A battery electrochemical intelligent analysis system and method based on charge and discharge curve data

Abstract

A battery electrochemical intelligent analysis system and method based on charge-discharge curve data includes: a user interaction layer that receives charge-discharge data files and parameter configurations, and obtains natural language analysis requirements; an intelligent analysis layer, based on the LangGraph framework, adopting a two-layer architecture of coordination and execution layers, managing and coordinating supervisory agents, data processing agents, electrochemical analysis agents, data plotting agents, quality review agents, analysis report agents, and question-answering agents through state graphs, completing a complete intelligent analysis task from data processing, electrochemical parameter analysis, chart plotting, quality review, report generation to question-answering interaction; and an output layer that receives charts, reports, and question-answering results generated by the intelligent analysis layer, converts the format, and transmits them to the user interaction layer for display or download. This invention achieves automated and intelligent battery electrochemical data analysis without programming, lowers the professional threshold, and improves analysis efficiency and standardization.

Description

Technical Field

[0001] This invention relates to battery electrochemical analysis technology, and in particular to a battery electrochemical intelligent analysis system and method based on charge-discharge curve data. Background Technology

[0002] Rechargeable batteries, as rechargeable chemical power sources, are core components in new energy vehicles, energy storage systems, and portable electronic devices. With the increasing urgency of global energy structure transformation and sustainable development, the research and development of new rechargeable batteries such as lithium-ion batteries, sodium-ion batteries, and solid-state batteries has become a key focus for scientific research and industry. The performance of battery materials directly determines the battery's energy density, power density, cycle life, and safety; therefore, the research and performance characterization of battery materials are particularly important.

[0003] A crucial method for characterizing the electrochemical performance of battery materials is charge-discharge cycle testing. Researchers use battery testing systems such as LAND, Neware, and Arbin to conduct charge-discharge tests on batteries, recording data such as voltage, current, and capacity. The traditional data analysis process includes the following steps: first, opening the test data file and parsing its format using the accompanying data analysis software from LAND or Neware; then, manually selecting the cycle data to be analyzed; copying the data to Excel for preliminary processing and calculations; next, importing the processed data into professional plotting software such as Origin or MATLAB for visualization; and finally, writing an analysis report based on the chart results.

[0004] However, traditional battery data analysis methods face the following dilemmas and challenges. First, they have high professional barriers. Researchers need to master the operation of multiple software programs, including data analysis software like Landon or Newway, data processing functions in Excel, and plotting functions in Origin or MATLAB. Some researchers also need Python programming skills to automate the analysis. The learning cycle can take several months, making analysis difficult for researchers without programming backgrounds or familiarity with professional plotting software like Origin. Second, the process is cumbersome and time-consuming. Analyzing a set of charge-discharge data containing 100 cycles typically takes 1 to 2 hours, including multiple steps such as data loading, cycle extraction, parameter calculation, data cleaning, chart creation, and format adjustment. The entire process requires switching between multiple software programs, resulting in low efficiency and a high risk of errors. Third, there is a lack of standardization and automation. Different researchers use different analysis methods and chart styles, making it difficult to guarantee the reproducibility of analysis results and the consistency of chart quality. This is especially true in team collaborations and paper publication, where charts often need repeated revisions to meet journal standards.

[0005] In recent years, large language models have made groundbreaking progress in the field of natural language processing. Models such as the GPT and Claude series have demonstrated superior language understanding and generation capabilities, accurately grasping user intent and generating expected responses. The emergence of large language models provides a new approach to solving the interactive challenges in traditional data analysis, allowing users to describe their analytical needs in natural language without needing to learn complex programming languages ​​or software operations. Simultaneously, the development of agent technology has provided an effective solution for automating complex tasks. By decomposing complex tasks into multiple relatively independent sub-tasks, each handled by a specialized agent equipped with dedicated tool components, specific functions are completed through tool calls, achieving a modular and scalable system design.

[0006] To effectively manage the execution flow of multi-agent systems, a robust framework is needed to define agent state transitions, manage message passing between agents, and control branching and looping. LangGraph, a framework within the LangChain ecosystem, is specifically designed for building multi-agent applications. Based on the concept of a StateGraph, it describes the agent execution flow by defining nodes and edges. In the LangGraph framework, nodes represent agents or processing functions, and edges represent state transitions. Edges can be unconditional, directly leading to the next node, or conditional, selecting different target nodes based on conditions. The LangGraph framework provides conditional routing to allow dynamic adjustment of the flow based on execution results, loop control to allow the flow to return to previous nodes for iterative optimization, and a checkpoint mechanism to save and restore the execution state, supporting long-running tasks.

[0007] In conclusion, while large language models, intelligent agent technology, and the LangGraph framework provide a technological foundation for intelligent data analysis, how to effectively combine these technologies and apply them to the field of secondary battery charge and discharge data analysis, and build an intelligent analysis system that allows researchers without programming skills or professional graphing software operation skills to directly obtain professional-grade charts and analysis reports from raw data, and to perform data question answering and basic knowledge question answering, remains an urgent technical problem to be solved.

[0008] It should be noted that the information disclosed in the background section above is only for understanding the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0009] The main objective of this invention is to overcome the deficiencies in the aforementioned background technology and provide a battery electrochemical intelligent analysis system and method based on charge-discharge curve data.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: A battery electrochemical intelligent analysis system based on charge-discharge curve data, comprising: The user interaction layer is used to receive charging and discharging data files and parameter configurations uploaded by users, and to obtain user analysis requirements input in natural language. The intelligent analysis layer, built around the LangGraph framework, adopts a two-layer architecture of coordination and execution. It manages and coordinates multiple agents through a state graph. These agents include at least a supervisory agent, a data processing agent, an electrochemical analysis agent, a data plotting agent, a quality review agent, an analysis report agent, and a question-answering agent. This layer is used to complete the entire intelligent analysis task, from data processing, electrochemical parameter analysis, chart creation, quality review, report generation to question-answering interaction. The output layer receives charts, reports, and Q&A results generated by the intelligent analysis layer, converts them into formats, and then passes them to the user interaction layer for display or download.

[0011] A battery electrochemical intelligent analysis method based on charge-discharge curve data, using the aforementioned system, the method comprising: Step A: Receive the charging and discharging data file and parameter configuration uploaded by the user through the user interaction layer, and obtain the user's analysis requirements input in natural language. Step B: The coordination layer of the intelligent analysis layer parses the user intent through a large language model, and has a built-in error recovery mechanism to repair or trigger retries for returned format errors, generating an execution plan containing an ordered sequence of functional units; Step C: According to the execution plan, the execution layer of the intelligent analysis layer first calls the data processing unit to load and verify the data file, extract the charge and discharge data of the specified cycle and calculate the specific capacity, and store the processed data into the first cache layer; Step D: If the execution plan includes advanced analysis, the data analysis unit is invoked to analyze the cached data from the perspective of median voltage and / or differential capacity, and a processing mechanism including adaptive adjustment of smoothing window, adaptive backoff of peak detection threshold and adaptive matching of material phase transition voltage range is implemented, and the analysis results are stored in the second cache layer. Step E: Call the plotting unit to read data and analysis results from the cache and generate scientific charts that conform to the preset journal style; Step F: Call the quality review unit to score the quality of the generated charts. If the score is below the threshold and it is determined to be a data selection strategy problem, select to generate a report branch containing a quality warning, and at the same time perform material characteristic verification and generate research suggestions. Step G: Invoke the report generation unit to integrate all analysis results, quality assessment information, and research recommendations to generate a structured analysis report; and Step H: The output layer converts the format of the generated charts, reports, and Q&A results generated by the Q&A unit, and displays or downloads them through the user interaction layer.

[0012] The intelligent battery electrochemical analysis system and method based on charge-discharge curve data provided by this invention have the following beneficial effects: First, this invention breaks through the dependence of traditional battery data analysis methods on professional software operation and programming skills. Through a multi-agent collaborative architecture built around the LangGraph framework, it modularizes and automates the entire process of tasks, including data processing, electrochemical parameter analysis, differential capacity analysis, chart plotting, quality review, report generation, and question-and-answer interaction. Users only need to describe their analysis needs in natural language, and the system can automatically complete intent parsing, task decomposition, and agent scheduling. This achieves one-stop battery data analysis without programming, significantly reducing the learning cost and operational threshold for researchers, and solving the problems of inefficiency and error-proneness caused by manually switching between multiple software programs in existing technologies.

[0013] Furthermore, this invention introduces multiple adaptive processing mechanisms at the electrochemical analysis algorithm level. The differential capacity analysis subunit can automatically adjust the smoothing window according to the actual number of data points, adopt a two-stage peak detection strategy, and dynamically match the phase transition voltage window based on the nickel content of the material, thereby ensuring the robustness and accuracy of dQ / dV characteristic peak identification under different data qualities and different material systems; the median voltage analysis subunit provides a quantitative evaluation basis for battery kinetic performance through polarization voltage level classification and trend identification. These mechanisms make the analysis results not only automated but also highly adaptable and reliable.

[0014] Furthermore, this invention implements intelligent process control and fault tolerance through the LangGraph state graph engine. The system defines state data structures including user input, execution plan, retry flags, and error fields, and utilizes routing functions to implement conditional routing, loop recursion, and early termination. When a format error occurs during intent parsing, the built-in error recovery mechanism can automatically repair it and trigger a retry; when quality review finds that the graph score is below the threshold, the system can intelligently select the "generate a warning report" branch instead of invalid redrawing based on the specific reason for the substandard score, avoiding process rigidity. This dynamic process adjustment and fault tolerance capability enables the system to output reliable results stably even when facing diverse user commands and abnormal data, improving the system's robustness and practicality.

[0015] Furthermore, this invention employs a three-layer caching mechanism to store raw cycle data, advanced analysis results, and quality assessment information respectively. The question-and-answer unit can directly retrieve data from the cache and generate natural language answers, supporting real-time queries for parameters such as capacity, coulombic efficiency, median voltage, and dQ / dV peak position for specific cycles. Simultaneously, it combines a large language model to implement basic battery knowledge question-and-answer, enhancing the system's interactivity and knowledge service capabilities. The output layer supports multi-format (PNG, PDF, Markdown) and multi-channel (Web, IDE, terminal, file system) adaptation. The generated charts strictly adhere to the academic standards of top-tier journals, enabling users without professional drawing skills to directly obtain scientific charts and structured analysis reports that meet publication requirements.

[0016] In summary, this invention effectively solves the technical challenges of high professional threshold, cumbersome operation, and lack of standardization and intelligence in the field of battery charge and discharge data analysis. It realizes full-process automation from raw data to professional-grade charts, analysis reports, and intelligent Q&A, significantly improving the efficiency and reliability of battery material research and development and performance characterization.

[0017] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0018] Figure 1 This is an overall architecture diagram of the battery electrochemical intelligent analysis system according to an embodiment of the present invention.

[0019] Figure 2 This is an architecture diagram of the intelligent analysis module in an embodiment of the present invention.

[0020] Figure 3 This is a flow logic diagram between various Agent nodes driven by the LangGraph state graph in an embodiment of the present invention.

[0021] Figure 4 This is a schematic diagram of the differential capacity curve of an embodiment of the present invention (showing the dQ / dV curves of the charging and discharging process).

[0022] Figure 5 This is a schematic diagram of the median voltage evolution curve in an embodiment of the present invention (showing the trend of the median voltage during charging and discharging with the number of cycles).

[0023] Figure 6 This is a schematic diagram of the charge-discharge curves of an embodiment of the present invention (showing the voltage-specific capacity curve of a single charge-discharge cycle).

[0024] Figure 7 This is a schematic diagram of the charge-discharge curves of an embodiment of the present invention (showing the voltage-specific capacity curves of multiple charge-discharge cycles). Detailed Implementation

[0025] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention.

[0026] The relevant terminology used in the embodiments of this application: LangChain is an application development framework based on Large Language Model (LLM). It standardizes and simplifies the use of LLM in two main ways: integration, which integrates external data (such as files, other applications, API data, etc.) into the LLM; and agent, which allows the LLM to interact with specific environments through decision-making and assists in determining the next step of the operation.

[0027] An agent acts as a multi-functional interface, capable of accessing and using a suite of tools. Based on user input, the agent decides which tools to invoke. It can not only use multiple tools simultaneously, but also use the output data of one tool as input data for another.

[0028] LangGraph is a framework within the LangChain ecosystem designed specifically for building multi-agent applications. Based on the concept of a StateGraph, it describes the execution flow of agents by defining nodes and edges. Nodes represent agents or processing functions, and edges represent state transitions, which can be unconditional or conditional.

