Efficient lithium battery life prediction system based on machine learning

Abstract

The invention provides an efficient lithium battery life prediction system based on machine learning, and relates to the technical field of safety monitoring and fault diagnosis of a power system. The prediction main system comprises a data acquisition and preprocessing module, a machine learning model construction module, a model training and optimization module, a life prediction module, a result evaluation and feedback module and a data storage and management module. By means of an advanced algorithm and accurate data preprocessing, limitation of a traditional empirical model is broken, reliable battery remaining life guidance is provided for equipment in different application scenes, operation stability and use planning rationality are improved, and meanwhile, by means of a pre-training model fine tuning technology and a scene specific optimization strategy, the service life of the battery is optimized. And the prediction model adaptive to each scene is quickly customized, an accurate result is output, and the universality and the practicability are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of safety monitoring and fault diagnosis of power systems, and in particular to a high-efficiency lithium battery life prediction system based on machine learning. Background Art

[0002] As the world accelerates its transition to clean energy and electronic devices explode, lithium batteries, as core energy storage components, shoulder important tasks in key areas such as electric vehicles, energy storage power stations, and consumer electronics. Their performance is closely related to the reliable operation, service life, and safety of various devices. Therefore, accurate prediction of lithium battery life has become the focus of attention in the current industry and academia.

[0003] In the early days, empirical models were popular. Such methods were mostly based on the number of battery charge and discharge cycles or the length of use to roughly estimate the lifespan. Just like in the field of consumer electronics, manufacturers often give consumers a relatively fixed battery life reference based on the test results of charge and discharge cycles under standard working conditions, informing them that the battery capacity of the product may decay to 80% of the initial capacity after several charge and discharge cycles.

[0004] Although existing methods can predict lithium battery life, the following problems still exist:

[0005] Problem 1: Traditional empirical models rely too much on simple charge and discharge cycles or usage time indicators, and fail to fully consider complex actual working conditions, resulting in significant deviations between the predicted results and the actual lifespan. This makes it impossible to provide users with reliable guidance on the remaining battery life, greatly reducing the stability and plannability of equipment use.

[0006] Problem 2: Although the physical model is highly theoretical, the internal process of lithium batteries is extremely complex, and the required measurement equipment is expensive and the computing resources are huge, which not only raises the research threshold, but also makes it difficult to popularize it on a large scale in the industry and cannot benefit many terminal applications;

[0007] Problem 3: Common traditional machine learning models are difficult to control the nonlinear characteristics of lithium battery life affected by complex interactions of multiple factors, and their accuracy is unsatisfactory. In addition, they lack the ability to flexibly adapt to different working conditions in different application scenarios and cannot optimize the prediction effect in a targeted manner, which limits their universality and practicality.

[0008] Question 4: Most existing technologies fail to organically integrate multi-source data. Key life information in unstructured data such as battery production processes and maintenance records is often overlooked, resulting in one-sided model training data and an inability to fully explore the potential laws of battery life, which in turn affects prediction accuracy.

[0009] Therefore, an efficient lithium battery life prediction system based on machine learning is needed to solve the above problems. Summary of the invention

[0010] Technical issues solved

[0011] In view of the deficiencies in the prior art, the present invention provides an efficient lithium battery life prediction system based on machine learning, which solves the problems in the above background technology.

[0012] Technical Solution

[0013] To achieve the above objectives, the present invention is implemented through the following technical solutions: an efficient lithium battery life prediction system based on machine learning, including a prediction main system, the prediction main system includes a data acquisition and preprocessing module, a machine learning model construction module, a model training and optimization module, a life prediction module, a result evaluation and feedback module, and a data storage and management module;

[0014] The data acquisition and preprocessing module uses an intelligent sensor array, which integrates voltage sensors, current sensors, temperature sensors, internal resistance sensors and charge and discharge rate sensors to collect the voltage, current, temperature, internal resistance and charge and discharge rate parameters of the lithium battery in real time, and calculates and statistics the usage time. The collected data is preprocessed by combining adaptive dynamic filtering technology with wavelet transform. The adaptive dynamic filtering dynamically adjusts the filtering parameters according to the real-time changes of the data to accurately remove high-frequency noise. The wavelet transform finely decomposes and reconstructs the data of each frequency band to efficiently eliminate abnormal values ​​and low-frequency interference.

[0015] The machine learning model building module builds an initial prediction model based on the factors affecting the life of lithium batteries, combined with long short-term memory networks, support vector regression and random forests, and designs the model architecture according to the characteristics of long short-term memory networks, support vector regression and random forest algorithms, and determines the number of nodes and connection methods of the input layer, hidden layer and output layer of the model;

[0016] The model training and optimization module divides the preprocessed data into a training set, a validation set, and a test set according to a specific ratio, uses the training set to train the constructed initial model, adopts gradient descent and adaptive learning rate adjustment optimization algorithms, and continuously adjusts the hyperparameters of the model in combination with the validation set, wherein the adjusted hyperparameters include learning rate, number of iterations, and regularization coefficient, to improve the overall performance of the model;

[0017] The life prediction module inputs the real-time collected and pre-processed lithium battery data into the trained and optimized model to quickly predict the current remaining life of the lithium battery and output the prediction results in an intuitive form of the number of charge and discharge cycles and the percentage of remaining available capacity;

[0018] The result evaluation and feedback module uses the test set to evaluate the accuracy of the prediction results, uses the mean square error, mean absolute error, and determination coefficient to quantitatively evaluate the prediction accuracy of the model, and feeds back the error information to the model training and optimization module based on the evaluation results to drive further optimization of the model;

[0019] The data storage and management module classifies and stores the collected raw data, pre-processed data, model training process data, and prediction results, and establishes data indexes to facilitate data query and tracing.

[0020] Preferably, the data acquisition and preprocessing module simultaneously captures the microscopic physical change signals of the lithium battery during the charging and discharging process, including the crystal structure changes of the electrode material and the ion concentration fluctuations of the electrolyte during operation. At the same time, the data acquisition and preprocessing module has an intelligent inspection function. Through the built-in fault diagnosis model, the health status of the lithium battery is monitored in real time based on the collected data. When potential fault hazards or abnormal data fluctuations are found, a multimodal warning is immediately issued, and the warning information is pushed to the operation and maintenance personnel in the form of sound, light, text messages, and system pop-up windows.

