Battery K value real-time prediction system and battery processing system

Abstract

The invention provides a battery K value real-time prediction system and a battery processing system, the prediction system comprises a plurality of artificial intelligence Agents respectively corresponding to different task types, and the artificial intelligence Agents are respectively a data acquisition Agent, a feature extraction Agent and a prediction Agent which are in communication connection in sequence; the data acquisition Agent is used for sensing and acquiring multi-dimensional state information in a battery operation environment so as to obtain original data; the feature extraction Agent is used for converting the original data into a feature set capable of representing battery performance so as to obtain feature data; and the prediction Agent is used for executing K value prediction based on the feature data and the learned prediction strategy so as to obtain a predicted K value. According to the battery K value real-time prediction system, a multi-agent cooperation mechanism is formed, each Agent focuses on a specific task, complex changes of a battery operation environment can be dynamically adapted through division of labor, cooperation and information transmission, the data processing flow is optimized, the calculation complexity is reduced, and the real-time requirement is met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery parameter prediction, and in particular to a battery K value real-time prediction system and a battery production system. BACKGROUND

[0002] With the promotion of global energy transformation and sustainable development, lithium batteries, as the core technology in the field of modern energy storage, are widely used in portable electronic devices, electric vehicles, and renewable energy storage systems. Their performance directly affects the efficiency, cost, and user experience in related application fields. Among the many performance indicators of lithium batteries, K value (voltage decay rate) as a key parameter reflecting battery performance degradation has important research and application value. K value can quantify the voltage decay rate of the battery during charging and discharging from a microscopic level, providing fine feature support for battery state of health (SOH), remaining useful life (RUL) prediction, and production quality control. However, existing K value prediction techniques face multiple challenges in industrial scenarios, limiting their accuracy, real-time performance, and interpretability.

[0003] Traditional prediction models have significant shortcomings when dealing with lithium battery K value prediction. Specifically, long short-term memory network (LSTM) as a classic time series model can handle long-term dependencies, but in industrial environments, time series data often appears discontinuous due to equipment failure or process adjustment, making it difficult for LSTM to accurately capture data correlations and reducing prediction accuracy. The Transformer model performs well in various fields with its self-attention mechanism, but it is limited in feature extraction ability when faced with complex noise interference in industrial data, making it difficult to accurately grasp the performance change rules of the battery. Models such as LightGBM based on tree ensemble have theoretical bottlenecks in feature interaction depth, making it difficult to fully explore the complex nonlinear relationships between voltage, current, temperature, and other factors, limiting the improvement of prediction performance.

[0004] Overall, traditional prediction methods rely on a single model or algorithm, lack collaboration mechanisms, and are difficult to adapt to dynamic changes in battery operating environments and complex working conditions, resulting in insufficient robustness and real-time performance of prediction results.

[0005] In addition, lithium battery production requires higher real-time monitoring and prediction. Production line real-time prediction needs to complete K value prediction in a very short time (usually less than 50ms) to support timely adjustment of production parameters. However, traditional computing architectures have computational complexity and data transmission bottlenecks when dealing with complex data, making it difficult to meet real-time requirements. SUMMARY

[0006] The purpose of the present application is to provide a battery K value real-time prediction system and a battery production system that improve the accuracy, real-time performance, and interpretability of battery K value prediction.

[0007] In order to achieve the above object, the application provides a battery K value real-time prediction system based on multi-agent cooperation, which comprises a plurality of artificial agents (Agents) corresponding to different task types, namely a data collection Agent, a feature extraction Agent and a prediction Agent connected in sequence in communication. The data collection Agent is used for perceiving and acquiring multi-dimensional state information in the battery operating environment to obtain original data. The feature extraction Agent is used for converting the original data into a feature set capable of representing the battery performance to obtain feature data. The prediction Agent is used for performing K value prediction based on the feature data and a learned prediction strategy to obtain a predicted K value.

[0008] Preferably, the multi-dimensional state information collected by the data collection Agent includes at least two of voltage, current, temperature and internal resistance, and the data collection Agent is further used for cleaning and denoising the collected data.

[0009] Preferably, for any task type, a plurality of Agents with the same function are created, and the plurality of Agents with the same function obtain tasks based on an auction algorithm.

[0010] Preferably, Agents with different functions belonging to different task types obtain computing resources based on an auction algorithm.

[0011] Preferably, the prediction Agent is constructed based on a reinforcement learning model of a Q-Learning algorithm.

[0012] Preferably, the prediction Agent adaptively selects a prediction model according to the characteristics of the current feature data.

