Online battery life prediction method based on big data cloud platform based on mechanism information

Abstract

The present invention relates to an online battery life prediction method based on a big data cloud platform based on mechanism information, and belongs to the field of battery performance testing. The method comprises the following steps: collecting battery charge and discharge cycle data under a variety of fast charging cycle protocols, and establishing a database; establishing a capacity-voltage curve through discharge cycle data, and extracting the capacity difference-voltage curve of the battery sample at different times, wherein the capacity difference-voltage curve contains aging mechanism information; performing dimension reduction processing on the curve to obtain feature data, taking the feature data as input, and taking the remaining life and capacity decay inflection point of the battery sample as output for regression model training to obtain a prediction model; taking the current battery as the battery to be predicted, reading the discharge cycle data of the battery to be predicted, and predicting the remaining life and capacity decay inflection point of the battery through the prediction model and performing life classification. The present invention establishes a mode combining the features based on mechanism information and artificial intelligence technology to accurately predict the battery life and inflection point online.

Description

Technical Field

[0001] The present invention belongs to the field of battery performance testing and relates to an online battery life prediction method based on a big data cloud platform based on mechanism information. Background Art

[0002] As a new type of energy, lithium-ion batteries have been rapidly developed in recent years due to their advantages such as high operating voltage, high specific energy, high charge and discharge efficiency, low self-discharge rate, no memory effect, and long cycle life. At present, in addition to being used in daily devices such as mobile phones and notebooks, lithium-ion batteries are also widely used in important fields such as electric vehicles, spacecraft, and artificial satellites. During long-term use, lithium-ion batteries will undergo a series of electrochemical reactions and physical changes inside, such as dissolution of positive electrode materials, phase changes of positive electrode materials, and decomposition of electrolytes. These changes will lead to battery performance and capacity degradation, which will cause the system that uses it as the main energy source to fail and collapse, and in severe cases will even cause property losses and casualties, such as mobile phone explosions or electric vehicles suddenly losing control. Therefore, it is of great significance to predict the life of the battery. The traditional prediction method is to establish an offline prediction model, assume that the future working conditions are similar to the past, and use the early partial capacity to predict the decay trajectory until the end of the battery life is reached. For example, after the automotive power battery is manufactured or designed by the manufacturer, a partial charge and discharge cycle test is carried out to evaluate the performance of the battery, including the battery life. The hypothetical principle of the above method is unrealistic in practice. After the prediction model is established, it will no longer be updated and changes will not be recorded in real time. However, factors such as the battery load condition of electric vehicles may change at any time, and the battery performance degradation mechanism is very complex. The offline prediction model cannot obtain accurate prediction results.

[0003] With the deepening of research, the prior art proposes to perform online prediction of battery life by establishing an online prediction model. At present, the main method is to predict the remaining service life of lithium-ion batteries by establishing a mechanism-based physical model or using a data-driven method. Since it is difficult to find a model that conforms to the battery aging mechanism in the prior art, it is generally believed that the prediction accuracy of the data-driven method is higher, such as the Chinese patent with publication number CN109977622B. A method for predicting the remaining life of a power battery is provided, which is based on the fusion drive of the historical data of the power battery body and the cloud big data. During the implementation of this method, the online historical data of the battery body to be predicted and the cloud big data containing non-body information can be fully utilized to predict the remaining service life of the battery. It has many beneficial effects such as simple steps, high prediction accuracy, can greatly improve the utilization rate of available information sources, and can support any amount of non-body information as prediction input, and can effectively adapt to big data application scenarios. However, data-driven methods usually only use capacity degradation information. Such methods lack a physical information basis and have little practical significance for the characterization of battery life. They require a lot of training and a large amount of data, are prone to data silos, and take a long time, making them less efficient in predicting the remaining battery life.

