Method for predicting service life of battery

Abstract

The invention relates to the technical field of automobile batteries, and discloses a battery service life prediction method, which comprises the following steps: S1, collecting use data of a battery; s2, cleaning and preprocessing the collected battery data, removing abnormal data, filling missing values, and performing standardization or normalization processing; s3, a nonlinear dynamic model is constructed, the change relation of the service life of the battery along with time is described, and the model considers multiple influence factors of the battery; and S4, based on battery historical data. According to the method, the prediction strategy is dynamically adjusted by combining Gaussian process regression and reinforcement learning, and the problem that a static model is poor in adaptability to battery state change in the prior art is solved; bayesian optimization is introduced to optimize hyper-parameters, so that prediction precision and stability are improved; and through a real-time data monitoring and dynamic adjustment mechanism, the real-time performance and the accuracy of the prediction model during battery state change are ensured.

Description

Technical Field

[0001] The invention relates to the technical field of automobile batteries, and in particular to a method for predicting battery service life. Background Art

[0002] In the current battery life prediction technology, most methods rely on static models or empirical formulas. For example, many methods use traditional statistical models or prediction models based on physical laws. Although these methods are simple and easy to implement, their accuracy often cannot meet the requirements of high precision. Especially when the battery usage environment changes, the prediction ability of the static model is significantly reduced. The battery usage status is constantly changing. For example, the battery's charging and discharging mode, temperature, and current will change with time and environmental changes, and the existing technology lacks sufficient adaptive capabilities in the face of these dynamic changes. Since there is no effective mechanism to adjust according to real-time changes, the deviation of battery life prediction increases and cannot effectively reflect the actual situation.

[0003] In addition, in existing battery life prediction technologies, model optimization mostly relies on manual experience or simple parameter adjustments. This method is not only inefficient, but also difficult to find the optimal solution in a complex environment. Especially in battery models, the relationships between many key parameters (such as battery voltage, current, temperature, etc.) are nonlinear, and traditional optimization methods cannot fully capture these complex relationships, resulting in a significant reduction in the accuracy of prediction results. Therefore, existing methods have great limitations when solving complex nonlinear problems.

[0004] In addition, existing technologies are often unable to cope with the problem of large-scale data and real-time data updates. Most traditional battery life prediction methods are based on static analysis of historical data, and as the battery status changes during use, the model cannot be updated in time. This method cannot respond quickly to the real-time status of the battery, resulting in a significant reduction in the timeliness and accuracy of the prediction results in rapidly changing application scenarios. Especially in dynamic applications such as electric vehicles, the state of the battery changes at any time, and existing technologies find it difficult to reflect these changes in real time and adjust the prediction strategy. Summary of the invention

[0005] In view of the deficiencies of the prior art, the present invention provides a battery life prediction method, which solves the problems of poor adaptability, low optimization efficiency and insufficient real-time prediction capability of the prior art battery life prediction technology in a dynamically changing environment.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: A battery service life prediction method comprises the following steps: S1. Collect battery usage data; S2. Clean and preprocess the collected battery data, remove abnormal data, fill missing values, and perform standardization or normalization; S3, constructing a nonlinear dynamic model to describe the relationship between battery life and time, wherein the model takes into account multiple influencing factors of the battery; S4. Based on the battery historical data, a nonlinear relationship between battery life and multiple state variables is established through Gaussian process regression method; S5. Optimizing the hyperparameters of the Gaussian process regression model based on Bayesian optimization; S6. Use reinforcement learning methods to dynamically adjust the battery life prediction strategy and optimize the model prediction results in real time; S7. Monitor the battery usage status in real time, continuously collect battery usage data, and update model parameters based on real-time data. By combining reinforcement learning and Bayesian optimization, dynamically adjust the prediction strategy so that the model can cope with dynamic changes during battery use.

[0007] Preferably, the step of collecting battery data in step S1 includes: S1.1. Use the battery management system to collect the battery voltage, current, temperature, charge and discharge status, charging current, charging mode, charging times and other parameters related to the battery status in real time through the CAN bus interface or other communication protocols; S1.2. Store the collected data as time series data.

[0008] Preferably, the data preprocessing step in step S2 includes: S2.1. Clean the collected battery usage data to remove outliers and noise in the data; S2.2. Interpolate the missing data using mean interpolation, linear interpolation or time series-based interpolation methods to fill in the data gaps; S2.3. Standardize or normalize different features in the battery data.

[0009] Preferably, the nonlinear dynamic model building step in step S3 includes: S3.1. Assume that the battery life has a nonlinear relationship with the battery voltage, current and temperature; S3.2. Construct a nonlinear dynamic equation to describe the change of battery life over time. Its mathematical expression is: in, L(t) Indicates battery life. f is a nonlinear function that describes the variation of battery life. v(t), i(t) and t is the input feature.

[0010] Preferably, the Gaussian process regression method steps in step S4 include: S4.1. Model the nonlinear relationship between battery life and battery state variables using a Gaussian process regression model. S4.2. During the training phase, the battery history data And the corresponding battery life data , learn the mean function of the Gaussian process m(x) and covariance function ; S4.3, prediction phase, using the trained Gaussian process model to predict new input data Make predictions and calculate the predicted values The mean and variance , predict the remaining life of the battery.

