Battery full life cycle intelligent management method and system based on large model

Abstract

The invention discloses a battery full life cycle intelligent management method and system based on a large model, and the method comprises the following steps: S1, collecting the operation data of a battery, and carrying out the preprocessing; s2, inputting the pre-processed sample into a pre-trained Transform encoder model, and extracting a time sequence and cross-stage characteristics; s3, executing model reasoning, outputting a state index, and generating a battery state vector; s4, identifying an abnormal category and a position in combination with the working condition information; s5, generating a dynamically optimized battery management strategy based on the battery state index and the abnormity identification result; s6, deploying a lightweight model at the edge device, executing local reasoning and uploading data; s7, the cloud updates the model through self-supervised training and issues the model; and S8, repeatedly executing the steps S1 to S7, and carrying out optimized closed-loop management. According to the invention, large model modeling and an edge cloud cooperation mechanism are fused, and intelligent sensing and dynamic management of the whole life cycle of the battery are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of battery intelligent management, and in particular to a method and system for intelligent management of the entire life cycle of a battery based on a large model. Background Art

[0002] In the existing technology, the Battery Management System (BMS) usually adopts rule-based or lightweight model-based strategies for state estimation and operation control, and its functions are mainly concentrated in aspects such as state of charge (SOC), state of health (SOH) and fault monitoring. Most of these systems use traditional methods such as Kalman filtering, support vector machines, fuzzy logic or shallow neural networks to realize data processing and state judgment. However, with the continuous expansion of battery application scenarios, especially in the fields of new energy vehicles, power storage and consumer electronics, the battery operating environment has shown characteristics such as strong nonlinearity, time-varying and multi-operating condition crossing. The traditional BMS method has exposed obvious performance bottlenecks in dealing with complex operating states, abnormal pattern recognition and strategy optimization.

[0003] On the one hand, existing methods generally rely on model structures and discrimination thresholds set by expert experience, which makes it difficult to accurately reflect the state changes of batteries at different stages of their life cycle (including manufacturing, transportation, use, storage and retirement), especially when faced with cross-scenario and cross-system deployments, and lack good generalization capabilities. On the other hand, most current BMS lack a systematic accumulation and dynamic learning mechanism for operating data. Model training is often based on static historical data, and cannot achieve adaptive updates of strategies as the battery operating environment changes, resulting in reduced state estimation accuracy and delayed policy control. In addition, most existing methods are deployed in a single center or local end, lacking edge computing capabilities and cloud-based collaborative mechanisms, making it difficult to achieve high-frequency data processing and large-scale model iterative updates, and difficult to ensure the system's response efficiency and long-term evolution capabilities.

[0004] The present invention proposes a method for intelligent management of the entire life cycle of batteries based on a large model, which has made systematic improvements to the above-mentioned problems. The method adopts a Transformer encoder neural network with strong generalization ability, and through deep modeling of the operating data collected by the battery at different stages of the entire life cycle, effectively mines the time-dependent structure and cross-stage semantic features, and realizes high-precision prediction of status indicators such as SOC, SOH and RUL, and supports timely identification of abnormal conditions such as thermal runaway, capacity drop, and voltage drift. At the same time, the method constructs an edge deployment architecture to achieve lightweight model reasoning and local real-time response, and continuously optimizes model parameters in the cloud through a self-supervised training mechanism to ensure that the system can achieve distributed collaborative updates under the premise of data security, thereby forming a closed-loop management capability for the entire battery operation process, significantly improving the adaptability and intelligence level of existing technologies in status assessment, abnormality perception and policy decision-making.

[0005] Therefore, how to provide a battery full life cycle intelligent management method and system based on a large model is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0006] One purpose of the present invention is to propose a battery full life cycle intelligent management method and system based on a large model. The present invention integrates Transformer large model modeling, edge lightweight deployment and cloud federated update mechanism to achieve high-precision prediction of battery operating status, intelligent identification of anomalies and dynamic optimization of strategies, and build an intelligent management closed loop covering the entire process from manufacturing to decommissioning, thereby improving the system's real-time, adaptability and life cycle evolution capabilities, and breaking through the technical bottlenecks of traditional BMS in cross-stage modeling and continuous optimization.

[0007] A battery life cycle intelligent management method based on a large model according to an embodiment of the present invention includes the following steps:

[0008] S1. Collecting battery operation data during manufacturing, transportation, use, storage, and retirement. The operation data includes voltage, current, temperature, internal resistance, capacity, charge and discharge cycles, and related maintenance records. The operation data is uniformly preprocessed, including missing value filling, outlier removal, time series alignment, multimodal normalization, and feature vector construction.

[0009] S2. Constructing input samples based on the preprocessed operational data and inputting the input samples into a pre-trained Transformer encoder neural network model. The Transformer encoder model includes a multi-layer multi-head self-attention mechanism, a feedforward neural network, and a residual connection structure to model the temporal dependencies and cross-stage features in the operational data.

[0010] S3. Execute the forward reasoning process of the Transformer model, output the corresponding battery status indicators, including state of charge (SOC), state of health (SOH), and remaining useful life (RUL), and construct a battery state vector.

[0011] S4. Perform abnormal pattern recognition based on the battery state vector and current operating condition information to determine whether there is thermal runaway, capacity drop, voltage drift, or abnormal cycle performance. If an abnormality is identified, record the abnormality type and corresponding battery cell location;

[0012] S5. Based on the battery status indicators and the abnormality identification results, generate a dynamically optimized battery management strategy, which includes charge and discharge current regulation, thermal management parameter configuration, battery module switching control, and operation mode adjustment;

[0013] S6. Deploy a lightweight version of the Transformer encoder model on the edge device, using edge computing power for local inference and real-time response. At the same time, the collected new operation data and model inference results are uploaded to the cloud server via encrypted communication.

[0014] S7. After receiving the uploaded data, the cloud server performs incremental updates on the Transformer encoder model parameters based on the self-supervised training method, and sends the updated model version to the edge device for deployment and replacement, provided that the system security policy is met.

[0015] S8. Repeat steps S1 to S7 to optimize closed-loop management of the battery's continuous state perception, abnormality prediction, and intelligent strategy throughout its entire life cycle.