[0004] The mechanism-based physical model mainly quantifies the factors affecting battery performance, and obtains the evolution process of lithium-ion battery performance degradation according to the chemical reactions occurring during the charging and discharging process of lithium-ion batteries and their specific physical properties. The biggest challenge currently faced is that there are many factors that affect each other in the battery performance evolution process, which leads to performance degradation, and it is difficult to reliably simulate the complex dynamic system of lithium-ion battery life aging. For example, the Chinese patent with publication number CN114814631 A discloses an online life prediction method for automotive lithium batteries based on cloud computing and feature selection, which collects data of the life of a new lithium battery and the constant current and constant voltage charging stages in the same preset charge and discharge cycle interval and stores them in the cloud; uses cloud computing to obtain the number distribution feature sets of the constant current and constant voltage charging mode stages of the current lithium battery; repeats the above steps to obtain the life and number distribution feature sets of each lithium battery; fuses the number distribution feature sets and then performs feature selection to obtain feature optimization training sets; trains and obtains the trained lithium battery life prediction regression model; when predicting online, collects and obtains the feature optimization selection set of the lithium battery to be predicted, and uses the trained lithium battery life prediction regression model to predict the cycle life of the current lithium battery to be predicted. This method solves the problem that online life prediction of lithium batteries in practical applications relies on specific discharge patterns and accurate capacity measurement. However, this method starts collecting data from the preset charge and discharge cycle interval of the battery, which is also time-consuming and highly dependent on data quality, and does not consider predicting the remaining life of the battery at any stage of battery use and at different cycle interval lengths.

[0005] Therefore, how to improve the efficiency of predicting the remaining service life of the battery and realize online battery life prediction while ensuring the accuracy of the prediction is an urgent problem to be solved. Summary of the invention

[0006] In view of this, the purpose of the present invention is to establish a model based on the combination of feature extraction containing physical information of battery aging and artificial intelligence technology on a cloud platform to achieve accurate prediction of battery performance.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] The online battery life prediction method based on the big data cloud platform of mechanism information includes the following steps:

[0009] S1: Collect battery charge and discharge cycle data under various fast charging cycle protocols and establish a battery online cloud platform database; the battery charge and discharge cycle data includes battery temperature, current, voltage and charge and discharge capacity;

[0010] S2: Establishing a capacity-voltage curve through the discharge cycle data, and extracting the capacity difference-voltage curve of the battery sample at different times from the capacity-voltage curve, wherein the capacity difference-voltage curve contains the sample aging mechanism information, and the capacity difference-voltage curve is calculated by subtracting the capacity at the corresponding voltage in the capacity-voltage curve of the mth cycle and the nth cycle of the moving window;

[0011] S3: Perform dimensionality reduction processing on the data, set the voltage sampling interval to 80mv, and take 20 points of characteristic data ΔQ_V from the capacity difference-voltage curve; take ΔQ_V as input, and take the remaining life and capacity decay inflection point of the battery sample as output, perform regression model training under the framework of the data-driven method, and obtain an online life prediction model and inflection point prediction model of the battery; the algorithm used in the regression model is a Gaussian process regression algorithm or a partial least squares regression algorithm;

[0012] S4: taking the current battery as the battery to be predicted, reading the discharge cycle data of the battery to be predicted during its service period, calculating the capacity-voltage curve of the battery to be predicted, and extracting characteristic data ΔQ_V from the capacity-voltage curve of the battery to be predicted;

[0013] S5: Substitute the ΔQ_V obtained in S4 into the online life prediction model to predict the remaining life of the battery to be predicted, and predict the capacity decay inflection point of the predicted battery through the inflection point prediction model.

[0014] Further, in S3, when the regression model is a Gaussian process regression algorithm, the probability distribution on the function is defined by Gaussian process regression, which is defined as:

[0015] f(x)~GP(m(x), k(x, x′))

[0016] Among them, m(x) is the mean function, k(x, x) is the covariance function, and its formula is as follows:

[0017] m(x)=E[f(x)]

[0018] k(x,x′)=E[(f(x)-m(x))(f(x′)-m(x′)) T ]

[0019] For any finite input data set, such as X = (x1, x2, ..., x p ) n×p , all have probability distribution, namely p(f(x1),f(x2),...f(x p )), where the mean function m(x) and the covariance function k(x, x′) are given by the kernel function K i，j =κ(x i , x j ) is given;

[0020] Using the Matérn kernel function:

[0021]

[0022] Where, v = 5 / 2, R v is the modified Bessel function, and the average function is defined as m(x) = 0;

[0023] Now we have a training set D containing input and output i ,y i )}, then in the X that needs to be predicted * Calculate the conditional distribution p(y * |X * , x, y), the Gaussian distribution is given by:

[0024] p(y * |X * , X, y) = N(y * |m * , σ * )

[0025] According to the conditional distribution properties of multidimensional Gaussian distribution, we have:

[0026] m * =K(X,X * )T K(X,X) -1 y

[0027] σ * =K(X * , X * )-K(X,X * ) T K(X,X) -1 K(X,X * ).