[0011] Preferably, the Bayesian optimization step in step S5 includes: S5.1. Use Bayesian optimization method to optimize the model's hyperparameters through Gaussian process regression model; S5.2. Use expected improvement as the acquisition function to calculate the optimal hyperparameters for each step.

[0012] Preferably, the reinforcement learning step in step S6 includes: S6.1. Model the battery life prediction process as a Markov decision process and define the system state The current state of the battery, action For prediction results, reward is the model prediction accuracy; S6.2. Use reinforcement learning methods, especially Q-learning or deep Q-network, to dynamically adjust the strategy of the battery life prediction model according to environmental changes and real-time data feedback.

[0013] Preferably, the calculation formula to be improved in step S5.2 is: in, is the current optimal hyperparameter, is the expected value.

[0014] The present invention also provides a battery life prediction device, comprising: a memory, a processor, and a battery life prediction program stored in the memory and executable on the processor, wherein the battery life prediction program is configured to implement a battery life prediction method as described above.

[0015] The present invention also provides a storage medium, on which a battery service life prediction program is stored. When the battery service life prediction program is executed by a processor, a battery service life prediction method as described above is implemented.

[0016] The present invention provides a method for predicting battery service life, which has the following beneficial effects: 1. The present invention adopts a battery life prediction technology solution based on the combination of Gaussian process regression and reinforcement learning, achieving the technical effect of dynamically adjusting the battery life prediction strategy. Compared with the static battery life prediction method in the prior art, the present invention can optimize and adjust the prediction strategy in real time according to the actual usage status of the battery, solving the problem of reduced prediction accuracy of the prior art when the battery status changes.

[0017] 2. The present invention optimizes the hyperparameters of the Gaussian process regression model by introducing the Bayesian optimization method, which significantly improves the prediction accuracy and stability of the model. Compared with the traditional manual parameter adjustment method, the Bayesian optimization can automatically select the optimal hyperparameters, avoid human intervention and experience dependence, solve the problem of the negative impact of hyperparameter selection on the model effect in the prior art, and improve the applicability and reliability of the system.

[0018] 3. The present invention uses a dynamic adjustment mechanism of real-time data monitoring and reinforcement learning to enable the battery life prediction model to quickly adapt to changes in the battery working state, ensuring the accuracy and real-time nature of the prediction results. Compared with the lag in processing battery usage data in the prior art, the present invention can update the model at the moment the battery state changes, effectively solving the defect of insufficient prediction ability of the existing method in a dynamic working environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of the method flow of the present invention; Figure 2 It is a schematic diagram of the structure of the battery service life prediction device of the present invention.

[0020] Among them, 40, battery life prediction device; 41, processor; 42, memory; 43, storage medium. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the specification 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.

[0022] Please refer to the attached Figure 1The present invention provides a method for predicting battery life, comprising the following steps: S1. Collect battery usage data; In this embodiment, the battery usage data is collected in real time by the battery management system (BMS). The BMS uses the CAN bus interface or other suitable communication protocols to exchange data with the battery control unit (BCU) to obtain the operating status of the battery in real time. In general, the collected battery data includes the battery voltage , Current ,temperature , charge and discharge status, charging current, charging mode and charging times. These parameters can reflect the performance and degradation characteristics of the battery under different working conditions.

[0023] As an option, the collected data is not limited to traditional battery status data, but can also include some external environmental data, such as temperature, humidity, and vehicle driving mode information. Specifically, in different application scenarios (such as electric vehicles, hybrid vehicles, etc.), different battery usage scenarios may require additional data collection to ensure that the model can accurately reflect the battery aging process in a complex environment.

[0024] In one possible implementation, the battery management system combines data collection and communication interfaces with the vehicle's diagnostic system to enable real-time updates of various status data during battery operation. In each time period, the collected data will be stored as a time series, which can be used for training and testing of battery life prediction models.

[0025] Battery voltage and current collection: The voltage and current of the battery are key parameters that reflect the battery performance. and current The change of can directly affect the life of the battery, especially in high-frequency charging and discharging scenarios. The battery management system collects the voltage and current signals on the battery terminals to reflect the energy transfer during the battery charging and discharging process.

[0026] Temperature and environmental factors: Battery temperature It is an important factor affecting the battery life. The electrochemical reaction rate of the battery is different under different temperature conditions. Too high or too low temperature will lead to a decrease in battery performance. Therefore, the battery management system will be equipped with a temperature sensor to monitor the temperature changes inside and outside the battery in real time.

[0027] Charge and discharge status and charging mode: Charging mode (such as fast charging, slow charging) and charge and discharge state (SOC) are important factors affecting battery life. By monitoring these states, it is possible to evaluate the health of the battery during the charging process, especially in scenarios with frequent charging and discharging, and understand the impact of charging mode on battery life.

[0028] Charging current and charging times: The size of the charging current and the number of charging times are one of the key indicators that determine the battery life. A higher charging current will cause the internal temperature of the battery to rise, increasing the speed of battery aging. Therefore, the battery management system needs to accurately record the current and charging times during each charge in order to subsequently evaluate the battery life.

