Lithium ion battery management method and system based on visual large language model

Abstract

The application belongs to the field of lithium ion battery management, and provides a lithium ion battery management method and system based on a visual large language model, obtains battery operation sampling data, and runs a pre-constructed battery physical model; based on the battery operation sampling data and the battery physical model, state estimation is performed, whether a battery management event is triggered is judged according to residual window statistics, instantaneous residual extreme value and operation condition anchor points, when the event is triggered, a residual panel chart of multiple trajectories is generated, and a structured numerical summary is formed; a query vector is constructed based on the current event, and similar historical cases are recalled from a pre-constructed retrieval enhanced knowledge base partitioned according to tasks; the residual panel chart, the structured numerical summary, the operation condition anchor points and the historical cases are used as multi-modal inputs, a visual large language model is used to determine a task route; and according to the task route, a corresponding execution branch is entered. The application can guarantee the real-time performance of state estimation, and avoids large model reasoning on each frame of high-frequency data.

Description

Technical Field

[0001] This invention belongs to the field of lithium-ion battery management, specifically relating to a lithium-ion battery management method and system based on a visual large language model. Background Technology

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

[0003] Lithium-ion battery management systems (BMS) need to simultaneously perform tasks such as state-of-charge (SOC) estimation, state-of-health (SOH) estimation, fault diagnosis, capacity calibration, and remaining lifetime prediction. Existing solutions typically use equivalent circuit models, extended Kalman filters, rule-based thresholding, machine learning models, or cloud-based knowledge bases to handle these tasks separately. While these conventional management solutions can handle SOC estimation or specific fault diagnosis tasks, their comprehensive ability to cope with the highly time-varying and nonlinear characteristics of lithium-ion batteries is limited. Furthermore, SOC estimation, fault diagnosis, parameter maintenance, and remaining lifetime prediction are usually treated as relatively independent processing chains, lacking reasonable scheduling logic for each function.

[0004] With the development of large language models, existing technologies have begun to explore their use in system modeling or digital twin interaction. However, directly applying large language models to battery management still faces the following challenges: when processing long-term battery data, large language models are limited by context length and time alignment capabilities, making it difficult to fully express subtle residual patterns; continuously calling large models on high-frequency sampled data will result in a significant computational burden; and the capabilities of large models typically rely on offline training or centralized updates, making it difficult to instantly precipitate new faults, new operating conditions, and new life trajectories that occur during operation into reusable online knowledge. Summary of the Invention

[0005] To address the aforementioned problems, this invention proposes a lithium-ion battery management method and system based on a visual large language model. This invention can guarantee the real-time performance of state estimation and avoids performing large model inference on each frame of high-frequency data.

[0006] According to some embodiments, the present invention adopts the following technical solution: A lithium-ion battery management method based on a visual large language model includes the following steps: Obtain battery operation sampling data, and run a pre-built battery physical model based on the operation sampling data; Based on the battery operation sampling data and battery physical model, state estimation is performed to obtain the state of charge, terminal voltage estimation and voltage residual sequence; Based on the residual window statistics, instantaneous residual extreme values ​​and operating condition anchor points, determine whether a battery management event has been triggered, and continue state estimation if the event has not been triggered. When an event is triggered, a residual panel plot containing voltage, current, state of charge and residuals is generated to form a structured numerical summary. Based on the current event, construct a query vector and recall similar historical cases from a pre-built task-partitioned enhanced knowledge base; The residual panel plot, structured numerical summary, operating condition anchor points, and historical cases are used as multimodal inputs, and a visual big language model is used to determine the task route. Based on the task route, enter the corresponding execution branch to perform selective parameter updates, fault diagnosis, capacity calibration, remaining lifetime prediction, or noise ignoring.

[0007] As an alternative implementation, the process of running a pre-built battery physical model based on the running sampled data includes: according to the sampling time... k Construct the running sampling vector: ; in, , and They represent the first k The terminal voltage, current, and temperature at each sampling time will The extended Kalman filter is input, and during the state prediction update process, the battery state of charge, polarization branch voltage, and terminal voltage estimates are updated synchronously using the battery physical model. Based on the terminal voltage estimates and terminal voltage, a voltage residual sequence is generated.

[0008] As an alternative implementation, the process of determining whether a battery management event is triggered based on residual window statistics, instantaneous residual extreme values, and operating condition anchor points includes: for a length of L sliding window Construct residual eigenvectors: ; Among them, RMSE r The mean square root of the window residuals. r and std r These represent the mean and standard deviation of the window residuals, respectively. Corr represents the maximum absolute value of the window residual. rI This indicates the correlation between the residual and the current, hf ratio This represents the proportion of high-frequency energy in the residual; when the eigenvector contains skew r and kurt r At that time, skew r Kurt represents the residual skewness. r This indicates the residual kurtosis.

[0009] Calculate the root mean square baseline of the residuals and the extreme value baseline of the instantaneous residuals within the preheating window, and then calculate the corresponding thresholds: ; ; in, (x, a, b) means restricting x to the interval [a, b], where x is... or , a is or , , respectively, are the corresponding lower limits of the threshold, and b is or , which are the corresponding upper limits of the threshold. and This is the sensitivity coefficient. The root mean square baseline of the residuals, The baseline for the instantaneous residual extreme value; The event trigger flag is represented as: ; in, This indicates whether a running condition anchor point appears within the window. If the value is 1, a battery management event is triggered. This represents the OR operation in logical operations; if If the value is zero, then the continuous state estimate is maintained.

[0010] As an alternative implementation, the operating condition anchor point includes at least one of the following: static anchor point, constant current anchor point, constant voltage anchor point, pulse anchor point, and equivalent full-cycle anchor point. As context input, it is used to assist in determining whether the current residual form corresponds to parameter drift, real fault, capacity calibration opportunity, or noise.