[0028] Further, in S3, when the regression model is a partial least squares regression algorithm, the principal components in the input and the output are extracted to maximize the correlation between the principal components extracted from the input and the output, respectively, that is:

[0029]

[0030] Among them, E0 and F0 are the data sets after input and output standardization respectively, t1 is the first component of E0, and u1 is the first component of F0, which are defined as: t1=E0ω1, u1=F0c1 respectively.

[0031] Furthermore, in S3, the capacity decay inflection point is obtained by a knee algorithm; the knee algorithm is to calculate the line distance from the beginning to the end of the battery life in the aging trajectory, and select the point with the largest line distance.

[0032] Furthermore, in the S3, life classifier training is also included, and the life classifier training includes: dividing the battery samples into two groups of long life and short life to perform classifier training to obtain an online battery life classifier; and also includes S6, grouping the batteries to be predicted through the life classifier.

[0033] Furthermore, when the method for training the lifespan classifier is a support vector machine (SVM) classifier, it specifically includes:

[0034] Find a hyperplane in the feature space. The hyperplane is defined as:

[0035] ω T x+b=0

[0036] The corresponding classification function is:

[0037]

[0038] Among them, α=(α1, α2,...α n ) is the Lagrange multiplier, and K is a kernel function operation, as follows:

[0039] K(x1,x2)=(<x1,x2> +R) d

[0040] Where R = 1, d = 2; the objective function of the quadratic SVM is:

[0041]

[0042] Furthermore, the remaining cycle period is obtained through the remaining life of the battery, and the battery life is classified according to the remaining cycle period and the life classification threshold; if the remaining cycle period is greater than the life classification threshold, it is a long life, and if the remaining cycle period is less than the life classification threshold, it is a short life.

[0043] Furthermore, the life classification threshold is the difference between the initial classification threshold and the cycle period, and the initial classification threshold is 550 cycles. The cycle period is included in the battery charge and discharge cycle data of the battery service period to be predicted.

[0044] The beneficial effects of the present invention are:

[0045] This method achieves accurate prediction of online battery life prediction and capacity decay inflection point prediction by establishing a model based on a combination of features containing battery aging mechanism information and artificial intelligence technology on a cloud platform, and classifies batteries according to their life. This method can play an important role in online battery management. Even if the battery has been used for many years, it can still accurately predict the remaining service life and capacity decay inflection point of the battery. Understanding the remaining service life and inflection point of the battery can not only provide clear guidance for predictive maintenance of the battery system, but also show the residual value of the battery required by the warranty company or the buyer's second-hand trade. In addition, online classification of "good (long life) / bad (short life)" batteries based on information from only a few cycles can greatly speed up the sorting / reorganization process of retired batteries for secondary use, which is conducive to improving energy utilization and promoting the development of new energy.

[0046] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0048] Figure 1 It is a flow chart of the method for online battery life prediction of a big data cloud platform based on mechanism information of the present invention;

[0049] Figure 2 is a ΔQ(V) curve diagram in an embodiment of the method;

[0050] Figure 3 is the ΔQ extracted in the embodiment of this method (100-10) (V) characteristic curve diagram;

[0051] Figure 4 A schematic diagram of a method for identifying an inflection point in an embodiment;

[0052] Figure 5 It is a schematic diagram of the inflection point results of some batteries identified by the knee algorithm;

[0053] Figure 6 It is a schematic diagram of cycle life prediction under different cycle periods with a cycle interval of 120 times;

[0054] Among them, (a) 130-10 window, (b) 240-120 window, (c) 370-250 window, (d) 490-370 window, (e) 600-480 window, (f) 720-600 window, this result is based on the GPR model;

[0055] Figure 7 is the root mean square error of the training and test data sets for Gaussian process regression and partial least squares regression under different windows;

[0056] Among them, (a) and (b) represent the root mean square error of the training and test data sets based on the Gaussian process regression method, and (c) and (d) represent the root mean square error of the training and test data sets based on the partial least squares regression method;

[0057] Figure 8 The average percentage errors of the training and testing data sets of Gaussian process regression and partial least squares regression models under different windows in life prediction;

[0058] Among them, polylines of different colors represent different windows, (a) and (b) represent the average percentage errors of the training and test data sets based on the Gaussian process regression method, and (c) and (d) represent the average percentage errors of the training and test data sets based on the partial least squares regression method;

[0059] Fig. 9 This is a schematic diagram of inflection point prediction for different windows with a window length of 120;

[0060] Among them, the lengths of the windows are (a) 130-10, (b) 200-80, (c) 270-150, (d) 340-220, (e) 400-280, and (f) 480-360.