[0021] Preferably, the data acquisition and preprocessing module supports multi-source heterogeneous data fusion. When processing structured data collected by sensors, it also connects to external unstructured data, where the unstructured data mainly includes battery production process documents and maintenance record logs, and mines implicit information related to life to provide more comprehensive data support for model training. In the data preprocessing stage, transfer learning technology is used to draw on existing data processing experience under similar battery models and working conditions to accelerate the preprocessing process of newly collected data.

[0022] Preferably, the machine learning model building module introduces adversarial generative network technology, which is combined with long short-term memory network, support vector regression and random forest. The generator in the adversarial generative network technology is responsible for simulating the distribution of lithium battery data, and the discriminator distinguishes the authenticity of the data. The two are trained adversarially to make the model adapt to complex actual working conditions. At the same time, when optimizing the model architecture, a dynamic pruning strategy is adopted. According to the parameter importance evaluation during the model training process, the connections and nodes that have little impact on the prediction results are automatically pruned, thereby reducing the complexity of the model and improving the operating efficiency.

[0023] Preferably, the model training and optimization module adopts a federated learning framework to achieve collaborative model training between local battery energy storage stations and vehicle manufacturer battery data centers while ensuring data privacy. The intermediate results of model parameters are exchanged through encryption algorithms to avoid original data transmission and reduce the risk of data leakage. In addition, the training strategy is dynamically adjusted based on the performance feedback of the model in actual prediction tasks in combination with a reinforcement learning algorithm.

[0024] Preferably, when outputting the prediction results, the life prediction module not only gives the basic number of charge and discharge cycles and the remaining available capacity percentage, but also uses visualization technology to generate a dynamic battery life decay curve to intuitively display the future performance change trend of the battery. At the same time, combined with the expert system, the prediction results are checked for rationality based on the built-in battery field knowledge rule base.

[0025] Preferably, when outputting the prediction results, the life prediction module not only gives the basic number of charge and discharge cycles and the remaining available capacity percentage, but also uses visualization technology to generate a dynamic battery life decay curve to intuitively display the future performance change trend of the battery. At the same time, combined with the expert system, the prediction results are checked for rationality based on the built-in battery field knowledge rule base.

[0026] Preferably, when outputting the prediction results, the life prediction module not only gives the basic number of charge and discharge cycles and the remaining available capacity percentage, but also uses visualization technology to generate a dynamic battery life decay curve to intuitively display the future performance change trend of the battery. At the same time, combined with the expert system, the prediction results are checked for rationality based on the built-in battery field knowledge rule base.

[0027] Preferably, the prediction main system also has the ability of self-evolutionary learning, based on the meta-learning framework of reinforcement learning, continuously optimizes its own learning strategy according to the feedback rewards of the actual prediction tasks, and is deeply integrated with the external intelligent operation and maintenance platform to update its own knowledge system in a timely manner.

[0028] Preferably, the prediction main system targets different application scenarios such as electric vehicles, energy storage power stations, and consumer electronics, and uses pre-trained model fine-tuning technology to quickly customize prediction models that adapt to the needs of different scenarios, and after the model outputs the prediction results, it combines scenario-specific optimization strategies.

[0029] Beneficial Effects

[0030] The present invention provides an efficient lithium battery life prediction system based on machine learning. It has the following beneficial effects:

[0031] 1. The present invention comprehensively considers the multi-dimensional parameters and microscopic physical information of lithium batteries in actual use, and with the help of advanced algorithms and precise data preprocessing, breaks the limitations of traditional empirical models, and provides reliable guidance on the remaining battery life for devices in different application scenarios (electric vehicles, energy storage power stations, consumer electronics), thereby improving operational stability and rationality of usage planning; at the same time, it uses pre-trained model fine-tuning technology and scenario-specific optimization strategies to quickly customize prediction models that are suitable for each scenario, output accurate results, and enhance universality and practicality.

[0032] 2. The present invention uses a low-cost integrated intelligent sensor array, which is easy to deploy. The model construction and training stages introduce adversarial generative networks and dynamic pruning strategies, and combine the federated learning framework to ensure privacy-protected collaborative training, reduce dependence on expensive equipment and super computing power, and solve the problems of high cost and difficulty in promoting physical models, thus helping to popularize lithium battery life prediction technology on a large scale in the industry and benefit end users.

[0033] 3. The data acquisition and preprocessing module of the present invention integrates sensor structured data and unstructured data such as battery production process and maintenance records, mines implicit life information, draws on experience with the help of transfer learning, and empowers model training in all aspects; combined with preprocessing methods such as adaptive dynamic filtering and wavelet transform, the model can accurately capture the potential laws of lithium battery life, significantly improve prediction accuracy, and lay a solid foundation for full life cycle management.

[0034] 4. The prediction main system in the present invention is self-evolving based on the meta-learning framework of reinforcement learning, and optimizes the learning strategy according to the actual prediction feedback reward, automatically adjusts the key factors to adapt to the changing working conditions; and is integrated with the intelligent operation and maintenance platform to update the knowledge system in real time; at the same time, the data acquisition module performs intelligent inspection and real-time monitoring, and instantly issues multi-modal warnings in case of fault hazards, ensuring battery safety and ensuring stable operation of the escort system. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is the overall framework diagram of the present invention;

[0036] Figure 2 This is an operation flow chart of the prediction main system of the present invention. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Specific embodiment one:

[0039] like Figure 1-2 As shown, the high-efficiency lithium battery life prediction system based on machine learning includes a prediction main system, which includes a data acquisition and preprocessing module, a machine learning model building module, a model training and optimization module, a life prediction module, a result evaluation and feedback module, and a data storage and management module;

[0040] The data acquisition and preprocessing module uses an intelligent sensor array, which integrates voltage sensors, current sensors, temperature sensors, internal resistance sensors and charge and discharge rate sensors to collect the voltage, current, temperature, internal resistance and charge and discharge rate parameters of lithium batteries in real time. At the same time, it calculates and statistics the usage time. The collected data is preprocessed by combining adaptive dynamic filtering technology with wavelet transform. Adaptive dynamic filtering dynamically adjusts the filtering parameters according to the real-time changes of data to accurately remove high-frequency noise. Wavelet transform finely decomposes and reconstructs the data of each frequency band to efficiently eliminate outliers and low-frequency interference.