[0011] As an alternative implementation, upon event triggering, the process of generating a residual panel plot containing voltage, current, state of charge, and residuals, and forming a structured numerical summary, includes: the multi-track panel plot is represented as follows: ; in, （） Indicates that the window W The operation of rendering multi-source sequences within a unified timeline into an image panel. For voltage trajectory, For current trajectory, For the trajectory of the charged state, For residual loci; Structured numerical summaries include at least two of the following metrics: ; in, The root mean square of the residuals, The mean of the residuals, The standard deviation of the residuals. For residual skewness, For residual kurtosis, The maximum absolute value of the residual. The correlation between residual and current, This represents the proportion of high-frequency residual energy.

[0012] As an alternative implementation, the process of constructing a query vector based on the current event and recalling similar historical cases from a pre-built task-partitioned enhanced retrieval knowledge base includes: dividing the pre-built task-partitioned enhanced retrieval knowledge base into multiple namespaces based on tasks, such as fault, parameter optimization, capacity, and remaining lifetime; and recalling K historical cases from the corresponding namespaces according to the current task or candidate tasks after the event is triggered. ; ; ; ; ; in, This represents a structured numerical summary of the event window W. This represents the residual statistical eigenvector within window W. This indicates the anchor point of the operating condition corresponding to window W. This represents the operating condition label corresponding to window W, Normalize(·) represents vector normalization, and E img (P W ) indicates the residual panel plot P W Visual features obtained through image encoding; DB param DB fault DB cap DB rul and DB noise These represent the parameter update library, fault case library, capacity calibration library, remaining lifetime trajectory library, and noise event library, respectively. c Let sim(·) represent the vectorized representation of historical case c, sim(·) represent the similarity function, K represent the number of recalled cases, and C K t Let C represent the set of K similar historical cases recalled from the namespace corresponding to task t, where T represents the candidate task set, and C represents the set of K similar historical cases recalled from the namespace corresponding to task t. K This represents the set of historical cases after merging across candidate tasks.

[0013] As an alternative implementation, the process of determining task routing using the residual panel plot, structured numerical summary, operational condition anchor points, and historical cases as multimodal inputs and a visual big language model includes: obtaining the event task routing result using the visual big language model based on the multi-trajectory panel plot, structured numerical summary, operational condition anchor points, and historical cases, as expressed as: ； ； in, (·) represents the multimodal inference function of the visual big language model. Event window W The corresponding residual panel plot, Represents a structured numerical summary. This represents the set of similar historical cases retrieved. Indicates preset routing constraints; Let represent the probability of candidate task t. The event routing result is represented by c, the routing confidence score, m, the parameter selection mask, u, the structured reason description, argmax, the candidate task with the highest probability of selection, and T, which represents the set of candidate tasks. The event routing result includes parameter updates, faults, capacity anchors, remaining lifetime predictions, noise, and unknowns.

[0014] As an alternative implementation, the process of entering the corresponding execution branch according to the task routing and performing fault diagnosis includes: constructing a fault feature vector. ; in, For the parameters of the equivalent circuit model, For statistical characteristics of operating conditions, The residual time-frequency characteristics include: the equivalent circuit model parameter fingerprint includes the current parameter value, parameter change amount, and parameter change rate; the operating condition statistical characteristics include current, temperature, and voltage window statistics; the residual time-frequency characteristics include residual root mean square, residual kurtosis, residual-current correlation, and high-frequency energy ratio. The fault namespace of the enhanced knowledge base is retrieved to find similar fault cases based on fault feature vectors. The visual big language model takes the current multi-trajectory panel map, similar fault cases and structured numerical summary as input to obtain fault type, fault severity level and handling suggestions. The process of predicting remaining battery life, based on the task routing and entering the corresponding execution branch, includes: constructing a lifetime feature vector based on the battery health state curve. ; in, This represents the curve compression feature obtained by segmenting and aggregating the battery health state curve. Indicates the capacity decay slope, Indicates the trend of impedance change. Indicates the inflection point position. Indicates the sharpness of the inflection point; Similar degradation trajectories are retrieved based on the lifetime feature vector, and the current battery health state curve and similar degradation trajectories are rendered into a trajectory comparison map. The battery health state outputs the remaining lifetime prediction result based on the current battery health state curve, similar degradation trajectories and structured numerical summary. The remaining lifetime prediction result is the remaining number of cycles, the remaining running time or the expected time to reach the preset capacity threshold.

[0015] As an alternative implementation, the following steps are also included: writing the event features, the output of the visual large language model, and the execution results back to the corresponding task namespace to form an online learning closed loop; if the nearest neighbor distance between the newly written case and the existing case meets the merging condition, the exponential moving average is used for merging. ; in, For the new case vector, For existing case vectors, This is the merged case vector. To consolidate weights, if the consolidation criteria are not met, the new case will be written as an independent case. In the fault namespace, set a ratio constraint between normal samples and faulty samples to prevent the number of normal samples from exceeding the set value.

[0016] A lithium-ion battery management system based on a visual large language model includes: The sampling data acquisition module is configured to acquire battery operation sampling data and run a pre-built battery physical model based on the operation sampling data; The digital twin module is configured to perform state estimation based on the battery operation sampling data and the battery physical model, and obtain the state of charge, terminal voltage estimation and voltage residual sequence. The event triggering module is configured to determine whether to trigger a battery management event based on residual window statistics, instantaneous residual extreme values, and operating condition anchor points. If no event is triggered, state estimation continues. The image and summary generation module is configured to generate a residual panel plot containing voltage, current, state of charge and residuals of multiple trajectories when an event is triggered, forming a structured numerical summary. The retrieval enhancement knowledge base is configured to be pre-built, partitioned by task, constructs query vectors based on the current event, and recalls similar historical cases. The visual big language model decision module is configured to use the residual panel plot, structured numerical summary, operating condition anchor point and historical cases as multimodal inputs to determine the task route using the visual big language model. The execution module is configured to enter the corresponding execution branch according to the task route and perform selective parameter updates, fault diagnosis, capacity calibration, remaining lifetime prediction, or noise ignoring.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides an event-triggered mechanism that allows the visual large language model to intervene only when residual abnormalities or anchor events occur, significantly reducing the computational overhead of point-by-point inference on high-frequency sampled data.