[0016] Optionally, the S1 specifically includes:

[0017] S11. During the battery manufacturing, transportation, use, storage, and retirement stages, use embedded sensors or remote monitoring devices to collect battery operating data, including voltage data, current data, temperature data, internal resistance value, capacity value, number of charges, number of discharges, charge and discharge duration, number of cycles, operating timestamps, and related maintenance logs;

[0018] S12. Perform missing value filling processing on the collected operating data, wherein the missing data is filled by sliding window interpolation or fixed value replacement to ensure the integrity of the time series;

[0019] S13. Perform outlier removal processing on the operating data, identifying and removing data points that exceed the normal operating range by setting a boundary threshold range to eliminate distortion caused by hardware drift or sudden interference;

[0020] S14. Perform time series alignment on the processed operating data, synchronize the timestamps of different data sources to a unified time axis, and ensure that multi-parameter data are aligned and matched according to the acquisition time;

[0021] S15. Normalize various types of operating data to convert feature values ​​of different dimensions into a unified numerical range to reduce scale differences between features;

[0022] S16. Construct a feature vector sample based on the normalized data. The feature vector is intercepted according to a fixed window length and encapsulated as structured input data for subsequent modeling processing.

[0023] Optionally, the S2 specifically includes:

[0024] S21, organizing the constructed feature vector samples into a multidimensional input sequence in chronological order, wherein the input sequence is divided into fixed time windows to form input samples of multiple continuous time periods;

[0025] S22. Perform position encoding on each input sample and introduce temporal position information in the feature dimension to preserve the time series characteristics of the running data.

[0026] S23, inputting the input sample with added position information into a Transformer encoder neural network model, wherein the model includes multiple encoder layers, each encoder layer includes a multi-head self-attention mechanism, a feedforward neural network structure, a residual connection and a normalization module;

[0027] S24. Inside the Transformer encoder model, multiple layers of processing are performed on the input samples in sequence. Each layer completes self-attention weight calculation, context feature extraction and nonlinear transformation, and outputs deep feature representation layer by layer.

[0028] S25. Extract the global feature output corresponding to each input sample to form a unified encoding result, and pass it to the state reasoning module for calculation of the battery state indicator.

[0029] Optionally, the S3 specifically includes:

[0030] S31, receiving the deep feature representation output by the Transformer encoder model in step S2 as input for battery state inference;

[0031] S32. Based on the deep feature representation, construct representation channels for the state of charge (SOC), state of health (SOH), and remaining useful life (RUL), respectively. Each channel consists of several fully connected layers to generate corresponding state prediction results.

[0032] S33, performing vector splicing on the prediction results output by each channel to form a complete battery state vector, wherein the battery state vector includes estimated information on the current battery performance, lifespan, and operating status;

[0033] S34: Output and cache the battery state vector for input into the subsequent abnormality identification module and strategy optimization module.

[0034] Optionally, the S4 specifically includes:

[0035] S41, receiving the battery state vector output in step S3 and current operating condition parameters, wherein the operating condition parameters include ambient temperature, charge and discharge rate, battery pack load state, and historical operating curve;

[0036] S42. Perform joint feature fusion on the battery state vector and the operating condition parameter to construct an abnormality recognition input sample, wherein the fusion method includes feature splicing and dimension matching processing;

[0037] S43. Input the abnormal identification input sample to the abnormality detection module, and the abnormality detection module performs matching judgment based on the predefined abnormal pattern to determine whether there is thermal runaway, capacity drop, voltage drift or abnormal cycle performance behavior;

[0038] S44. When an abnormality is detected, mark the abnormality category, record the corresponding abnormality occurrence time, abnormality duration, and the battery module or cell location to which the abnormality belongs;

[0039] S45. Output the abnormality identification result and pass it to the strategy optimization module for adjusting the battery operation control strategy.

[0040] Optionally, the S5 specifically includes:

[0041] S51, receiving the battery state vector output in step S3 and the abnormality identification result generated in step S4 as input basis for strategy generation;

[0042] S52. Construct a strategy decision input vector based on the state of charge, health status, and remaining service life in the battery status indicators, combined with the current abnormality category and operating condition characteristics;

[0043] S53. Based on the policy decision input vector, calling a preset policy optimization rule set or policy generation model to generate a dynamic management policy for the current operating state;

[0044] S54. Setting charging current limit parameters, discharging current adjustment parameters, thermal management start threshold and adjustment gear, battery module start and stop sequence and switching conditions, and operating mode adjustment range in the management strategy;

[0045] S55: Output the generated management strategy to the control interface module to guide the battery management system to adjust the operating parameters in real time, and record the strategy execution log for subsequent optimization iterations.

[0046] Optionally, the S6 specifically includes:

[0047] S61. Lightweight the Transformer encoder neural network model and generate a model version suitable for edge devices using parameter pruning, channel compression, or knowledge distillation.

[0048] S62. Deploy a lightweight model version on the edge device, configure the local inference engine, load the model weights and input interface module, and ensure that state reasoning and anomaly identification tasks can be completed independently on the edge side;

[0049] S63: Collect new battery operation data in real time, and form new input samples after completing missing value filling, data alignment and normalization processing through the local preprocessing module;

[0050] S64: Input the newly generated input sample into the edge model, perform local reasoning, and output the updated battery status indicator and anomaly recognition result;

[0051] S65: Packaging the local reasoning results of each round with the corresponding running data to form an upload data packet with a timestamp;

[0052] S66. Send the uploaded data packet to the cloud server via an encrypted communication protocol, wherein the encrypted communication protocol includes a data signature, encrypted transmission, and integrity verification mechanism.

[0053] Optionally, the S7 specifically includes:

[0054] S71. The cloud server receives data packets uploaded from multiple edge devices, including edge-collected operation data, model inference output results, acquisition timestamps, edge device identifiers, and system operation logs, and performs parsing, verification, and archiving operations on the data packets to construct a structured training sample set.

[0055] S72. Divide the training sample set by index based on device source, and extract data with clear labels from each group as supervised training samples, and extract data with missing labels or weak labels as self-supervised training samples to form a mixed data set;

[0056] S73. Build an edge collaborative update framework, determine the set of edge nodes participating in this round of training, download their corresponding local model parameter snapshots, and use a secure multi-party computing protocol to perform weighted aggregation on each edge model parameter in the cloud to obtain candidate global model parameters.