[0029] Data storage and subsequent use: The collected battery data is usually stored in a time series format. In some application scenarios, the collected data is not only used to predict the battery life in real time, but also needs to be stored in a database for subsequent use, such as long-term trend analysis, model training, etc. The data storage method should be an efficient system that supports big data processing, such as a relational database or a time series database.

[0030] In order to ensure the accuracy of the collected data in subsequent processing, the battery management system also needs to pre-process the real-time collected data. This process includes data cleaning and supplementation. For example, abnormal values ​​in the collection process (such as equipment failure or short-term battery fluctuations) can be eliminated through specific algorithms; missing data can be supplemented through linear interpolation or mean filling to ensure data integrity.

[0031] In this embodiment, the battery management system needs to make a preliminary estimate of the battery performance based on the real-time data to provide basic data for subsequent life prediction. In particular, when calculating the impact of battery voltage, current and temperature on life, a benchmark model can be established based on historical data, and then the model prediction results can be adjusted through the continuously accumulated real-time data.

[0032] For example, based on historical battery data, a preliminary prediction model for battery life can be established through nonlinear regression methods or machine learning algorithms. The model can be dynamically adjusted according to real-time collected data to more accurately predict the remaining battery life.

[0033] In practical applications, in addition to common battery voltage, current, temperature and other parameters, more factors related to battery aging can be further introduced, such as: Environmental conditions: such as the impact of the temperature and humidity conditions on the battery life; Usage mode: impact and loss of battery such as rapid acceleration and sudden braking; The battery's own health status: The battery's health status (SOH) is periodically detected, and the remaining life of the battery is predicted by the changes in SOH.

[0034] With this extended information, the battery life prediction model can more accurately reflect the actual usage of the battery, thereby improving the reliability and adaptability of the prediction results.

[0035] S2. Clean and preprocess the collected battery data, remove abnormal data, fill missing values, and perform standardization or normalization; In this embodiment, the steps of data preprocessing mainly include the following aspects: outlier detection and removal, missing value filling, data standardization and normalization processing, etc.

[0036] Generally, the collected data may contain noise or outliers due to sensor failure, battery abnormality, etc. To ensure the accuracy of the data, it is necessary to first detect outliers on the data. Outlier detection can be based on statistical methods (such as standard deviation-based detection methods) or use machine learning algorithms for anomaly identification. If the values ​​of some data points are significantly beyond the normal range, they can be considered outliers and removed from the data set to prevent them from affecting model training.

[0037] As an option, outlier detection can also be based on the physical laws of battery usage characteristics. For example, changes in battery current and voltage should follow certain physical relationships. If the collected data violates the known characteristics of the battery, it can be considered as abnormal data. Specifically, for abnormal points in battery voltage and current data, thresholds can be set according to actual working conditions. If the collected current and voltage values ​​exceed a reasonable range, they will be marked as abnormal values ​​and removed.

[0038] In one possible implementation, the processing of missing values ​​is also an important part of data preprocessing. Missing data often affects the training and prediction accuracy of the model. Usually, interpolation methods can be used to fill in the missing data. Common interpolation methods include linear interpolation, Lagrange interpolation, etc. If the data is less missing, you can also choose a simple mean or median to fill in the missing data. For example, in the time series data of battery temperature and voltage, if the battery temperature data at certain moments is missing, it can be filled in by linear interpolation of the previous and next adjacent data. Specifically, for time series data, assuming that at time point and There are battery temperature data on and , the missing data is in the middle of the moment , the missing data can be estimated by interpolation: in, For missing moments The predicted temperature.

[0039] When dealing with missing values, the type and distribution of data also need to be considered. For example, missing data for battery state of charge and discharge (SOC) may need to be filled based on the battery's charging mode and historical discharge data, rather than simply through interpolation methods. This filling method based on physical laws and state models helps to ensure the rationality of the filled data.

[0040] Furthermore, in view of the dimensional differences of different features, it is usually necessary to standardize or normalize the data. Battery usage data involves parameters of multiple dimensions, and the numerical ranges and units of these parameters are different. In order to avoid deviations in model training due to dimensional differences of certain features, it is often necessary to standardize the data. Generally speaking, standardization converts the data into a standard normal distribution with a mean of zero and a variance of one. For example, for battery voltage data For standardization, the following formula can be used: in, is the mean value of the voltage data, is the standard deviation of the voltage data. The standardized data will make each feature have the same scale, avoiding a feature that is too large or too small to dominate the model training.

[0041] In another possible implementation, normalization can also be used as an alternative to standardization, especially when dealing with feature data with non-Gaussian distribution. Normalization maps the data to a specified range (such as [0,1]), avoiding data distortion that may be caused by over-standardization. For features such as battery temperature, the following normalization formula can be used: in, and They are the minimum and maximum values ​​of the battery temperature data respectively.

[0042] In some embodiments, to further improve data quality, anomaly detection algorithms such as anomaly detection based on isolation forests may be used to detect the collected battery data and automatically filter out abnormal values ​​that do not conform to the expected pattern. Once these abnormal values ​​are detected, they can be replaced by interpolation, deletion, or rules.