[0029] A charge-discharge curve is a curve plotted with specific capacity as the abscissa and voltage as the ordinate during the charging and discharging process of a secondary battery. It is an important means of characterizing the electrochemical performance of a battery.

[0030] Median voltage refers to the voltage value at which the capacity reaches 50% during charging or discharging. It can reflect the average operating voltage and polarization degree of the battery.

[0031] The dQ / dV differential capacity curve is a curve plotted by differentiating the charge-discharge curve, with voltage as the abscissa and differential capacity dQ / dV as the ordinate. Its peak position corresponds to the phase transition voltage of the material and can be used to analyze the phase transition behavior and structural evolution of electrode materials.

[0032] Coulombic efficiency is the ratio of discharge capacity to charge capacity, reflecting the reversibility of the battery and the degree of side reactions.

[0033] Unless otherwise defined, 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 application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0034] This invention provides a battery electrochemical intelligent analysis system and method based on charge-discharge curve data. The system uses the LangGraph framework as its core to construct a multi-agent collaborative analysis workflow. Through natural language interaction, it enables researchers without programming skills, such as those lacking Python programming skills or the ability to operate professional graphing software like Origin, to directly obtain professional-grade scientific charts and structured analysis reports from raw charge-discharge data.

[0035] See Figures 1 to 7 This invention provides a battery electrochemical intelligent analysis system based on charge-discharge curve data, comprising: The user interaction layer is used to receive charging and discharging data files and parameter configurations uploaded by users, and to obtain user analysis requirements input in natural language. The intelligent analysis layer, built around the LangGraph framework, adopts a two-layer architecture of coordination and execution. It manages and coordinates multiple agents through a state graph. These agents include at least a supervisory agent, a data processing agent, an electrochemical analysis agent, a data plotting agent, a quality review agent, an analysis report agent, and a question-answering agent. This layer is used to complete the entire intelligent analysis task, from data processing, electrochemical parameter analysis, chart creation, quality review, report generation to question-answering interaction. The output layer receives charts, reports, and Q&A results generated by the intelligent analysis layer, converts them into their correct formats, and then passes them to the user interaction layer for display or download. The output layer includes chart output units, report output units, and Q&A output units, automatically selecting the appropriate format according to the requirements of different output channels, and supports exporting from web interfaces, integrated development environments (IDEs), terminal interfaces, and file systems.

[0036] This application's embodiments are guided by the user's battery data analysis needs. They guide the user to conduct electrochemical data analysis through natural language interaction, and combine large language models to achieve intelligent decision-making, automated processes, and fault-tolerant handling, realizing one-stop battery data analysis and reducing the burden on battery R&D personnel.

[0037] In some embodiments, the user interaction layer includes: a data upload unit for receiving and verifying charge and discharge data files in various formats, supporting data formats from various battery testing devices; a parameter configuration unit for receiving material parameters, experimental parameters, test conditions, and journal styles set by the user; a task triggering unit for receiving user analysis instructions via either quick operation or natural language input; a result display unit for displaying the execution process and results of the analysis task in the form of dialogue history and chart preview; and a status monitoring unit for displaying and confirming the current system's key parameter configuration status in real time.

[0038] In some embodiments, the intelligent analysis layer includes: a coordination layer comprising a task coordination unit that implements the functions of the supervisory agent; the task coordination unit further comprising: an intent parsing subunit, used to analyze natural language instructions through a large language model to identify task types and extract key parameters; the intent parsing subunit has a built-in error recovery mechanism, which automatically expands the returned loop number parameter into a list of consecutive integers and performs format validation when it detects that the loop number parameter is a range string; if the repair fails, the error message is appended to the prompt word and a retry flag is set, and the process is looped back to the coordination layer for re-execution through the routing function of the state graph until the retry limit is reached and the task is terminated; a plan generation subunit, used to generate an ordered sequence of functional units as an execution plan according to the task type; and a task scheduling subunit, used to schedule the various units of the execution layer according to the execution plan. The execution layer includes: a data processing unit, which implements the functions of the data processing agent, used for loading files, verifying cycle numbers, extracting data, and performing specific capacity calculations; the tool components configured in the data processing unit include: a data reading tool, used for reading test data files and supporting multiple data formats including Neware format, LAND format, and Arbin format; a data cleaning tool, used for identifying and processing outliers, missing values, and duplicate values ​​in charge-discharge curve data; a cycle extraction tool, used for extracting data for a specified cycle from a complete data file; and a specific capacity calculation tool, used for calculating specific capacity based on the active substance. The system includes: a quality calculation unit for specific capacity per cycle; a capacity statistics tool for calculating charge capacity, discharge capacity, and capacity retention rate per cycle; a coulombic efficiency calculation tool for calculating coulombic efficiency per cycle; a data analysis unit integrating the functions of a median voltage analysis agent and a dQ / dV differential analysis agent for calculating electrochemical parameters and identifying phase transition characteristics; and a plotting unit implementing the functions of the data plotting agent to generate charts conforming to academic standards. The system supports the automatic generation of various scientific charts, including charge-discharge curves, dQ / dV differential capacity curves, and median voltage evolution curves, all strictly adhering to Nature standards. The system adheres to the academic standards of top journals such as *Nature Energy*, *Science*, and *ACS Energy Letters*, including detailed requirements for fonts, color schemes, line styles, and annotations. The plotting unit utilizes the following tools: a Matplotlib plotting library calling tool for creating chart objects; a chart configuration tool for setting basic chart parameters; an academic style rendering tool for applying the chart standards of top journals such as *Nature Energy*, *Science*, and *ACS Energy Letters*; a multi-subplot layout tool for generating complex charts containing multiple subplots; a chart annotation tool for adding annotation information; a quality review unit, implementing the functions of the quality review agent, for evaluating chart quality and verifying material characteristics; a report generation unit, implementing the functions of the analysis report agent, for integrating results and generating structured reports; and a question-and-answer unit, implementing the functions of the question-and-answer agent, for reading data from the cache and generating natural language answers. The system supports expert-level question-and-answer sessions based on the current data. Users can ask data-related questions such as capacity values ​​for specific cycles, dQ / dV peak information, and capacity decay trends, as well as basic knowledge questions such as fundamental concepts of battery materials, charging and discharging principles, and testing methods.

[0039] In some embodiments, the data analysis unit includes: The median voltage analysis subunit is used to calculate the median voltage and polarization voltage during the charge and discharge process, and to classify the severity level according to the polarization voltage value to evaluate the battery kinetic performance. The tool components used by the median voltage analysis subunit include: a median voltage calculation tool, an average voltage calculation tool, a polarization voltage calculation tool, and a polarization evaluation tool; and The differential capacity analysis subunit is used to calculate the dQ / dV differential capacity curve and implement an adaptive adjustment mechanism, including: automatically adjusting the smoothing window size according to the actual number of data points; employing a two-stage peak detection strategy, automatically lowering the threshold and re-detecting if no peak is detected in the first stage; shifting each phase transition voltage window according to the nickel content resolved from the material name using a preset offset formula to match different series of materials; and dynamically setting an outlier truncation threshold based on the curve mean and standard deviation, and determining whether to perform truncation based on the comparison between the upper quantile of the curve (e.g., the 99th quantile) and the truncation threshold, with truncation only performed when the former exceeds the threshold. The tool components used by the differential capacity analysis subunit include: a data preprocessing tool for six-step data cleaning, a Savitzky-Golay smoothing tool for smoothing, a dQ / dV calculation tool for calculating the differential capacity curve, and a phase transition peak identification tool for identifying the position and intensity of phase transition peaks.

[0040] In some embodiments, the quality review unit includes: The chart quality assessment subunit is used to score the quality of charts based on predefined scoring criteria including font, label, legend, data point density, color contrast and resolution, and to set a passing threshold. The materials characteristic verification subunit combines the analysis results with benchmark parameters preset based on industry standards. Through benchmark verification and deviation quantification, it verifies whether key performance indicators are within reasonable ranges, identifies abnormal data, and provides warning information and research suggestions. The process branch decision function is activated when the chart quality assessment subunit detects that the overall score is lower than the qualified threshold. Based on the specific reason for the score failure, if it is determined that the data selection strategy is the cause of the failure to improve through redrawing, the "Generate a report containing quality warnings" branch is selected, and the score details and branch decision conclusion are written to the task status instead of triggering the redrawing branch.

[0041] In some embodiments, the system employs a three-layer caching mechanism, including: a first caching layer for storing cyclic data extracted by the data processing unit, including charge / discharge capacity, coulombic efficiency, and specific capacity; a second caching layer for storing the analysis results of the data analysis unit, including median voltage evolution data, polarization voltage assessment, and dQ / dV peak information; and a third caching layer for storing the evaluation results of the quality review unit and the report content of the report generation unit; the question-and-answer unit retrieves data from the corresponding caching layer according to the user query through the cached data reading subunit, and the natural language generation subunit converts it into a natural language answer.

[0042] In some embodiments, the intelligent analysis layer implements multi-agent collaboration through the LangGraph framework, which includes: The state graph definition mechanism defines the execution flow of an agent through nodes and edges. Nodes represent agents or processing functions, and edges represent state transition paths. The node configuration mechanism is used to configure the corresponding agent unit or processing function for each node. The edge configuration mechanism is used to define the transition rules between nodes. The task status data structure should at least include user input, file path, execution plan, current step, extracted data, analysis results, chart path, retry flag, and error field; The routing function is used to read the retry flag, current node identifier, execution plan and error field in the task status after each node execution is completed, so as to dynamically determine the next node to be executed: if the retry flag is true, it loops back to the coordination layer node; if the error field is not empty, it checks whether it is an unrecoverable error or other conditions that do not meet the conditions for further normal operation. If so, it routes to the termination node; otherwise, it routes to the next node according to the execution plan. Conditional routing mechanism is used to dynamically determine the next node to be executed based on the current state and execution result; A loop control mechanism is used to support the process returning to the previous node for re-execution under specific conditions; and The checkpoint mechanism is used to save a snapshot of the state during process execution.

[0043] In some embodiments, the plurality of intelligent agent units cooperate with each other through the following mechanisms: a message passing mechanism for passing task information and execution results between intelligent agent units; a state sharing mechanism for maintaining a global task state data structure; a data transfer mechanism for sharing data between intelligent agent units through a global cache; and an error propagation mechanism for transmitting error information and triggering an error recovery process.

[0044] In some embodiments, the system also has dynamic process adjustment capabilities, including: The intermediate result evaluation function is used to evaluate the quality of intermediate results during task execution; The workflow branching decision function dynamically selects the execution path based on task type or intermediate result quality. When a chart quality score is detected to be below a threshold and determined to be a data selection strategy issue, it chooses to generate a report branch containing a quality warning instead of triggering a redrawing branch. An early termination function is used to terminate task execution when an unrecoverable error or other conditions that prevent further normal operation are detected. In some embodiments, the system also has fault-tolerant processing capabilities, including: Error detection functionality is used to monitor the execution results of each functional unit in order to detect format errors or missing data; The automatic retry function is used to automatically retry when a format error is detected. The error message and format example are appended to the prompt word, and the process is looped back to the coordination layer node for re-execution through the routing function of the state diagram. A maximum of a preset number of retries are allowed. The checkpoint saving function is used to automatically save a snapshot of the task state after execution at each node through the checkpoint mechanism of the LangGraph framework, ensuring the traceability of the task execution process; and The error reporting function is used to record and report error information to the user.

[0045] In some embodiments, the system supports the following functions: plotting single-cycle and multi-cycle charge-discharge curves, supporting comparison of charge-discharge performance across multiple cycles in the same coordinate system; plotting dQ / dV differential capacity curves, supporting identification of phase transition characteristic peaks and their voltage positions during charge-discharge processes; plotting median voltage evolution curves, supporting analysis of battery polarization characteristics; electrochemical data Q&A, supporting querying capacity values, coulombic efficiency, and dQ / dV peak position information for specific cycles; and battery fundamentals Q&A, supporting querying basic concepts of battery materials, charge-discharge principles, and testing methods.

[0046] In some embodiments, the system supports user-defined parameters, including: analysis parameters, such as setting the number of activation cycles, voltage range, capacity normalization method, and dQ / dV smoothing window size; style parameters, such as selecting plot style, report template, modifying chart titles, legend content, and axis names; file formats, supporting multiple data file formats such as NDA and CSV files; material system parameters, used to set the material type and active substance mass; experimental parameters, used to set the charge / discharge rate and voltage range; test condition parameters, used to set the test temperature and charging mode; activation cycle setting parameters, used to set the number of activation cycles; and journal style parameters, used to set plot styles required by different journals.