[0057] S74. Introduce a self-supervised learning mechanism to perform auxiliary training tasks on the candidate model, wherein the training tasks include sequence reconstruction based on temporal masking, contrast discrimination based on feature perturbation, or fragment reasoning based on context prediction, so as to enhance the model's semantic extraction ability and structural generalization ability on unlabeled samples;

[0058] S75. Perform integrity verification and performance evaluation on the model optimized through self-supervised training. Evaluation metrics include state prediction accuracy, anomaly detection accuracy, model inference latency, and communication overhead on a standard validation set. If all metrics meet the set thresholds, the model update is considered successful.

[0059] S76. The updated model that has passed verification is version numbered, digitally signed, and encrypted, and sent to the corresponding edge device through a secure channel based on the identity authentication protocol. After receiving the updated model, the edge device executes a hot replacement or grayscale loading strategy to achieve online update and uninterrupted deployment of model parameters.

[0060] Optionally, the S8 specifically includes:

[0061] S81. After the edge device completes the model update deployment, it automatically reinitializes the local operation status monitoring process, reloads the latest version of the Transformer encoder model, and configures its inference interface and data flow channel to ensure that the inference service is synchronized with the local control strategy module;

[0062] S82. During system operation, continuously repeat steps S1 to S7 to periodically or event-drivenly collect operational data, perform status assessments, identify anomalies, generate control strategies, upload cloud training data, and accept model updates to achieve dynamic model iteration and adaptive strategy adjustment.

[0063] S83. Log and analyze the policy feedback for each round of closed-loop execution. The log content includes the state prediction value, policy execution result, system load information, fault frequency, model version number, and edge device operation status.

[0064] S84. Build a closed-loop data management module for the entire lifecycle, uploading operational data, inference results, exception records, and strategic behaviors generated at each stage to a cloud database to form a data accumulation system that supports continuous optimization.

[0065] S85. Set up model evolution tracking and lifecycle mapping logic in the cloud server to archive and classify battery operation characteristics in different periods and industry scenarios as a pre-selected sample pool source for subsequent model training;

[0066] S86. The key performance indicators and model performance results of the closed-loop management process are synchronously displayed on the operation and maintenance platform interface for remote review and manual intervention by engineering and technical personnel, forming a full-life cycle intelligent decision support system driven by large models.

[0067] According to an embodiment of the present invention, a large-scale model-based intelligent management system for the entire battery life cycle includes the following modules:

[0068] A data acquisition and preprocessing module is used to collect operational data during the battery manufacturing, transportation, use, storage, and retirement stages, and perform missing value filling, outlier removal, time series alignment, multimodal normalization, and feature vector construction on the data;

[0069] The Transformer state modeling module is used to input preprocessed input samples into the pre-trained Transformer encoder neural network model, perform multi-layer self-attention mechanism and feedforward neural calculation, output battery state indicators and construct a battery state vector;

[0070] The anomaly detection module is used to determine whether there is thermal runaway, capacity drop, voltage drift or cycle performance abnormality based on the battery state vector and the current operating conditions, and record the abnormality type and battery location;

[0071] A strategy generation module, which combines battery status indicators with anomaly detection results to generate battery management strategies for charge and discharge regulation, thermal management configuration, battery module switching, and dynamic adjustment of operating modes;

[0072] Edge inference and communication module, which is used to deploy lightweight model versions on edge devices to achieve local inference and real-time response, while uploading running data and inference results to cloud servers through encrypted communication;

[0073] The model update module is used to aggregate and optimize model parameters based on self-supervised training methods on the cloud server, and after passing performance evaluation, the updated model is securely sent to the edge device for deployment and replacement;

[0074] The closed-loop management module is used to continuously execute state perception, inference calculation, anomaly identification, strategy optimization and model update processes in the system, realizing continuous intelligent management and decision support for the entire battery operation process.

[0075] The beneficial effects of the present invention are:

[0076] This paper introduces a large-scale neural network model based on the Transformer structure to perform deep semantic modeling and state feature extraction on battery operation data, significantly improving the accuracy and stability of battery state recognition and life prediction.

[0077] First, compared with the shallow models and rule-driven methods commonly used in existing BMS systems, the present invention can construct long-distance dependent feature expressions based on multidimensional time series data, thereby achieving high-precision reasoning of key indicators such as state of charge (SOC), state of health (SOH) and remaining life (RUL), which is suitable for intelligent perception needs under complex working conditions and dynamic environments.

[0078] Secondly, the present invention introduces an edge-cloud collaboration mechanism into the system architecture design. This allows for local reasoning and rapid response by deploying lightweight model versions on edge devices, while utilizing cloud servers to perform self-supervised training, enabling periodic model optimization and version updates. This architecture not only ensures low latency and high availability at runtime, but also overcomes the limitations of traditional methods that cannot balance real-time performance and continuous learning capabilities, effectively supporting the lifecycle evolution of the model and its adaptation to multiple scenarios.

[0079] In addition, the present invention also constructs a complete closed-loop data management and intelligent decision-making process, organically integrating state recognition, anomaly detection, strategy generation, model update and operation and maintenance feedback into a unified management system. By continuously collecting and dynamically learning key data during the battery operation process, the system can realize adaptive adjustment of strategies and real-time optimization of operating behaviors, and has the self-evolutionary characteristics of "the more you use it, the more accurate it becomes, and the more you use it, the more you understand it." Overall, the present invention not only significantly improves the intelligence level and operating efficiency of the battery management system, but also provides a technical foundation for building a safe, efficient and sustainable battery life cycle management platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0081] Figure 1 This is an overall flow chart of a large-scale model-based intelligent management method for the entire battery life cycle proposed by the present invention;

[0082] Figure 2 This is a schematic diagram of the structure of a battery full life cycle intelligent management system based on a large model proposed by the present invention; DETAILED DESCRIPTION

[0083] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0084] refer to Figure 1 , a battery life cycle intelligent management method based on a large model, including the following steps:

[0085] S1. Collecting battery operation data during manufacturing, transportation, use, storage, and retirement. The operation data includes voltage, current, temperature, internal resistance, capacity, charge and discharge cycles, and related maintenance records. The operation data is uniformly preprocessed, including missing value filling, outlier removal, time series alignment, multimodal normalization, and feature vector construction.