[0043] For the processing of battery usage data, this embodiment also includes some common technical details, such as noise removal. During data preprocessing, signals such as battery voltage and current may be interfered by external noise. In order to improve the reliability of data, noise removal is necessary. The influence of noise on battery data can be effectively reduced by filter-based denoising algorithms (such as Kalman filtering, mean filtering, etc.).

[0044] Kalman filter is a widely used method in signal processing. It can effectively reduce the noise in battery voltage and current signals and retain the real battery change trend by estimating and updating the state of dynamic systems. The estimation and update process of battery voltage signal using Kalman filter can be expressed as: in, is the state of the system (such as battery voltage), is the observed value, A. B and C is the system matrix, and represent process noise and observation noise respectively.

[0045] S3, constructing a nonlinear dynamic model to describe the relationship between battery life and time, wherein the model takes into account multiple influencing factors of the battery; In this embodiment, after completing data cleaning, filling and standardization in the data preprocessing stage, the battery life prediction enters the stage of building a nonlinear dynamic model. The model will consider various influencing factors of the battery under different working conditions and describe the attenuation process of the battery life through mathematical formulas.

[0046] In general, the key goal of nonlinear dynamic models is to model the relationship between battery life and its operating conditions and state parameters (such as current, voltage, temperature, etc.), and predict its change law through mathematical expressions. The mathematical form of the model can be expressed by the following dynamic equation: in, L(t) represents the remaining life of the battery at time t, v(t), i(t) and t Represent the voltage, current and temperature of the battery respectively. It is a nonlinear function that describes how battery life changes over time.

[0047] Specifically, the nonlinear function in the model It can be defined based on factors such as the battery's operating environment, charge and discharge mode, and battery health status. These factors may have different effects on battery life under different operating conditions, so It is necessary to consider the degradation rate of the battery under different operating conditions. For example, at high temperature or high load, the battery will generally age faster, while at low temperature or low load, the battery aging process may be slower.

[0048] As an option, the nonlinear dynamic model can use a model based on physical laws, such as the electrochemical reaction model of the battery. This model can accurately simulate the changes in the battery life during the charging and discharging process by describing the electrochemical reaction process inside the battery. This physical model usually involves the battery's electrochemical reaction equations, thermodynamic equations, and material aging laws.

[0049] In another possible implementation, the nonlinear dynamic model can be implemented through a data-driven approach, such as using a machine learning algorithm or a deep learning network to train the battery's historical data. By learning from a large amount of historical data, the model can automatically identify the complex nonlinear relationship between battery life and multiple factors. In this way, the model can provide personalized predictions for different batteries and application scenarios by learning the changing patterns of battery life.

[0050] In this embodiment, in the process of constructing the nonlinear dynamic model, considering the complexity of battery life, a segmented model is usually used to describe the life change under different working conditions. For example, when the battery is in a charging state, temperature and current have a greater impact on the life; while in a discharging state, the battery life change is related to factors such as load current and temperature. Therefore, the design of the nonlinear dynamic model should dynamically adjust the parameters of the model according to the behavior of the battery in different states.

[0051] In some embodiments, in order to further improve the prediction accuracy of the model, it is possible to consider introducing more state variables, such as the battery's state of health (SOH). Battery health is an indicator that reflects the degree of battery aging, and is usually related to factors such as the battery's internal resistance and capacity. By adding battery health status to the model, the prediction of battery life can be made more accurate, especially in scenarios of long-term use or multiple charge and discharge.

[0052] In order to make the model adaptable to the characteristics of different batteries, feature selection and model training are essential steps. According to the actual usage data of the battery, the features that have a greater impact on the battery life can be selected through Gini coefficient analysis or other statistical methods, thereby reducing the impact of irrelevant features and improving the training efficiency of the model. For example, battery voltage, temperature, and charge and discharge current usually have a greater impact on battery life, so they need to be given priority.

[0053] In addition, in the process of battery life prediction, the calculation of nonlinear dynamic models can be solved by numerical simulation methods. Generally, numerical methods such as Euler method and Runge-Kutta method are used to approximate continuous dynamic equations to predict battery life. In some cases, more efficient numerical algorithms, such as parallel computing or GPU acceleration, can speed up the calculation of the model to meet the needs of real-time prediction.

[0054] In order to accurately model the nonlinear relationship between battery life and operating parameters, in addition to the basic dynamic equations, a state space model can also be established to describe the battery aging process. In this model, the battery state variables (such as voltage, current, temperature) will be used as input to predict the battery life through the evolution of the multidimensional state space.

[0055] For example, in some embodiments, the battery life prediction model may adopt a state space model of the following form: in, represents the state vector of the battery (such as voltage, current and temperature), Indicates control input (such as charging mode), and are process noise and observation noise, respectively. A. B and C is the system matrix, is the output (battery life prediction).

[0056] These formulas can be solved using numerical methods such as Kalman filtering to achieve accurate predictions of battery life.

[0057] S4. Based on the battery historical data, a nonlinear relationship between battery life and multiple state variables is established through Gaussian process regression method; In this embodiment, the battery life prediction model is constructed based on the historical data of the battery through the Gaussian process regression method. The historical data of the battery includes multiple state variables such as voltage, current, temperature, and corresponding battery remaining life data. These data are pre-processed as input and enter the Gaussian process regression model for training.