[0013] Preferably, when the data amount of the feature data is small and the features are simple, a support vector machine model is selected as the prediction model; when the data amount is large and there are complex time series features, a recurrent neural network model is selected as the prediction model.

[0014] Preferably, the prediction Agent fuses a recurrent neural network model and a support vector machine model for prediction processing, the recurrent neural network model is used for processing the feature data to capture dynamic features and long-term dependencies in the data, and the support vector machine model is used for classifying and predicting adjacent features output by the recurrent neural network model to obtain a predicted K value.

[0015] The application further provides a battery process system based on battery K value real-time prediction, which comprises the battery K value real-time prediction system and a decision Agent. The decision Agent generates a production optimization decision according to the predicted K value output by the prediction Agent, to drive real-time adjustment and quality control of the battery process.

[0016] Compared with the prior art, the battery K value real-time prediction system provided by the above technical solution of the application forms a multi-agent collaboration mechanism through the sequential communication connection of the data collection Agent, the feature extraction Agent and the prediction Agent. Each Agent focuses on a specific task (data collection, feature extraction, K value prediction), and through division of labor and information transmission, it can dynamically adapt to the complex changes of the battery operating environment. Based on the learned prediction strategy, the prediction Agent can effectively cope with challenges such as data discontinuity and noise interference in industrial scenarios, significantly improving the robustness and stability of the prediction result; in addition, through the division of labor mechanism of multi-agent collaboration, the data collection, feature extraction and prediction task modules are processed modularly, optimizing the data processing process and reducing the computational complexity. The independent operation and efficient communication of each Agent reduce the bottleneck of data transmission, so that the system can realize fast response in industrial scenarios and meet the real-time requirements, providing reliable support for timely adjustment of production parameters. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The battery K value real-time prediction system principle structure diagram in the embodiment of the application. DETAILED DESCRIPTION

[0018] To explain the technical content, structural features, achieved purposes and effects of the application in detail, the following will be described in detail in combination with the embodiments and the accompanying drawings.

[0019] The embodiment discloses a battery K value real-time prediction system based on multi-agent collaboration, to solve the technical problem that the existing battery K value prediction technology is limited in precision, real-time performance and interpretability in industrial scenarios.

[0020] As shown in the figure, the battery K value real-time prediction system in the embodiment includes a plurality of artificial intelligence Agents (Agents) corresponding to different task types, which are sequentially connected in communication, namely a data collection Agent, a feature extraction Agent and a prediction Agent.

[0021] For the data collection Agent, it is used to perceive and acquire multi-dimensional state information in the battery operating environment to obtain raw data. Specifically, the data collection Agent interacts with various sensors to collect key parameters in the battery charging and discharging process in real time.

[0022] For the feature extraction Agent, it is used to transform the original data into a set of features that can represent the battery performance, to obtain feature data, and provide high-resolution input information for prediction. Specifically, the feature extraction Agent uses various data processing techniques (such as time domain statistics, frequency domain analysis) to extract key features (such as voltage mean, variance, frequency energy distribution, etc.) that reflect the battery performance.

[0023] For the prediction Agent, it is used to perform K value prediction based on the feature data and the learned prediction strategy, to obtain the predicted K value.

[0024] The embodiment forms an efficient multi-agent collaboration mechanism through the sequential communication connection of the data collection Agent, the feature extraction Agent, and the prediction Agent, significantly improving the robustness and real-time performance of the battery K value real-time prediction. Each Agent focuses on a specific task, achieving effective task division. The data collection Agent can dynamically adapt to the complex changes of the battery operating environment, ensuring the accuracy of the original data. The feature extraction Agent can transform the original data into a high-quality feature set, effectively dealing with industrial scene challenges such as data discontinuity and noise interference. The prediction Agent can output more accurate and stable predicted K values based on these high-quality feature data. This collaboration mechanism optimizes the data processing process and reduces the computational complexity. The independent operation and efficient communication of each Agent reduce the bottleneck of data transmission, enabling the system to respond quickly in industrial scenarios and meet the demand of less than 50ms for battery K value production line real-time prediction, providing reliable support for timely adjustment of production parameters, thereby improving production efficiency and product quality.

[0025] On the other hand, the multi-dimensional state information collected by the data collection Agent includes at least two of voltage, current, temperature, and internal resistance. The data collection Agent is also used to clean and denoise the collected data.