[0047] In some embodiments, the system is scalable, including: scalable material system, scalable journal style, scalable data file format, scalable analysis function, scalable output format, scalable front-end interaction, and scalable tool components.

[0048] This invention also provides a battery electrochemical intelligent analysis method based on charge-discharge curve data. Using the system described above, the method includes: Step A: Receive the charging and discharging data file and parameter configuration uploaded by the user through the user interaction layer, and obtain the user's analysis requirements input in natural language. Step B: The coordination layer of the intelligent analysis layer parses the user intent through a large language model, and has a built-in error recovery mechanism to repair or trigger retries for returned format errors, generating an execution plan containing an ordered sequence of functional units; Step C: According to the execution plan, the execution layer of the intelligent analysis layer first calls the data processing unit to load and verify the data file, extract the charge and discharge data of the specified cycle and calculate the specific capacity, and store the processed data into the first cache layer; Step D: If the execution plan includes advanced analysis, the data analysis unit is invoked to analyze the cached data from the perspective of median voltage and / or differential capacity, and a processing mechanism including adaptive adjustment of smoothing window, adaptive backoff of peak detection threshold and adaptive matching of material phase transition voltage range is implemented, and the analysis results are stored in the second cache layer. Step E: Call the plotting unit to read data and analysis results from the cache and generate scientific charts that conform to the preset journal style; Step F: Call the quality review unit to score the quality of the generated charts. If the score is below the threshold and it is determined to be a data selection strategy problem, select to generate a report branch containing a quality warning, and at the same time perform material characteristic verification and generate research suggestions. Step G: Invoke the report generation unit to integrate all analysis results, quality assessment information, and research recommendations to generate a structured analysis report; and Step H: The output layer converts the format of the generated charts, reports, and Q&A results generated by the Q&A unit, and displays or downloads them through the user interaction layer.

[0049] The following further describes the implementation methods and advantages of specific embodiments of the present invention.

[0050] A battery electrochemical intelligent analysis system based on charge-discharge curve data is disclosed. The system comprises a user interaction layer, an intelligent analysis layer, and an output layer. The user interaction layer receives charge-discharge data files uploaded by the user, obtains analysis requirements through natural language, and displays the analysis results to the user. The intelligent analysis layer uses LangGraph as its core framework and adopts a two-layer architecture of coordination and execution layers. It manages and coordinates supervisory agents, data processing agents, median voltage analysis agents, dQ / dV differential analysis agents, data plotting agents, quality review agents, analysis report agents, and question-and-answer agents to complete the complete intelligent analysis tasks of data processing, electrochemical parameter analysis, differential capacity analysis, chart plotting, quality review, report generation, and question-and-answer interaction. Each agent is equipped with dedicated tool components to perform specific functions through tool calls. The output layer receives the charts, reports, and question-and-answer results generated by the intelligent analysis layer, performs format conversion, and then transmits them to the user interaction layer for display or download. This invention realizes intelligent and automated analysis of secondary battery charge-discharge curve data.

[0051] Analysis of secondary battery charge and discharge data may involve various operations such as data reading, format parsing, cyclic extraction, parameter calculation, differential analysis, chart creation, and chart or data analysis. Traditional analysis methods require switching between multiple software programs. Different battery testing equipment (such as Neware, LAND, and Arbin) outputs data in different formats and standards, existing in various file formats such as Excel, CSV, and TXT, making it difficult to support automated analysis.

[0052] This application provides a battery electrochemical intelligent analysis system based on charge-discharge curve data. It can combine large language models and multi-agent technology to realize automated processing and intelligent analysis of multi-format charge-discharge data, automatically generate professional-grade scientific charts and analysis reports, perform result review and error handling, and interact with users to ask and answer questions about the data or charts.

[0053] Figure 1 This is a system architecture diagram of a battery electrochemical intelligent analysis system based on charge-discharge curve data provided in an embodiment of this application. Figure 1 As shown, in order to achieve intelligent analysis of charging and discharging data, the system provided in this application includes a user interaction layer, used to receive charging and discharging curve data files and related parameter configurations uploaded by users, and to obtain user analysis requirements input in natural language; an intelligent analysis layer, used to use the LangGraph framework as the core construction framework to coordinate multiple intelligent agents to complete data analysis tasks, and is divided into a coordination layer and an execution layer; and an output layer, used to provide the analysis and user query results to users through different channels such as front-end interface display, terminal display, and IDE image file download.

[0054] The user interaction module receives user data input and analysis commands, displays analysis results through a visual interface, and provides multiple interaction methods to meet the research needs of different scenarios. The user interaction module includes a data upload unit 21, a parameter configuration unit 22, a task triggering unit 23, a result display unit 24, and a status monitoring unit 25.

[0055] The data upload unit 21 is used to receive battery test data files in various formats, including BlueDian.nda format, CSV text format, TXT text format, etc., and performs format verification and preprocessing on the uploaded data files.

[0056] In some implementations, the data upload unit 21 receives the data file uploaded by the user through the file upload component, saves the file to a specified path (such as the data directory), and returns the file path identifier to the session state manager for subsequent analysis tasks. The data upload unit 21 also provides real-time feedback on the file upload status, displaying the file name and a success message to ensure that the user clearly perceives that the data has been successfully loaded.

[0057] The parameter configuration unit 22 is used to receive material parameters, experimental parameters, test conditions and journal styles set by the user, and to verify the validity of the parameters.

[0058] In some implementations, the parameter configuration unit 22 includes a material system configuration subunit, an experimental parameter configuration subunit, a test condition configuration subunit, and a journal style configuration subunit.

[0059] The material system configuration subunit is used to select the type of cathode material for the battery. This subunit provides a variety of predefined material options, offering corresponding material type selections based on different battery systems (such as lithium-ion batteries, sodium-ion batteries, solid-state batteries, etc.). The system automatically applies the corresponding material parameters, including theoretical capacity, operating voltage range, and elemental composition, based on the user-selected material type.

[0060] For example, for ternary cathode material systems in lithium-ion batteries, the material system configuration sub-unit can provide material options such as NCM523, NCM622, NCM811, and NCM111. Each material type corresponds to a preset Ni / Co / Mn composition ratio and expected capacity range. For example, the composition ratio of NCM523 is Ni:50%, Co:20%, and Mn:30%, with an expected capacity range of 155-165 mAh / g (4.3V cutoff voltage); the composition ratio of NCM811 is Ni:80%, Co:10%, and Mn:10%, with an expected capacity range of 190-210 mAh / g (4.3V cutoff voltage). For sodium-ion battery systems, the material system configuration sub-unit can provide options such as layered oxide materials and Prussian blue-based materials.

[0061] The experimental parameter configuration subunit is used to receive the active material mass parameter and set the number of activation cycles. The active material mass parameter is mandatory, expressed in milligrams (mg), with a value range of 0.0-100.0 and a step value of 0.1. If the user does not enter this parameter, the system displays a prominent error message, preventing the analysis task from executing; if the user enters this parameter, the system displays a confirmation prompt, ensuring the parameter is set correctly. The activation cycle number parameter ranges from 0-10, with a step value of 1. Activation cycles refer to the pretreatment cycles performed on the battery before formal testing, used to activate the electrode materials and stabilize battery performance. When the activation cycle number is greater than 0, the system will automatically add an activation cycle offset when accepting "post-activation" related tasks.

[0062] For example, when the number of activation circles is set to 3, if the user requests the data for the actual 4th circle corresponding to "the 1st circle after activation", the system will automatically perform the cycle number conversion.

[0063] The test condition configuration subunit receives test condition parameters such as charge / discharge rate, voltage range, test temperature, and charging mode. The charge and discharge rates range from 0.1 to 5.0C, with a step value of 0.1C; the lower and upper voltage limits can be adjusted according to different battery systems, with a step value of 0.1V; the test temperature ranges from -20 to 80°C, with a step value of 1°C; and the charging mode supports both constant current / constant voltage (CC-CV) and constant current (CC) options. These test condition parameters are used to generate a complete analysis report.

[0064] For example, for lithium-ion battery NCM materials, the typical range for the lower voltage limit is 2.0-3.5V, and the typical range for the upper voltage limit is 4.0-4.8V. For sodium-ion battery materials, the typical range for the lower voltage limit is 1.5-2.5V, and the typical range for the upper voltage limit is 3.5-4.5V.

[0065] The Journal Style Configuration subunit is used to select the plot style for the target journal. It offers a variety of predefined journal style options, including style standards from top journals such as Nature Energy, Science, and Advanced Energy Materials. Each style defines plotting parameters such as font type, font size, resolution, and curve color to ensure that the generated charts meet the submission requirements of the target journal. Users can also customize the chart style by entering corresponding standards.

[0066] The task triggering unit 23 is used to receive the user's analysis instructions and pass the instructions to the intelligent analysis module for execution. The task triggering unit 23 includes a quick example triggering subunit and a natural language triggering subunit.

[0067] The quick example trigger subunit provides several predefined quick operation buttons, each corresponding to a common analysis task. These quick operation buttons include, but are not limited to: initial coulomb efficiency query button, charge / discharge curve plotting for a specific number of cycles button, differential capacity analysis button, multi-cycle comparison button, post-activation evolution comparison button, and median voltage evolution analysis button. When a user clicks a quick operation button, the system automatically generates the corresponding analysis command and considers the activation cycle offset to convert the cycle number.

[0068] For example, when the user clicks the "Analysis of Differential Capacity in the First Activation Cycle" button, the system automatically generates the instruction "Plot the dQ / dV curve for the fourth cycle" based on the activation cycle number parameter (e.g., 3 cycles), ensuring that the cycle number analyzed correctly corresponds to the data after activation. When the user clicks the "Median Voltage of Cycles 1-100 After Activation" button, the system converts it to "Analyze the evolution of the median voltage from cycle 4 to cycle 103".

[0069] The natural language triggering subunit provides a text input box, allowing users to input analysis commands in natural language. The natural language triggering subunit supports both Chinese and English commands and can understand multiple expressions.

[0070] For example, users can input natural language commands such as "plot the charge-discharge curve of the 100th cycle", "plot the charge-discharge comparison chart of the 5th, 50th and 100th cycles", "analyze the median voltage of the 1st to 100th cycles after activation", and "how many oxidation peaks are there in the 1st cycle", and the system can understand and execute the corresponding analysis tasks.

[0071] In some implementations, after receiving a user instruction, the natural language triggering subunit first verifies whether the required parameters (such as the mass of the active substance) have been filled in. If the required parameters are not filled in, an error message is displayed and the task execution is blocked. If the required parameters are filled in, the user instruction, data file path, material parameters, experimental parameters, test condition parameters, journal style parameters, and other information are encapsulated into a task request and passed to the intelligent analysis module for execution.

[0072] Results display unit 24 is used to display the execution process and results of the analysis task, including a dialogue history display sub-unit and a chart preview sub-unit.

[0073] The dialogue history display sub-unit presents user commands and system responses in a chat-like format. This sub-unit maintains a message list, recording user-sent command messages and system-returned result messages in chronological order. User messages display the original text of the user-input command; system messages display different response content based on the task type. For question-and-answer tasks (such as the first Coulomb efficiency query), system messages display the specific text answer; for plotting tasks, system messages display a message indicating successful plot generation, including key information such as material system, plot path, and quality score.

[0074] In some implementations, the dialogue history display sub-unit displays a loading prompt "AI is processing, please wait..." during task execution. After the task is completed, the loading prompt is replaced with the actual result message to ensure that the user clearly understands the execution status of the task.

[0075] The chart preview sub-unit displays different types of analysis charts in the form of tabs, including charge / discharge curve tab, differential capacity analysis tab, and median voltage analysis tab.

[0076] The Charge / Discharge Curve tab displays charge / discharge curve graphs and supports PNG image previews. Once a charge / discharge curve graph is generated, the Charge / Discharge Curve tab automatically loads and displays the graph image; if the graph has not yet been generated, the tab displays a user guide to help users complete the steps of data preparation, parameter settings, and task triggering.

[0077] The Differential Capacity Analysis tab displays differential capacity graphs. These graphs include differential capacity curves (dQ / dV or dV / dQ) for both charging and discharging processes, along with corresponding electrochemical characterization. When the Differential Capacity Analysis option is enabled and a graph is generated, the tab automatically loads and displays the graph image. When the Differential Capacity Analysis option is disabled, the tab displays a description of its function, introducing its research value, analytical content, generation methods, and applicable scenarios.

[0078] For example, for NCM ternary materials, such as Figure 4 As shown, the differential capacity curve chart includes two sub-charts, left and right. The left chart shows the dQ / dV curve during the charging process (note: the horizontal line segment at the beginning is the result after noise is zeroed out and does not reflect the actual phase transition information), and the right chart shows the dQ / dV curve during the discharging process.