[0061] Fig.10 Schematic diagram of classification prediction results, indicating the classification accuracy under different moving windows within 5 period intervals. (a) and (b) respectively represent the classification accuracy of the training and test data sets based on the support vector machine classifier, and (c) and (d) represent the classification accuracy of the training and test data sets based on the logistic regression classifier. DETAILED DESCRIPTION

[0062] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0063] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0064] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0065] See also Figure 1 to Figure 5 , which is an online battery life prediction method based on big data cloud platform based on mechanism information, including the following steps:

[0066] S1: Collect battery charge and discharge cycle data under various fast charging cycle protocols and establish a battery online cloud platform database; the battery charge and discharge cycle data includes battery temperature, current, voltage and charge and discharge capacity;

[0067] S2: Establishing a capacity-voltage curve through the discharge cycle data, and extracting the capacity difference-voltage curve of the battery sample at different times from the capacity-voltage curve, wherein the capacity difference-voltage curve contains the sample aging mechanism information, and the capacity difference-voltage curve is calculated by subtracting the capacity at the corresponding voltage in the capacity-voltage curve of the mth cycle and the nth cycle of the moving window;

[0068] S3: Perform dimensionality reduction processing on the data, set the voltage sampling interval to 80mv, and take 20 points of characteristic data ΔQ_V from the capacity difference-voltage curve; take ΔQ_V as input, and take the remaining life and capacity decay inflection point of the battery sample as output, perform regression model training under the framework of the data-driven method, and obtain an online life prediction model and inflection point prediction model of the battery; the algorithm used in the regression model is a Gaussian process regression algorithm or a partial least squares regression algorithm;

[0069] S4: taking the current battery as the battery to be predicted, reading the discharge cycle data of the battery to be predicted during its service period, calculating the capacity-voltage curve of the battery to be predicted, and extracting characteristic data ΔQ_V from the capacity-voltage curve of the battery to be predicted;

[0070] S5: Substitute the ΔQ_V obtained in S4 into the online life prediction model to predict the remaining life of the battery to be predicted, and predict the capacity decay inflection point of the predicted battery through the inflection point prediction model.

[0071] Specifically, in S1, the publicly available "A123" dataset is used. This dataset contains 124 commercial lithium iron phosphate (LFP) / graphite batteries. These battery samples are cycled under various fast charging conditions until the battery reaches the end of battery life (EOL). In this method, EOL is defined as the number of cycles required for each battery to reach 80% of its initial capacity. Typically, in the dataset, the battery is cycled under fast charging conditions ranging from 3.6C to 6C, and then discharged at 4C until the lower cutoff voltage is reached. It contains a total of 72 charging protocols (almost all batteries have a charging time between 9 and 15 minutes), and the cycle life of the batteries in the dataset ranges from 150 to 2300 times (the average cycle life is 806 and the standard deviation is 377). The dataset includes the temperature, current, voltage, charge capacity, and discharge capacity of each battery.

[0072] In this scheme, the data set is divided into a training data set and a test data set, where the training data set is used to train the model and the test data set is used to evaluate the model. The number of data sets in different moving windows is different. Overall, the ratio of the training data set to the test data set is roughly 3:2.

[0073] The capacity-voltage curve contains a wealth of information about the battery aging characteristics. Figure 2 As shown in Figure 1, the capacity-voltage curve moves as the number of cycles increases, and the area between adjacent voltage curves that can represent the battery discharge energy decreases as the battery ages. (m-n) (V), such as ΔQ (100-10) (V) Variance characteristics of (capacitance difference-voltage curve), ΔQ (100-10) (V) represents the energy change of the battery during different cycles, which is used to indicate the battery life. (100-10) (V) is the voltage difference between the capacity-voltage curves of the 100th cycle and the 10th cycle. The typical shape of this feature is as follows: Figure 3 shown.

[0074] ΔQ for different values ​​of m and n (m-n) (V) also has a strong correlation with battery life, which indicates that this feature within the moving window (defined by m and n) can robustly represent the remaining useful life (RUL) of the battery.

[0075] In order to reduce the input dimension of the data, the dimension can be reduced by changing different sampling intervals without affecting the characteristic curve. For example, Figure 3 The ΔQ under different sampling points is given (100-10) (V)-voltage effect sampling results. It can be clearly observed that when the sampling interval is 80mV, the basic shape of the curve remains unchanged. That is, in this case, only 20 points are needed to capture ΔQ (100-10) The main features of the (V) curve greatly reduce the dimensionality and storage space of the input data. Therefore, only 20 points in the ΔQ(mn)(V) curve need to be sampled and input into the machine learning (ML) model to predict RUL. The reduction of the input data dimensionality is conducive to effective online battery life prediction without affecting the physical information contained.