[0058] In general, the Gaussian process regression method can learn the nonlinear relationship between battery life and various state variables through input feature data (such as voltage, current, temperature, etc.). Specifically, the GPR model assumes that the relationship between battery life and input features follows a Gaussian process. The Gaussian process is defined as a random process with the following properties: Mean function m(x) , used to express the expected value of battery life; Covariance function , used to represent the similarity between input features.

[0059] The goal of Gaussian process regression is to use the training data to learn these two functions so that the model can make accurate predictions based on new input data.

[0060] As an option, the historical data of the battery may contain multiple different types of features (such as battery voltage, current, temperature, etc.), and the relationship between these features is often nonlinear. In order to better simulate this complex relationship, the GPR model captures the interdependence between features through the covariance function. The covariance function is usually selected as the RBF (Radial Basis Function) kernel, and the formula is as follows: in, is the variance of the signal, is the length scale parameter that determines the similarity between data points.

[0061] Specifically, the input of the GPR model includes the characteristic vectors of the battery voltage, current and temperature. , and the output is the remaining life of the battery The trained model can provide life prediction for new battery data. : in, is the new input feature to be predicted, X is the input feature data in the training set, y is the corresponding battery life data, is the predicted mean, is the predicted variance. Through these formulas, the GPR model can not only give the predicted value of battery life, but also provide the uncertainty of the predicted results.

[0062] In order to improve the accuracy of the model, in this embodiment, in addition to the basic characteristics of the battery such as voltage, current and temperature, other variables that affect the battery life may be added, such as battery health status (SOH), charging mode, and usage frequency. These variables can more comprehensively describe the operating status of the battery and further improve the accuracy of life prediction.

[0063] In some embodiments, the prediction accuracy can be further improved by optimizing the hyperparameters of the model. For example, Bayesian optimization methods can be used to optimize the hyperparameters in the GPR model, such as the length scale parameter Sum signal variance Bayesian optimization automatically selects the optimal hyperparameters by minimizing the prediction error of the model, thereby improving the generalization ability and accuracy of the model.

[0064] In another possible implementation, the training process of GPR may involve a lot of calculations, especially when the amount of data is large. Therefore, the training process can be accelerated by parallel computing or distributed computing. In addition, when making real-time predictions, the computational complexity of the GPR model is high, and the inference process may need to be further optimized to meet the real-time requirements in practical applications.

[0065] In the training and prediction process of the GPR model, the core calculation is to solve the covariance matrix. Specifically, the training process of GPR requires calculating a n×n The covariance matrix of K(X,X) , and perform an inverse operation on it. The computational complexity of this process is ,in n is the number of samples of the training data. In order to improve computational efficiency, especially in the case of large-scale datasets, a low-rank approximation or a sparse GPR method can be used.

[0066] For the prediction stage, the computational complexity is This also means that as the amount of training data increases, the model's reasoning speed may be affected. Therefore, in some high-frequency prediction scenarios, you can consider improving prediction efficiency by optimizing algorithms or hardware acceleration.

[0067] S5. Optimizing the hyperparameters of the Gaussian process regression model based on Bayesian optimization; In this embodiment, after predicting the battery life by Gaussian process regression (GPR) method in step S4, a preliminary model output is obtained. However, the effect of the GPR model is highly dependent on the hyperparameters therein, such as the parameters of the covariance function (e.g., the length scale Sum signal variance ), etc. The selection of these hyperparameters has a direct impact on the prediction results of the model. Therefore, optimizing these hyperparameters through Bayesian optimization is a key step to improve the prediction accuracy of the model.

[0068] In general, the basic idea of ​​Bayesian optimization is to approximate the expensive objective function by building a surrogate model, and then select the hyperparameters to be explored next through the acquisition function (such as expected improvement EI). The surrogate model usually uses a Gaussian process to represent the objective function because the Gaussian process can provide uncertainty estimates about the prediction results.

[0069] As an option, the goal of Bayesian optimization is to minimize or maximize the objective function, where the objective function is the negative log-likelihood function of the GPR model. Specifically, the goal is to automatically adjust the hyperparameters of the Gaussian process through Bayesian optimization so that the model exhibits the minimum prediction error on a given data set. The core formula of the Bayesian optimization process is: in, is a hyperparameter of the model, is the expected improvement, which is used to select the hyperparameters that are most likely to improve the current best solution. The expected improvement function can be expressed as: in, is the best known point at present. Indicates expected value.

[0070] In another possible implementation, during the Bayesian optimization process, the acquisition function may use different strategies, such as probability improvement (PI) or knowledge gradient (KG). However, expected improvement (EI) usually performs better in practical applications, so in this embodiment, EI is selected as the acquisition function.

[0071] In the Bayesian optimization process, the initial hyperparameter values ​​need to be determined first. Generally, you can choose a uniform distribution or set the initial value based on prior knowledge. Next, the hyperparameters are adjusted iteratively to optimize the performance of the GPR model. Each iteration requires using a Bayesian optimized surrogate model (usually a Gaussian process model) to predict the value of the objective function and use EI to decide which hyperparameters to explore next.