[0041] The machine learning model building module builds an initial prediction model based on the factors affecting the life of lithium batteries, combined with long short-term memory networks, support vector regression and random forests. It also designs the model architecture based on the characteristics of long short-term memory networks, support vector regression and random forest algorithms, and determines the number of nodes and connection methods of the model's input layer, hidden layer and output layer;

[0042] The model training and optimization module divides the preprocessed data into training set, validation set and test set according to a specific ratio, uses the training set to train the constructed initial model, adopts gradient descent and adaptive learning rate adjustment optimization algorithm, and continuously adjusts the model's hyperparameters in combination with the validation set. The hyperparameters adjusted include learning rate, number of iterations and regularization coefficient to improve the overall performance of the model;

[0043] The life prediction module inputs the real-time collected and pre-processed lithium battery data into the trained and optimized model to quickly predict the current remaining life of the lithium battery and output the prediction results in an intuitive form of the number of charge and discharge cycles and the percentage of remaining available capacity;

[0044] The result evaluation and feedback module uses the test set to evaluate the accuracy of the prediction results, and uses indicators such as mean square error, mean absolute error, and determination coefficient to quantitatively evaluate the prediction accuracy of the model. Based on the evaluation results, the error information is fed back to the model training and optimization module to drive further optimization of the model;

[0045] The data storage and management module classifies and stores the collected raw data, preprocessed data, model training process data, and prediction results, and establishes data indexes to facilitate data query and tracing.

[0046] During operation, the data acquisition and preprocessing module simultaneously captures the microscopic physical change signals of the lithium battery during the charging and discharging process, including the crystal structure changes of the electrode material and the ion concentration fluctuations of the electrolyte. At the same time, the data acquisition and preprocessing module has an intelligent inspection function. Through the built-in fault diagnosis model, it monitors the health status of the lithium battery in real time based on the collected data. Once potential fault hazards or abnormal data fluctuations are found, a multi-modal warning is immediately issued, and the warning information is pushed to the operation and maintenance personnel in the form of sound, light, text messages, and system pop-up windows.

[0047] The data acquisition and preprocessing module supports multi-source heterogeneous data fusion. When processing structured data collected by sensors, it also connects to external unstructured data. The unstructured data mainly includes battery production process documents and maintenance record logs, and mines the implicit information related to life to provide more comprehensive data support for model training. In the data preprocessing stage, transfer learning technology is used to draw on existing data processing experience under similar battery models and working conditions to accelerate the preprocessing process of newly collected data.

[0048] The machine learning model building module introduces adversarial generative network technology, which is combined with long short-term memory networks, support vector regression and random forests. The generator in the adversarial generative network technology is responsible for simulating the distribution of lithium battery data, and the discriminator is responsible for distinguishing the authenticity of the data. The two are trained adversarially to make the model adapt to complex actual working conditions. At the same time, when optimizing the model architecture, a dynamic pruning strategy is adopted. According to the parameter importance evaluation during the model training process, the connections and nodes that have little impact on the prediction results are automatically pruned, thereby reducing the complexity of the model and improving operating efficiency.

[0049] The model training and optimization module adopts a federated learning framework to achieve collaborative model training between local battery energy storage stations and vehicle manufacturers' battery data centers while ensuring data privacy. It exchanges the intermediate results of model parameters through encryption algorithms to avoid original data transmission and reduce the risk of data leakage. It also combines reinforcement learning algorithms to dynamically adjust training strategies based on performance feedback of the model in actual prediction tasks.

[0050] When outputting the prediction results, the life prediction module not only gives the basic number of charge and discharge cycles and the percentage of remaining available capacity, but also uses visualization technology to generate a dynamic battery life decay curve to intuitively display the future performance change trend of the battery. At the same time, it combines the expert system and the built-in battery field knowledge rule base to verify the rationality of the prediction results.

[0051] When outputting the prediction results, the life prediction module not only gives the basic number of charge and discharge cycles and the percentage of remaining available capacity, but also uses visualization technology to generate a dynamic battery life decay curve to intuitively display the future performance change trend of the battery. At the same time, it combines the expert system and the built-in battery field knowledge rule base to verify the rationality of the prediction results.

[0052] When outputting the prediction results, the life prediction module not only gives the basic number of charge and discharge cycles and the percentage of remaining available capacity, but also uses visualization technology to generate a dynamic battery life decay curve to intuitively display the future performance change trend of the battery. At the same time, it combines the expert system and the built-in battery field knowledge rule base to verify the rationality of the prediction results.

[0053] The main prediction system also has the ability of self-evolutionary learning. Based on the meta-learning framework of reinforcement learning, it continuously optimizes its own learning strategy according to the feedback rewards of actual prediction tasks, and is deeply integrated with the external intelligent operation and maintenance platform to update its own knowledge system in a timely manner.

[0054] The main prediction system targets different application scenarios such as electric vehicles, energy storage power stations, and consumer electronics. Through pre-trained model fine-tuning technology, it quickly customizes prediction models that adapt to the needs of different scenarios, and combines scenario-specific optimization strategies after the model outputs the prediction results. Specific embodiment 2:

[0056] like Figure 1-2 As shown, the following is the workflow of the entire prediction main system;

[0057] 1. Data collection and preprocessing process

[0058] Smart sensor array data acquisition

[0059] The intelligent sensor array integrates voltage, current, temperature, internal resistance and charge and discharge rate sensors to collect key parameters of the lithium battery operation process in real time and accurately. At the same time, it accurately calculates the usage time of the lithium battery based on the system's built-in clock.

[0060] For example, when an electric vehicle is driving, sensors collect data several times per second to fully capture the battery's real-time status under different operating conditions.

