An Online Prediction Method for Battery Health State under Hybrid Working Conditions

Abstract

The present invention relates to a machine learning method for online prediction of battery health state and life driven by mechanism and data fusion, belonging to the field of battery detection. The method includes the following steps: obtaining voltage-capacity data of the battery when it is charged to the cut-off voltage at a certain current rate under different working conditions, then continuing to charge at a constant voltage until full charge, and then standing still for a period of time; based on the extracted voltage curve during standing still, using a second-order RC equivalent circuit model of the battery to identify six parameters including open-circuit voltage, resistance, and capacitance containing battery mechanism information as model inputs to obtain an online battery health state prediction model, an online remaining useful life prediction model of the battery, and an online remaining life classification model of the battery, so as to realize fast and accurate online prediction of the battery health state, remaining useful life, and battery life classification.

Description

Technical Field

[0001] The present invention belongs to the field of battery detection, and relates to an online prediction method for the state of health of a battery under hybrid operating conditions. Background Art

[0002] For the online performance prediction of batteries, the main methods are to construct prediction models based on mechanism-based physical models or data-driven models. Due to the complexity of the battery aging mechanism, it is currently difficult to find a battery aging mechanism model that conforms to multiple operating conditions. For example, in the Chinese patent with the publication number CN112881916A, an empirical model of battery aging at different temperatures is constructed using battery charge and discharge cycle data, and different methods and weight methods are used to comprehensively predict the state of health and remaining useful life of the battery to improve the prediction accuracy. However, this method overly relies on charge and discharge cycle data, assumes that future working conditions are similar to the past, predicts the capacity decay trajectory, and does not fully consider the real-time changes in the battery load conditions and different charge and discharge conditions. In addition, with the development of computer technology and artificial intelligence, data-driven methods based on statistical and machine learning technologies have attracted much attention in the research of the battery field due to their powerful data processing and nonlinear fitting capabilities. For example, in the Chinese patent with the publication number CN111007417A, a feature parameter extracted at a local charging voltage change node is provided, and a health factor is calculated as an input to train a multi-input data-driven regression model. In the implementation process of this method, complete charging voltage data is not required, and the prediction of the state of health and remaining useful life of the battery is realized based on the inconsistent health factor, which conforms to practical applications. However, the data extracted from the charging voltage node is easily affected by experimental conditions, resulting in data fluctuations and thus affecting the prediction accuracy. Moreover, it only considers the constant current discharge and the same temperature conditions, and the extrapolation ability of its model is limited. At the same time, the key of the data-driven method lies in the extraction of degradation features. The relationship between the input and output features largely determines the estimation performance, and there is usually a lack of corresponding mechanism information basis for the actual characterization of battery life, and the interpretability is poor.

[0003] To address the above problems, it is promising to combine the mechanism-based physical model and the data-driven method. Therefore, how to effectively integrate the mechanism model and the data-driven method under multiple operating conditions, improve the efficiency and reliability of the online prediction model, and realize the online estimation of the state of health of the battery, the prediction and classification of the remaining useful life, while ensuring high accuracy is an urgent problem to be solved. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide an online prediction method for the state of health of a battery under hybrid operating conditions.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] An online prediction method for the state of health of a battery under mixed working conditions, the method comprising the following steps:

[0007] S1: Obtain the voltage-capacity data of the battery after charging at a certain current rate to the cut-off voltage under different working conditions, then continuing to charge at a constant voltage until full charge, and then standing for a period of time.

[0008] S2: Based on the extracted voltage curve during standing, use a mechanism model to identify the characteristic parameters containing battery mechanism information as the model input.

[0009] S3: After the standing is over, let the battery discharge at a constant current to the discharge cut-off voltage, and the discharged capacity is used as the current capacity. The ratio of it to the initial capacity is the state of health of the battery. Use the state of health of this battery as the output to train a regression prediction model based on a machine learning method to obtain an online state of health prediction model for the battery.

[0010] S4: Use the remaining service life at any cycle within the entire life cycle of the battery before the capacity decays to 80% of the initial capacity as the output to train a regression prediction model based on a machine learning method to obtain an online remaining service life prediction model for the battery.

[0011] S5: Set classification dynamic thresholds based on the different states of health at the battery cycle, classify the battery into long-life and short-life groups as the output, and train a regression classifier based on a machine learning method to obtain an online remaining life classification model for the battery.

[0012] S6: For the battery under test under different working conditions, also extract the voltage data during the standing period to identify the parameters of the mechanism model, and input the identified parameters into the trained online state of health prediction model for the battery, the online remaining service life prediction model for the battery, and the online remaining life classification model for the battery.