[0072] In general, the convergence speed of Bayesian optimization is closely related to the complexity of the objective function. When the objective function is relatively smooth and the parameter space is small, Bayesian optimization can find the global optimal solution in fewer experimental steps. In order to improve the efficiency of Bayesian optimization, parallel computing or GPU-based acceleration technology can be introduced, which is particularly important when processing large-scale battery data.

[0073] As an option, Bayesian optimization can be combined with an early stopping strategy to avoid wasting computing resources. If the performance of the model has stabilized at a good level after several iterations, Bayesian optimization can stop early to reduce unnecessary calculations and model training time.

[0074] In the Bayesian optimization process, the hyperparameters are updated based on the following formula: in, represents the current hyperparameter value, is the step size parameter, is the gradient of the expected improvement. After each update, Bayesian optimization re-evaluates the surrogate model and selects the next hyperparameter value to explore. This process continues until a stopping condition is met (such as the prediction error is small enough, or the maximum number of iterations is reached).

[0075] In practical applications, the hyperparameters selected during Bayesian optimization include: Length scale : Used to control the similarity between data points. Smaller values ​​indicate more differences between data points, while larger values ​​indicate less differences between data points.

[0076] Signal variance : Used to control the overall variance of the data, reflecting the fluctuation range of battery life prediction.

[0077] Noise variance : Used to measure the noise level of the model on the training data.

[0078] The optimization of these hyperparameters will directly affect the prediction results of the GPR model and can improve the accuracy and generalization ability of the model.

[0079] In some embodiments, in addition to traditional hyperparameter optimization, multi-objective optimization can be used to further improve the performance of the model. For example, in practical applications, in addition to reducing prediction errors, factors such as computational efficiency and model complexity can also be considered. In this case, the Bayesian optimization method can be extended to multi-objective optimization to balance the weights and trade-offs between multiple objectives.

[0080] In addition, in the case of large-scale data sets, in order to speed up the calculation of Bayesian optimization, sparse Gaussian process can be used to replace the traditional full Gaussian process method. This method effectively reduces the computational complexity by reducing the matrix size in the calculation, thereby speeding up the process of hyperparameter optimization.

[0081] S6. Use reinforcement learning methods to dynamically adjust the battery life prediction strategy and optimize the model prediction results in real time; In this embodiment, based on the established Gaussian process regression (GPR) model and Bayesian optimization method, reinforcement learning is further introduced into the battery life prediction model as a dynamic adjustment mechanism. Generally speaking, in battery life prediction, the working state of the battery (such as voltage, current, temperature, etc.) will change over time, which requires the model to have the ability to adapt in real time. Through the reinforcement learning algorithm, the model can continuously adjust its prediction strategy according to the actual use of the battery to improve the prediction accuracy.

[0082] Specifically, the application of reinforcement learning methods in battery life prediction models can be modeled by Markov decision processes (MDPs). In this model, the state of the system Indicates the current working status of the battery (such as voltage, current and temperature, etc.), and the action It represents the prediction strategy adopted by the system (such as a method for predicting battery life). The system is evaluated by its prediction accuracy at a certain moment. The smaller the prediction error, the higher the reward value. In this way, the reinforcement learning model can continuously learn and optimize the battery life prediction strategy.

[0083] As an option, Q-learning or Deep Q Network (DQN) can be used for reinforcement learning implementation. In Q-learning, the intelligent agent chooses the best action based on the current state to maximize the long-term reward. The agent's behavior and policy are updated by the following equation: in, Indicates in status Next select action The value of is the learning rate, is the discount factor, is the current reward value, is in the next state The maximum expected reward for all possible actions.

[0084] Specifically, the model adjusts the strategy for battery life prediction based on actual feedback. For example, if the battery voltage is too low or the temperature is too high, the reinforcement learning agent may dynamically adjust the weights of the prediction method to reduce the prediction error in such extreme cases. Reinforcement learning can effectively cope with dynamic changes in the battery aging process and make the model more adaptable.

[0085] In one possible implementation, the Deep Q Network (DQN) approach can be used in complex battery life prediction models, especially in high-dimensional state spaces. Through neural networks, DQN is able to learn complex nonlinear relationships in battery life prediction tasks, thereby enhancing the robustness and accuracy of the system.

[0086] In the process of combining reinforcement learning with GPR, the task of reinforcement learning is to adjust the prediction strategy of the GPR model so that it can make the best decision in the new battery usage environment. For example, reinforcement learning can optimize the kernel function parameters in Gaussian process regression (such as the length scale parameter and signal variance parameter ) to ensure that the model provides more accurate battery life prediction in actual use.

[0087] Specifically, the goal of reinforcement learning is to continuously update the model's parameters through environmental feedback so that the model can provide the best prediction results under different battery conditions. For example, when the battery temperature is too high or the current is too large, reinforcement learning can adjust the model's weights and hyperparameters based on historical data of battery life prediction to improve the accuracy and reliability of the prediction results.

[0088] In general, the model that combines GPR and reinforcement learning can adjust the prediction strategy through continuous learning in real-time prediction, so as to effectively cope with the complex changes in the battery aging process and adapt to different battery usage scenarios. For example, in the application scenario of electric vehicles, the driving mode of the vehicle will directly affect the service life of the battery. Reinforcement learning can automatically adjust the prediction strategy so that the model can adapt to such changes.