[0061] Adaptive dynamic filtering and wavelet transform preprocessing

[0062] The collected data first enters the adaptive dynamic filtering link. This technology dynamically adjusts the filtering parameters according to the real-time changes in the data to accurately remove high-frequency noise, such as interference signals caused by instantaneous circuit fluctuations.

[0063] Next, the data flows into the wavelet transform processing flow. The wavelet transform finely decomposes and reconstructs the data in each frequency band, efficiently eliminating outliers and low-frequency interference. Occasional abnormal situations such as sensor reading deviations can be effectively identified and corrected.

[0064] After these two steps of preprocessing, the purity and quality of the data are greatly improved, laying a solid foundation for subsequent model training and life prediction.

[0065] Microscopic physical change signal capture and intelligent inspection

[0066] During operation, the data acquisition and preprocessing module uses advanced microscopic detection technology to simultaneously capture the microscopic physical change signals of lithium batteries during the charging and discharging process, focusing on monitoring the crystal structure changes of electrode materials and the ion concentration fluctuations of the electrolyte. These microscopic information are crucial for in-depth understanding of the internal chemical reactions of the battery and predicting the battery performance degradation in advance.

[0067] The built-in fault diagnosis model monitors the health status of the lithium battery in real time based on all collected data (including macro parameters and micro signals). Once potential fault hazards are found, such as abnormal increase in battery internal resistance, obvious signs of microstructure degradation, or abnormal fluctuations in data, the intelligent inspection program is immediately initiated.

[0068] Through multi-modal warning methods such as sound and light, text messages, system pop-up windows, etc., warning information will be pushed to operation and maintenance personnel in a timely manner so that they can take corresponding measures, such as checking the battery connection lines and adjusting the charging and discharging strategies.

[0069] Multi-source heterogeneous data fusion and transfer learning

[0070] The module supports multi-source heterogeneous data fusion. When processing structured data collected by sensors, it actively connects to external unstructured data, mainly including battery production process documents and maintenance record logs. Using natural language processing and data mining technology, it extracts implicit information related to life from these unstructured data, such as special process parameters in the battery production process, causes of failures and repair measures in past maintenance cases, etc., and converts them into structured data to provide more comprehensive data support for model training.

[0071] In the data preprocessing stage, transfer learning technology is introduced. When faced with newly collected data, the system first retrieves the existing data processing experience of similar battery models and working conditions, draws on previous learning results, quickly adjusts preprocessing parameters and methods, accelerates the cleaning and conversion of new data, and reduces computing resource consumption and time costs.

[0072] 2. Machine Learning Model Construction Process

[0073] Combining multiple algorithms to build an initial model

[0074] According to the factors affecting the life of lithium batteries, such as charge and discharge rate, temperature environment, frequency of use, etc., the three algorithms of long short-term memory network, support vector regression and random forest are organically combined to construct an initial prediction model.

[0075] In view of the fact that the long short-term memory network algorithm is good at processing time series data, when designing the model architecture, the input layer accepts lithium battery parameter data over a period of time in time series, such as the voltage, current and other parameters collected every minute in the past hour. The hidden layer adopts a carefully designed gated recurrent unit (GRU) structure to accurately control the input, forgetting and output of information, so that the model can effectively capture the dynamic changes of battery performance over time. The output layer outputs the preliminary prediction results that integrate the time series features.

[0076] For the support vector regression algorithm, the input layer uses the kernel function to map the collected parameters to a high-dimensional space and mine potential features. The hidden layer optimizes the decision boundary by adjusting the weight of the support vector so that it can accurately fit the complex nonlinear relationship between battery life and parameters under limited data. The output layer gives accurate prediction values.

[0077] With the advantages of random forest algorithm in processing high-dimensional data and resisting overfitting, the input layer performs random subspace sampling on the original data, constructs multiple decision trees as hidden layers, and each tree is independently trained based on different sampled data. A comprehensive and robust prediction result is obtained at the output layer through a voting mechanism, effectively reducing the risk of overfitting that may be caused by a single algorithm.

[0078] Introducing Generative Adversarial Networks and Dynamic Pruning Optimization

[0079] Generative Adversarial Network (GAN) technology is introduced and combined with the above three core algorithms. The generator in GAN is responsible for simulating the data distribution of lithium batteries under different working conditions, such as simulating the changes in battery parameters under high temperature and high-rate charge and discharge scenarios, and the discriminator determines the authenticity of the data. The adversarial training of the two enables the model to adapt to more complex and changeable actual working conditions, and improves the robustness and generalization ability of the model.

[0080] When optimizing the model architecture, a dynamic pruning strategy is adopted. During the model training process, the importance of model parameters is evaluated in real time, and connections and nodes that have little impact on the prediction results are automatically pruned based on the evaluation results, such as some neuron connections whose weights are close to zero and contribute little to the final output, thereby reducing model complexity, reducing computing resource requirements, and improving operating efficiency.

[0081] 3. Model training and optimization process

[0082] Data partitioning and basic training

[0083] The preprocessed data is divided into training set, validation set and test set according to a specific ratio (such as 70% training set, 15% validation set, and 15% test set). The constructed initial model is trained with the training set, and the gradient descent and adaptive learning rate adjustment optimization algorithms are used. The hyperparameters of the model are continuously adjusted in combination with the validation set. The key hyperparameters to be adjusted are the learning rate, number of iterations and regularization coefficient.

[0084] In the early stage of training, a relatively large learning rate is set. As the number of training rounds increases, the learning rate is adaptively reduced to avoid missing the global optimal solution and prevent oscillation near the local optimal solution. The model complexity is controlled by the regularization coefficient to prevent overfitting. The performance of the model on unseen data is monitored using the validation set, and the training strategy is adjusted in time according to the validation results.

[0085] Co-optimization of Federated Learning and Reinforcement Learning

[0086] By adopting a federated learning framework, the model is trained collaboratively with local battery energy storage stations and vehicle manufacturers' battery data centers. Under the premise of ensuring data privacy, all participants exchange the intermediate results of model parameters through encryption algorithms to avoid the transmission of original data and effectively reduce the risk of data leakage.