[0018] This invention renders voltage, current, state of charge, and residuals as multi-trajectory panel plots, enabling the visual large language model to identify the causes corresponding to residual morphology by means of image understanding capabilities. This invention integrates parameter updates, fault diagnosis, capacity calibration, and remaining lifetime prediction into the same digital twin closed loop through a unified event routing mechanism, avoiding the separation of these tasks. This invention features a task-partitioned enhanced knowledge base and an online write-back mechanism, enabling the retrieval, accumulation, and reuse of various cases, thereby enhancing the engineering value of historical experience.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0020] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0021] Figure 1 This is a schematic diagram of a lithium-ion battery management system based on a visual large language model, according to one embodiment. Figure 2 This is a schematic diagram of an event-driven management process based on a visual large language model, as exemplified by one embodiment. Figure 3(a) is a multi-track panel diagram of an event triggering window according to an embodiment; Figure 3(b) is a schematic diagram of a numerical summary of one embodiment; Figure 4(a) is a schematic diagram of the lifetime feature vector construction in one embodiment; Figure 4(b) is a schematic diagram of the similar degradation trajectory retrieval distance in one embodiment; Figure 5 This is a schematic diagram illustrating remaining lifetime prediction based on similar degradation trajectory retrieval, according to one embodiment. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0025] Where there is no conflict, the embodiments and features described in this application may be combined with each other.

[0026] Example 1 A lithium-ion battery management method based on a visual big language model is proposed. This method uses residual panel visual routing as its core mechanism, constructing a closed loop consisting of continuous physical twin estimation, event triggering, residual panel visual representation, retrieval and enhanced case comparison, visual big language model routing decision, and result writing back. In this closed loop, state estimation is performed continuously; the visual big language model only intervenes when an event is triggered, undertaking cross-modal cause discrimination and task routing. This dual-speed architecture of high-frequency physical data and low-frequency visual data maintains the real-time performance of state estimation while avoiding large-model inference for each frame of high-frequency data.

[0027] Specifically, such as Figure 2 As shown, it includes the following steps: S1. Obtain battery operation sampling data; S2. The digital twin module continuously outputs the state of charge, terminal voltage estimate, and voltage residual sequence based on the battery physical model and state estimator. S3. The event triggering module determines whether to trigger a battery management event based on the residual window statistics, instantaneous residual extreme values ​​and operating condition anchor points. If the event is not triggered, the digital twin module continues to run at high frequency without calling the visual big language model. S4. When the event is triggered, generate a residual panel plot containing multiple trajectories such as voltage, current, state of charge and residuals, and form a structured numerical summary. S5. Construct a query vector based on the current event and recall similar historical cases from the task-partitioned enhanced knowledge base; S6. The residual panel plot, structured numerical summary, operating condition anchor point and historical cases are used as multimodal inputs and submitted to the visual big language model. The visual big language model outputs task routing, confidence, parameter selection mask and structured reasons. S7. Enter the corresponding execution branch according to the task route, including selective parameter update, fault diagnosis, capacity calibration, remaining lifetime prediction or noise ignoring; S8. Write the event features, visual language model output, and execution results back to the corresponding task namespace to form an online learning closed loop.

[0028] The following is a detailed introduction to each step.

[0029] First, obtain the operational sampling data of the lithium-ion battery and construct the sampling vector: ; in, , and These represent the terminal voltage, current, and temperature at the k-th sampling time, respectively.

[0030] In this embodiment, the digital twin module is based on a second-order RC equivalent circuit model and the voltage estimate V generated by the extended Kalman filter. hat (k), State of charge (SOC)(k) and voltage residual r(k): ; Where r(k) represents the voltage residual at the k-th sampling time, V meas (k) represents the actual measured terminal voltage. This represents the estimated terminal voltage obtained from the battery physical model and state estimator.

[0031] Then, the residual feature vector is calculated within the sliding window W, and a structured numerical summary corresponding to the operating condition anchor point, operating condition label, and battery identifier is generated. : ; ; The event triggering mechanism is an essential component of this embodiment, and not merely a display or alarm logic selectively employed in the embodiment. The residual root mean square baseline is determined within the warm-up window. and instantaneous residual extreme baseline Based on this, the root mean square threshold of the residuals and the instantaneous residual threshold are formed: ; ; in, (x, a, b) means restricting x to the interval [a, b], where x is... or , a is or , , respectively, are the corresponding lower limits of the threshold, and b is or , which are the corresponding upper limits of the threshold. and This is the sensitivity coefficient. The root mean square baseline of the residuals, The baseline is the extreme value of the instantaneous residual.

[0032] Furthermore, by combining operational condition anchor points such as static, constant current, constant voltage, pulse, and equivalent full cycle, event trigger judgments are formed: ; in, This indicates whether a running condition anchor point appears within the window. If the value is 1, a battery management event is triggered. This represents the OR operation in logical operations; if If the value is zero, then the continuous state estimate is maintained.

[0033] For example, confirmation may be granted only if multiple consecutive sliding windows meet the residual threshold condition, or if the instantaneous residual exceeds the limit and continues to exceed the preset delay window. Set to 1 to avoid frequent triggering of the visual large language model due to sampling noise. If If the value is 0, continue performing high-frequency continuous estimation and updating the residual sequence, without generating a visual large language model request; if The system constructs an event object containing window number, unit number, anchor point type, residual statistics, and operating condition label, and sets the window... The voltage, current, state of charge, and residuals within are rendered as time-aligned residual panel plots. The residual panel plot is not a regular curve output, but rather an event evidence plot used by the visual large language model to identify morphological features such as abrupt changes, shifts, oscillations, tails, inflection points, and current correlations. ; ; ; Normalize(·) represents the operation of normalizing and fusing residual statistical features, visual features, operating condition anchors, and operating condition labels.