[0089] In some embodiments, the training process of the model can be accelerated by parallelizing the reinforcement learning method, especially in the case of large-scale data sets. Through multi-threading or GPU acceleration, the training speed of reinforcement learning can be greatly improved, thereby meeting the needs of real-time battery life prediction.

[0090] In addition, the prediction accuracy of the model can be further improved through multi-task learning. By sharing the parameters of the model in multiple related tasks, reinforcement learning can share experience in different battery types or usage environments, thereby improving the generalization ability of the prediction strategy. For example, for different types of batteries (such as lead-acid batteries, lithium batteries, etc.), through multi-task learning, the model can simultaneously optimize the life prediction strategy of multiple battery types, improve efficiency and reduce training time.

[0091] In some highly complex scenarios, deep reinforcement learning can also be combined to process more complex input-output relationships through deep neural networks and optimize the strategy of the battery life prediction model. This method is particularly suitable for processing high-dimensional state spaces and complex nonlinear relationships.

[0092] S7. Monitor the battery usage status in real time, continuously collect battery usage data, and update model parameters based on real-time data. By combining reinforcement learning and Bayesian optimization, dynamically adjust the prediction strategy so that the model can cope with dynamic changes during battery use.

[0093] In this embodiment, real-time monitoring of the battery usage status is achieved based on real-time data collection of the battery management system (BMS) and sensors. Generally, the BMS continuously collects information such as the battery voltage, current, temperature, charge and discharge status, charging current, and charging mode, which directly reflects the working status of the battery. By collecting this data in real time, the model can reflect the current status of the battery, thereby providing basic data for subsequent life prediction.

[0094] Specifically, during the use of the battery, the state of the battery changes at any time. In order to accurately predict the remaining life of the battery, the model needs to continuously update the parameters based on real-time data. At this time, the dynamic adjustment strategy based on reinforcement learning and Bayesian optimization is particularly important. Through this dynamic adjustment mechanism, the model can continuously learn new state information during the use of the battery, automatically adjust the prediction strategy, and ensure that the prediction results of the battery life always maintain high accuracy.

[0095] In this mechanism, the combination of reinforcement learning methods and Bayesian optimization methods plays a crucial role. Reinforcement learning can adjust the model's prediction strategy based on real-time feedback, while Bayesian optimization is used to optimize the model's hyperparameters. Generally speaking, reinforcement learning methods optimize the prediction results of battery life by selecting the best action strategy in the state space to maximize long-term rewards. Specifically, the model uses reinforcement learning to adjust the battery life prediction strategy by monitoring the battery status in real time and calculating the prediction error.

[0096] As an alternative, reinforcement learning uses Markov decision process (MDP) to model the battery life prediction process. By defining the state of the battery (including battery voltage, current, temperature, etc.), and define actions (i.e., the model’s prediction strategy), and the reward (Based on the prediction accuracy), reinforcement learning can dynamically adjust the prediction strategy to maximize the reward. Specifically, the model learns based on new battery data and continuously optimizes the prediction strategy to improve the accuracy of battery life prediction.

[0097] In another possible implementation, Bayesian optimization is used to optimize hyperparameters in reinforcement learning. In the reinforcement learning process, the prediction strategy of the model is not only affected by the state of the environment, but also by hyperparameters (such as prediction error tolerance, learning rate, etc.). Bayesian optimization methods can optimize these hyperparameters through acquisition functions such as expected improvement (EI), thereby improving the prediction ability of the model. Through Bayesian optimization, the optimal hyperparameter combination can be quickly found with a limited number of experiments, further improving the accuracy of battery life prediction and the generalization ability of the model.

[0098] In this embodiment, the process of real-time data update is as follows. Whenever the battery usage status changes, the BMS will collect new data in real time and pass it to the battery life prediction model. The model will then adjust the prediction strategy based on the new data. For example, if the battery temperature is too high or the voltage is too low, the model may adjust the prediction strategy based on historical experience or real-time feedback to ensure that accurate remaining life predictions can be provided under these extreme conditions.

[0099] Specifically, the model will affect key parameters in the GPR model, such as the length scale parameter of the kernel function and the signal variance, through the real-time update process. For each new battery state, reinforcement learning will dynamically adjust these hyperparameters so that the GPR model can maintain high prediction accuracy when the battery working state changes. This self-adjusting mechanism ensures that the battery life prediction can accurately reflect the actual battery usage through continuous feedback and adjustment in each state cycle.

[0100] In some embodiments, in order to ensure the computational efficiency and real-time responsiveness of the model, data processing and model parameter updates can be accelerated through parallel computing or distributed computing. These technologies can effectively reduce the delay of data collection and model training, thereby achieving real-time prediction.

[0101] In the dynamic adjustment process, the expected improvement (EI) as the acquisition function is the core part of Bayesian optimization, which guides the optimization of hyperparameters by evaluating the uncertainty of the current model. Specifically, the expected improvement calculation formula is as follows: in, is the current optimal prediction value, is the expected operation, and the maximum improvement is calculated. Bayesian optimization automatically adjusts the hyperparameters through this method, so that the battery life prediction gradually approaches the optimal solution after multiple iterations.