[0079] The Median Voltage Analysis tab is used to display median voltage analysis charts, such as... Figure 5 As shown in the figure, this example illustrates the trends in median charging voltage, median discharging voltage, and polarization voltage (taking the cycle from the 5th to the 100th cycle as an example). When the median voltage analysis option is enabled and the chart is generated, the median voltage analysis tab automatically loads and displays the chart image; when the median voltage analysis option is not enabled, the median voltage analysis tab displays a function description, introducing the content, evaluation criteria, generation method, and applicable scenarios of median voltage analysis.

[0080] In some implementations, the tab of the chart preview sub-unit is always visible, regardless of whether the chart has been generated, ensuring that users can view the status or usage guidelines of various types of charts at any time.

[0081] The user interaction module also includes a status monitoring unit 25, which displays the current values ​​of key parameters in real time. The status monitoring unit 25 displays key parameters such as material system, journal format, active substance quality, and number of activation zones in the form of indicator cards, allowing users to quickly confirm the current system configuration status.

[0082] Figure 2 This is a diagram of the intelligent analysis layer architecture of the battery electrochemical intelligent analysis system provided in this application embodiment. Figure 2 As shown, the intelligent analysis layer receives analysis tasks from the user interaction layer and can automatically complete the entire intelligent analysis process, including data processing, electrochemical parameter analysis, differential capacity analysis, chart plotting, quality review, report generation, and question-and-answer interaction. The intelligent analysis layer is built on the LangGraph state graph engine and adopts a two-layer architecture of coordination and execution layers. The coordination layer 31 implements the functions of the supervisory agent, responsible for intent parsing, execution plan generation, and task scheduling. The execution layer 32 includes a data processing unit 321, a data analysis unit 322, a plotting unit 323, a quality review unit 324, a report generation unit 325, and a question-and-answer unit 326, respectively implementing the functions of the data processing agent, median voltage analysis agent, dQ / dV differential analysis agent, data plotting agent, quality review agent, analysis report agent, and question-and-answer agent. Each unit is coordinated through the LangGraph state graph, and the coordination layer 31 dynamically schedules the various agent units in the execution layer according to user needs to complete the end-to-end intelligent analysis task.

[0083] In some implementations, the coordination layer 31 includes an intent parsing subunit 311, a plan generation subunit 312, and a task scheduling subunit 313.

[0084] The intent parsing subunit 311 is used to understand the user's analytical needs. This subunit analyzes the user's natural language input using a large language model to identify task types, extract key parameters, and detect analytical requirements. Task types include, but are not limited to: charge / discharge curve plotting, differential capacity analysis, median voltage analysis, data querying, and style adjustment. Key parameters include, but are not limited to: cycle number, material type, plot type, and journal style. Advanced analytical requirements (differential capacity analysis and median voltage analysis) are identified through keyword detection.

[0085] In some implementations, the intent parsing subunit 311 has a built-in error recovery mechanism for detecting and correcting format errors returned by the large language model. When the parameter format returned by the large language model does not meet expectations, the intent parsing subunit automatically attempts multiple correction strategies. For a loop number parameter, if a range format (such as "1-100", "1 to 100", "1 to 100") is returned, the intent parsing subunit calls a range expansion function to convert it into a complete list; if a comma-separated format (such as "1,10,100") is returned, all numbers are extracted and converted into a list; if a single number is returned, it is converted into a single-element list. If the correction is successful, the parsing result continues to be processed; if the correction fails, the default value is used or the task is terminated and the error is reported to the user.

[0086] For example, when a user inputs "Analyze the charge-discharge curves of NCM811 from cycle 1 to 100", the intent parsing sub-unit extracts the material type as NCM811, the cycle number range as 1-100 (expanded into a complete list [1,2,3, ..., 100]), and the task type as charge-discharge curve plotting. When the user inputs "How many oxidation peaks are there in cycle 1?", the intent parsing sub-unit identifies the task type as a data query, and the query content as the number of peaks in differential capacity analysis. When the large language model returns the cycle number parameter as the string format "1-100", the error recovery mechanism of the intent parsing sub-unit detects the format error, automatically calls the range expansion function to convert "1-100" into a complete list [1, 2, 3, 4, ..., 99, 100], a total of 100 elements, and then continues to process the parsing results.

[0087] The plan generation subunit 312 is used to generate an execution plan based on the task type and parameters. An execution plan is an ordered sequence of functional units that defines the task's execution flow. The plan generation subunit determines which functional units need to be called and in what order, based on the task requirements.

[0088] For example, for a charge / discharge curve plotting task, the execution plan might be: ["data_processor", "plotter", "reviewer","report"], indicating that data processing, plotting, and quality review are performed sequentially. For a task that includes differential capacity analysis, the execution plan might be: ["data_processor", "dq / dv_analyzer", "plotter", "reviewer","report"], performing differential capacity analysis first after data processing, and then plotting. For a combined task that includes both differential capacity analysis and median voltage analysis, the execution plan might be: ["data_processor", "median_voltage_analyzer", "dq / dv_analyzer", "plotter", "reviewer","report"], ensuring that all analyses are completed before plotting.

[0089] For example, consider a user data query task. If the system has cached the charge / discharge data for cycle 10, when a user asks "What is the specific capacity for cycle 10?", the intent parsing subunit detects that the data is cached, generates an execution plan ["qa_agent"], and directly reads the data from the cache to answer the question without needing to retrieve it again. If the system has cached the differential capacity analysis results, when a user asks "How many oxidation peaks are there?", the intent parsing subunit generates an execution plan ["qa_agent"] and directly reads the peak information from the cache to answer the question. If no data is cached, a complete execution plan including data extraction and analysis steps is generated.

[0090] The plan generation sub-unit 312 follows these principles when generating the execution plan: the data processing unit must be executed before all analysis and plotting units to ensure that the data is ready; all analysis units must be executed before plotting units to ensure that the analysis results can be used for plotting; the quality review unit must be executed after the plotting unit to evaluate the quality of the charts and generate analysis reports; and the data query unit must be completed after the analysis and plotting units to ensure that there are accurate data analysis results and plotting results to reply to the user.

[0091] The task scheduling subunit 313 is used to schedule each functional unit sequentially according to the execution plan. The task scheduling subunit 313 maintains the current execution state, including information such as the current step, the current functional unit, and completed steps. After each functional unit completes execution, the task scheduling subunit determines the next functional unit based on the execution plan and the current state, and passes the task context information.

[0092] In some implementations, the task scheduling subunit 313 uses a LangGraph state graph with nodes and edges for task scheduling. The system defines a task state data structure, including fields such as user input, file path, material parameters, execution plan, current step, extracted data, analysis results, graph path, and error information. Each functional unit acts as a state transition node, receiving the current state as input, executing the corresponding processing logic, updating the state, and returning. The task scheduling subunit 313 determines the next state transition based on a routing function.

[0093] The data processing unit 321 is used to load test data files, verify cycle numbers, extract charge and discharge data for specified cycles, and convert the data into the format required for analysis and plotting.

[0094] In some implementations, the data processing unit 321 employs a tool-based agent architecture. The data processing unit 321 is configured with a set of data processing tools, including file loading tools, loop query tools, loop verification tools, and data extraction tools. The data processing unit 321 understands task requirements through a large language model, autonomously selects and invokes appropriate tools to complete the data processing task.

[0095] The file loading tool is used to read battery test data files. It supports multiple file formats, including .nda format, CSV text format, and TXT text format. The tool parses the file content to extract raw battery test data, including data columns such as time, voltage, current, and capacity.

[0096] The loop lookup tool is used to retrieve a list of available loop numbers from a file. It analyzes the loop identifiers in the file and returns all available loop numbers along with their basic information (such as loop type, start and end times).

[0097] The loop validation tool verifies whether a user-requested loop number exists in a file. It takes a list of loop numbers as input, checks the availability of each number, and returns the validation result. If some loop numbers are missing, the tool returns a warning message suggesting the user adjust the loop number range.

[0098] The data extraction tool is used to extract charge and discharge data for a specified cycle. It receives a list of cycle numbers and active material quality parameters, extracts the raw data for the corresponding cycle from the file, performs data cleaning and preprocessing, calculates specific capacity, distinguishes between charging and discharging processes, and stores the processed data in a global cache. The tool supports flexible cycle number formats, capable of parsing single cycle numbers, comma-separated multiple cycle numbers, continuous ranges, and other input formats.

[0099] For example, for NCM811 material, the data extraction tool calculates the specific capacity based on an active material mass of 5.2 mg. Assuming the total charge capacity in the fourth cycle is 1.067 mAh, then the charge specific capacity = 1.067 mAh / 5.2 mg = 205.2 mAh / g. The data extraction tool stores the charge and discharge data in a voltage-specific capacity format, with charging data sorted from low voltage to high voltage and discharging data sorted from high voltage to low voltage.

[0100] The data analysis unit 322 is used to perform in-depth electrochemical data analysis. The data analysis unit 322 includes a differential capacity analysis subunit and a median voltage analysis subunit.

[0101] The differential capacity analysis subunit is used to calculate differential capacity curves (dQ / dV or dV / dQ) to identify the electrochemical characteristics of materials. The subunit reads charge-discharge data for a specified cycle from the global cache, performs numerical differentiation calculations on the voltage-capacity curves to obtain the differential capacity curves, smooths the curves to reduce noise, detects characteristic peaks in the curves, identifies parameters such as peak position (voltage value), peak intensity, and peak area, matches characteristic peaks with corresponding electrochemical processes according to the material type, and stores the analysis results in the global cache for use by the plotting unit.

[0102] For example, for NCM series ternary materials, the differential capacity analysis subunit identifies the phase transition characteristic peaks of the material. NCM materials undergo multiple crystal structure phase transitions during charge and discharge, which are represented by characteristic peaks on the dQ / dV curve. The differential capacity analysis subunit matches the detected peaks to the corresponding phase transition types based on the voltage range of the peak positions. For NCM811 material, during charging, the peak in the voltage range of 3.55-3.75V matches the H1→M phase transition (hexagonal phase 1 to monoclinic phase), and the peak in the voltage range of 3.90-4.10V matches the M→H2 phase transition (monoclinic phase to hexagonal phase 2). During discharging, the peak in the voltage range of 3.80-4.00V matches the H2→M phase transition, and the peak in the voltage range of 3.50-3.70V matches the M→H1 phase transition. For NCM622 material, the phase transition peak voltage range is between that of NCM523 and NCM811, and the position of each phase transition peak decreases slightly overall with increasing nickel content. For NCM523 material, due to its lower nickel content, the phase transition behavior is more stable. The charging H1→M peak is about 3.60-3.80V, and the M→H2 peak is about 3.95-4.15V. The position and intensity of the peaks change little during cycling.

[0103] In some implementations, the differential capacity analysis subunit uses a Savitzky-Golay filter to smooth the dQ / dV curve, and uses the scipy.signal.find_peaks function to detect peaks, setting the prominence threshold to 12% of the maximum value of the dQ / dV curve and the peak height threshold to 20% of the maximum value of the dQ / dV curve to filter out significant characteristic peaks.

[0104] The median voltage analysis subunit is used to calculate parameters such as median voltage, polarization voltage, average voltage, and voltage efficiency during the charge and discharge process, and to evaluate the battery's internal resistance and kinetic performance. The median voltage analysis subunit reads charge and discharge data from multiple cycles in the global cache, calculates the median charging voltage (voltage value when capacity reaches 50%), the median discharging voltage (voltage value when capacity reaches 50%), and the polarization voltage (median charging voltage minus median discharging voltage) for each cycle, stores the calculation results as structured data, and generates an analysis report.

[0105] In some implementations, the median voltage analysis subunit assesses the severity of polarization based on the polarization voltage value and classifies it into multiple levels. A polarization voltage less than 50mV is rated as "Excellent," indicating excellent kinetic performance; a polarization voltage in the range of 50-100mV is rated as "Good," indicating good performance; a polarization voltage in the range of 100-200mV is rated as "Average," indicating the need for optimization; a polarization voltage in the range of 200-300mV is rated as "Poor," indicating high internal resistance; and a polarization voltage greater than 300mV is rated as "Severe," indicating significant limitations. The median voltage analysis subunit describes the polarization level for each cycle in the analysis report, identifies stages of polarization deterioration, and provides optimization suggestions.

[0106] For example, for NCM811 material, the median voltage analysis subunit adjusts the evaluation criteria according to the material characteristics. Because the kinetics of high-nickel NCM811 material are relatively slow, the polarization voltage is typically in the range of 50-120 mV, and a rating of "Good" to "Average" is considered normal performance. The median voltage analysis subunit states in its analysis report: "The polarization voltage of the NCM811 material used at 0.5C rate is 58-70 mV, rated 'Good,' which is consistent with the typical performance of high-nickel materials."

