoded and injected into the domain knowledge enhancement prompt information in the form of natural language text or structured fields to constrain the output of the large language model to conform to the physical boundary of battery degradation. The temporal semantic mapping module is used to map the effective charging segment into a continuous semantic vector in the large language model embedding space via the temporal semantic mapper; The large language model inference module is used to receive the domain knowledge-enhanced prompt information and the continuous semantic vector, and train the large language model with the sum of physical consistency constraint loss and regression loss as the objective function, and output the battery health status estimate; the physical consistency constraint loss includes a monotonicity constraint term that constrains the health status estimate not to increase abnormally with the aging process and a consistency constraint term that constrains the deviation between the model prediction result and the reference health status obtained based on ampere-hour integration. The uncertainty assessment and adaptive fusion module is used to independently estimate the effective charging segment by the physical model to obtain the physical model output, make multiple predictions and calculate the uncertainty index for the same input sample, and weight and fuse the battery health state estimate with the physical model output based on the uncertainty index to obtain the final battery health state estimate; the physical model is at least one of the ampere-hour integration method, equivalent circuit model method or electrochemical model method.

9. The battery health status assessment system based on a knowledge-enhanced large language model according to claim 8, characterized in that, The data acquisition and segment filtering module obtains the battery monitoring data from the vehicle terminal or cloud battery management platform, and filters it based on current stability constraints and temperature and voltage range constraints. The health factors constructed and standardized by the health factor construction and standardization module include time-series health factors and statistical health factors, and the time-series health factors and statistical health factors are standardized. The temporal semantic mapper in the temporal semantic mapping module is composed of a trainable neural network; the large language model in the large language model inference module adopts the method of freezing all parameters or freezing some parameters during the training phase to achieve efficient parameter adaptation; the large language model outputs a hidden layer embedding that represents the battery degradation state; the hidden layer embedding is mapped to the battery health state estimate through a trainable numerical regression output layer. The domain knowledge enhancement prompt generation module's prompt generation function is implemented using text templates, continuous soft prompt vectors, key-value prefix vectors, or a combination of the above forms; the temporal semantic mapper of the temporal semantic mapping module is any one of a one-dimensional convolutional network, a recurrent neural network, a Transformer encoder, an attention matching network, or a combination of the above networks.

10. The battery health status assessment system based on a knowledge-enhanced large language model according to claim 9, characterized in that, The long-term monitoring data acquired by the data acquisition and segment screening module includes individual cell voltage, total voltage, total current, battery temperature, state of charge, cumulative mileage, charging time, and vehicle identification information. The statistical health factors constructed by the health factor construction and standardization module include cumulative mileage, rated capacity, charging duration, average charging current, charging temperature statistics, individual cell voltage difference statistics, and charging capacity statistics within the effective voltage window determined based on the battery chemistry system. The effective voltage window is the same as the voltage range constraint interval used to screen effective charging segments. The uncertainty assessment and adaptive fusion module uses any one of the following to measure uncertainty indicators: prediction variance, prediction entropy, quantile interval width, model integration divergence, or confidence score. It also uses any one of the following methods: hard switching or weighted fusion processing to achieve the fusion of the output of the large language model and the output of the physical model.

Images