[0086] S2. Constructing input samples based on the preprocessed operational data and inputting the input samples into a pre-trained Transformer encoder neural network model. The Transformer encoder model includes a multi-layer multi-head self-attention mechanism, a feedforward neural network, and a residual connection structure to model the temporal dependencies and cross-stage features in the operational data.

[0087] S3. Execute the forward reasoning process of the Transformer model, output the corresponding battery status indicators, including state of charge (SOC), state of health (SOH), and remaining useful life (RUL), and construct a battery state vector.

[0088] S4. Perform abnormal pattern recognition based on the battery state vector and current operating condition information to determine whether there is thermal runaway, capacity drop, voltage drift, or abnormal cycle performance. If an abnormality is identified, record the abnormality type and corresponding battery cell location;

[0089] S5. Based on the battery status indicators and the abnormality identification results, generate a dynamically optimized battery management strategy, which includes charge and discharge current regulation, thermal management parameter configuration, battery module switching control, and operation mode adjustment;

[0090] S6. Deploy a lightweight version of the Transformer encoder model on the edge device, using edge computing power for local inference and real-time response. At the same time, the collected new operation data and model inference results are uploaded to the cloud server via encrypted communication.

[0091] S7. After receiving the uploaded data, the cloud server performs incremental updates on the Transformer encoder model parameters based on the self-supervised training method, and sends the updated model version to the edge device for deployment and replacement, provided that the system security policy is met.

[0092] S8. Repeat steps S1 to S7 to optimize closed-loop management of the battery's continuous state perception, abnormality prediction, and intelligent strategy throughout its entire life cycle.

[0093] This invention builds a neural network inference mechanism centered around a Transformer encoder, leveraging real-time edge sensing to achieve a highly integrated management process for battery state identification, anomaly diagnosis, and dynamic strategy optimization throughout the entire battery manufacturing and retirement process. Compared to existing BMS systems driven by static thresholds or rule-based logic, this approach offers dynamic adaptability to time-varying operating conditions and across lifecycle phases. It can continuously evolve to improve state estimation accuracy, response control precision, and system intelligence, significantly enhancing the safety, stability, and cost-effectiveness of battery systems under large-scale deployment and long-term operation.

[0094] In this embodiment, S1 specifically includes:

[0095] S11. During the battery manufacturing, transportation, use, storage, and retirement stages, use embedded sensors or remote monitoring devices to collect battery operating data, including voltage data, current data, temperature data, internal resistance value, capacity value, number of charges, number of discharges, charge and discharge duration, number of cycles, operating timestamps, and related maintenance logs;

[0096] S12. Perform missing value filling processing on the collected operating data, wherein the missing data is filled by sliding window interpolation or fixed value replacement to ensure the integrity of the time series;

[0097] S13. Perform outlier removal processing on the operating data, identifying and removing data points that exceed the normal operating range by setting a boundary threshold range to eliminate distortion caused by hardware drift or sudden interference;

[0098] S14. Perform time series alignment on the processed operating data, synchronize the timestamps of different data sources to a unified time axis, and ensure that multi-parameter data are aligned and matched according to the acquisition time;

[0099] S15. Normalize various types of operating data to convert feature values ​​of different dimensions into a unified numerical range to reduce scale differences between features;

[0100] S16. Construct a feature vector sample based on the normalized data. The feature vector is intercepted according to a fixed window length and encapsulated as structured input data for subsequent modeling processing.

[0101] Through the unified collection and hierarchical cleaning and merging strategy of multi-stage and multi-type sensor data, combined with sliding window interpolation, anomaly threshold removal, cross-source timestamp alignment and scale normalization processing, this method constructs a unified input tensor system suitable for Transformer modeling requirements, significantly improving the integrity and temporal consistency of operating data, laying a high-quality input foundation for cross-cycle health modeling and semantic extraction, and enhancing the system's robust processing capabilities against complex environmental fluctuations, sensor accuracy degradation and abnormal data interference.

[0102] In this embodiment, S2 specifically includes:

[0103] S21, organizing the constructed feature vector samples into a multidimensional input sequence in chronological order, wherein the input sequence is divided into fixed time windows to form input samples of multiple continuous time periods;

[0104] S22. Perform position encoding on each input sample and introduce temporal position information in the feature dimension to preserve the time series characteristics of the running data.

[0105] S23, inputting the input sample with added position information into a Transformer encoder neural network model, wherein the model includes multiple encoder layers, each encoder layer includes a multi-head self-attention mechanism, a feedforward neural network structure, a residual connection and a normalization module;

[0106] S24. Inside the Transformer encoder model, multiple layers of processing are performed on the input samples in sequence. Each layer completes self-attention weight calculation, context feature extraction and nonlinear transformation, and outputs deep feature representation layer by layer.

[0107] S25. Extract the global feature output corresponding to each input sample to form a unified encoding result, and pass it to the state reasoning module for calculation of the battery state indicator.

[0108] This method significantly enhances the Transformer encoder's ability to model temporal evolution structures by introducing positional encoding and a unified temporal window mechanism during the input modeling phase. It also achieves unified embedding and distribution mapping of lifecycle samples within the model structure. Its combination of a multi-head attention mechanism and a residual feedforward architecture effectively overcomes the inadequate modeling of long-term, nonlinear degradation patterns by shallow networks, improving generalization accuracy and convergence efficiency on heterogeneous battery samples and providing core technical support for unified health state modeling.

[0109] In this embodiment, S3 specifically includes:

[0110] S31, receiving the deep feature representation output by the Transformer encoder model in step S2 as input for battery state inference;

[0111] S32. Based on the deep feature representation, construct representation channels for the state of charge (SOC), state of health (SOH), and remaining useful life (RUL), respectively. Each channel consists of several fully connected layers to generate corresponding state prediction results.

[0112] S33, performing vector splicing on the prediction results output by each channel to form a complete battery state vector, wherein the battery state vector includes estimated information on the current battery performance, lifespan, and operating status;

[0113] S34: Output and cache the battery state vector for input into the subsequent abnormality identification module and strategy optimization module.