[0107] The plotting unit 323 is used to generate professional charts that conform to the standards of scientific research journals based on the processed data and analysis results. The plotting unit 323 includes a charge-discharge curve plotting subunit, a differential capacity curve plotting subunit, and a median voltage curve plotting subunit.

[0108] The charge / discharge curve plotting subunit is used to generate charge / discharge curve charts. This subunit supports two modes: single-cycle charge / discharge curves and multi-cycle comparison curves. For single-cycle mode, the subunit reads the charge / discharge data for a specified cycle from the global cache, plots the charging and discharging curves in the same coordinate system, labels key data points, configures plotting parameters such as font, line width, and color according to the journal's style, and saves the data in both PNG and PDF formats. For multi-cycle comparison mode, the subunit reads data from multiple cycles from the global cache, configures plotting parameters such as font, line width, and color according to the journal's style, uses different colors or line styles to distinguish different cycles, adds a legend to indicate the cycle number corresponding to each curve, and saves the data in both PNG and PDF formats. The journal style settings are shown in the Nature Energy journal example below.

[0109] JOURNAL_STYLES= {

[0110] "nature_energy": {

[0111] "figure_size": {

[0112] "single_column": (90 / 25.4, 90 / 25.4 0.75), # 90mm × 67.5mm (Nature standard single column) "double_column": (180 / 25.4, 135 / 25.4) # 180mm × 135mm (Nature standard double column) }, "fonts": { "family": "Arial", # Nature requires Arial, while Liberation Sans is a Linux rollback. "axis_label": 7, "tick_label": 6, "legend": 6 }, "lines": { "width": 1.0, "axis_width": 0.75 }, "colors": { # Nature's recommended colorblind-friendly color scheme (Wong color scheme) "charge": "#0173B2", # Nature Blue (deep blue, colorblind friendly) "discharge": "#DE8F05", # Nature Orange (Dark Orange, Colorblind Friendly) "polarization": "#029E73", # Nature Green (cyan-green, used for polarization voltage) "multi_cycle": "viridis", # Gradient color scheme (colorblind friendly) "highlight": "#CC78BC", # magenta (for emphasis) "background": "#ECE133" # Yellow (used for background annotation) }, "dpi": 300 }} The differential capacity curve plotting subunit is used to generate differential capacity curve charts. It reads the differential capacity analysis results from the global cache, including differential capacity curve data and characteristic peak information. The subunit creates a chart layout containing two subplots: the left subplot displays the differential capacity curve during the charging process, and the right subplot displays the differential capacity curve during the discharging process.

[0113] The median voltage plotting subunit is used to generate median voltage analysis charts. It reads the median voltage analysis results from the global cache, creates a chart layout containing multiple subplots, and displays information such as the evolution of median charging and discharging voltage and changes in polarization voltage.

[0114] The quality review unit 324 is used to evaluate the quality of the generated charts and graphs, verify the rationality of the analysis results, and provide quality evaluation information for report generation. The quality review unit 324 includes a chart quality assessment subunit and a material characteristic verification subunit.

[0115] The chart quality assessment subunit is used to score the quality of charts according to predefined scoring criteria. These criteria include, but are not limited to: font size and clarity, axis label integrity, legend position and readability, data point density and distribution, color contrast, line thickness, and resolution. The chart quality assessment subunit scores each item, calculates the total score, and assigns a quality level (e.g., Excellent, Good, Needs Improvement). The subunit stores the assessment results in a global cache for use by the report generation unit 325.

[0116] The material characteristic verification subunit is used to verify whether the analysis results conform to the known characteristics of the material. Based on the material type, the subunit checks whether key performance indicators are within a reasonable range. This reasonable range is combined with pre-set hard-coded parameters and existing knowledge from the large model to identify abnormal data or analysis results, providing warning information. The material characteristic verification subunit stores the verification results and research suggestions in a global cache for use by the report generation unit 325.

[0117] The aforementioned hard-coded parameters are all derived from publicly published electrochemical literature and industry standard test specifications. Taking NCM523 as an example, its theoretical specific capacity of 278 mAh / g is calculated based on Faraday's law for the lithium intercalation reaction of layered oxides; the actual specific capacity range (155–165 mAh / g, cutoff voltage 4.3 V) and the first-cycle irreversible capacity loss (approximately 22 mAh / g) are typical values ​​obtained by referring to numerous experimental reports of similar materials; the first-cycle coulombic efficiency range (84%–90%) and the lower limit of normal cycle coulombic efficiency (99%) are derived from the factory specifications and cycle performance test data of commercial NCM523 materials; the phase transition voltage range (H1→M: 3.60–3.80 V, M→H2: 3.95–4.15 V, H2→H3: 4.15–4.25 V) and the characteristic peak positions of dQ / dV (charging peaks approximately 3.70 V, 3.90 V, 4.20 V, discharge peaks approximately 3.65 V, 3.85 V, 4.15 V) are also analyzed. V) The parameters were determined with reference to publicly available research results from X-ray diffraction and in-situ electrochemical testing. These parameters are stored as ranges rather than single values ​​to accommodate reasonable fluctuations between different batches of materials and under different testing conditions. The parameters for NCM622 and NCM811 were determined using the same methodology, with the upper limit of capacity and the phase transition voltage position adjusted accordingly as the nickel content increases.

[0118] In practical analysis, the parameters from the aforementioned literature sources serve only as initial reference benchmarks. The system employs a three-step mechanism of "benchmark verification—deviation quantification—combined judgment" to integrate them with the current measured data. First, the material characteristic verification subunit uses the voltage range and capacity interval in the parameter configuration as benchmarks to perform compliance verification on the measured data, determining whether the current test conditions fall within the reasonable range of known material properties. Second, for cases where the measured values ​​deviate from the benchmark interval, the system does not directly classify them as abnormal but quantifies the deviation magnitude and assesses the reasonableness of the deviation in conjunction with the user-configured test conditions (such as cutoff voltage, charge / discharge rate, and test temperature). Finally, when generating analysis reports and research suggestions, the large language model uses the measured data as the primary basis and the benchmark interval in the parameter configuration as the comparison background, clearly explaining the differences between the measured values ​​and typical values ​​in the literature and the possible reasons in the report, avoiding directly outputting literature parameters as analysis conclusions.

[0119] The report generation unit 325 integrates analysis results, quality assessment information, and research recommendations to generate a complete structured analysis report. The report generation unit 325 includes a content integration subunit and a report writing subunit.

[0120] The content integration subunit is used to collect and organize various information required for report generation. It obtains test condition parameters (material system, active material mass, charge / discharge rate, voltage range, test temperature, charging mode, etc.) from the user interaction module, reads cycle data statistics (number of cycles, capacity range, coulombic efficiency, etc.) from the cache of the data processing unit 321, reads advanced analysis results (median voltage evolution data, polarization voltage assessment, dQ / dV peak information, phase transition characteristics, etc.) from the cache of the data analysis unit 322, obtains a list of chart file paths from the plotting unit 323, and reads chart quality scores and material characteristic verification results from the cache of the quality review unit 324. The content integration subunit categorizes and organizes this information according to the report's chapter structure, providing structured data input for report writing.

[0121] The report writing sub-unit is used to generate formatted analytical report text according to a predefined report template. Based on a large language model, the report writing sub-unit automatically writes the text content for each chapter according to the structured data provided by the content integration sub-unit. The report content includes, but is not limited to: a data overview chapter (basic information such as material system, test cycle range, active material mass, and test conditions); an electrochemical performance analysis chapter (first-cycle charge / discharge capacity, first coulombic efficiency, capacity-voltage characteristics, etc.); a cycle performance evaluation chapter (capacity decay trend, capacity retention, cycle stability, etc.); an advanced analysis results chapter (dQ / dV differential capacity analysis results including phase transition peak information, median voltage, and polarization analysis results including polarization evolution trend); a figure and graph quality evaluation chapter (quality scores and ratings for each figure and graph); and a research suggestion chapter (journal submission suggestions, figure caption suggestions, and discussion points).

[0122] For example, for the comprehensive analysis of NCM811 material, the report writing sub-unit generates the following chapters: Chapter 1: Data Overview (Material system: NCM811; test cycles from 1 to 100 after activation; active material mass: 5.2 mg; test conditions: 0.5C charge / discharge, 3.0-4.3V voltage range, 25°C temperature, CC-CV charging mode); Chapter 2: Electrochemical Performance Analysis (First-cycle charge specific capacity: 205.2 mAh / g; Discharge specific capacity: 188.3 mAh / g; Initial coulombic efficiency: 91.7%); Chapter 3: Cyclic Performance Evaluation (Final-cycle charge specific capacity: 198.6 mAh / g; Discharge specific capacity: 179.5 mAh / g). The data is presented in Chapter 4, which analyzes the differential capacity of NCM811 materials (charge process: H1→M peak 3.712V and M→H2 peak 4.063V; discharge process: H2→M peak 3.952V and M→H1 peak 3.668V; all phase transition peak positions are consistent with typical characteristics of NCM811 materials). Chapter 5 analyzes the median voltage and polarization (initial polarization 58 mV rated good, final polarization 70 mV rated good, polarization increase 20.7% within the normal range). Chapter 6 evaluates the quality of the graphs (charge-discharge comparison graph 95 points rated excellent, dQ / dV curve 92 points rated excellent, median voltage evolution graph 94 points rated excellent). Chapter 7 provides research recommendations (recommendation to Nature Energy or Energy Storage Materials, with suggestions for figure captions and discussion points).

[0123] The question-and-answer unit 326 is used to process user data queries and battery knowledge question-and-answer requests, providing data information query services through natural language interaction. The question-and-answer unit 326 includes a cached data reading subunit and a natural language generation subunit.

[0124] The cached data reading subunit is used to retrieve the data required by the user's query from the global cache. Based on the cycle number and parameter type of the user's query, the cached data reading subunit locates the corresponding cached data and reads the raw data, analysis results, or chart information. The cached data reading subunit supports querying various data types, including cycle charge / discharge capacity, coulombic efficiency, specific capacity, voltage plateau, median voltage, polarization voltage, dQ / dV peak information, phase transition characteristics, etc.

[0125] For example, when a user asks "What is the Coulomb efficiency for the 100th cycle?", the cached data reading subunit checks if data with the key "cycle_100_data" exists in the global cache. If it exists, the cached data reading subunit reads the relevant data and returns it to the natural language generation subunit for further processing. If it does not exist, the cached data reading subunit returns a null value, and the task coordination unit routes the task to the data processing unit 321 to extract the data first.

[0126] The Natural Language Generation (NLP) subunit is used to convert data and computation results into natural language responses. Based on a large language model, the NLP subunit generates response text that conforms to natural conversational habits, according to the type and tone of the user's question. For simple numerical queries, the NLP subunit generates concise numerical responses; for analytical questions, it provides explanatory responses based on data trends; and for basic battery knowledge questions, it uses the knowledge base of the large language model to generate informative responses.

[0127] For example, when a user asks "What is the coulombic efficiency on the 100th cycle?", the natural language generation subunit generates the answer: "The coulombic efficiency on the 100th cycle is 97.6%." When a user asks "What is the capacity decay trend of this material?", the natural language generation subunit reads capacity data from multiple cycles in the cache, calculates the capacity retention rate, analyzes the decay rate, and generates a descriptive answer: "From cycle 1 to cycle 100, the material's capacity retention rate is 95.2%, with slow overall decay. It is basically stable in the first 50 cycles, and shows slight decay in the last 50 cycles, which is normal cycle performance." When a user asks "What is coulombic efficiency?", the natural language generation subunit generates an educational answer: "Coulombic efficiency is the ratio of battery discharge capacity to charge capacity, reflecting how much electricity can be effectively released during charging. The higher the coulombic efficiency, the better the battery's reversibility and the fewer side reactions."

[0128] The intelligent analysis module utilizes the LangGraph state graph engine to achieve multi-agent collaboration. The LangGraph state graph engine defines a state graph where nodes represent functional units and edges represent task flow relationships. The engine supports conditional edges, dynamically determining the next node based on the task state and execution result. It also supports cyclic edges, allowing tasks to return to previous nodes for re-execution under specific conditions, such as returning to the task coordination unit during error recovery. The engine maintains a task state data structure, passing states between nodes to ensure the continuity of the task context. Table 1 shows the configuration table of LangGraph state graph nodes and edges, and the flow states between internal nodes (i.e., each agent) are as follows. Figure 3 As shown.