[0034] This embodiment sets up a retrieval-enhanced knowledge base partitioned by task, which includes at least a parameter update database, a fault case database, a capacity calibration database, a remaining lifetime trajectory database, and a noise event database. ; ; ; in, ( ) represents the similarity function. This indicates that the retrieval of the current event is enhanced by generating a query vector. DB represents the historical case vector. param DB fault DB cap DB rul and DB noise These represent the parameter update library, fault case library, capacity calibration library, remaining lifetime trajectory library, and noise event library, respectively. This represents the set of K similar historical cases recalled from the namespace corresponding to task t. Let K represent the merged candidate case set, K represent the number of recalled cases, and T represent the candidate task set, where task t includes selective parameter updates. Fault diagnosis Capacity calibration Remaining life prediction Noise ignored .

[0035] Therefore, Retrieval-augmented Generation (RAG) does not simply add explanatory text to the model, but rather provides a constrained comparison between the current residual event and historical parameter drift, real failures, capacity anchors, lifetime degradation, and noise perturbation cases, reducing the risk of visual large language models making isolated judgments based solely on the current image.

[0036] The residual panel plot Structured numerical summarization RAG search case The visual large language model is input with a preset routing constraint G, and outputs task probability, event routing, confidence score, parameter selection mask, and structured explanation of the reasons. ； ； in, (·) represents the multimodal inference function of the visual big language model. Event window W The corresponding residual panel plot, Represents a structured numerical summary. This represents the set of similar historical cases retrieved. Indicates preset routing constraints; Let represent the probability of candidate task t. The event routing result is represented by c, the routing confidence score, m, the parameter selection mask, u, the structured reason description, argmax, the candidate task with the highest probability of selection, and T, which represents the set of candidate tasks. The event routing result includes parameter updates, faults, capacity anchors, remaining lifetime predictions, noise, etc.

[0037] With the help of the above structured output, the visual big language model can simultaneously incorporate visual morphology, structured numerical summaries and historical cases from the residual panel plot into the judgment, thereby distinguishing different event semantics such as slow residual shift caused by parameter drift, residual abrupt change or high-frequency oscillation caused by fault, the appearance of capacity anchors that allow the battery state of health (SOH) to be calibrated, and degradation trajectories supporting the prediction of remaining lifetime.

[0038] like For parameter updates, restricted optimization is performed only on the parameters selected by the mask; unselected parameters remain locked. ; ; in, For the updated parameters, The parameters before the update. To represent the change in parameters, To represent element-wise multiplication, , , , , These are parameter selection masks for resistors R0, R1, C1, R2, and C2, respectively. R0, R1, and R2 represent the internal resistance in ohms, the fast-polarization resistor, and the slow-polarization resistor, respectively. C1 and C2 are the capacitors connected in parallel with resistors R1 and R2, respectively. , , , , The values ​​are all When the corresponding parameter selection mask is equal to 1, it means that the corresponding parameter needs to be updated; when the corresponding parameter selection mask is equal to 0, it means that the corresponding parameter remains unchanged.

[0039] like For fault diagnosis, the fault type, severity level, and handling recommendations are output based on the fault case library. If For remaining lifetime prediction, a weighted estimate is performed based on similar degradation trajectories obtained from RAG retrieval: ; ; in, These are the weighting coefficients. Let be the distance between the current event and the i-th similar degenerate trajectory. Let be the distance between the current event and the j-th similar degenerate trajectory. This represents a temperature parameter used to adjust the degree to which distance differences affect weight allocation; The smaller the value, the more concentrated the weight distribution, and the more the model tends to select trajectories that are closer together. The larger the value, the smoother the weight distribution, and the smaller the weight difference between different trajectories. This represents the lifetime termination cycle number for this similar trajectory. This represents the current loop number. Finally, the event features, residual panel plot, RAG retrieval results, routing results from the visual large language model, and execution results are presented. Write back to the corresponding task namespace: .

[0040] in, Indicates the task namespace. This indicates that the retrieval of the current event is enhanced by generating a query vector. Event window W The corresponding residual panel plot, Represents a structured numerical summary. Indicates the event routing result. This indicates the execution result, and u represents the structured reason explanation.

[0041] Example 2 A lithium-ion battery management system based on a visual large language model, such as Figure 1 As shown, it includes: The system includes a lithium-ion battery pack 100, a sampling data acquisition module 101, a digital twin module 102, an event triggering module 103, an image and summary generation module 104, a retrieval enhancement knowledge base 105, a visual large language model decision module 106, an execution module 107, and a write-back module 108.

[0042] The sampling data acquisition module 101 is used to acquire the operational sampling data of the lithium-ion battery pack 100. The operational sampling data may include data such as terminal voltage, current, temperature, timestamp, and battery cell number.

[0043] The digital twin module 102 is used to run a battery physical model and a filter based on the running sampled data. In one embodiment, the battery physical model employs a second-order RC equivalent circuit model, and the filter employs an extended Kalman filter. The digital twin module 102 outputs the state of charge (SOC), estimated terminal voltage, and voltage residual sequence based on the running sampled data. The voltage residual sequence serves as a common chain of evidence for subsequent event triggering, residual panel plot generation, and visual large language model routing, rather than being an intermediate quantity used solely for monitoring filter error.

[0044] The event triggering module 103 determines whether to trigger a battery management event based on the sliding window statistics of the voltage residual sequence and the operating condition anchor point. The image and summary generation module 104 generates a multi-trajectory panel plot and a structured numerical summary after the event is triggered. The retrieval enhancement knowledge base 105 retrieves historical cases. The visual large language model decision module 106 outputs the event routing results. The execution module 107 performs parameter updates, fault diagnosis, capacity calibration, remaining lifetime prediction, or noise ignoring based on the event routing results. The write-back module 108 writes the event results back to the retrieval enhancement knowledge base 105.