[0114] This invention decouples the SOC, SOH, and RUL status indicators into independent modeling channels and constructs a state vector structure, improving the model's expression accuracy and multidimensional parsing capabilities in complex state spaces. This structure supports structured references and expanded judgment conditions for state information in the strategy generation module, shifting battery management from single-objective driven to multi-objective collaborative optimization. This enhances the system's responsiveness to various management tasks, such as operating mode switching and battery life determination.

[0115] In this embodiment, the S4 specifically includes:

[0116] S41, receiving the battery state vector output in step S3 and current operating condition parameters, wherein the operating condition parameters include ambient temperature, charge and discharge rate, battery pack load state, and historical operating curve;

[0117] S42. Perform joint feature fusion on the battery state vector and the operating condition parameter to construct an abnormality recognition input sample, wherein the fusion method includes feature splicing and dimension matching processing;

[0118] S43. Input the abnormal identification input sample to the abnormality detection module, and the abnormality detection module performs matching judgment based on the predefined abnormal pattern to determine whether there is thermal runaway, capacity drop, voltage drift or abnormal cycle performance behavior;

[0119] S44. When an abnormality is detected, mark the abnormality category, record the corresponding abnormality occurrence time, abnormality duration, and the battery module or cell location to which the abnormality belongs;

[0120] S45. Output the abnormality identification result and pass it to the strategy optimization module for adjusting the battery operation control strategy.

[0121] This method constructs an anomaly recognition discriminator system through the multi-dimensional fusion of battery state vectors and external operating parameters, enabling the joint identification of multi-source anomaly patterns such as voltage drift, capacity jump, and thermal runaway. The system not only supports static anomaly judgment but also models anomaly trends based on dynamic time series characteristics. It has the ability to proactively detect fault trends and issue pre-warnings, effectively enhancing the battery system's proactive safety management and proactive response capabilities in critical load scenarios.

[0122] In this embodiment, the S5 specifically includes:

[0123] S51, receiving the battery state vector output in step S3 and the abnormality identification result generated in step S4 as input basis for strategy generation;

[0124] S52. Construct a strategy decision input vector based on the state of charge, health status, and remaining service life in the battery status indicators, combined with the current abnormality category and operating condition characteristics;

[0125] S53. Based on the policy decision input vector, calling a preset policy optimization rule set or policy generation model to generate a dynamic management policy for the current operating state;

[0126] S54. Setting charging current limit parameters, discharging current adjustment parameters, thermal management start threshold and adjustment gear, battery module start and stop sequence and switching conditions, and operating mode adjustment range in the management strategy;

[0127] S55: Output the generated management strategy to the control interface module to guide the battery management system to adjust the operating parameters in real time, and record the strategy execution log for subsequent optimization iterations.

[0128] The strategy generation module breaks away from the traditional BMS control approach that relies on empirical thresholds and fixed logic diagrams. Instead, it employs a deep state-driven mechanism to generate a set of adjustable parameters, integrating current state characteristics with historical abnormal behavior to achieve personalized, multi-dimensional, and multi-objective control strategy formulation. The system supports dynamic adjustment of charge and discharge parameters, thermal management strategies, and module configuration during runtime, achieving a flexible balance between battery life maintenance, safety assurance, and energy efficiency optimization, thereby improving energy utilization and system health maintenance throughout the entire life cycle.

[0129] In this embodiment, S6 specifically includes:

[0130] S61. Lightweight the Transformer encoder neural network model and generate a model version suitable for edge devices using parameter pruning, channel compression, or knowledge distillation.

[0131] S62. Deploy a lightweight model version on the edge device, configure the local inference engine, load the model weights and input interface module, and ensure that state reasoning and anomaly identification tasks can be completed independently on the edge side;

[0132] S63: Collect new battery operation data in real time, and form new input samples after completing missing value filling, data alignment and normalization processing through the local preprocessing module;

[0133] S64: Input the newly generated input sample into the edge model, perform local reasoning, and output the updated battery status indicator and anomaly recognition result;

[0134] S65: Packaging the local reasoning results of each round with the corresponding running data to form an upload data packet with a timestamp;

[0135] S66. Send the uploaded data packet to the cloud server via an encrypted communication protocol, wherein the encrypted communication protocol includes a data signature, encrypted transmission, and integrity verification mechanism.

[0136] By performing pruning, channel compression, and knowledge distillation on the Transformer model, the system can deploy approximately equivalent models on computing-constrained edge devices, achieving millisecond-level local inference capabilities. Combined with a data encryption upload mechanism, this ensures the system can autonomously perform state recognition and emergency response assessments locally, while also guaranteeing secure data flow for cloud-based learning. This enables resource collaboration, response latency optimization, and local closed-loop fault tolerance, meeting the edge deployment requirements of large-scale heterogeneous devices and enabling edge-cloud collaborative evolution.

[0137] In this embodiment, the S7 specifically includes:

[0138] S71. The cloud server receives data packets uploaded from multiple edge devices, including edge-collected operation data, model inference output results, acquisition timestamps, edge device identifiers, and system operation logs, and performs parsing, verification, and archiving operations on the data packets to construct a structured training sample set.

[0139] S72. Divide the training sample set by index based on device source, and extract data with clear labels from each group as supervised training samples, and extract data with missing labels or weak labels as self-supervised training samples to form a mixed data set;

[0140] S73. Build an edge collaborative update framework, determine the set of edge nodes participating in this round of training, download their corresponding local model parameter snapshots, and use a secure multi-party computing protocol to perform weighted aggregation on each edge model parameter in the cloud to obtain candidate global model parameters.

[0141] S74. Introduce a self-supervised learning mechanism to perform auxiliary training tasks on the candidate model, wherein the training tasks include sequence reconstruction based on temporal masking, contrast discrimination based on feature perturbation, or fragment reasoning based on context prediction, so as to enhance the model's semantic extraction ability and structural generalization ability on unlabeled samples;

[0142] S75. Perform integrity verification and performance evaluation on the model optimized through self-supervised training. Evaluation metrics include state prediction accuracy, anomaly detection accuracy, model inference latency, and communication overhead on a standard validation set. If all metrics meet the set thresholds, the model update is considered successful.