[0129] Table 1: LangGraph State Graph Configuration Table

[0130] For example, the basic workflow of the intelligent analysis module is as follows: The coordination layer 31 receives tasks from the user interaction module, parses the user's intent, and generates an execution plan; The task scheduling subunit 313 determines the task type based on the execution plan: if it is a question-and-answer query task, the task is directly routed to the question-and-answer unit 326; if it is a data analysis task, the task is routed to the data processing unit 321. Data processing unit 321 loads the file, verifies the loop number, extracts the data, stores the results in the global cache, and updates the task status. If the execution plan includes advanced analysis, the task scheduling subunit 313 routes the task to the data analysis subunit 322. The analysis subunit reads data from the cache, performs analysis, stores the results in the cache, and updates the task status. The task scheduling subunit 313 routes the task to the plotting unit 323. The plotting unit reads data and analysis results from the cache, generates a chart, saves the chart file, and updates the task status. Task scheduling subunit 313 routes the task to quality review unit 324. The quality review unit evaluates the quality of the chart, verifies the material characteristics, stores the evaluation results and research suggestions in the cache, and updates the task status. The task scheduling subunit 313 routes the task to the report generation unit 325. The report generation unit integrates all analysis results and quality assessment information, generates a structured analysis report, saves the report file, and updates the task status. The coordination layer 31 determines that the task is completed and returns the result to the user interaction module.

[0131] Within the architecture of the intelligent analysis module, the system can automatically complete the entire intelligent analysis process from data loading to report generation and question-and-answer interaction without human intervention. Through multi-agent collaboration and error recovery mechanisms, the system possesses strong robustness and adaptability, capable of processing test data from various battery systems, supporting a variety of advanced analysis functions, generating professional charts and reports that meet the standards of scientific research journals, and providing natural language question-and-answer services, significantly improving the efficiency and quality of battery data analysis.

[0132] The output module receives the analysis results generated by the intelligent analysis module, converts the results into various formats, and outputs them to different interfaces of the user interaction module through various channels. The output module includes a chart output unit 41, a report output unit 42, and a question-and-answer output unit 43.

[0133] The chart output unit 41 is used to process the chart files generated during the analysis process, supports multiple image formats and vector formats, and provides a unified chart access interface.

[0134] In some implementations, the chart output unit 41 receives the chart file path list generated by the drawing unit 323 of the intelligent analysis module, reads the binary data of the chart file, performs format conversion or compression processing according to the requirements of the target output channel, generates chart metadata (including file name, file size, resolution, creation time, quality score, etc.), encapsulates the chart binary data and metadata into a unified chart output object, and returns it to the user interaction module.

[0135] The chart output unit 41 supports different image formats. For the web interface, PNG format is provided for preview and PDF format for download; for the IDE interface, PNG format can be embedded in the document; for the terminal interface, a file path string is provided; for file system export, both PNG and PDF formats are provided.

[0136] For example, Figure 6 This is a schematic diagram of the 100th charge-discharge curve generated by the system. For example, Figure 7 The diagram shows a comparison of the multi-cycle charge-discharge curves generated by the system, illustrating the comparison of charge-discharge curves for different cycles such as the 5th, 50th, and 100th cycles in the same coordinate system.

[0137] The report output unit 42 is used to receive the report content generated by the intelligent analysis layer, generate report files in different formats, and pass them to the user interaction layer.

[0138] In some implementations, the report output unit 42 reads the report content and metadata stored in the report generation unit 325 from the global cache, supporting conversion between Markdown and PDF formats.

[0139] After completing the format conversion, the report output unit 42 saves the generated report file to the specified path, generates file metadata (including file name, file size, file format, generation time, etc.), encapsulates the report file path and metadata into a unified report output object, and returns it to the user interaction layer. Users can preview or download the report on the interface.

[0140] The question-and-answer output unit 43 is used to receive the question-and-answer results generated by the intelligent analysis layer, format them, and pass them to the user interaction layer for display.

[0141] In some implementations, the question-and-answer output unit 43 reads the natural language answer text and related data generated by the question-and-answer unit 326 from the global cache and performs format processing. The formatted question-and-answer result is then encapsulated into a unified output object, containing metadata such as message text, message type, timestamp, and related data references, and passed to the result display unit 24 of the user interaction layer. The result is then presented in the dialogue history display subunit in the form of a dialogue, achieving a natural and smooth question-and-answer interaction experience.

[0142] Example 1: Comparative Analysis of Multi-Cycle Charge-Discharge Curves

[0143] Example 1 provides a process for applying the battery electrochemical intelligent analysis system to multi-cycle charge-discharge comparison. The user uploads the NCM811 battery test data file (Neware format .nda file) through the user interaction layer, and inputs "generate charge-discharge curve comparison charts for the 5th, 20th, 50th, and 90th cycles" via natural language.

[0144] After receiving the instruction, the task triggering unit of the user interaction layer passes it to the intelligent analysis layer. The coordination layer 31 of the intelligent analysis layer parses the user intent through the large language model and identifies the task type as "multi-cycle charging and discharging comparison" with the target cycle number as [5, 20, 50, 90].

[0145] The data processing unit 321 extracts charge and discharge data for a specified cycle from the uploaded .nda file. For each cycle, the data processing unit 321 performs the following operations: identifies the charging segment (current > 0) and the discharging segment (current < 0), calculates the specific capacity based on the active material mass set by the user, and sorts the voltage-specific capacity data, with the charging segment sorted in ascending order of voltage and the discharging segment sorted in descending order of voltage.

[0146] The plotting unit 323 receives the processed data and creates a single coordinate system, with the horizontal axis representing specific capacity (mAh / g) and the vertical axis representing voltage (V). Based on the user-selected journal style (e.g., Nature Energy), the plotting unit 323 automatically applies appropriate parameters such as font, line width, and color scheme to generate a chart that conforms to the journal's submission standards. The plotting unit 323 sequentially plots four charge-discharge curves, using different colors to distinguish each cycle; solid lines are used for charging, and dashed lines are used for discharging.

[0147] The chart quality assessment subunit of Quality Review Unit 324 evaluates the generated charts, checking items including: axis label integrity, legend clarity, curve discrimination, and resolution compliance, giving a quality score of 95 points and a rating of "Excellent". The material characteristic verification subunit of Quality Review Unit 324 verifies whether the capacity values ​​of the four cycles are within the reasonable range of NCM811 material, generating a research recommendation: "The material has excellent cycling stability and is recommended for submission to Nature Energy journal".

[0148] The report generation unit 325's content integration subunit collects test condition parameters, cycle data statistics, chart file paths, and the evaluation results from the quality review unit 324. The report writing subunit integrates this information to generate a structured analysis report. The report content includes: a data overview describing the material system NCM811, test cycles 5, 20, 50, and 90, and test conditions; a performance analysis comparing key parameters such as charge / discharge capacity, coulombic efficiency, and voltage plateau across the four cycles, analyzing the capacity evolution trend; a chart quality evaluation indicating a quality score of 95 and compliance with Nature Energy journal standards; and research recommendations providing suggestions for future research directions based on the cycle performance characteristics. The report generation unit 325 stores the report content in a global cache.

[0149] The chart output unit 41 in the output layer receives the chart file generated by the plotting unit 323, encapsulates it in format, and supports PNG and PDF output. The report output unit 42 reads the report content generated by the report generation unit 325 from the cache, performs format conversion, generates a Markdown file, and converts it to PDF format as needed. The output layer passes the processed charts and reports to the user interaction layer, which displays chart previews and report summaries on the interface and provides download buttons for the charts and the full report. Users can download the chart files in PNG or PDF format, and the report files in Markdown or PDF format.

[0150] In this embodiment, the system achieves automated generation of data files into high-quality comparative charts and structured analysis reports through the collaboration of various levels and units, without requiring users to manually write code or adjust plotting parameters.

[0151] Example 2: Median Voltage Evolution Analysis

[0152] This embodiment illustrates the process of a battery electrochemical intelligent analysis system performing median voltage analysis over a specified cycling range. The user inputs the natural language command "Analyze the median voltage evolution from cycle 11 to cycle 100" through the user interaction layer.

[0153] The coordination layer 31 of the intelligent analysis layer parses the instruction, determines the analysis type as "median voltage evolution", generates an execution plan including two steps: data processing and median voltage analysis, and routes the task to the data processing unit 321.

[0154] After receiving the task, the data processing unit 321 checks whether data for cycles 11 to 100 already exists in the global cache. If the cache is empty, the data processing unit 321 extracts the specified range of cyclic data from the original data file. The data processing unit 321 performs the following operations: a file loading tool reads the data file and parses the format; a cycle query tool confirms the availability of cycles 11 to 100 in the file; a cycle verification tool checks the data integrity of each cycle (e.g., whether there are charging and discharging segments, and whether the number of data points is reasonable); and a data extraction tool extracts the charging and discharging data of each cycle and calculates the specific capacity. The data processing unit 321 cleans the extracted data, including removing duplicate data points, interpolating and filling missing values, and removing abnormal voltage or current abrupt changes. The data processing unit 321 stores the processed data of 90 cycles in the global cache and updates the task status.

[0155] According to the execution plan, the coordination layer 31 routes the task to the median voltage analysis subunit of the data analysis unit 322.

[0156] The median voltage analysis subunit reads the charge / discharge data from cycles 11 to 100 from the buffer of the data processing unit 321. For each cycle, the median voltage analysis subunit performs the following calculation steps: (1) Interpolate the charging specific capacity-voltage data to determine the voltage value when the charging specific capacity reaches 50%, and use it as the charging median voltage; (2) Interpolate the discharge specific capacity-voltage data to determine the voltage value when the discharge specific capacity reaches 50%, and use it as the median discharge voltage; (3) Calculate the polarization voltage, which is equal to the difference between the midpoint charging voltage and the midpoint discharging voltage; (4) Determine the severity level of polarization based on the range of polarization voltage.

[0157] The median voltage analysis subunit performs trend analysis on the polarization voltage over 90 cycles to identify the evolution stages of the polarization voltage. The median voltage analysis subunit stores the analysis results in a global cache.

[0158] The plotting unit 323 generates evolution curves based on the median voltage analysis results. The charts include: curves showing the change of charging median voltage with cycle number, the change of discharging median voltage with cycle number, and the change of polarization voltage with cycle number. The plotting unit 323 uses different colors to distinguish the three curves, generates a chart file, and saves it.

[0159] The chart quality assessment subunit of quality review unit 324 evaluates the generated charts, checking items including: completeness of axis labels, clarity of legends, curve discrimination, and resolution compliance, and provides a quality score. The material characteristic verification subunit verifies whether the polarization voltage evolution trend conforms to the typical characteristics of the material and generates research suggestions.

[0160] The report generation unit 325's report writing subunit integrates the analysis results and quality assessment results to generate a structured analysis report. The report content includes: a data table (containing median charge and discharge values, polarization voltage, and severity level for 90 cycles), an evolution trend description (cycle range and polarization voltage changes at each stage), a graph quality evaluation (quality score and rating), and a comprehensive evaluation (dynamic performance assessment of the material and research recommendations). The report generation unit 325 stores the report content in a global cache.

[0161] The output layer's chart output unit 41 receives chart files generated by the plotting unit 323 in PNG and PDF formats. The report output unit 42 reads report content from the cache, performs format conversion, and generates report files in Markdown and PDF formats. The output layer passes the processed charts and reports to the user interaction layer, which displays chart previews and report summaries on the interface and provides download buttons for the charts and the full report. Users can download chart files in PNG or PDF format, and report files in Markdown or PDF format.

[0162] In this embodiment, through the collaboration of various levels and units, the system can automatically complete the batch calculation, trend identification, and visualization output of median voltage, providing users with in-depth analysis of battery polarization characteristics.

[0163] Example 3: Single-loop data query

[0164] This embodiment illustrates the process by which a battery electrochemical intelligent analysis system queries a specific parameter for a single cycle. The user inputs a natural language question, "What is the coulombic efficiency in cycle 100?" through the user interaction layer. The task triggering unit of the user interaction layer receives this question and passes it to the intelligent analysis layer. The coordination layer 31 of the intelligent analysis layer parses the user's intent using a large language model, identifying the query type as "single cycle parameter query," the target cycle number as 100, and the query parameter as "coulombic efficiency."

[0165] Coordination layer 31 checks the global cache to determine whether the data for the 100th lap has been retrieved. If the data already exists in the cache, coordination layer 31 routes the task to question-and-answer unit 326; if the data does not exist, coordination layer 31 first routes the task to data processing unit 321, which retrieves the data for the 100th lap from the original data file and stores it in the global cache, and then routes it to question-and-answer unit 326.