[0087] At the same time, combined with the reinforcement learning algorithm, the training strategy is dynamically adjusted according to the performance feedback of the model in the actual prediction task. For example, if the model frequently has large errors in the battery life prediction task of a certain type of electric vehicle, the reinforcement learning algorithm automatically adjusts the data division ratio, increases the proportion of such battery data in the training set, or switches to a more suitable optimization algorithm combination to accelerate model convergence and improve the accuracy of the model in actual applications.

[0088] 4. Lifespan prediction process

[0089] Real-time data input and model prediction

[0090] The real-time collected and pre-processed lithium battery data is input into the trained and optimized model. The model quickly predicts the current remaining life of the lithium battery based on the complex relationship between battery life and various parameters learned in the early stage.

[0091] The prediction results are output in two intuitive forms: the number of charge and discharge cycles and the percentage of remaining available capacity, allowing users to quickly understand the health status of the battery. For example, the percentage of the battery's remaining available capacity corresponding to the remaining mileage and the estimated number of charge and discharge cycles that can be sustained are directly displayed on the electric vehicle dashboard, providing a basis for users' usage decisions.

[0092] Visual display and expert verification

[0093] When outputting the prediction results, visualization technology is used to generate a dynamic battery life decay curve to intuitively display the future performance trend of the battery. Users can view the curve through the mobile phone APP or the car's central control screen to understand the expected capacity decay of the battery over a period of time in the future (such as the next six months or a year).

[0094] At the same time, combined with the expert system, the prediction results are checked for rationality based on the built-in battery field knowledge rule base. The expert system judges whether the prediction results conform to physical laws and actual usage scenarios based on a large amount of battery experimental data and industry experience. If an abnormality is found, such as a sudden and significant decrease in the predicted remaining life without obvious external inducements, correction suggestions are immediately given, and the abnormal situation is fed back to the upstream module for further investigation to ensure the reliability of the prediction results.

[0095] V. Result Evaluation and Feedback Process

[0096] Quantitative indicator evaluation

[0097] The test set is used to evaluate the accuracy of the prediction results, and the mean square error, mean absolute error, and determination coefficient are used to quantitatively evaluate the prediction accuracy of the model. The mean square error measures the square average of the deviation between the predicted value and the true value, reflecting the overall accuracy of the prediction; the mean absolute error directly reflects the degree of average absolute deviation between the predicted value and the true value; the determination coefficient evaluates the model's ability to explain the data from the perspective of goodness of fit. These indicators are used to comprehensively and accurately evaluate the performance of the model.

[0098] User feedback drives adaptive evaluation

[0099] In addition to traditional indicator evaluation, we build an adaptive evaluation system driven by user feedback. We also collect feedback from users in the actual use of prediction results, such as the degree of help in decision-making and accuracy perception of prediction results. We quantify these subjective feedbacks through questionnaire surveys and user behavior analysis. For example, we integrate quantified user feedback into the evaluation model based on whether users replace batteries in time according to the prediction results and their satisfaction with the prediction results, and dynamically adjust the weights of evaluation indicators.

[0100] If it is found that users generally report that the prediction results in a certain scenario have large deviations, the weight of the relevant indicators in that scenario will be increased accordingly in subsequent evaluations to make the model evaluation more in line with actual application needs.

[0101] Fault diagnosis and retraining trigger

[0102] When the model prediction accuracy continues to be lower than the set threshold, in addition to triggering the retraining process, the fault diagnosis process is automatically started. The deep learning model is used to troubleshoot the model itself, looking for possible reasons for the decline in accuracy from multiple aspects such as data quality, model architecture, and algorithm parameters, such as whether there is a deviation in the recently accessed data or whether the model is overfitting.

[0103] Based on the fault diagnosis results, the data preprocessing method can be adjusted, the model architecture can be optimized, or the training parameters can be readjusted to ensure that the model always maintains a high prediction accuracy.

[0104] 6. Data Storage and Management Process

[0105] Classification storage and index creation

[0106] The data storage and management module classifies and stores the collected raw data, preprocessed data, model training process data, and prediction results, and establishes data indexes according to multiple dimensions such as data type, time series, battery number, etc. to facilitate data query and traceability.

[0107] For example, operation and maintenance personnel can quickly retrieve all the original collected data, preprocessing logs and corresponding life prediction results of the battery within a specific time period by entering the battery serial number and time range, providing strong support for battery performance analysis and troubleshooting.

[0108] Distributed storage and blockchain security

[0109] The use of a distributed storage architecture combined with blockchain technology can, on the one hand, disperse data and store it in multiple nodes, thereby improving the reliability and scalability of data storage. Even if some nodes fail, the data is still complete and available.

[0110] On the other hand, the tamper-proof nature of blockchain is used to ensure the authenticity and integrity of the data. In the data sharing process, automatic authorization and billing management of data access are achieved through smart contracts to ensure the rights and interests of data providers and users. For example, when battery manufacturers provide battery quality data to car companies, automatic billing and authorized access are carried out based on smart contracts. At the same time, the data update history is recorded to facilitate tracing back the data change process.

[0111] 7. System Self-Evolution and Scenario Adaptation Process

[0112] Self-evolution based on reinforcement learning

[0113] The main prediction system has the ability of self-evolutionary learning. Based on the meta-learning framework of reinforcement learning, it continuously optimizes its own learning strategy according to the feedback rewards of actual prediction tasks.

[0114] For example, when the system performs well in a battery life prediction task for a certain type of energy storage power station and obtains high user satisfaction and accuracy scores, the system automatically strengthens the learning path and model configuration in that scenario; conversely, if errors occur frequently in a new application scenario, the exploration mode is started to try different model architectures, algorithm combinations, and hyperparameter settings, and the model is automatically evolved through continuous trial and error and feedback learning.

[0115] Scenario adaptation and optimization strategy output

[0116] For different application scenarios such as electric vehicles, energy storage power stations, and consumer electronics, we use the pre-trained model fine-tuning technology to quickly customize prediction models that adapt to the needs of different scenarios. In the early stage of model construction, the basic model is trained based on a large amount of general battery data, and then the model is fine-tuned using a small amount of scenario-specific data based on the characteristics of different scenarios, such as the high-precision requirements of electric vehicles for mileage prediction and the emphasis of energy storage power stations on long-term stability prediction.