[0076] In addition, since there are few ultra-long life batteries in the data set (more than 1500 cycles), in order to better predict ultra-long life batteries, the difference in discharge capacity between the current charge and discharge cycle and the initial discharge cycle is selected as the calibration feature. The discharge capacity of the 10th cycle is used as the initial value, and the additional feature is recorded as ΔQ m-10 .

[0077] In S3, the Kneedle algorithm is used to identify the inflection point. The inflection point of the battery degradation trajectory is the point where the battery degradation rate changes from slow to fast, such as Figure 4 As shown, the inflection point is calculated as the point in the aging trajectory where the line distance from the beginning to the end of the battery life is the largest.

[0078] Figure 5 The results of inflection point identification for some batteries in the dataset using the Kneedle algorithm are shown.

[0079] This method extracts features containing information about the battery aging mechanism from the discharge voltage data to indicate the battery's health status. These features are constructed based on the data within a moving window, and the boundary is defined by two different aging states in the battery decay trajectory. As the battery ages, the moving window can capture the latest aging and operating information of the battery, and use this to reasonably predict the battery's aging trajectory.

[0080] Machine learning in S3 specifically includes:

[0081] Regression Model:

[0082] Gaussian process regression defines a probability distribution over a function, defined as:

[0083] f(x)~GP(m(x), k(x, x′))

[0084] Among them, m(x) is the mean function, k(x, x′) is the covariance function, and its formula is as follows:

[0085] m(x)=E[f(x)]

[0086] k(x,x′)=E[(f(x)-m(x))(f(x′)-m(x′)) T ]

[0087] For any finite input data set, such as X = (x1, x2, ..., x p ) n×p , all have probability distribution, namely p(f(x1),f(x2),...f(x p )), where the mean function m(x) and the covariance function k(x, x′) are given by the kernel function K i，j =κ(x i , x j ) is given;

[0088] In this method, the Matérn kernel function is used:

[0089]

[0090] Where, v = 5 / 2, R v is a modified Bessel function. The average function is usually defined as m(x)=0. This method follows this convention.

[0091] Now we have a training set D containing input and output i ,y i )}, then in the X that needs to be predicted * Calculate the conditional distribution p(y * |X *, x, y), the Gaussian distribution is given by:

[0092] p(y * |X * , X, y) = N(y * |m * , σ * )

[0093] According to the conditional distribution properties of multidimensional Gaussian distribution, we have:

[0094] m * =K(X,X * ) T K(X,X) -1 y

[0095] σ * =K(X * , X * )-K(X,X * ) T K(X,X) -1 K(X,X * )

[0096] The value of the covariance hyperparameter θ is determined by minimizing the negative log marginal likelihood (defined as Minimizing NLML automatically trades off between model complexity and data fitting, thereby improving overfitting to the data. Given an expression for NLML and its derivative with respect to θ, gradient-based optimization can be used to estimate θ. MatlabRegression learner is used to implement these algorithms.

[0097] In another embodiment, the regression model may also select the PLSR algorithm. PLSR is a multiple linear regression algorithm that considers both extracting the principal components from Y and X as much as possible and maximizing the correlation between the principal components extracted from X and Y, that is:

[0098]

[0099] Where X is the input (ΔQ_V), Y is the output (remaining life of the battery and capacity decay inflection point), E0 and F0 are the data sets after input and output standardization, t1 is the first component of E0, and u1 is the first component of F0, which are defined as:

[0100] t1=E0ω1,u1=F0c1。

[0101] In order to illustrate the prediction accuracy of this method, an evaluation model is used for testing. The evaluation model uses a test data set. The specific prediction model evaluation indicators use the root mean square error (RMSE) and the average percentage error (%err), that is, the evaluation model is:

[0102]

[0103]

[0104] In addition to the above basic methods, in practical applications, online life classification can provide the necessary sorting reference for battery second life applications. In order to improve the consistency of battery packs, retired batteries need to be recalibrated and reorganized so that they can be used reliably and safely in their second battery life stage.