[0166] The cache data reading subunit of question-answering unit 326 reads relevant values ​​from the cache. The natural language generation subunit encapsulates the result into a natural language answer: "The Coulomb efficiency of the 100th cycle is 97.6%", and stores the answer text in the cache.

[0167] The question-and-answer output unit 43 in the output layer reads the answer text generated by the question-and-answer unit 326 from the cache, formats it into a dialogue message object, and includes metadata such as the message text "The Coulomb efficiency of the 100th cycle is 97.6%", the message type "numerical query", and a timestamp, and passes it to the user interaction layer. The user interaction layer displays the answer in the dialogue history display sub-unit, and the user can view the result directly on the interface.

[0168] If the user continues to ask related questions (such as "What is the charging capacity on the 100th cycle?"), the intelligent analysis layer can directly read the data from the cache without repeated extraction, thereby improving the response speed.

[0169] In this embodiment, the system can quickly respond to users' single-loop parameter query requests through the collaboration of each layer and unit, realizing data access in a natural language question-and-answer manner.

[0170] Example 4: dQ / dV Differential Capacity Analysis

[0171] This embodiment illustrates the process of a battery electrochemical intelligent analysis system performing dQ / dV differential capacity analysis on a specified cycle. The user uploads NCM622 battery test data through the user interaction layer and inputs the command "Perform dQ / dV analysis for cycle 1".

[0172] The coordination layer 31 of the intelligent analysis layer parses the instruction, determines the analysis type as "dQ / dV differential capacity analysis", the target cycle number as cycle 1, and the material type as NCM622. The coordination layer 31 generates an execution plan, which includes five steps: data processing, dQ / dV analysis, plotting, quality review, and report generation.

[0173] Data processing unit 321 extracts the charge / discharge data for the first cycle from the global cache or the original data file. If the data is not in the cache, data processing unit 321 performs file loading, loop verification, and data extraction operations, cleans the data, and stores the processed data in the global cache. Coordination layer 31 routes the task to the differential capacity analysis subunit of data analysis unit 322.

[0174] The differential capacity analysis subunit reads the charge and discharge data of the first cycle from the cache. The following steps are performed on the charging and discharging processes respectively: (1) Numerical differentiation: The voltage-specific capacity data is numerically differentiated to obtain the dQ / dV curve. The central difference method is used to calculate dQ / dV, that is, for the intermediate data point i, the formula for calculating dQ / dV is: dQ / dV = (Q[i+1] - Q[i-1]) / (V[i+1] - V[i-1]); the first and last endpoints are treated by forward or backward difference as boundaries; (2) Smoothing: The original dQ / dV curve is smoothed and filtered, for example, the Savitzky-Golay filter is used, the default window size is 51 and the polynomial order is 3. When the actual number of data points is insufficient, the window size is automatically reduced to remove high-frequency noise; (3) Peak detection: Extreme points are detected on the smoothed dQ / dV curve as candidate peaks. The peak detection algorithm sets a threshold parameter to filter out peaks with sufficiently high significance; (4) Peak screening: Further screening of the detected candidate peaks removes false peaks with too low peak intensity or too small peak area, and retains the real phase transition characteristic peaks; (5) Peak matching: According to the material type (e.g., NCM622) and the voltage position of the peak, the detected peaks are matched to the corresponding phase transition types. For example, for NCM622 material, the peak matching in the range of 3.59-3.79V during charging is H1→M phase transition, and the peak matching in the range of 3.94-4.14V is M→H2 phase transition.

[0175] The differential capacity analysis subunit calculates the following parameters for each identified peak: peak voltage (the voltage value corresponding to the peak apex), peak intensity (the dQ / dV value of the peak apex), peak area (the integral area around the peak), and peak width (half-peak width). The differential capacity analysis subunit evaluates the significance of the peak based on these parameters and generates a significance score. The differential capacity analysis subunit stores the analysis results in a global cache.

[0176] During the execution of the above steps, the system implements an adaptive adjustment mechanism for key parameters such as the smoothing window size and peak detection threshold to ensure the robustness of the analysis under different data quality and material systems.

[0177] For adaptive adjustment of smoothing parameters: Before performing Savitzky-Golay filtering, the numerical micro-unit first obtains the actual number of data points n_raw after data cleaning. If n_raw is less than the current smoothing window smooth_window, the system automatically adjusts smooth_window to max(5, n_raw / / 2), and further selects the nearest odd number to meet the filter requirements; if the adjusted smooth_window is not greater than the polynomial order polyorder plus 2, then the polyorder is simultaneously reduced to smooth_window minus 1 to ensure that the parameter combination is always valid. The above adjustments only take effect in the current analysis, the system default parameters are unaffected, and each adjustment is logged for reference in subsequent reports.

[0178] Adaptive rollback for peak detection thresholds: The peak detection subunit employs a two-stage detection strategy. In the first stage, the peak detection algorithm is invoked using default thresholds (peak height ratio 0.20, significance ratio 0.12, minimum peak spacing 50 data points). If no candidate peaks are detected in the first stage, the system automatically enters the second stage, reducing the peak height threshold to 15% of the maximum dQ / dV value, the significance threshold to 8%, and the minimum peak spacing halved (not less than 10 data points), and re-executes peak detection. If no peaks are detected in either stage, the system returns an empty list and notes this in the analysis report, without interrupting the overall analysis process. When rollback is triggered, the quality review unit will correspondingly lower the rating of the peak detection confidence sub-item in the chart quality score.

[0179] For adaptive matching of material phase transition voltage range: The peak matching subunit calculates the voltage offset based on the Ni content (ni_content) obtained from the material name using the formula v_offset = -0.15 × (ni_content - 0.5), and shifts each phase transition voltage window (H1→M, M→H2, H2→H3, etc.) by this offset. This offset pattern is derived from the statistical analysis of phase transition voltages of NCM materials with different Ni contents, allowing the system to automatically adapt to different series of materials such as NCM523, NCM622, and NCM811 without requiring manual configuration of the voltage range by the user. For high-nickel materials such as NCM811, the system also increases the maximum number of retained peaks to 5 to match additional phase transition peaks that may appear at high cutoff voltages.

[0180] For statistical adaptive handling of outliers in the dQ / dV curve: The system dynamically sets the outlier truncation threshold to the mean plus three times the standard deviation, based on the mean and standard deviation of the current curve. Truncation is only performed when the 99th percentile value exceeds this threshold; otherwise, the original curve amplitude is preserved intact. This strategy avoids the over-truncation problem caused by using a fixed threshold for curves with significant differences in dQ / dV amplitudes under different material systems and testing magnifications.

[0181] Coordination layer 31 routes the task to plotting unit 323. Plotting unit 323 generates a bi-subplot based on the dQ / dV analysis results: the left subplot displays the charging dQ / dV curve, and the right subplot displays the discharging dQ / dV curve. Plotting unit 323 creates a coordinate system with voltage (V) on the horizontal axis and dQ / dV (mAh / g / V) on the vertical axis. In each subplot, plotting unit 323 draws a smoothed dQ / dV curve (using a solid blue line for charging and a solid orange line for discharging). Plotting unit 323 automatically applies parameters such as font, line width, and color scheme according to the journal style selected by the user.

[0182] The chart quality assessment subunit of quality review unit 324 evaluates the generated dQ / dV chart, checking items including curve smoothness, axis label standardization, and resolution compliance. Quality review unit 324 scores each item according to preset scoring criteria, generating an overall quality score and rating. The material characteristic verification subunit verifies whether the identified phase transition peaks conform to the known characteristics of NCM622 material and provides material analysis suggestions.

[0183] The report generation unit 325's report writing subunit integrates the analysis results and quality assessment results to generate a structured analysis report. The report content includes: analysis cycle information (material type NCM622, cycle number 1), phase transition characteristics during charging (number of detected peaks, peak voltage, peak intensity, peak area, phase transition type, and significance score for each peak), phase transition characteristics during discharging (also including detailed parameters for each peak), electrochemical reaction mechanism analysis (assessing the material's reversibility and kinetic properties based on peak symmetry and voltage difference), graph quality evaluation (quality score of the dQ / dV graph and compliance with journal standards), and research suggestions (journal submission suggestions and discussion points). The report generation unit 325 stores the report content in a global cache.

[0184] The chart output unit 41 in the output layer receives the dQ / dV chart file generated by the plotting unit 323, processes its format, and supports PNG and PDF output formats. The report output unit 42 reads the report content from the cache, performs format conversion, generates a Markdown format file, and converts it to PDF format as needed. The output layer passes the processed charts and reports to the user interaction layer. Users can download the charts in PNG or PDF format, and the report files in Markdown or PDF format.

[0185] In this embodiment, through the collaboration of various layers and units, the system can automatically complete the entire process from data extraction, dQ / dV curve calculation, phase transition characteristic peak identification to visualization output and analysis report generation, providing users with in-depth analysis of the electrochemical reaction mechanism of materials.

[0186] Example 5: Process Branch Decision and Error Recovery Task

[0187] This embodiment illustrates the automatic handling process of the battery electrochemical intelligent analysis system when two types of anomalies occur: intent parsing format errors and substandard chart quality. It demonstrates the system's node scheduling, state propagation, conditional branching, and error recovery mechanisms based on the LangGraph state graph engine. Users upload NCM811 battery test data files through the user interaction layer and input "Draw a comparison chart from cycle 1 to 100, and provide median voltage and differential capacity analysis" via natural language.

[0188] After receiving the instruction, the task triggering unit 23 of the user interaction layer passes it to the intelligent analysis layer 3. The coordination layer 31 of the intelligent analysis layer (i.e., the SupervisorAgent node) calls the intent parsing subunit 311 to construct a prompt word by combining the user instruction with the current cached state information (all three caches are empty), and calls the large language model for intent parsing. Since the user instruction contains the range expression "cycle 1 to cycle 100", the large language model returns the cycle number field as a range string "1-100" instead of the integer list format required by the system. When the intent parsing subunit performs format validation on the returned result, it detects that the cycle number field type is incorrect, immediately triggers the built-in error recovery mechanism, tries various format matching modes on the returned content, identifies the start value 1 and the end value 100 in the range string, expands it into a complete list containing 100 consecutive integers, and performs defensive validation on the corrected result again to confirm that the format is correct.

[0189] If the above repair fails, the intent parsing subunit 311 appends the error message and format example to the prompt for the next call and writes the retry flag to the task status. The routing function of the state graph reads the retry flag from the task status after each node execution: if the retry flag is true, the process loops back to the coordinating layer node for re-execution; if the retry flag is false and the error field is empty, it routes to the next node according to the execution plan; if the error field is not empty, it routes to the termination node. The system allows a maximum of three retries. After exceeding this limit, the error reason is written to the task status and routed to the termination node, reporting the parsing failure reason to the user. In this example, the format repair succeeded on the first attempt, the retry flag was set to false, and the error field remained empty.

[0190] After the coordination layer 31 completes intent parsing, it writes the following content into the task status: the task type is charge / discharge curve plotting and joint analysis; the loop number list consists of 100 integers from loop 1 to loop 100; the dQ / dV analysis flag and the median voltage analysis flag are both set to true; and the execution plan consists of the following nodes in sequence: Data ProcessorAgent, MedianVoltageAnalyzerAgent, dQ / dV Analysis Node, PlotterAgent, ReviewerAgent, and ReportAgent. The routing function reads the first node in the execution plan and routes the process to the data processing unit 321.

[0191] The specific execution processes of the data processing unit 321, the median voltage analysis subunit, and the differential capacity analysis subunit are as described in Examples 1, 2, and 4, respectively, and will not be repeated here. After each unit completes its execution, the current node identifier is written into the task status. The routing function determines the next node and completes the routing based on the current node identifier in the task status and the execution plan. After the above three units complete their execution, the key fields in the task status are updated as follows: the first cache layer stores the raw charging and discharging data from rounds 1 to 100, and the second cache layer stores the median voltage and polarization analysis results for 100 cycles and the dQ / dV peak identification results for round 1, respectively. The routing function routes the process to the plotting unit 323.

[0192] The plotting unit 323 reads the dQ / dV analysis flag and the median voltage analysis flag from the task status. After confirming that both are true, it reads the above two analysis results from the second cache layer and generates a multi-turn charge-discharge comparison chart, a median voltage evolution curve, and a dQ / dV differential capacity curve in sequence, and applies the journal style specification. The plotting unit 323 writes the generated chart file paths to the task status and performs an existence check on each chart file. If all charts are successfully generated, the routing function routes the process to the quality review unit 324, which performs a quality assessment on all charts. Then, the report generation unit 325 integrates all analysis results to generate a complete report. If some chart files are missing, the system writes the path information of the failed charts to the error field of the task status, while retaining the paths of the successfully generated charts. When the routing function detects that the error field is not empty and the current node is the plotting unit, it routes the process to the quality review unit 324 to perform a quality assessment on the successfully generated charts. Then, the report generation unit 325 integrates the analysis results of the successful charts and the error information of the failed charts into the analysis report, explaining to the user the reasons for the failure of some charts and suggesting handling methods. If all charts fail to generate, the routing function directly routes the process to the report generation unit 325, which integrates all error information into the analysis report and explains the reasons for the failure to the user. After the report is generated, it routes to the termination node, and the task ends.