[0026] The working principle of the data collection Agent is that it not only perceives multi-dimensional state information in the battery operating environment through sensors, such as voltage, current, temperature, internal resistance, etc., but also has a data preprocessing module built in. When the sensors collect these raw data in real time, the data cleaning module will identify and process the outliers, missing values and redundant information in the data. For example, for voltage or current readings that are outside the normal range, the system will mark or correct them; for intermittent data interruptions, interpolation method will be used for filling. The data denoising module uses filtering algorithms (such as Kalman filtering, wavelet denoising or moving average filtering) to eliminate sensor noise and environmental interference, ensuring the smoothness and authenticity of the data. Through cleaning and denoising, the data collection Agent can provide purer and more reliable raw data, providing high-quality input for the feature extraction Agent and the prediction Agent, thereby significantly improving the accuracy and robustness of the entire K-value prediction system.

[0027] In addition, the data cleaning and denoising functions can be partially or entirely deployed on edge computing devices to achieve localized preprocessing of data, reducing data transmission delay and network bandwidth occupancy.

[0028] On the other hand, for any task type, multiple functionally identical Agents are created, and the multiple functionally identical Agents obtain tasks based on an auction algorithm.

[0029] In complex industrial environments, a single Agent may face a bottleneck in computing resources or task processing delays, especially under high concurrency and high real-time requirements. To improve the concurrent processing capability and robustness of the system, this embodiment introduces a multi-Agent parallel processing mechanism and uses an auction algorithm for task allocation. For any task type, such as data collection, feature extraction or K-value prediction, the system creates multiple functionally identical Agents. These Agents register their capabilities and available states when the system starts. When a new task arrives, the task scheduling module will start an "auction" process. Each functionally identical Agent acts as a "bidder" and submits a "bid" (i.e., indicates its willingness and ability to handle the task) to the task scheduling module based on its current load, processing capacity and historical performance. The task scheduling module acts as an "auctioneer" and allocates tasks to the "winning" Agent according to the pre-set auction rules (e.g., selecting the Agent with the fastest response, lowest load or highest historical success rate). This task acquisition mechanism based on the auction algorithm can dynamically allocate tasks to the most suitable Agent, thereby avoiding overloading of a single Agent and improving the parallel processing capability of the task and the overall throughput of the system. At the same time, when a certain Agent fails, other functionally identical Agents can quickly take over its tasks, ensuring the fault tolerance and high availability of the system.

[0030] In another aspect, functionally different Agents belonging to different task types acquire computing resources based on an auction algorithm.

[0031] In a multi-agent system, Agents of different task types (e.g., data collection Agent, feature extraction Agent, prediction Agent) have different demands for computing resources, and the resources are limited. Traditional static resource allocation can lead to uneven resource utilization or bottlenecks, affecting the overall performance of the system. The present embodiment introduces a resource acquisition mechanism based on an auction algorithm, enabling functionally different Agents to dynamically and efficiently compete for and acquire the required computing resources.

[0032] When the system starts or an Agent needs additional computing resources, each functionally different Agent submits a "resource request" to a central resource manager (or resource scheduling Agent) according to its current task demand and priority.

[0033] The resource manager will start a "resource auction" process. Each Agent acts as a "bidder" and "bids" according to its urgency for resources, task priority, and expected resource usage efficiency. For example, a prediction Agent with high real-time requirements may submit a higher "virtual price" to bid for CPU and memory resources, while a background data cleaning Agent may have a higher demand for storage resources. The resource manager, as an "auctioneer", allocates computing resources to the "winning" Agents according to the pre-set resource allocation strategy (e.g., maximizing system throughput, minimizing task delay, or guaranteeing resources for critical tasks). This mechanism ensures that critical tasks have priority resources, optimizes overall resource utilization, avoids resource waste and conflicts, and improves the response speed and stability of the system under complex working conditions.

[0034] On the other hand, a reinforcement learning model based on Q-Learning algorithm is used to construct the prediction Agent.

[0035] In this embodiment, the prediction Agent is a reinforcement learning model based on the Q-Learning algorithm. The model includes a state perception module, an action selection module, and a Q-value update module. The function of the state perception module is to receive the feature data provided by the feature extraction Agent, which constitutes the state s of the reinforcement learning model. The action selection module selects a prediction action a based on the current state s and the current Q-value table (or Q-network) through a certain strategy (such as the ε-greedy strategy). The action can be to select a specific prediction model (such as a support vector machine model or a recurrent neural network model), adjust the model parameters, or directly output a predicted K value. For example, if the current feature data indicates that the battery is in a specific aging stage, the action selection module may select a model that is more accurate for predicting this stage. The Q-value update module calculates the reward r based on the error between the actual K value and the predicted K value after the Agent performs the action a and obtains the prediction result. Then, using the Bellman equation and parameters such as learning rate α and discount factor γ, the Q(s, a) value in the Q-value table is updated.