[0045] In one implementation, the sampling data acquisition module 101 constructs a running sampling vector according to the sampling time k: ; Digital twin module 102 will The input is an extended Kalman filter, and the estimated values ​​of SOC(k), polarization branch voltage, and terminal voltage are updated synchronously during state prediction and measurement updates.

[0046] In one implementation, the terminal voltage of the second-order RC equivalent circuit model can be expressed as: ; in, This is the open-circuit voltage corresponding to the battery's state of charge (SOC). For current, For ohmic internal resistance, and These are the voltages of the two polarization branches, respectively.

[0047] The polarization branch can be updated according to the following recursive formula: ; ; Among them, V i (k) represents the polarization voltage of the i-th polarization branch at the k-th sampling time, a i Let I(k) represent the discrete attenuation coefficient of the i-th polarization branch, and let R represent the current at the k-th sampling time.i Let τ represent the resistance of the i-th polarization branch. i Let represent the time constant of the i-th polarization branch, and Δt represent the sampling interval, i=1,2. The digital twin module 102 generates the voltage residual sequence: ; in, This represents the actual measured terminal voltage. The voltage residual sequence is also used for state estimation quality assessment, event triggering, and visual large language model decision-making.

[0048] The event triggering module 103 performs sliding window statistics on the voltage residual sequence. For a length of... L sliding window The event triggering module 103 can construct the residual feature vector: ; Among them, RMSE r The mean square root of the window residuals. r and std r These represent the mean and standard deviation of the window residuals, respectively. Corr represents the maximum absolute value of the window residual. rI This indicates the correlation between the residual and the current, hf ratio This represents the proportion of high-frequency energy in the residual; when the eigenvector contains skew r and kurt r At that time, skew r Kurt represents the residual skewness. r This indicates the residual kurtosis.

[0049] In one implementation, the root mean square baseline of the residuals is first calculated in the preheating window. and instantaneous residual extreme baseline Then calculate the threshold: ; ; in, (x, a, b) means restricting x to the interval [a, b], where x is... or , a is or , , respectively, are the corresponding lower limits of the threshold, and b is or , which are the corresponding upper limits of the threshold. and This is the sensitivity coefficient.

[0050] The event trigger flag can be represented as: ; in, This indicates whether a running condition anchor point appears within the window. If the value is 1, a battery management event is triggered. This represents the OR operation in logical operations; if If the value is zero, the digital twin module 102 will continue to run continuously without calling the visual large language model decision module 106.

[0051] Operating condition anchors can include static anchors, constant current anchors, constant voltage anchors, pulse anchors, and equivalent full-cycle anchors. Operating condition anchors do not determine the event type independently, but rather serve as contextual inputs to the visual large language model decision module 106 to assist in determining whether the current residual form corresponds to parameter drift, a real fault, a capacity calibration opportunity, or noise.

[0052] As shown in Figure 3(a), the image and abstract generation module 104 generates a multi-track panel plot after the event is triggered. The multi-track panel plot includes at least a voltage track, a current track, a SOC track, and a residual track, and all tracks are aligned on the time axis. The multi-track panel plot can be represented using polylines, region shading, event markers, and threshold lines, but does not depend on the experimental images in the paper. The multi-track panel plot can be represented as follows: ; in, （） Indicates to the window W The operation of rendering multi-source sequences within a unified timeline into an image panel. For voltage trajectory, For current trajectory, For the trajectory of the charged state, For residual loci; As shown in Figure 3(b), a structured numerical summary may include at least two of the following metrics: ; in, The root mean square of the residuals, The mean of the residuals, The standard deviation of the residuals. For residual skewness, For residual kurtosis, The maximum absolute value of the residual. The correlation between residual and current, This represents the proportion of high-frequency residual energy.

[0053] The enhanced knowledge base 105 is divided into namespaces based on tasks, such as fault, parameter optimization, capacity, and remaining lifetime. After an event is triggered, the system retrieves K historical cases from the corresponding namespace according to the current task or candidate tasks. ; ; ; ; ; in, φ represents a structured numerical summary of the event window W. r (W) represents the residual statistical characteristic vector within window W. This indicates the anchor point of the operating condition corresponding to window W. This represents the operating condition label corresponding to window W, and Normalize(·) represents vector normalization. This indicates the residual panel plot P W Visual features obtained through image encoding; DB param DB fault DB cap DB rul and DB noise These represent the parameter update library, fault case library, capacity calibration library, remaining lifetime trajectory library, and noise event library, respectively. c Let sim(·) represent the vectorized representation of historical case c, sim(·) represent the similarity function, K represent the number of recalled cases, and C K t Let C represent the set of K similar historical cases recalled from the namespace corresponding to task t, where T represents the candidate task set, and C represents the set of K similar historical cases recalled from the namespace corresponding to task t. K This represents the set of historical cases after merging across candidate tasks.

[0054] The visual large language model decision module 106 receives multi-trajectory panel plots, structured numerical summaries, runtime anchor points, and RAG historical cases, and outputs event routing results. This process can be represented as: ； ； Among them, F VLM (·) represents the multimodal inference function of the visual large language model, P W This represents the residual panel plot corresponding to event window W, S W Represents structured numerical summaries, C K p represents the set of similar historical cases retrieved, and G represents the preset routing constraints; tThe output represents the probability of candidate task t, route represents the event routing result, c represents the route confidence, m represents the parameter selection mask, u represents the structured reasoning, argmax represents the candidate task with the highest probability of selection, and T represents the set of candidate tasks. The event routing result includes parameter updates, faults, capacity anchors, remaining lifetime predictions, noise, etc.