[0105] Battery capacity, as a direct indicator of battery health, is crucial in the screening and reorganization process. However, battery capacity is not the only important factor in determining battery performance. Even with similar initial capacities, battery degradation trajectories can vary greatly. Therefore, other characteristics such as internal resistance, electrochemical impedance spectroscopy (EIS), etc. have been proposed as important supplements to improve the consistency of battery sorting. Although the sorting performance has been improved, at least two problems have arisen. First, the relationship between the additional sorting indicators and battery degradation and EOL is unclear, and second, the testing of these health characteristics (EIS, etc.) is usually cost and time intensive.

[0106] In order to solve these important problems, while training the prediction model, that is, in S3, life classifier training is also performed simultaneously. The features containing the aging information of the battery samples are taken as input, and the battery samples are divided into two groups of long life and short life for classifier training to obtain an online battery life classifier. When predicting the remaining life of the battery to be predicted, the battery to be predicted is also grouped by the life classifier.

[0107] The battery life is divided into long life and short life with the life classification threshold as the dividing line. The battery with a remaining cycle value greater than the life classification threshold is a long life, and the battery with a remaining cycle value less than the life classification threshold is a short life. The life classification threshold is the difference between the initial classification threshold and the cycle value. The cycle is included in the battery charge and discharge cycle data of the battery service period to be predicted. The initial classification threshold is 550 cycles.

[0108] The training of the lifespan classifier specifically includes:

[0109] In the classification test, the SVM method is selected. When solving the binary classification problem, the goal of SVM is to find a hyperplane in the feature space.

[0110] The hyperplane is defined as:

[0111] ω T x +b=0

[0112] The corresponding classification function is:

[0113]

[0114] Among them, α=(α1, α2,...α n ) is the Lagrange multiplier, K is a kernel function operation,

[0115] as follows:

[0116] K(x1,x2)=(<x1,x2> +R) d

[0117] Where R = 1, d = 2; the objective function of the quadratic SVM is:

[0118]

[0119] In another embodiment, a logistic regression classification algorithm (logistic) can also be selected. The purpose of logistic regression is to learn a 0 / 1 classification model from features. This model uses a linear combination of features as independent variables and uses a logistic function (also called a sigmoid function) to map the independent variables to (0, 1). The mapped values ​​are considered to be the probabilities of belonging to the 0 / 1 classification, which can also achieve similar effects.

[0120] The prediction effect of this method is:

[0121] 1. Lifespan prediction

[0122] Gaussian process regression (GPR) and partial least squares regression (PLSR) methods are applied to predict the RUL of the battery, and the moving windows start from the 10th, 120th, 250th, 370th, 480th and 600th cycles (a total of 6 moving windows), and the window lengths are set to 30, 60, 90 and 120 cycles, respectively. If the design life of the battery is 10 years and the service life before EOL is 1000 times, then they are 30 times and 120 times, respectively, which is equivalent to a battery service life of one season and one year.

[0123] like Figure 6 As shown in the figure, (a), (b), (c), (d), (e) and (f) are the prediction results of the training and test data sets using the GPR method under a cycle window length of 120. The results show that the RUL value of the battery can be accurately predicted online at different aging stages, which is crucial for reliably evaluating the performance of the battery during operation.

[0124] Table 1 - Remaining useful life prediction results of different machine learning models

[0125]

[0126]

[0127] Table 1 shows the prediction results of the RMSE and mean percentage error results of the remaining useful life of the battery for the GPR algorithm and the PLSR algorithm under a 120-cycle window length. For both algorithms, the RMSE and range error (MPE) decrease as the window moves forward, indicating that the extracted features can more accurately capture the aging characteristics of the battery. This increasing prediction accuracy is favored in practice because an accurate battery RUL can still be predicted even if the battery has been in use for many years. As shown in Table 1, when GPR predicts RUL, the RMSE and MPE are mostly less than 100 times and 10%, respectively, and when the battery approaches EOL, the errors of the two are reduced to 55 times and 3.55%, respectively. PLSR performs slightly worse, with most RMSE and MPE greater than GPR.

[0128] like Figure 7 and Figure 8 As shown in the figure, the RMSE and MPE of the training and test data sets for online RUL prediction using the GPR model and the PLSR model are shown respectively, and the broken lines of different colors represent different windows. The results show that the accuracy of GPR and PLSR in predicting battery RUL is similar, which indicates that effective mechanism information features are extracted for reliable battery RUL prediction. However, the test results based on GPR show that both RMSE and MPE decrease with the increase of window length. In addition, as the window size increases (more operational information is obtained), the RMSE and MPE decrease when GPR predicts RUL. Both downward trends of RMSE and MPE are beneficial in the practical application of aging batteries, but they are not obvious in the RUL prediction results based on the PLSR method.