[0143] S76. The updated model that has passed verification is version numbered, digitally signed, and encrypted, and sent to the corresponding edge device through a secure channel based on the identity authentication protocol. After receiving the updated model, the edge device executes a hot replacement or grayscale loading strategy to achieve online update and uninterrupted deployment of model parameters.

[0144] This method effectively meets battery data privacy, industry data isolation, and data sovereignty requirements by enabling multi-terminal distributed model updates when raw data cannot be transmitted. Combined with self-supervisory task design, the system can perform auxiliary training based on historical operating data and unlabeled anomaly fragments, continuously optimizing the model's perception boundaries and anomaly detection capabilities. The update process includes parameter verification, performance evaluation, and secure delivery, ensuring the credibility and high controllability of model iterations.

[0145] In this embodiment, the S8 specifically includes:

[0146] S81. After the edge device completes the model update deployment, it automatically reinitializes the local operation status monitoring process, reloads the latest version of the Transformer encoder model, and configures its inference interface and data flow channel to ensure that the inference service is synchronized with the local control strategy module;

[0147] S82. During system operation, continuously repeat steps S1 to S7 to periodically or event-drivenly collect operational data, perform status assessments, identify anomalies, generate control strategies, upload cloud training data, and accept model updates to achieve dynamic model iteration and adaptive strategy adjustment.

[0148] S83. Log and analyze the policy feedback for each round of closed-loop execution. The log content includes the state prediction value, policy execution result, system load information, fault frequency, model version number, and edge device operation status.

[0149] S84. Build a closed-loop data management module for the entire lifecycle, uploading operational data, inference results, exception records, and strategic behaviors generated at each stage to a cloud database to form a data accumulation system that supports continuous optimization.

[0150] S85. Set up model evolution tracking and lifecycle mapping logic in the cloud server to archive and classify battery operation characteristics in different periods and industry scenarios as a pre-selected sample pool source for subsequent model training;

[0151] S86. The key performance indicators and model performance results of the closed-loop management process are synchronously displayed on the operation and maintenance platform interface for remote review and manual intervention by engineering and technical personnel, forming a full-life cycle intelligent decision support system driven by large models.

[0152] This method builds a complete closed-loop lifecycle management mechanism. After each round of data collection, state reasoning, anomaly identification, strategy generation, and model update, the results of each module are summarized and mapped to the full lifecycle management map through log analysis and feedback mechanism, forming a structured system evolution view and parameter linkage mechanism. It supports the backtracking of strategy effects, clustering of anomaly frequencies, and tracking of model convergence processes, enabling the operation and maintenance side to perform higher-level strategy audits and system-level health assessments.

[0153] refer to Figure 2 , a battery life cycle intelligent management system based on a large model, including the following modules:

[0154] A data acquisition and preprocessing module is used to collect operational data during the battery manufacturing, transportation, use, storage, and retirement stages, and perform missing value filling, outlier removal, time series alignment, multimodal normalization, and feature vector construction on the data;

[0155] The Transformer state modeling module is used to input preprocessed input samples into the pre-trained Transformer encoder neural network model, perform multi-layer self-attention mechanism and feedforward neural calculation, output battery state indicators and construct a battery state vector;

[0156] The anomaly detection module is used to determine whether there is thermal runaway, capacity drop, voltage drift or cycle performance abnormality based on the battery state vector and the current operating conditions, and record the abnormality type and battery location;

[0157] A strategy generation module, which combines battery status indicators with anomaly detection results to generate battery management strategies for charge and discharge regulation, thermal management configuration, battery module switching, and dynamic adjustment of operating modes;

[0158] Edge inference and communication module, which is used to deploy lightweight model versions on edge devices to achieve local inference and real-time response, while uploading running data and inference results to cloud servers through encrypted communication;

[0159] The model update module is used to aggregate and optimize model parameters based on self-supervised training methods on the cloud server, and after passing performance evaluation, the updated model is securely sent to the edge device for deployment and replacement;

[0160] The closed-loop management module is used to continuously execute state perception, inference calculation, anomaly identification, strategy optimization and model update processes in the system, realizing continuous intelligent management and decision support for the entire battery operation process.

[0161] Through system modularization, the core functions are broken down into seven submodules: data acquisition, state modeling, anomaly diagnosis, edge reasoning and communication, strategy optimization, model update, and closed-loop control. This allows for independent deployment and collaborative collaboration of each subsystem in terms of physical structure, functional logic, and deployment, resulting in excellent hardware and software decoupling. This system architecture supports flexible integration of multi-vendor, multi-type, and multi-purpose battery platforms, offering high scalability, high maintainability, and engineering feasibility.

[0162] Example 1:

[0163] In order to verify the feasibility of the present invention in implementation, the present invention was applied to the new energy vehicle operation and maintenance platform of a domestic power battery manufacturer in East China. Focusing on the large lithium-ion power battery packs that are equipped with the company's newly launched commercial vehicles, the engineering deployment and practical application verification of the large-scale model-based battery full life cycle intelligent management method proposed in the present invention were carried out.

[0164] During the mass production phase, the first batch of 1,320 vehicles were deployed in three locations, covering three typical operating conditions: municipal commuting, freight delivery, and short-distance passenger transport. Each vehicle is equipped with a lithium iron phosphate battery pack with a capacity of 73.5kWh and a nominal voltage of 614V. The battery pack comprises 12 modules and 192 cells, and is equipped with high-precision temperature, current, voltage, and internal resistance sensors. Full lifecycle operating data is synchronously uploaded to the onboard TBOX module via the CAN bus and to the operations and maintenance platform every five minutes.

[0165] The method described in the present invention is deployed with this batch of vehicles as the pilot object. First, the sensor sampling frequency, abnormal drift threshold, data structure definition, etc. are standardized in the battery offline link. During the operation stage, the vehicle's battery pack operation data is collected in real time through the on-board terminal, and missing values ​​are filled, abnormal points are eliminated, and multi-channel normalization is performed locally on the TBOX side, and uniformly encapsulated as input tensors in the form of sliding time windows. The data is locally inferred by the lightweight Transformer model version running on the edge, and the SOC, SOH and RUL indicators are output in real time, while judging whether there are risks such as thermal runaway and abnormal capacity attenuation.