[0055] When the event routing result is a parameter update, the visual large language model decision module 106 outputs a parameter selection mask. For a second-order RC equivalent circuit model, the parameter selection mask can be expressed as: ; in, , , , , Resistors R0, R1, and R2, respectively, and the fast time constant. and slow time constant The parameter selection mask, whose values ​​are all A value of 1 indicates that the corresponding parameter needs to be updated, while a value of 0 indicates that the corresponding parameter remains unchanged. Resistors R0, R1, and R2 represent the internal resistance in ohms, the fast polarization resistor, and the slow polarization resistor, respectively. Capacitor C1 is the capacitor connected in parallel with resistor R1, and capacitor C2 is the capacitor connected in parallel with resistor R2. The fast time constant... =R1C1, slow time constant =R2C2.

[0056] Execution module 107 performs restricted optimization on the selected parameters. An example objective function is: ; And satisfy the following constraints: ; ; Where θ represents the parameter vector of the second-order RC equivalent circuit model, θ* represents the optimized parameter vector, and ρ δ (·) denotes the robust Huber loss function, V meas (k) represents the measured terminal voltage at the k-th sampling time, V model (k;θ) represents the terminal voltage calculated by the model under parameter θ, where θ j Let θ represent the j-th parameter to be optimized. j,current This represents the current value of the j-th parameter. and These represent the lower and upper bounds set around the current parameter value, respectively, m jThis represents the value corresponding to the j-th parameter in the parameter selection mask. By using the parameter selection mask m, the execution module 107 will not blindly refit all parameters, thereby reducing the risk of parameter oscillation and fault absorption.

[0057] After the parameters are updated, the digital twin module 102 can expand the covariance of the extended Kalman filter to absorb the uncertainty caused by parameter jumps. For example, a larger expansion coefficient can be used for the polarization voltage dimension and a smaller expansion coefficient for the SOC dimension.

[0058] When the event routing result indicates a fault, execution module 107 enters the fault diagnosis branch. The fault diagnosis branch first constructs a fault feature vector: ; in, For the parameters of the equivalent circuit model, For statistical characteristics of operating conditions, The residual time-frequency characteristics include: the equivalent circuit model parameter fingerprint can include the current parameter value, parameter change amount, and parameter change rate; the operating condition statistical characteristics can include current, temperature, and voltage window statistics; and the residual time-frequency characteristics can include residual root mean square, residual kurtosis, residual-current correlation, and high-frequency energy proportion.

[0059] The retrieval enhancement knowledge base 105 retrieves similar fault cases based on fault feature vectors within its fault namespace. The visual large language model decision module 106 takes the current multi-track panel diagram, similar fault cases, and structured numerical summaries as input, and outputs the fault type, fault severity level, and handling recommendations. Fault types may include poor contact, internal short circuit, sensor drift, connection anomalies, and noise interference. Handling recommendations may include continued monitoring, derating operation, isolating battery cells, or triggering maintenance alarms.

[0060] The fault diagnosis branch, as one of the execution branches of this invention, shares the same event-triggered evidence chain as the parameter update branch. Therefore, when the residual abnormality more closely matches the fault pattern, the system can avoid incorrectly processing the abnormality as a parameter update.

[0061] As shown in Figure 4(a), when the event routing result is a capacity anchor or remaining lifetime prediction, execution module 107 enters the remaining lifetime prediction branch. This branch constructs a lifetime feature vector based on the battery health state curve: ； in, This represents the curve compression feature obtained by segmenting and aggregating the battery health state curve. Indicates the capacity decay slope, Indicates the trend of impedance change. Indicates the inflection point position. Indicates the sharpness of the inflection point.

[0062] The remaining lifetime namespace of the enhanced knowledge base 105 is retrieved based on lifetime feature vectors to find similar degradation trajectories, as shown in Figure 4(b). The image and summary generation module 104 can render the current SOH curve and similar degradation trajectories into a trajectory comparison map, and as shown in Figure 4(b). Figure 5 The remaining lifetime prediction result is shown. The visual large language model decision module 106 outputs the remaining lifetime prediction result based on the current SOH curve, similar degradation trajectory, and structured numerical summary. The remaining lifetime prediction result can be the remaining number of cycles, the remaining running time, or the expected time to reach a preset capacity threshold.

[0063] The role of this branch is to provide a comparable lifetime reference for the current battery by utilizing historical degradation trajectories, without requiring the visual large language model to infer lifetime point by point from the raw high-frequency sampled data.

[0064] like Figure 1 As shown, the enhanced retrieval knowledge base 105 includes multiple namespaces. These namespaces can include various categories such as fault parameter updates, fault cases, capacity calibration, remaining lifetime trajectory, and noise events.

[0065] After each event is processed, the write-back module 108 writes the event characteristics, event routing results, and execution results to the corresponding namespace. If the nearest neighbor distance between the newly written case and an existing case meets the merging condition, an exponential moving average can be used for merging. ; in, For the new case vector, For existing case vectors, This is the merged case vector. This determines the weighting for merging. If the merging criteria are not met, the new case will be written as a separate case.

[0066] Within the fault namespace, constraints can be set on the ratio of normal samples to faulty samples to prevent an excessive number of normal samples from diluting fault cases. Through the aforementioned partitioning and write-back mechanisms, the system can continuously accumulate parameter update cases, fault cases, capacity calibration cases, and lifetime prediction cases during operation.

[0067] In some cases, even multiple consecutive parameter updates fail to reduce the residual, or the decision confidence of the visual large language model is low. In these situations, gray area detection can be triggered. Gray area detection integrates the visual large language model confidence, the difference in probabilities between the first two types of routes, recent calibration gains, out-of-distribution distance, and the degree of numerical audit conflict to obtain a gray area score. ; in, For the confidence of the visual large language model, The difference between the probabilities of the first two types of routes. For recent calibration benefits, For the distance outside the distribution, For the degree of conflict in numerical audits, to As weight.