[0117] After the model outputs the prediction results, it combines the scenario-specific optimization strategies to provide users with more targeted decision support. For example, in the electric vehicle scenario, the optimal charging route is planned for the user based on the remaining battery life and road condition information, and nearby charging stations are recommended; in the energy storage power station scenario, the charging and discharging scheduling plan is optimized based on the battery life prediction results to improve energy utilization efficiency. Specific embodiment three:

[0119] like Figure 1-2 As shown, the key algorithm mentioned in Example 1 is analyzed in detail below, including its core mathematical formula and explanation:

[0120] 1. Adaptive dynamic filtering technology

[0121] Adaptive dynamic filtering dynamically adjusts the filtering parameters according to the real-time changes of data to accurately remove high-frequency noise. Assuming that the input signal is x(n), the output signal after filtering is y(n);

[0122] It can be expressed as:

[0123]

[0124] Where y(n) is the output signal after filtering at time n; x(n) is the input signal at time n; w i (n) is the i-th weight coefficient of the filter at time n; M is the order of the filter.

[0125] 2. Wavelet transform

[0126] Wavelet transform precisely decomposes and reconstructs the data in each frequency band, effectively eliminating outliers and low-frequency interference. The formula for continuous wavelet transform is:

[0127]

[0128] Where W f (a, b) are the wavelet transform coefficients of the signal f(t) at scale a and displacement b; f(t) is the original signal; a is the scale parameter, which controls the expansion and contraction of the wavelet function; b is the displacement parameter, which controls the translation of the wavelet function; ψ(t) is the wavelet mother function; ψ * (t) represents its conjugate function.

[0129] 3. Support Vector Regression (SVR)

[0130] Given a training sample (x i ,y i ),i=1,…,n, the goal of SVR is to find a regression function f(x) such that:

[0131]

[0132] Among them, K(x i ,x) is the kernel function. Commonly used kernel functions include linear kernel Gaussian Kernel etc.; i , is the Lagrange multiplier; b is the bias term; x i is the input of the i-th training sample; y i is the output of the i-th training sample; x is the input of the sample to be predicted.

[0133] 4. Random Forest

[0134] Random forest is an integrated learning model composed of multiple decision trees. For a regression problem, the prediction result of random forest is is the average of all decision tree predictions:

[0135]

[0136] in is the prediction result of random forest; T is the number of decision trees; is the prediction result of the tth decision tree.

[0137] 5. Generative Adversarial Networks (GANs)

[0138] The goal of the generator G is to minimize the output of the discriminator D, that is:

[0139]

[0140] Where G is the generator, which maps the noise z to the data space; D is the discriminator, which determines whether the input data is real data or generated data;data (x) is the distribution of real data; p z (z) is the distribution of noise; z is the noise vector; x is the real data.

[0141] 6. Gradient Descent

[0142] Gradient descent is used to update the model parameters θ. In each iteration, the parameter update formula is:

[0143]

[0144] Where θ is the parameter vector of the model; η is the learning rate; is the gradient of the loss function J(θ) with respect to the parameter θ.

[0145] 7. Mean Square Error (MSE)

[0146] Used to evaluate the accuracy of the model prediction results, the formula is:

[0147]

[0148] Where n is the number of samples; y i is the true value of the i-th sample; is the predicted value of the i-th sample.

[0149] 8. Mean Square Error (MSE)

[0150] Used to evaluate the accuracy of the model prediction results, the formula is:

[0151]

[0152] Where n is the number of samples; y i is the true value of the i-th sample; is the predicted value of the i-th sample. Specific embodiment four:

[0154] like Figure 1-2 As shown, the following is a description of the specific application logic steps of each module and algorithm in the high-efficiency lithium battery life prediction system based on machine learning:

[0155] 1. Data acquisition and preprocessing module

[0156] Data collection:

[0157] An intelligent sensor array is used, which integrates voltage sensors, current sensors, temperature sensors, internal resistance sensors and charge and discharge rate sensors to collect the voltage, current, temperature, internal resistance and charge and discharge rate parameters of the lithium battery in real time during the charge and discharge process.

[0158] At the same time, the built-in timing function is used to calculate and count the usage time of the lithium battery.

[0159] During this process, the smart sensor array can also capture the microscopic physical change signals of lithium batteries during charging and discharging, such as changes in the crystal structure of electrode materials and fluctuations in the ion concentration of the electrolyte. It also supports docking with external unstructured data, such as battery production process documents and maintenance record logs, to prepare for the subsequent mining of implicit information related to life.

[0160] Data preprocessing:

[0161] A method combining adaptive dynamic filtering technology and wavelet transform is adopted.

[0162] Adaptive dynamic filtering: Dynamically adjust the filtering parameters according to the real-time changes of the collected data. For the input signal, the filtered output signal is calculated by the formula, where is the weight coefficient of the filter at the moment and is the filter order, so as to accurately remove high-frequency noise.

[0163] Wavelet transform: The continuous wavelet transform formula is used to process the data, where is the wavelet transform coefficient of the signal under scale and displacement, is the wavelet mother function, and is its conjugate function. By adjusting and finely decomposing and reconstructing the data in each frequency band, outliers and low-frequency interference can be efficiently eliminated.

[0164] By using transfer learning technology and drawing on existing data processing experience under similar battery models and working conditions, the preprocessing process of newly collected data can be accelerated.

[0165] The built-in fault diagnosis model monitors the health status of the lithium battery in real time based on the collected data. Once potential fault hazards or abnormal data fluctuations are found, a multi-modal warning is immediately issued. The warning information is pushed to the operation and maintenance personnel through sound and light, text messages, and system pop-up windows.

[0166] 2. Machine Learning Model Building Modules

[0167] Initial model building:

[0168] According to the influencing factors of lithium battery life, the initial prediction model is constructed by combining long short-term memory network (LSTM), support vector regression (SVR) and random forest (RF).

[0169] LSTM: Based on the input (input at the moment), the hidden state at the previous moment and the state of the memory unit, the forget gate is used to decide which information in the memory unit to retain or forget. The input gate controls the input of new information, the candidate memory unit generates possible memory content, the memory unit updates the memory state, and the output gate determines the output information. Finally, the hidden state at the current moment is obtained, where is the weight matrix, is the bias vector, is the Sigmoid activation function, and is the hyperbolic tangent activation function.