[0013] S7: Online predict the state of health of the battery, the remaining service life, the long-life classification of the battery, and the short-life classification of the battery.

[0014] Optionally, the mechanism model is a second-order RC equivalent circuit model in the battery equivalent circuit model.

[0015] Optionally, in S2, for the data collected on vehicle, the standing period after the battery is fully charged becomes a promising stage for extracting reliable health-related characteristics [1,2]; the voltage curve during the standing period shows a regular downward trend as the battery ages; extract the characteristic parameters reflecting the internal electrochemical aging process of the battery through the equivalent circuit model ECM, namely the open-circuit voltage, resistance, and capacitance.

[0016] Based on machine learning methods, the physical features extracted from the ECM are used as inputs, and the battery health state / life is used as the output; lightweight machine learning methods, including Gaussian Process Regression (GPR) and Gaussian Process Classification (GPC), are used for state-of-health prediction, life prediction, and classification purposes respectively.

[0017] Optionally, the machine learning method is a physics-informed machine learning (ML) method.

[0018] Optionally, in S2, for the data collected on vehicles, the discharging process is complex and variable, and the charging process usually starts from an uncertain state of charge, making the rest period after the battery is fully charged a promising stage for extracting reliable health-related features [1,2]; the voltage curve in the rest period shows a regular downward trend as the battery ages; characteristic parameters reflecting the internal electrochemical aging process of the battery are extracted through an equivalent circuit model (ECM).

[0019] Based on a physics-informed machine learning (ML) method, the physical features extracted from the ECM are used as inputs, and the battery health state / life is used as the output; lightweight machine learning methods, including Gaussian Process Regression (GPR) and Gaussian Process Classification (GPC), are used for state-of-health prediction, life prediction, and classification purposes respectively.

[0020] Optionally, in S2, the equivalent circuit consists of an ohmic resistance R0 and two RC networks R1, R2, C1, and C2 connected in series; the battery terminal voltage U t (t) is calculated as:

[0021] U t (t) = OCV - i(t)R0 - U1(t) - U2(t)

[0022] where i(t) is the load current at time t, positive for discharging and negative for charging, and U1(t) and U2(t) are the terminal voltages of the two RC networks at time t respectively.

[0023] During the rest process after charging, when i(t) = 0, the differential forms of the terminal voltages of the RC networks with respect to time are:

[0024]

[0025]

[0026] i(0) = i at t = 0, where i is the cut-off current during the constant voltage (CV) charging process, and U1(t) and U2(t) are derived as:

[0027]

[0028]

[0029] The battery terminal voltage is expressed as:

[0030]

[0031] During the rest process, i(t)=0. When t>0, there is:

[0032]

[0033] Using the non-linear fitting method to fit the battery voltage with the experimental data to obtain the model parameters of OCV, R1, R2, C1, and C2 at different cycles; assuming t = 0, the parameter R0 at different cycles is calculated as follows:

[0034]

[0035] Complete the identification of the six parameters of the equivalent circuit model to obtain the input features.

[0036] Optionally, in S6, Gaussian process regression, output y i = f(x i ) + ε i , ε i ~ N(0, σ 2 ), and assume that both f(x) and ε follow Gaussian processes; and all the outputs in the training set are Y = (y1,..., y n ) T , and all the inputs are X = (x1,..., x n ) T , there is F = (f(x1),..., f(x n )) ~ N(0, K), K ij = k(x i , x j ), k(.) is the kernel function; the kernel function is the exponential covariance function:

[0037]

[0038] where l j represents the length scale of the jth feature, σ f is the signal standard deviation; θ = (σ, σ f , l1,..., l m ) are the hyperparameters of Gaussian process regression, and their values are obtained by minimizing the log-likelihood function NLML = -logp(Y|X, θ); in Gaussian process regression, the joint probability distribution of the training set and the prediction set is:

[0039]

[0040] Where \(I\) is an \(n\)-order identity matrix, \(x^*\) and \(y^*\) are the input and output of the prediction set respectively; on the premise that the input and output of the training set and the input of the prediction set are known, the conditional probability distribution of the output of the prediction set is:

[0041] y^*|x^*,Y,X\(\sim\)N(\(\mu,\sigma^*\) 2 )

[0042] \(\mu = K(x^*,X)(K(X,X)+\sigma\) 2 I) -1 Y

[0043] \(\sigma^*\) 2 \(= K(x^*,x^*)-K(x^*,X)(K(X,X)+\sigma\) 2 I) -1 K(X,x^*)

[0044] Where \(\mu\) is the predicted value to be obtained, \(\sigma^*\) 2 is the prediction variance;