[0068] When the gray area score exceeds a threshold, execution module 107 can perform an A / B test among candidate structural models. Candidate structural models can include equivalent circuit models with bias compensation and equivalent circuit models with hysteresis compensation. If a candidate structure reduces residuals during the observation period without triggering rollback conditions, the candidate structure is adopted; otherwise, it is rolled back to the original structure.

[0069] In one embodiment, the system described in this invention can be deployed as a web-based visual operation and maintenance terminal. This terminal receives the outputs of the aforementioned modules through a communication interface, and organizes and displays digital twin estimation, residual panel evidence, retrieval enhancement cases, visual large language model inference, and execution results using the same event as an index.

[0070] The maintenance terminal may include a battery pack summary area, a single-unit card area, a single-unit details area, and a fault event pipeline area, which are used for global status summary, triggering single-unit location, time-series evidence display, and single event full-process tracing, respectively. Those skilled in the art can choose deployment forms such as browsers, industrial tablets, vehicle-side diagnostic screens, cloud consoles, or edge server consoles according to the application scenario, and design specific page layouts, operation controls, and interaction methods; the above specific implementations do not constitute a limitation on the scope of protection of this invention.

[0071] The above embodiments illustrate that the present invention can complete event triggering, visual large language model routing, selective parameter updating, fault diagnosis, remaining lifetime prediction, online knowledge write-back, and operation and maintenance visualization within a unified closed loop. Fault diagnosis and remaining lifetime prediction are not separate functions independent of the main invention, but rather two execution branches under the same event routing result; the operation and maintenance visualization terminal is a deployment form of the above closed loop in vehicles, power plants, energy storage systems, or experimental platforms.

[0072] In summary, this invention renders the voltage, current, state of charge, and residuals within the event window as a multi-trajectory panel plot, using image modalities to compensate for the difficulty in fully aligning long-time-series data within the context of a large model; it enhances the knowledge base through task-partitioned retrieval, allowing the model to refer to similar historical cases when making judgments; and it continuously accumulates new experiences generated during operation into reusable online knowledge through a case write-back mechanism, thereby achieving collaborative online updates.

[0073] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. CD - ROM It takes the form of a computer program product implemented on (such as optical memory, etc.).

[0074] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0075] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0076] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0077] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A lithium-ion battery management method based on a visual large language model, characterized in that, Includes the following steps: Obtain battery operation sampling data, and run a pre-built battery physical model based on the operation sampling data; Based on the battery operation sampling data and battery physical model, state estimation is performed to obtain the state of charge, terminal voltage estimation and voltage residual sequence; Based on the residual window statistics, instantaneous residual extreme values ​​and operating condition anchor points, determine whether a battery management event has been triggered, and continue state estimation if the event has not been triggered. When an event is triggered, a residual panel plot containing voltage, current, state of charge and residuals is generated to form a structured numerical summary. Based on the current event, construct a query vector and recall similar historical cases from a pre-built task-partitioned enhanced knowledge base; The residual panel plot, structured numerical summary, operating condition anchor points, and historical cases are used as multimodal inputs, and a visual big language model is used to determine the task route. Based on the task route, enter the corresponding execution branch to perform selective parameter updates, fault diagnosis, capacity calibration, remaining lifetime prediction, or noise ignoring.

2. The lithium-ion battery management method based on a visual large language model as described in claim 1, characterized in that, The process of running the pre-built battery physics model based on the aforementioned sampling data includes: according to the sampling time... k Construct the running sampling vector: ； in, , and They represent the first k The terminal voltage, current, and temperature at each sampling time will The extended Kalman filter is input, and during the state prediction update process, the battery state of charge, polarization branch voltage, and terminal voltage estimates are updated synchronously using the battery physical model. Based on the terminal voltage estimates and terminal voltage, a voltage residual sequence is generated.

3. The lithium-ion battery management method based on a visual large language model as described in claim 1, characterized in that, The process of determining whether a battery management event is triggered based on residual window statistics, instantaneous residual extreme values, and operating condition anchor points includes: for a length of L sliding window Construct the residual eigenvectors: ； Calculate the root mean square baseline of the residuals and the extreme value baseline of the instantaneous residuals within the preheating window, and then calculate the corresponding thresholds: ； ； in, (x, a, b) means restricting x to the interval [a, b], where x is... or , a is or , , respectively, are the corresponding lower limits of the threshold, and b is or , which are the corresponding upper limits of the threshold. and This is the sensitivity coefficient. The root mean square baseline of the residuals, The baseline for the instantaneous residual extreme value; The event trigger flag is represented as: ； in, This indicates whether a running condition anchor point appears within the window. If the value is 1, a battery management event is triggered. This represents the OR operation in logical operations; if If the value is zero, then the continuous state estimate is maintained.

4. The lithium-ion battery management method based on a visual large language model as described in claim 1, characterized in that, The operating condition anchor points include at least one of the following: static anchor points, constant current anchor points, constant voltage anchor points, pulse anchor points, and equivalent full-cycle anchor points. These serve as context inputs to help determine whether the current residual form corresponds to parameter drift, a real fault, a capacity calibration opportunity, or noise.

5. The lithium-ion battery management method based on a visual large language model as described in claim 1, characterized in that, When an event is triggered, the process of generating a residual panel plot containing voltage, current, state of charge, and residuals across multiple trajectories to form a structured numerical summary includes: The multi-trajectory panel plot is represented as follows: ； in, （） Indicates to the window W The operation of rendering multi-source sequences within a unified timeline into an image panel. For voltage trajectory, For current trajectory, For the trajectory of the charged state, For residual loci; Structured numerical summaries include at least two of the following metrics: ； in, The root mean square of the residuals, The mean of the residuals, The standard deviation of the residuals. For residual skewness, For residual kurtosis, The maximum absolute value of the residual. The correlation between residual and current, This represents the proportion of high-frequency residual energy.