[0170] SVR: Given a training sample, the goal is to find a regression function, where is a kernel function (such as a linear kernel, a Gaussian kernel, etc.), is a Lagrange multiplier, and is a bias term. By optimizing these parameters, the model fits the data.

[0171] RF: It consists of multiple decision trees, each of which is trained based on a different subset of the training data and a different subset of features. For regression problems, the prediction result of the random forest is the average of the prediction results of all decision trees, that is, is the number of decision trees, and is the prediction result of the th decision tree.

[0172] Model architecture optimization:

[0173] The model architecture is designed based on the characteristics of the above three algorithms to determine the number of nodes and connection methods of the model's input layer, hidden layer, and output layer.

[0174] The generative adversarial network (GAN) technology is introduced. The generator is responsible for simulating the distribution of lithium battery data, and the discriminator is responsible for distinguishing the authenticity of the data. The two are trained adversarially, and the goal is to minimize the output of the discriminator, that is, where is the distribution of real data, is the distribution of noise, and is the noise vector. Through this adversarial training, the model can adapt to complex actual working conditions.

[0175] A dynamic pruning strategy is adopted to automatically prune connections and nodes that have little impact on the prediction results based on the parameter importance evaluation during the model training process, thereby reducing the model complexity and improving operation efficiency.

[0176] 3. Model training and optimization module

[0177] Data partitioning:

[0178] The preprocessed data is divided into training set, validation set and test set in a specific proportion. The training set is used for model training, the validation set is used to adjust the model hyperparameters, and the test set is used for final model performance evaluation.

[0179] Model training:

[0180] The constructed initial model is trained using the training set and the gradient descent algorithm. In each iteration, the model parameters are updated according to the formula based on the gradient of the loss function with respect to the parameters, where is the learning rate. At the same time, an adaptive learning rate is used to adjust the optimization algorithm. The value changes dynamically according to the training situation. Combined with the validation set, the model's hyperparameters, such as learning rate, number of iterations, and regularization coefficient, are continuously adjusted to improve the overall performance of the model.

[0181] By adopting a federated learning framework, local battery energy storage stations and vehicle manufacturers' battery data centers can collaborate in training models while ensuring data privacy. The intermediate results of model parameters are exchanged through encryption algorithms to avoid the transmission of original data and reduce the risk of data leakage.

[0182] Combined with the reinforcement learning algorithm, the training strategy is dynamically adjusted according to the performance feedback of the model in the actual prediction task. For example, when it is found that the model has a large prediction error under certain working conditions, the reinforcement learning algorithm will guide the model to pay more attention to the data characteristics of these working conditions and adjust the training direction.

[0183] 4. Life Prediction Module

[0184] Real-time prediction:

[0185] The real-time collected and pre-processed lithium battery data is input into the trained and optimized model. The model quickly predicts the current remaining life of the lithium battery based on its own prediction logic (such as the combined prediction method of LSTM, SVR, RF and other algorithms).

[0186] Result output:

[0187] The prediction results are output in an intuitive form of the number of charge and discharge cycles and the percentage of remaining available capacity.

[0188] At the same time, visualization technology is used to generate a dynamic battery life attenuation curve to intuitively display the future performance change trend of the battery. The specific approach is to draw a curve that changes over time based on the remaining life data at different time points predicted by the model through the drawing library.

[0189] Combined with the expert system, based on the built-in battery domain knowledge rule base, the rationality of the prediction results is checked. For example, it is checked whether the predicted remaining life is consistent with the general life range under the battery model and operating conditions. If any unreasonable situation is found, it may trigger further analysis or adjustment mechanism.

[0190] 5. Result Evaluation and Feedback Module

[0191] Accuracy Assessment:

[0192] The test set is used to evaluate the accuracy of the prediction results, and the mean square error, mean absolute error, and determination coefficient are used to quantitatively evaluate the prediction accuracy of the model, where is the number of samples, is the true value of the th sample, is the predicted value of the th sample, and is the mean of the true values.

[0193] Feedback optimization:

[0194] Based on the evaluation results, the error information is fed back to the model training and optimization module to drive further optimization of the model. For example, if the MSE is large, the model training and optimization module may increase the number of training iterations, adjust the learning rate, or optimize the model architecture to reduce the prediction error.

[0195] 6. Data Storage and Management Module

[0196] Data classification storage:

[0197] The collected raw data, preprocessed data, model training process data, and prediction results are classified and stored, and data indexes are established to facilitate data query and traceability. Different types of data are stored in different database tables or file directories, and the required data can be quickly located and obtained through indexes. For example, the raw data is stored in the "raw data" table in the order of collection time, the preprocessed data is stored in the "preprocessed data" table, the model training process data records the hyperparameters, loss values ​​and other information of each training and is stored in the "training process data" table, and the prediction results are stored separately and associated with the corresponding metadata such as battery identification.