[0036] This process enables the prediction Agent to gradually optimize its prediction strategy through continuous trial and error and learning, so that it can select the optimal prediction action under the characteristics of the given feature data, thereby obtaining the most accurate predicted K value. Through the adaptive learning ability of the Q-Learning algorithm, the prediction Agent can better adapt to the dynamic changes of the battery operating environment, improving the robustness and real-time performance of the prediction.

[0037] On the other hand, the prediction Agent adaptively selects a prediction model according to the characteristics of the current feature data.

[0038] Specifically, when the amount of data in the feature data is small and the features are relatively simple, a support vector machine model (SVM) is selected as the prediction model; when the data amount is large and has complex time series features, a recurrent neural network model is selected as the prediction model.

[0039] When the amount of data in the feature data is small and the features are relatively simple, for example, in the initial stage of battery performance degradation, the model adaptive selection module of the prediction Agent will preferentially select a support vector machine model. The support vector machine model has good generalization ability on small sample and high-dimensional data, can effectively capture the nonlinear relationship in the data, and has relatively low computational complexity, which can quickly give the prediction result.

[0040] When the data volume of feature data is large and has complex time series characteristics, for example, during the long-term operation of a battery, its performance degradation presents complex dynamic changes and long-term dependencies, the model adaptive selection module of the prediction Agent will select a recurrent neural network model. Recurrent neural network models, especially long short-term memory networks (LSTM) or gated recurrent units (GRU), are good at processing sequence data and can capture time dependencies and long-term patterns in data, thus showing stronger prediction ability on complex time series data.

[0041] This strategy of dynamically selecting a prediction model based on data characteristics ensures that the prediction Agent can perform optimally in different data scenarios, improving the accuracy and efficiency of K value prediction.

[0042] More specifically, the model adaptive selection module of the prediction Agent includes a data characteristic analyzer and a model switcher. The data characteristic analyzer performs real-time evaluation on the feature data transmitted by the feature extraction Agent, analyzing its data volume (for example, by counting the number of data points) and feature complexity (for example, by calculating the correlation, nonlinearity, or time dependence between features). When the data characteristic analyzer determines that the current feature data has a small data volume and simple features, for example, the number of data points is below a threshold and the feature dimension is low and changes smoothly, the model switcher will route the prediction task to a support vector machine model.

[0043] When the data characteristic analyzer determines that the current data volume is large and has complex time series characteristics, for example, the number of data points exceeds a threshold and the data presents obvious time dependence and volatility, the model switcher will route the prediction task to a recurrent neural network model.

[0044] In addition, the challenge of battery K value prediction is that the data not only contains dynamic changes and long-term dependencies (suitable for recurrent neural network models), but also may have complex nonlinear relationships and boundary problems (suitable for support vector machine models). A single model is difficult to capture these complex characteristics simultaneously, resulting in limited prediction accuracy. To solve this problem, the prediction Agent combines a recurrent neural network model and a support vector machine model for prediction processing. The recurrent neural network model is used to process the feature data to capture dynamic characteristics and long-term dependencies in the data. The support vector machine model is used to classify and predict the features adjacent to the output of the recurrent neural network model to obtain the predicted K value.

[0045] The working principle of the above-mentioned fusion of recurrent neural network model and support vector machine model for prediction is as follows: First, the feature data provided by the feature extraction Agent is input into a recurrent neural network model. Recurrent neural network models (such as LSTM or GRU) are good at processing sequence data and can effectively capture dynamic features (such as instantaneous fluctuations in voltage and current) and long-term dependencies (such as slow decay trends in the battery aging process) in the battery K-value data. The output of the recurrent neural network model is no longer the final K-value prediction, but "high-level features" or "feature neighbors" extracted after deep learning on the original feature data. These high-level features contain time series patterns and potential correlations that are difficult to observe directly in the original data.

[0046] Subsequently, these high-level features output by the recurrent neural network model are input into a support vector machine model. Support vector machines are good at handling non-linear classification and regression problems, and can further classify or predict based on these high-level features. The support vector machine model models the features extracted by the recurrent neural network model by finding the optimal hyperplane in high-dimensional space, and finally outputs an accurate predicted K-value.

[0047] This fusion mechanism allows the recurrent neural network model to fully exploit the temporal information in the data, while the support vector machine model performs fine classification or regression based on this information, thereby combining the strengths of both models and significantly improving the accuracy and robustness of battery K-value prediction.