[0129] In addition, the PLSR model requires artificially finding the optimal number of principal components, the value of which varies with the experimental data within each moving window. This variation increases the difficulty and infeasibility of the PLSR algorithm in practical applications. In addition, the GPR model has an important advantage over PLSR in that it can provide uncertainty predictions, which makes online battery life predictions more reliable. Therefore, in practical applications, GPR is a potential tool of choice for accurately predicting battery RUL than PLSR.

[0130] 2. Inflection point prediction

[0131] Based on the RUL prediction, the GPR model and PLSR model are also used to perform online prediction of the inflection point. The moving windows start from the 10th, 80th, 150th, 220th, 280th and 360th cycles, respectively, and the window lengths are set to 30, 60, 90 and 120 cycles, respectively.

[0132] Fig. 9 The prediction result of the inflection point of the GPR model when the window length is 120 cycles. The cell data set selected here is before the inflection point. Overall, the inflection point prediction is similar to the RUL prediction mode, which proves the correlation between the inflection point and life span from the side, and also illustrates the feasibility of online prediction of the inflection point.

[0133] Table 2 - Inflection point prediction results of different machine learning models

[0134]

[0135] As can be seen from Table 2, the RMSE and MPE of the inflection point prediction based on GPR are within 80 cycles and 8% respectively, and this accuracy can be used for online prediction. The prediction results of PLSR are also listed in Table 2, and its RMSE and MPE are mostly greater than GPR. Therefore, compared with PLSR, the GPR model can more accurately predict the RUL and inflection point of the battery while providing uncertainty, and is more suitable and feasible for online prediction scenarios.

[0136] In summary, Gaussian process regression (GPR) based on different moving windows can more accurately predict the battery RUL and inflection point than partial least squares regression (PLSR), and its root mean square error (RMSE) and mean percentage error (MPE) are mostly less than 100 times and 8%, respectively. As the window moves, the RUL and inflection point prediction accuracy are improving, and the RMSE and MPE under the 120-cycle window length are reduced to less than 85 cycles and less than 5%, respectively, near the battery EOL or inflection point. This result shows that the battery's RUL and inflection point can still be accurately predicted after the battery has been put into use for many years, verifying the high reliability and feasibility of the proposed method.

[0137] 3. Lifespan classification

[0138] In the lifetime classification, four moving windows were set at the 100th, 200th, 300th and 400th cycles respectively using features containing information about the battery aging mechanism, and different window lengths were considered, ranging from a length of 5 cycles to a length of only 1 cycle. The battery cells were divided into two groups: a long-life group and a short-life group, and the lifetime classification threshold was set to 550 cycles. Two typical classification methods including support vector machine classifier and logistic regression classifier were used for online classification.

[0139] like Fig.10As shown in Figure 2, the experimental results show that the support vector machine can accurately classify the battery life with only one cycle of data, and the classification accuracy is mainly above 85%. As the single-cycle window moves forward, the accuracy further rises to around 90%, once again demonstrating the high performance of online battery life classification. This breakthrough has the potential to greatly improve the consistency of classification by covering battery life evaluation, but also significantly reduce the cost and time required for traditional classification by utilizing only one cycle of online data.

[0140] Table 3 - Classification results of different machine learning models

[0141]

[0142] Table 3 shows the classification accuracy of the SVM classifier and the logistic regression classifier when the window length is 1 cycle. Only one cycle of data is needed to classify the battery as "good / bad", and the classification accuracy of the SVM classifier is above 85%, while the effect of the logistic regression classifier is slightly worse.