[0166] When a vehicle model deployed in a certain area reached its 36th cycle day and a total mileage of 4,186 km, the edge model determined that the battery's SOH was declining at a significantly higher rate than the average for the same period. The short-term predicted RUL was less than 350 days (the average for all vehicles was 630 days), and the "capacity degradation risk" anomaly label was simultaneously triggered. After the data was encrypted and uploaded to the cloud, the platform's federated training task automatically completed a new round of global model aggregation in the early morning of the following day, constructed a disturbance feature subtask based on the anomaly, and issued a new version of the model to the vehicle via OTA that evening. The updated model improved the SOH prediction accuracy of the same scenario by 12.4%, effectively achieving early fault identification and model adaptive updates.

[0167] In terms of strategic response, the system automatically lowered the thermal management control parameters based on the SOH prediction results, adjusting the temperature control targets of modules 4 and 6 from the original 25°C to 21°C, and lowering the maximum charging current limit to 85A (originally set at 105A). This adjustment effectively suppressed the module temperature difference fluctuation amplitude during the subsequent 12 days of operation, with the maximum temperature difference dropping from 8.6°C to 3.1°C, avoiding further performance deterioration.

[0168] The system has been deployed on 1,320 vehicles, monitoring a cumulative 1.43 million hours of battery operation, triggering 67.21 million edge-side prediction requests and identifying 3,126 potential anomalies, including 37 early warnings of thermal runaway trends, 18 voltage spikes, and 112 cycle tripping anomalies. Compared to the company's traditional BMS system, within the same data collection dimensions and sample range, this system's prediction accuracy increased by approximately 15.2%, and its policy response latency was reduced to 0.6 seconds, significantly improving the proactiveness of operational interventions and the intelligence of policy regulation.

[0169] This embodiment shows that the present invention not only solves the problems of the existing BMS system in rough state modeling, reliance on fixed thresholds for abnormal judgment, and lack of adaptive optimization capabilities, but also realizes distributed model deployment, automatic strategy generation, and lifecycle closed-loop management. It is suitable for core scenarios in large-scale vehicle operations, power grid energy storage systems, and industrial energy systems that are highly sensitive to battery status, and has extremely high engineering promotion value and industrial implementation prospects.

[0170] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A battery life cycle intelligent management method based on a large model, characterized in that: The steps include: S1. Collect battery operation data and perform unified preprocessing on the operation data; S2. Construct input samples based on the preprocessed running data and input the input samples into the pre-trained Transformer encoder neural network model; S3, executes the forward reasoning process of the Transformer encoder neural network model, outputs the corresponding battery status indicator, and constructs the battery status vector; S4. Perform abnormal pattern recognition based on the battery state vector and the current operating condition information to determine whether an abnormality exists. If an abnormality is identified, record the abnormality type and the corresponding battery cell location; S5. Generate a dynamically optimized battery management strategy based on battery status indicators and abnormality identification results; S6. Deploy a lightweight version of the Transformer encoder model on the edge device, using edge computing power for local inference and real-time response. At the same time, the collected new operation data and model inference results are uploaded to the cloud server via encrypted communication. S7. After receiving the uploaded data, the cloud server performs incremental updates on the Transformer encoder model parameters based on the self-supervised training method, and sends the updated model version to the edge device for deployment and replacement, provided that the system security policy is met. S8. Repeat steps S1 to S7 to optimize closed-loop management of the battery's continuous state perception, abnormality prediction, and intelligent strategy throughout its entire life cycle.

2. The method for intelligent management of the battery life cycle based on a large model according to claim 1 is characterized in that: Said S1 specifically includes: S11. During the battery manufacturing, transportation, use, storage, and retirement stages, use embedded sensors or remote monitoring devices to collect battery operating data, including voltage data, current data, temperature data, internal resistance value, capacity value, number of charges, number of discharges, charge and discharge duration, number of cycles, operating timestamps, and related maintenance logs; S12. Performing missing value filling processing on the collected operating data, wherein the missing data is filled by sliding window interpolation or fixed value replacement; S13, performing outlier elimination processing on the operating data, identifying and removing data points that exceed the normal operating range by setting a boundary threshold range; S14. Perform time series alignment on the processed operating data to synchronize timestamps from different data sources to a unified time axis; S15. Normalize various types of operating data to convert characteristic values ​​of different dimensions into a unified numerical range; S16. Construct a feature vector sample based on the normalized data. The feature vector is intercepted according to a fixed window length and encapsulated as structured input data.

3. The method for intelligent management of the battery life cycle based on a large model according to claim 1 is characterized in that: The S2 specifically includes: S21, organizing the constructed feature vector samples into a multidimensional input sequence in chronological order, wherein the input sequence is divided into fixed time windows to form input samples of multiple continuous time periods; S22. Perform position encoding processing on each input sample and introduce temporal position information in the feature dimension; S23, inputting the input sample with added position information into a Transformer encoder neural network model, wherein the model includes multiple encoder layers, each encoder layer includes a multi-head self-attention mechanism, a feedforward neural network structure, a residual connection and a normalization module; S24. Inside the Transformer encoder model, multiple layers of processing are performed on the input samples in sequence. Each layer completes self-attention weight calculation, context feature extraction and nonlinear transformation, and outputs deep feature representation layer by layer. S25. Extract the global feature output corresponding to each input sample to form a unified representation encoding result, and pass it to the state reasoning module.

4. The method for intelligent management of the battery life cycle based on a large model according to claim 1, characterized in that: The S3 specifically includes: S31, receiving the deep feature representation output by the Transformer encoder model in step S2; S32. Based on the deep feature representation, construct representation channels for the state of charge (SOC), state of health (SOH), and remaining useful life (RUL), respectively. Each channel consists of several fully connected layers. S33, performing vector splicing on the prediction results output by each channel to form a complete battery state vector, wherein the battery state vector includes estimated information on the current battery performance, lifespan, and operating status; S34: Output and cache the battery state vector.