[0048] In another preferred embodiment of the present application, a battery process system based on real-time battery K-value prediction is also disclosed, which includes a battery K-value real-time prediction system and a decision Agent as described above.

[0049] The decision Agent generates production optimization decisions based on the predicted K-value output by the prediction Agent to drive real-time adjustments and quality control of the battery process.

[0050] Specifically, the decision Agent contains a decision logic module and an execution instruction generation module. The decision logic module receives the predicted K-value and combines pre-set production rules, quality standards, cost models, and historical production data for intelligent analysis and judgment. For example, if the predicted K-value indicates that the battery performance may not meet the standards, the decision logic module will trigger the corresponding optimization strategy. The execution instruction generation module generates specific and executable production adjustment instructions based on the judgment of the decision logic module, such as adjusting the charging current, changing the baking temperature, optimizing the formula ratio, or triggering the quality detection process. These instructions are then sent to the automated equipment or operators of the battery process to achieve real-time adjustment and quality control of the production line.

[0051] With the introduction of the decision Agent, the system achieves seamless integration of battery K-value prediction and actual production control, ensuring intelligent and efficient battery production processes.

[0052] In addition, the decision logic of the decision Agent can employ intelligent decision algorithms, such as decision models based on multi-objective optimization, reinforcement learning, expert systems, or fuzzy logic, to handle more complex production constraints and optimization objectives. The decision Agent can be integrated with industrial control systems such as MES (Manufacturing Execution System) or SCADA (Supervisory Control and Data Acquisition System) to achieve deeper automation and control. The granularity of production optimization decisions can be finer, for example, customized adjustment plans can be generated for individual battery cells or specific production batches. The decision Agent can have learning and adaptive capabilities, continuously optimizing its decision strategies by analyzing the effects of historical decisions. In addition, a human-machine interaction interface can be added to allow operators to review and intervene in the suggestions of the decision Agent.

[0053] The above only discloses the preferred embodiments of the present application, of course, cannot be limited to the scope of the patent rights of the present application, therefore, the equivalent changes made in the scope of the patent application of the present application, still belongs to the scope covered by the present application.

Claims

1. A battery K value real-time prediction system based on multi-agent cooperation, characterized in that, The system comprises a plurality of artificial intelligence agents (Agents) corresponding to different task types, which are sequentially connected in communication and include a data collection Agent, a feature extraction Agent, and a prediction Agent; The data collection Agent is configured to perceive and acquire multi-dimensional state information in a battery operating environment to obtain raw data; The feature extraction Agent is configured to convert the raw data into a feature set capable of representing battery performance to obtain feature data; The prediction Agent is configured to perform K-value prediction based on the feature data and a learned prediction strategy to obtain a predicted K-value.

2. The battery K-value real-time prediction system of claim 1, wherein, The multi-dimensional state information collected by the data collection Agent includes at least two of voltage, current, temperature, and internal resistance, and the data collection Agent is further configured to clean and denoise the collected data.

3. The battery K-value real-time prediction system of claim 1, wherein, For any task type, a plurality of Agents with the same function are created, and the plurality of Agents with the same function obtain tasks based on an auction algorithm.

4. The battery K-value real-time prediction system of claim 1, wherein, Agents with different functions belonging to different task types obtain computing resources based on an auction algorithm.

5. The battery K-value real-time prediction system of claim 1, wherein, The prediction Agent is constructed based on a reinforcement learning model of a Q-Learning algorithm.

6. The battery K-value real-time prediction system of claim 1, wherein, The prediction Agent adaptively selects a prediction model according to the characteristics of the current feature data.

7. The battery K-value real-time prediction system of claim 6, wherein, When the data amount of the feature data is small and the features are simple, a support vector machine model is selected as the prediction model; when the data amount is large and the features have complex time series characteristics, a recurrent neural network model is selected as the prediction model.

8. The battery K-value real-time prediction system of claim 1, wherein, The prediction Agent performs prediction processing by fusing a recurrent neural network model and a support vector machine model, the recurrent neural network model is used to process the feature data to capture dynamic characteristics and long-term dependencies in the data, and the support vector machine model is used to classify and predict adjacent features output by the recurrent neural network model to obtain a predicted K-value.

9. A battery process system based on real-time prediction of battery K value, characterized in that, The system comprises a plurality of artificial intelligence agents (Agents) corresponding to different task types, which are sequentially connected in communication and include a data collection Agent, a feature extraction Agent, and a prediction Agent; The decision Agent generates production optimization decisions based on the predicted K-value output by the prediction Agent to drive real-time adjustment and quality control of the battery manufacturing process.

Images