[0198] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprising a reference structure" do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0199] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An efficient lithium battery life prediction system based on machine learning, characterized by: It includes a prediction main system, which includes a data acquisition and preprocessing module, a machine learning model building module, a model training and optimization module, a life prediction module, a result evaluation and feedback module, and a data storage and management module; The data acquisition and preprocessing module uses an intelligent sensor array, which integrates voltage sensors, current sensors, temperature sensors, internal resistance sensors and charge and discharge rate sensors to collect the voltage, current, temperature, internal resistance and charge and discharge rate parameters of the lithium battery in real time, and calculates and statistics the usage time. The collected data is preprocessed by combining adaptive dynamic filtering technology with wavelet transform. The adaptive dynamic filtering dynamically adjusts the filtering parameters according to the real-time changes of the data to accurately remove high-frequency noise. The wavelet transform finely decomposes and reconstructs the data of each frequency band to efficiently eliminate abnormal values ​​and low-frequency interference. The adaptive dynamic filtering formula is as follows: Assume that the input signal is x(n), and the output signal after filtering is y(n); It can be expressed as: Where y(n) is the output signal after filtering at time n; x(n) is the input signal at time n; w i (n) is the i-th weight coefficient of the filter at time n; M is the order of the filter; The wavelet transform formula is as follows: Where W f (a, b) are the wavelet transform coefficients of the signal f(t) at scale a and displacement b; f(t) is the original signal; a is the scale parameter, which controls the expansion and contraction of the wavelet function; b is the displacement parameter, which controls the translation of the wavelet function; ψ(t) is the wavelet mother function; ψ * (t) represents the conjugate function; The machine learning model building module builds an initial prediction model based on the factors affecting the life of lithium batteries, combined with long short-term memory networks, support vector regression and random forests, and designs the model architecture according to the characteristics of long short-term memory networks, support vector regression and random forest algorithms, and determines the number of nodes and connection methods of the input layer, hidden layer and output layer of the model; The model training and optimization module divides the preprocessed data into a training set, a validation set, and a test set according to a specific ratio, uses the training set to train the constructed initial model, adopts gradient descent and adaptive learning rate adjustment optimization algorithms, and continuously adjusts the hyperparameters of the model in combination with the validation set, wherein the adjusted hyperparameters include learning rate, number of iterations, and regularization coefficient, to improve the overall performance of the model; The life prediction module inputs the real-time collected and pre-processed lithium battery data into the trained and optimized model to quickly predict the current remaining life of the lithium battery and output the prediction results in an intuitive form of the number of charge and discharge cycles and the percentage of remaining available capacity; The result evaluation and feedback module uses the test set to evaluate the accuracy of the prediction results, uses the mean square error, mean absolute error, and determination coefficient to quantitatively evaluate the prediction accuracy of the model, and feeds back the error information to the model training and optimization module based on the evaluation results to drive further optimization of the model; The data storage and management module classifies and stores the collected raw data, pre-processed data, model training process data, and prediction results, and establishes data indexes to facilitate data query and tracing.

2. The high-efficiency lithium battery life prediction system based on machine learning according to claim 1 is characterized in that: During operation, the data acquisition and preprocessing module simultaneously captures the microscopic physical change signals of the lithium battery during the charging and discharging process, including the crystal structure changes of the electrode material and the ion concentration fluctuations of the electrolyte. At the same time, the data acquisition and preprocessing module has an intelligent inspection function. Through the built-in fault diagnosis model, the health status of the lithium battery is monitored in real time based on the collected data. When potential fault hazards or abnormal data fluctuations are found, a multi-modal warning is immediately issued, and the warning information is pushed to the operation and maintenance personnel in the form of sound, light, text messages, and system pop-up windows.

3. The high-efficiency lithium battery life prediction system based on machine learning according to claim 1 is characterized in that: The data acquisition and preprocessing module supports multi-source heterogeneous data fusion. When processing structured data collected by sensors, it also connects to external unstructured data, where the unstructured data mainly includes battery production process documents and maintenance record logs, and mines implicit information related to life to provide more comprehensive data support for model training. In the data preprocessing stage, transfer learning technology is used to draw on existing data processing experience under similar battery models and working conditions to accelerate the preprocessing process of newly collected data.

4. The high-efficiency lithium battery life prediction system based on machine learning according to claim 1 is characterized in that: The machine learning model building module introduces adversarial generative network technology, combined with long short-term memory network, support vector regression and random forest. The generator in the adversarial generative network technology is responsible for simulating the distribution of lithium battery data, and the discriminator discriminates the authenticity of the data. The two are trained against each other to make the model adapt to complex actual working conditions. At the same time, when optimizing the model architecture, a dynamic pruning strategy is adopted. According to the parameter importance evaluation during the model training process, the connections and nodes with less impact on the prediction results are automatically pruned, the model complexity is reduced, and the operation efficiency is improved. The formula of the random forest is as follows: in is the prediction result of random forest; T is the number of decision trees; is the prediction result of the tth decision tree.

5. The high-efficiency lithium battery life prediction system based on machine learning according to claim 1 is characterized in that: The model training and optimization module adopts a federated learning framework to achieve collaborative model training between local battery energy storage stations and vehicle company battery data centers while ensuring data privacy. It exchanges intermediate results of model parameters through encryption algorithms to avoid original data transmission and reduce the risk of data leakage. It also combines reinforcement learning algorithms to dynamically adjust training strategies based on performance feedback of the model in actual prediction tasks.

6. The high-efficiency lithium battery life prediction system based on machine learning according to claim 1 is characterized in that: When outputting the prediction results, the life prediction module not only gives the basic number of charge and discharge cycles and the remaining available capacity percentage, but also uses visualization technology to generate a dynamic battery life decay curve to intuitively display the future performance change trend of the battery. At the same time, combined with the expert system, the prediction results are checked for rationality based on the built-in battery field knowledge rule base.

7. The high-efficiency lithium battery life prediction system based on machine learning according to claim 1 is characterized in that: The result evaluation and feedback module constructs an adaptive evaluation system driven by user feedback. In addition to traditional indicator evaluation, it also collects feedback from users during the actual use of prediction results, including the degree of helpfulness and accuracy of the prediction results for decision-making. These subjective feedbacks are quantified and integrated into the evaluation model to dynamically adjust the evaluation indicator weights.

8. The high-efficiency lithium battery life prediction system based on machine learning according to claim 1 is characterized in that: The data storage and management module adopts a distributed storage architecture combined with blockchain technology to store data in a decentralized manner, while utilizing the tamper-proof nature of blockchain to ensure the authenticity and integrity of the data.

9. The high-efficiency lithium battery life prediction system based on machine learning according to claim 1, characterized in that: The prediction main system also has the ability of self-evolutionary learning. Based on the meta-learning framework of reinforcement learning, it continuously optimizes its own learning strategy according to the feedback rewards of actual prediction tasks, and is deeply integrated with the external intelligent operation and maintenance platform to update its own knowledge system in a timely manner.

10. The high-efficiency lithium battery life prediction system based on machine learning according to claim 1, characterized in that: The prediction main system uses pre-trained model fine-tuning technology to quickly customize prediction models that adapt to the needs of different scenarios for different application scenarios such as electric vehicles, energy storage power stations, and consumer electronics.

Images