[0193] The chart quality assessment subunit of quality review unit 324 evaluates the multi-cycle charge-discharge comparison chart according to predefined scoring criteria. The inspection items include: axis label integrity, legend clarity, data point density and distribution, color contrast, line thickness and resolution, etc. Due to the overlay of 100 charge-discharge curves, the data point density is too high and the curves overlap severely, resulting in low scores for data point density and color contrast. The overall quality score is 58 points, lower than the system's preset acceptable threshold of 75 points. Upon detecting that the score is below the threshold, quality review unit 324 activates the process branch decision function, judging based on the specific reasons for the substandard score: the root cause of the low score is that the user selected 100 consecutive cycles for overlay plotting, which is a data selection strategy problem. Re-executing the plotting unit cannot improve the chart quality. Therefore, the system selects the "generate a report containing quality warnings" branch instead of the "trigger re-plotting" branch, and writes the overall score, detailed scores, and branch decision conclusions into the third cache layer of the task status. The material characteristic verification subunit verifies whether the dQ / dV peak position and polarization voltage evolution trend conform to the known characteristics of NCM811 material, and generates research suggestions. The routing function routes the process to the report generation unit 325.

[0194] The processing steps of the report generation unit 325, the output layer, and the user interaction layer are as described in Example 1, and will not be repeated here. Users can view previews of the three charts, quality warning prompts, and the complete report on the interface, and resubmit the task after adjusting the cycle number selection based on the improvement suggestions in the report.

[0195] In this embodiment, the system achieves automatic correction and retry control of large language model format errors through the built-in error recovery mechanism and node loopback routing of the intent parsing subunit; it achieves dynamic scheduling and status transmission between nodes by continuously reading the retry flag, current node identifier, execution plan, and error field in the task status through the routing function; it ensures the orderly flow of data between nodes through a three-layer caching mechanism; and it achieves dynamic process adjustment based on intermediate result evaluation through the process branch decision function of the quality review unit, automatically selecting the output branch with a warning report instead of triggering invalid redraw when the comprehensive quality score does not meet the threshold. The above mechanisms correspond to and support the dynamic process adjustment capability and the fault tolerance capability, respectively.

[0196] In summary, the battery electrochemical intelligent analysis system and method proposed in this invention, driven by the LangGraph state graph engine, constructs a fully closed-loop automated analysis system encompassing natural language intent parsing, adaptive electrochemical analysis, intelligent quality review, and structured report generation. Its intent parsing subunit's built-in error recovery and retry mechanism, combined with the conditional branching and loop control of the state graph routing function, endows the system with strong robustness against unstructured inputs and abnormal processes. The differential capacity analysis subunit's data-dependent smoothing window adjustment, two-stage peak detection, and dynamic phase transition voltage matching based on the nickel content of the material ensure accurate feature identification across material systems and data quality levels. The scoring threshold and branch decision logic introduced by the quality review unit intelligently selects report warnings rather than invalid redraws based on the reasons for non-compliance, avoiding process rigidity. The collaborative design of the three-layer caching architecture and the question-answering unit enables immediate reuse of intermediate results and natural language interactive queries. Therefore, this invention not only completely solves the fundamental problems of switching between multiple software programs, cumbersome manual operation, and high professional threshold in traditional battery data analysis, but also significantly improves the efficiency, standardization level, and analytical depth of battery material research and development and performance characterization with an intelligent, adaptive, and fault-tolerant systematic solution.

Claims

1. A battery electrochemical intelligent analysis system based on charge-discharge curve data, characterized in that, include: The user interaction layer is used to receive charging and discharging data files and parameter configurations uploaded by users, and to obtain user analysis requirements input in natural language. The intelligent analysis layer, built around the LangGraph framework, adopts a two-layer architecture of coordination and execution. It manages and coordinates multiple agents through a state graph. These agents include at least a supervisory agent, a data processing agent, an electrochemical analysis agent, a data plotting agent, a quality review agent, an analysis report agent, and a question-and-answer agent. This layer is used to complete the entire intelligent analysis task from data processing, electrochemical parameter analysis, chart plotting, quality review, report generation to question-and-answer interaction. as well as The output layer receives charts, reports, and Q&A results generated by the intelligent analysis layer, converts them into formats, and then passes them to the user interaction layer for display or download.

2. The system according to claim 1, characterized in that, The user interaction layer includes: The data upload unit is used to receive and verify charge and discharge data files in various formats; The parameter configuration unit is used to receive user-set material parameters, experimental parameters, test conditions, and journal styles. The task triggering unit supports receiving user analysis commands through either quick operations or natural language input. The results display unit is used to show the execution process and results of the analysis task in the form of dialogue history and chart previews; and The status monitoring unit is used to display and confirm the current configuration status of key parameters of the system in real time.

3. The system according to claim 1, characterized in that, In the intelligent analysis layer: The coordination layer includes a task coordination unit that implements the functions of the supervisory agent, and the task coordination unit further includes: The intent parsing subunit is used to analyze natural language instructions through a large language model to identify task types and extract key parameters. The intent parsing subunit has a built-in error recovery mechanism. When the returned loop number parameter is detected to be a range string, it is automatically expanded into a list of consecutive integers and format validation is performed. If the repair fails, the error message is appended to the prompt word and a retry flag is set. The process is looped back to the coordination layer for re-execution through the routing function of the state graph until the retry limit is reached and the task is terminated. The plan generation subunit is used to generate an ordered sequence of functional units as an execution plan based on the task type; and The task scheduling subunit is used to schedule the various units of the execution layer according to the execution plan. The execution layer includes: The data processing unit implements the functions of the data processing intelligent agent, and is used to load files, verify loop numbers, extract data, and perform volume comparison calculations. The data analysis unit integrates the functions of the median voltage analysis agent and the dQ / dV differential analysis agent, and is used to calculate electrochemical parameters and identify phase transition characteristics. The drawing unit implements the functions of the data drawing agent and is used to generate charts that conform to academic standards. The quality review unit implements the functions of the quality review agent, and is used to evaluate the quality of charts and verify material characteristics. The report generation unit implements the functions of the analysis report agent, used to integrate results and generate a structured report; and The question-answering unit implements the functions of the question-answering agent, and is used to read data from the cache and generate natural language answers.

4. The system according to claim 3, characterized in that, The data analysis unit includes: The median voltage analysis subunit is used to calculate the median voltage and polarization voltage during the charge and discharge process, and to classify the severity level according to the polarization voltage value to evaluate the battery kinetic performance; and The differential capacity analysis subunit is used to calculate the dQ / dV differential capacity curve and implement an adaptive adjustment mechanism, including: automatically adjusting the smoothing window size according to the actual number of data points; adopting a two-stage peak detection strategy, automatically lowering the threshold and re-detecting if no peak is detected in the first stage; shifting each phase transition voltage window according to the nickel content resolved by the material name using a preset offset formula to match different series of materials; and dynamically setting an outlier truncation threshold based on the curve mean and standard deviation, and determining whether to perform truncation based on the comparison result between the upper tail quantile of the curve and the truncation threshold, and only performing truncation when the value exceeds the threshold.

5. The system according to claim 3, characterized in that, The quality review unit includes: The chart quality assessment subunit is used to score the quality of charts based on predefined scoring criteria including font, labels, legend, data point density, color contrast, and resolution, and to set a passing threshold. The materials characteristic verification subunit combines the analysis results with benchmark parameters preset based on industry standards. Through benchmark verification and deviation quantification, it verifies whether key performance indicators are within reasonable ranges and generates research recommendations. The process branch decision function is activated when the chart quality assessment subunit detects that the overall score is lower than the qualified threshold. Based on the specific reason for the score failure, if it is determined that the data selection strategy is the cause of the failure to be improved by redrawing, the "Generate a report containing quality warnings" branch is selected, and the score details and branch decision conclusion are written to the task status instead of triggering the redrawing branch.

6. The system according to claim 3, characterized in that, The system employs a three-layer caching mechanism, including: The first cache layer is used to store the cyclic data extracted by the data processing unit, including charge / discharge capacity, coulombic efficiency, and specific capacity. The second cache layer is used to store the analysis results of the data analysis unit, including median voltage evolution data, polarization voltage assessment, and dQ / dV peak information; and The third cache layer is used to store the evaluation results of the quality review unit and the report content of the report generation unit; The question-and-answer unit retrieves data from the corresponding cache layer based on the user's query through the cache data reading subunit, and the natural language generation subunit converts it into a natural language answer.

7. The system according to claim 1, characterized in that, The intelligent analysis layer achieves multi-agent collaboration through the LangGraph framework, and its technical features include: The state graph definition mechanism defines the execution flow of an agent through nodes and edges. Nodes represent agents or processing functions, and edges represent state transition paths. The task status data structure should at least include user input, file path, execution plan, current step, extracted data, analysis results, chart path, retry flag, and error field; The routing function is used to read the retry flag, current node identifier, execution plan and error field in the task status after each node execution is completed, so as to dynamically determine the next node to be executed: if the retry flag is true, it loops back to the coordination layer node; if the error field is not empty, it checks whether it is an unrecoverable error or other conditions that do not meet the conditions for further normal operation. If so, it routes to the termination node; otherwise, it routes to the next node according to the execution plan. Conditional routing mechanism is used to dynamically determine the next node to be executed based on the current state and execution result; A loop control mechanism is used to support the process in returning to the previous node for re-execution under specific conditions; and The checkpoint mechanism is used to save a snapshot of the state during process execution.

8. The system according to claim 1, characterized in that, The system also has dynamic process adjustment capabilities, including: The intermediate result evaluation function is used to evaluate the quality of intermediate results during task execution; The workflow branching decision function dynamically selects the execution path based on task type or intermediate result quality. When a chart quality score is detected to be below a threshold and determined to be a data selection strategy issue, it chooses to generate a report branch containing a quality warning instead of triggering a redrawing branch. The early termination function is used to terminate task execution when an unrecoverable error is detected and the conditions for further normal operation are no longer met.

9. The system according to claim 1, characterized in that, The system also has fault tolerance capabilities, including: Error detection functionality is used to monitor the execution results of each functional unit in order to detect format errors or missing data; The automatic retry function is used to automatically retry when a format error is detected. The error message and format example are appended to the prompt word, and the process is looped back to the coordination layer node for re-execution through the routing function of the state diagram. A maximum of a preset number of retries are allowed. The checkpoint saving function is used to automatically save a snapshot of the task state after execution at each node through the checkpoint mechanism of the LangGraph framework, ensuring the traceability of the task execution process; and The error reporting function is used to record and report error information to the user.

10. A battery electrochemical intelligent analysis method based on charge-discharge curve data, using the system described in any one of claims 1 to 9, characterized in that, The method includes: A: Receives user-uploaded charging / discharging data files and parameter configurations through the user interaction layer, and obtains user analysis requests input in natural language. B: The coordination layer of the intelligent analysis layer parses the user's intent through a large language model, and has a built-in error recovery mechanism to repair or trigger retries for returned format errors, generating an execution plan containing an ordered sequence of functional units; C: According to the execution plan, the execution layer of the intelligent analysis layer first calls the data processing unit to load and verify the data file, extract the charge and discharge data of the specified cycle and calculate the specific capacity, and store the processed data into the first cache layer; D: If the execution plan includes advanced analysis, the data analysis unit is invoked to analyze the cached data from the perspective of median voltage and / or differential capacity, and a processing mechanism including adaptive adjustment of smoothing window, adaptive backoff of peak detection threshold and adaptive matching of material phase transition voltage range is implemented, and the analysis results are stored in the second cache layer. E: Call the plotting unit to read data and analysis results from the cache and generate scientific charts that conform to the preset journal style; F: Call the quality review unit to score the quality of the generated charts. If the score is below the threshold and it is determined to be a data selection strategy problem, select to generate a report branch containing a quality warning, and at the same time perform material feature verification and generate research suggestions. G: Invoke the report generation unit to integrate all analysis results, quality assessment information, and research recommendations to generate a structured analysis report; and H: The output layer converts the format of the generated charts, reports, and Q&A results generated by the Q&A unit, and displays or downloads them through the user interaction layer.

Images