[0102] Through the combination of reinforcement learning, the model's prediction strategy can be optimized according to the latest battery status information after each adjustment, ensuring that the battery life prediction can change with the battery usage status. Reinforcement learning Q-learning or deep Q network (DQN) can further help the model learn the optimal decision-making strategy from historical data and avoid over-reliance on past prediction experience.

[0103] In some applications, the real-time monitoring and model adjustment system may be integrated into a more complex intelligent battery management system (BMS). In this case, the system can not only predict the battery life, but also respond immediately when the battery state changes, for example, by adjusting the charging strategy or taking thermal management measures to extend the battery life. In addition, combined with edge computing technology, real-time data collection and processing can be more efficient, ensuring that the battery management system can respond in real time and provide accurate battery life predictions.

[0104] Please refer to the attached Figure 2 The present invention also provides a battery life prediction device 40, comprising: a processor 41 and a battery life prediction program stored in the memory 42 and executable on the processor 41, wherein the battery life prediction program is configured to implement a battery life prediction method as described above.

[0105] The present invention further provides a storage medium 43 on which a battery life prediction program is stored. When the battery life prediction program is executed by the processor 41, a battery life prediction method as described above is implemented.

[0106] Among them, the storage medium 43 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read-Only Memory, referred to as EPROM), programmable read-only memory (Programmable Red-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic storage, flash memory, disk or optical disk.

[0107] 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. A method for predicting battery life, characterized in that: The following steps are involved: S1. Collect battery usage data; S2. Clean and preprocess the collected battery data, remove abnormal data, fill missing values, and perform standardization or normalization; S3, constructing a nonlinear dynamic model to describe the relationship between battery life and time, wherein the model takes into account multiple influencing factors of the battery; S4. Based on the battery historical data, a nonlinear relationship between battery life and multiple state variables is established through Gaussian process regression method; S5. Optimizing the hyperparameters of the Gaussian process regression model based on Bayesian optimization; S6. Use reinforcement learning methods to dynamically adjust the battery life prediction strategy and optimize the model prediction results in real time; S7. Monitor the battery usage status in real time, continuously collect battery usage data, and update model parameters based on real-time data. By combining reinforcement learning and Bayesian optimization, dynamically adjust the prediction strategy so that the model can cope with dynamic changes during battery use.

2. A battery service life prediction method according to claim 1, characterized in that: The step of collecting battery data in step S1 includes: S1.

1. Use the battery management system to collect the battery voltage, current, temperature, charge and discharge status, charging current, charging mode, charging times and other parameters related to the battery status in real time through the CAN bus interface or other communication protocols; S1.

2. Store the collected data as time series data.

3. A battery service life prediction method according to claim 1, characterized in that: The data preprocessing step in step S2 includes: S2.

1. Clean the collected battery usage data to remove outliers and noise in the data; S2.

2. Interpolate the missing data using mean interpolation, linear interpolation or time series-based interpolation methods to fill in the data gaps; S2.

3. Standardize or normalize different features in the battery data.

4. A battery service life prediction method according to claim 1, characterized in that: The nonlinear dynamics model building step in step S3 includes: S3.

1. Assume that the battery life has a nonlinear relationship with the battery voltage, current and temperature; S3.

2. Construct a nonlinear dynamic equation to describe the change of battery life over time. Its mathematical expression is: in, L(t) Indicates battery life. f is a nonlinear function that describes the variation of battery life. v(t), i(t) and t is the input feature.

5. A battery service life prediction method according to claim 1, characterized in that: The Gaussian process regression method steps in step S4 include: S4.

1. Model the nonlinear relationship between battery life and battery state variables using a Gaussian process regression model. S4.

2. During the training phase, the battery history data And the corresponding battery life data , learn the mean function of the Gaussian process m(x) and covariance function ; S4.3, prediction phase, using the trained Gaussian process model to predict new input data Make predictions and calculate predicted values The mean and variance , predict the remaining life of the battery.

6. A battery service life prediction method according to claim 1, characterized in that: The Bayesian optimization step in step S5 includes: S5.

1. Use Bayesian optimization method to optimize the model's hyperparameters through Gaussian process regression model; S5.

2. Use expected improvement as the acquisition function to calculate the optimal hyperparameters for each step.

7. A battery service life prediction method according to claim 1, characterized in that: The reinforcement learning step in step S6 includes: S6.

1. Model the battery life prediction process as a Markov decision process and define the system state The current status of the battery, action For prediction results, reward is the model prediction accuracy; S6.

2. Use reinforcement learning methods, especially Q-learning or deep Q-network, to dynamically adjust the strategy of the battery life prediction model according to environmental changes and real-time data feedback.

8. A battery service life prediction method according to claim 6, characterized in that: The calculation formula for the expected improvement in step S5.2 is: in, is the current optimal hyperparameter, is the expected value.

9. A battery life prediction device, characterized in that: include: A memory, a processor, and a battery service life prediction program stored in the memory and executable on the processor, wherein the battery service life prediction program is configured to implement a battery service life prediction method as claimed in any one of claims 1 to 8.

10. A storage medium, characterized in that: The storage medium stores a battery life prediction program, and when the battery life prediction program is executed by the processor, a battery life prediction method according to any one of claims 1 to 8 is implemented.

Images