[0143] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. An online battery life prediction method based on big data cloud platform based on mechanism information, characterized in that: The following steps are involved: S1: Collect battery charge and discharge cycle data under various fast charging cycle protocols and establish a battery online cloud platform database; the battery charge and discharge cycle data includes battery temperature, current, voltage and charge and discharge capacity; S2: Establishing a capacity-voltage curve through the discharge cycle data, and extracting the capacity difference-voltage curve of the battery sample at different times from the capacity-voltage curve, wherein the capacity difference-voltage curve contains the sample aging mechanism information, and the capacity difference-voltage curve is calculated by subtracting the capacity at the corresponding voltage in the capacity-voltage curve of the mth cycle and the nth cycle of the moving window; S3: Perform dimensionality reduction processing on the data, set the voltage sampling interval to 80mv, and take 20 points of characteristic data ΔQ_V from the capacity difference-voltage curve; take ΔQ_V as input, and take the remaining life and capacity decay inflection point of the battery sample as output, perform regression model training under the framework of the data-driven method, and obtain an online life prediction model and inflection point prediction model of the battery; the algorithm used in the regression model is a Gaussian process regression algorithm or a partial least squares regression algorithm; S4: taking the current battery as the battery to be predicted, reading the discharge cycle data of the battery to be predicted during its service period, calculating the capacity-voltage curve of the battery to be predicted, and extracting characteristic data ΔQ_V from the capacity-voltage curve of the battery to be predicted; S5: Substitute the ΔQ_V obtained in S4 into the online life prediction model to predict the remaining life of the battery to be predicted, and predict the capacity decay inflection point of the predicted battery through the inflection point prediction model.

2. The method for online battery life prediction based on big data cloud platform based on mechanism information according to claim 1 is characterized in that: In S3, when the regression model is a Gaussian process regression algorithm, the probability distribution on the function is defined by Gaussian process regression, which is defined as: f(x)~GP(m(x), k(x, x′)) Among them, m(x) is the mean function, k(x,x′) is the covariance function, and its formula is as follows: m(x)=E[f(x)] k(x,x′)=E[(f(x)-m(x))(f(x′)-m(x′)) T ] For any finite input data set, such as X = (x1, x2, ..., x p ) n×p , all have probability distribution, namely p(f(x1), f(x2), …f(x p )), where the mean function m(x) and the covariance function k(x, x′) are given by the kernel function K i，j =κ(x i , x j ) is given; Using the Matérn kernel function: Where, v = 5 / 2, R v is the modified Bessel function, and the average function is defined as m(x) = 0; Now we have a training set D containing input and output i ,y i )}, then in the X that needs to be predicted * Calculate the conditional distribution p(y * |X * ,X,y), the Gaussian distribution is given by: p(and * |X * ,X,y)=N(y * |m * ,σ * ) According to the conditional distribution properties of multidimensional Gaussian distribution, where: m * =K(X,X * ) T K(X,X) -1 y σ * =K(X * ,X * )-K(X,X * ) T K(X,X) -1 K(X,X * )。 3. The method for online battery life prediction based on big data cloud platform based on mechanism information according to claim 1 is characterized in that: In S3, when the regression model is a partial least squares regression algorithm, the principal components in the input and the output are extracted to maximize the correlation between the principal components extracted from the input and the output, respectively, that is: Among them, E0 and F0 are the data sets after input and output standardization respectively, t1 is the first component of E0, and u1 is the first component of F0, which are defined as: t1=E0ω1, u1=F0c1 respectively.

4. The method for online battery life prediction based on big data cloud platform based on mechanism information according to claim 1 is characterized in that: In S3, the capacity decay inflection point is obtained by a knee algorithm; the knee algorithm is to calculate the line distance from the beginning to the end of the battery life in the aging trajectory, and select the point with the largest line distance.

5. The method for online battery life prediction based on big data cloud platform based on mechanism information according to claim 1, characterized in that: In the S3, it also includes life classifier training, and the life classifier training includes: dividing the battery samples into two groups of long life and short life to perform classifier training to obtain an online battery life classifier; and also includes S6, grouping the batteries to be predicted through the life classifier.

6. The method for online battery life prediction based on big data cloud platform based on mechanism information according to claim 1, characterized in that: When the method for training the lifespan classifier is a support vector machine classifier, it specifically includes: Find a hyperplane in the feature space. The hyperplane is defined as: oh T x+b=0 The corresponding classification function is: Where α=(α1,α2,…α n ) is the Lagrange multiplier, and K is a kernel function operation, as follows: K(x1,x2)=(<x1,x2>+R) d Where R = 1, d = 2; the objective function of the quadratic SVM is:

7. The method for online battery life prediction based on big data cloud platform based on mechanism information according to claim 1 is characterized in that: The remaining recyclable period is obtained through the remaining life of the battery, and the battery life is classified according to the remaining recyclable period and the life classification threshold; if the remaining recyclable period is greater than the life classification threshold, it is a long life, and if the remaining recyclable period is less than the life classification threshold, it is a short life.

8. The method for online battery life prediction based on big data cloud platform based on mechanism information according to claim 1, characterized in that: The life classification threshold is the difference between the initial classification threshold and the cycle period, and the initial classification threshold is 550 cycles.

Images