5. The method for intelligent management of the battery life cycle based on a large model according to claim 1 is characterized in that: The S4 specifically includes: S41, receiving the battery state vector output in step S3 and current operating condition parameters, wherein the operating condition parameters include ambient temperature, charge and discharge rate, battery pack load state, and historical operating curve; S42. Perform joint feature fusion on the battery state vector and the operating condition parameter to construct an abnormality recognition input sample, wherein the fusion method includes feature splicing and dimension matching processing; S43. Input the abnormal identification input sample to the abnormality detection module, and the abnormality detection module performs matching judgment based on the predefined abnormal pattern to determine whether there is thermal runaway, capacity drop, voltage drift or abnormal cycle performance behavior; S44. When an abnormality is detected, mark the abnormality category, record the corresponding abnormality occurrence time, abnormality duration, and the battery module or cell location to which the abnormality belongs; S45. Output the abnormality identification result and pass it to the strategy optimization module for adjusting the battery operation control strategy.

6. The method for intelligent management of the battery life cycle based on a large model according to claim 1 is characterized in that: The S5 specifically includes: S51, receiving the battery state vector output in step S3 and the abnormality identification result generated in step S4; S52. Construct a strategy decision input vector based on the state of charge, health status, and remaining service life in the battery status indicators, combined with the current abnormality category and operating condition characteristics; S53. Based on the policy decision input vector, calling a preset policy optimization rule set or policy generation model to generate a dynamic management policy for the current operating state; S54. Setting charging current limit parameters, discharging current adjustment parameters, thermal management start threshold and adjustment gear, battery module start and stop sequence and switching conditions, and operating mode adjustment range in the management strategy; S55: Output the generated management policy to the control interface module and record the policy execution log.

7. The method for intelligent management of the battery life cycle based on a large model according to claim 1 is characterized in that: The S6 specifically includes: S61. Lightweight the Transformer encoder neural network model and generate a model version suitable for edge devices using parameter pruning, channel compression, or knowledge distillation. S62. Deploy the lightweight model version in the edge device, configure the local inference engine, and load the model weight and input interface module; S63: Collect new battery operation data in real time, and form new input samples after completing missing value filling, data alignment and normalization processing through the local preprocessing module; S64: Input the newly generated input sample into the edge model, perform local reasoning, and output the updated battery status indicator and anomaly recognition result; S65: Packaging the local reasoning results of each round with the corresponding running data to form an upload data packet with a timestamp; S66. Send the uploaded data packet to the cloud server via an encrypted communication protocol, wherein the encrypted communication protocol includes a data signature, encrypted transmission, and integrity verification mechanism.

8. The method for intelligent management of the battery life cycle based on a large model according to claim 1 is characterized in that: The S7 specifically includes: S71. The cloud server receives data packets uploaded from multiple edge devices, including edge-collected operation data, model inference output results, acquisition timestamps, edge device identifiers, and system operation logs, and performs parsing, verification, and archiving operations on the data packets to construct a structured training sample set. S72. Divide the training sample set by index based on device source, and extract data with clear labels from each group as supervised training samples, and extract data with missing labels or weak labels as self-supervised training samples to form a mixed data set; S73. Build an edge collaborative update framework, determine the set of edge nodes participating in this round of training, download their corresponding local model parameter snapshots, and use a secure multi-party computing protocol to perform weighted aggregation on each edge model parameter in the cloud to obtain candidate global model parameters. S74. Introducing a self-supervised learning mechanism to perform auxiliary training tasks on the candidate model, wherein the training tasks include sequence reconstruction based on temporal masking, contrast discrimination based on feature perturbation, or fragment reasoning based on context prediction; S75. Perform integrity verification and performance evaluation on the model optimized through self-supervised training. Evaluation metrics include state prediction accuracy, anomaly detection accuracy, model inference latency, and communication overhead on a standard validation set. If all metrics meet the set thresholds, the model update is considered successful. S76. The updated model that has passed the verification is version numbered, digitally signed, and encrypted, and sent to the corresponding edge device through a secure channel based on the identity authentication protocol. After receiving the updated model, the edge device executes a hot replacement or grayscale loading strategy.

9. The method for intelligent management of the battery life cycle based on a large model according to claim 1, characterized in that: The S8 specifically includes: S81. After the edge device completes the model update deployment, it automatically reinitializes the local operation status monitoring process, reloads the latest version of the Transformer encoder model, and configures its inference interface and data flow channel; S82. During system operation, continuously repeat steps S1 to S7 to periodically or event-drivenly collect operational data, perform status assessments, identify anomalies, generate control strategies, upload cloud training data, and accept model updates. S83. Log and analyze the policy feedback for each round of closed-loop execution. The log content includes the state prediction value, policy execution result, system load information, fault frequency, model version number, and edge device operation status. S84. Build a closed-loop data management module for the entire lifecycle, uploading operational data, inference results, exception records, and strategic behaviors generated at each stage to a cloud database to form a data accumulation system that supports continuous optimization. S85. Set up model evolution tracking and lifecycle mapping logic in the cloud server to archive and classify battery operation characteristics in different periods and industry scenarios; S86. The key performance indicators and model performance results of the closed-loop management process are synchronously displayed on the operation and maintenance platform interface to form a full-life cycle intelligent decision support system driven by a large model.

10. A battery full life cycle intelligent management system based on a large model, implementing a battery full life cycle intelligent management method based on a large model according to any one of claims 1 to 9, characterized in that: Includes the following modules: A data acquisition and preprocessing module is used to collect operational data during the battery manufacturing, transportation, use, storage, and retirement stages, and perform missing value filling, outlier removal, time series alignment, multimodal normalization, and feature vector construction on the data; The Transformer state modeling module is used to input preprocessed input samples into the pre-trained Transformer encoder neural network model, perform multi-layer self-attention mechanism and feedforward neural calculation, output battery state indicators and construct a battery state vector; The anomaly detection module is used to determine whether there is thermal runaway, capacity drop, voltage drift or cycle performance abnormality based on the battery state vector and the current operating conditions, and record the abnormality type and battery location; A strategy generation module, which combines battery status indicators with anomaly detection results to generate battery management strategies for charge and discharge regulation, thermal management configuration, battery module switching, and dynamic adjustment of operating modes; Edge inference and communication module, which is used to deploy a lightweight model version on edge devices and upload running data and inference results to the cloud server through encrypted communication; The model update module is used to aggregate and optimize model parameters based on self-supervised training methods on the cloud server, and after passing performance evaluation, the updated model is securely sent to the edge device for deployment and replacement; The closed-loop management module is used to continuously execute state perception, reasoning calculation, anomaly identification, strategy optimization and model update processes in the system.

Images