6. The lithium-ion battery management method based on a visual large language model as described in claim 1, characterized in that, The process of constructing a query vector based on the current event and recalling similar historical cases from a pre-built task-partitioned enhanced retrieval knowledge base includes: the pre-built task-partitioned enhanced retrieval knowledge base is divided into multiple namespaces according to tasks, such as parameter update library, fault case library, capacity calibration library, remaining lifetime trajectory library, and noise event library. After an event is triggered, K historical cases are recalled from the corresponding namespace according to the current task or candidate task. ； ； ； ； ； Among them, S W φ represents a structured numerical summary of the event window W. r (W) represents the residual statistical characteristic vector within window W. This indicates the anchor point of the operating condition corresponding to window W. This represents the operating condition label corresponding to window W, Normalize(·) represents vector normalization, and E img (P W ) indicates the residual panel plot P W Visual features obtained through image encoding; DB param DB fault DB cap DB rul and DB noise These represent the parameter update library, fault case library, capacity calibration library, remaining lifetime trajectory library, and noise event library, respectively. c Let sim(·) represent the vectorized representation of historical case c, sim(·) represent the similarity function, K represent the number of recalled cases, and C K t Let C represent the set of K similar historical cases recalled from the namespace corresponding to task t, where T represents the candidate task set, and C represents the set of K similar historical cases recalled from the namespace corresponding to task t. K This represents the set of historical cases after merging across candidate tasks.

7. The lithium-ion battery management method based on a visual large language model as described in claim 1, characterized in that, The process of determining task routing using the residual panel plot, structured numerical summary, operational condition anchor points, and historical cases as multimodal inputs and a visual big language model includes: obtaining the event task routing result using the visual big language model based on the multi-trajectory panel plot, structured numerical summary, operational condition anchor points, and historical cases, as expressed as: ； ； Among them, F VLM (·) represents the multimodal inference function of the visual large language model, P W This represents the residual panel plot corresponding to event window W, S W Represents structured numerical summaries, C K p represents the set of similar historical cases retrieved, and G represents the preset routing constraints; t Let represent the probability of candidate task t, route represent the event routing result, c represent the route confidence, m represent the parameter selection mask, u represent the structured reason explanation, argmax represent the candidate task with the highest selection probability, and T represent the set of candidate tasks; the event routing result includes several of the following: parameter update, fault, capacity anchor, remaining lifetime prediction, noise, and unknowns.

8. The lithium-ion battery management method based on a visual large language model as described in claim 1, characterized in that, The process of entering the corresponding execution branch based on the task route and performing fault diagnosis includes: constructing a fault feature vector. ； in, For the parameters of the equivalent circuit model, For statistical characteristics of operating conditions, The residual time-frequency characteristics include: the equivalent circuit model parameter fingerprint includes the current parameter value, parameter change amount, and parameter change rate; the operating condition statistical characteristics include current, temperature, and voltage window statistics; the residual time-frequency characteristics include residual root mean square, residual kurtosis, residual-current correlation, and high-frequency energy ratio. The fault namespace of the enhanced knowledge base is retrieved to find similar fault cases based on fault feature vectors. The visual big language model takes the current multi-trajectory panel map, similar fault cases and structured numerical summary as input to obtain fault type, fault severity level and handling suggestions. Alternatively, the process of entering the corresponding execution branch based on the task routing and performing the remaining lifetime prediction includes: constructing a lifetime feature vector based on the battery health state curve. ； in, This represents the curve compression feature obtained by segmenting and aggregating the battery health state curve. Indicates the capacity decay slope, Indicates the trend of impedance change. Indicates the inflection point position. Indicates the sharpness of the inflection point; Similar degradation trajectories are retrieved based on the lifetime feature vector, and the current battery health state curve and similar degradation trajectories are rendered into a trajectory comparison map. The battery health state outputs the remaining lifetime prediction result based on the current battery health state curve, similar degradation trajectories and structured numerical summary. The remaining lifetime prediction result is the remaining number of cycles, the remaining running time or the expected time to reach the preset capacity threshold.

9. A lithium-ion battery management method based on a visual large language model as described in claim 1, characterized in that, It also includes the following steps: The event features, the output of the visual large language model, and the execution results are written back to the corresponding task namespace to form an online learning loop. If the nearest neighbor distance between the newly written case and the existing case meets the merging condition, an exponential moving average is used for merging. ； in, For the new case vector, For existing case vectors, This is the merged case vector. To consolidate weights, if the consolidation criteria are not met, the new case will be written as an independent case. In the fault namespace, set a ratio constraint between normal samples and faulty samples to prevent the number of normal samples from exceeding the set value.

10. A lithium-ion battery management system based on a visual large language model, characterized in that, include: The sampling data acquisition module is configured to acquire battery operation sampling data and run a pre-built battery physical model based on the operation sampling data; The digital twin module is configured to perform state estimation based on the battery operation sampling data and the battery physical model, and obtain the state of charge, terminal voltage estimation and voltage residual sequence. The event triggering module is configured to determine whether to trigger a battery management event based on residual window statistics, instantaneous residual extreme values, and operating condition anchor points. If no event is triggered, state estimation continues. The image and summary generation module is configured to generate a residual panel plot containing voltage, current, state of charge and residuals of multiple trajectories when an event is triggered, forming a structured numerical summary. The retrieval enhancement knowledge base is configured to be pre-built, partitioned by task, constructing query vectors based on the current event, and recalling similar historical cases. The visual big language model decision module is configured to use the residual panel plot, structured numerical summary, operating condition anchor point and historical cases as multimodal inputs to determine the task route using the visual big language model. The execution module is configured to enter the corresponding execution branch according to the task route and perform selective parameter updates, fault diagnosis, capacity calibration, remaining lifetime prediction, or noise ignoring.

Images