[0045] In Gaussian process classification, the outputs of all samples are assumed to follow a Bernoulli distribution, and the output has only two possibilities, +1 and -1. The probability that the output is +1 is:

[0046] p(y = 1|x)=\(\varPhi(f(x))\)

[0047] Here, \(f(.)\) is the latent function, \(\varPhi(.)\) is the cumulative probability density function of the standard Gaussian distribution. This method uses the sigmoid function, and its formula is:

[0048]

[0049] The purpose of this function is to convert \(f(x)\) into a value in the interval \([0,1]\), and the conditional probability \(p(y|f(x))=\varPhi(yf(x))\); let \(F=(f(x_1),\cdots,f(x\) n ),(F|X\(\sim\)N(0,K)), Y=(y_1,\cdots,y n ) T , X=(x_1,\cdots,x n ) T , where \(x\) i and \(y\) i are the input and output of the training set respectively, \(K\) ij =k(x i ,x j ), \(k(.)\) is the kernel function, and the same kernel function is used in this method for classification and prediction; According to Bayes' rule:

[0050]

[0051]

[0052] p(Y|X) = ∫p(F|X)p(Y|F)df

[0053] In Gaussian process classification, the joint probability distribution of the training set and the prediction set is as follows:

[0054]

[0055] Where x* and y* are the input and output of the prediction set respectively; find the conditional probability distribution of the latent function of the prediction set:

[0056] f(x*)|X, x*, F ~ N(μ, σ 2 )

[0057] μ = K(x*, X)K(X, X) -1 F

[0058] σ 2 = K(x*, x*) - K(x*, X)K(X, X) -1 K(X, x*)

[0059] Combining f(x*)|X, x*, F and p(F|X, Y) gives:

[0060] p(f(x*)|X, Y, x*) = ∫p(f(x*)|X, x*, F)p(F|X, Y)df

[0061] According to the above formula, the output probability formula of the prediction set is:

[0062] p(y*|X, Y, x*) = ∫p(y*|f(x*))p(f(x*)|X, Y, x*)df(x*)

[0063] When solving, the Laplace numerical approximation method is used, that is:

[0064] f(x*)|X, x*, F ~ N(μ*, σ* 2 )

[0065]

[0066] σ* 2 = K(x*, x*) - K(x*, X)(K(X, X) + W -1 ) -1 K(X, x*)

[0067]

[0068]

[0069] The output probability of the prediction set is obtained by the following formula:

[0070]

[0071] If p(y*|X,Y,x*) > 0.5, the classification result of the prediction set is +1; otherwise, it is -1.

[0072] The beneficial effects of the present invention are as follows:

[0073] 1. Data is extracted using the open-circuit voltage of the battery, without the need for complete charge-discharge cycle data, and is not affected by different charging protocols and discharge strategies. For on-vehicle data, it is more in line with actual applications.

[0074] 2. A solution for battery performance prediction is proposed by combining a physical model based on mechanism information and a data-driven method. The characteristic parameters of the second-order RC equivalent circuit model are identified using the open-circuit voltage data.

[0075] 3. This method takes into account different temperatures, different charge-discharge conditions, and different types of lithium-ion batteries, and uses the Gaussian process regression method to obtain the relationship between the identified parameters and the battery health state and life. It more accurately realizes the online battery health state estimation, remaining useful life prediction, and life classification under multiple working conditions.

[0076] 4. The proposed mechanism-data fusion-driven online prediction method provides an important guarantee for the safe and reliable operation of the battery, helps to develop personalized battery management technologies, provides clear guidance for the predictive maintenance of the battery system, and at the same time, the high-accuracy online classification speeds up the sorting / recombination process of retired batteries, which is beneficial to improving energy utilization efficiency.

[0077] Other advantages, objectives, and features of the present invention will be described to some extent in the subsequent specification, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be learned from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following specification. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in detail preferably with reference to the accompanying drawings, where:

[0079] Figure 1 is the flow chart of the present invention;

[0080] Figure 2 is the modeling framework of the present invention;

[0081] Figure 3 is the equivalent circuit model;

[0082] Figure 4 It is a result graph for predicting the state of health of the battery;

[0083] Figure 5 It is a result graph for predicting the remaining useful life of the battery. Specific implementation manners

[0084] The following uses specific specific examples to illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners. The details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0085] Among them, the drawings are only for illustrative purposes, showing only schematic diagrams, rather than physical diagrams, and cannot be understood as a limitation to the present invention; in order to better illustrate the embodiments of the present invention, some components in the drawings will be omitted, enlarged or reduced, and do not represent the dimensions of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0086] In the drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "rear", etc. indicating the orientation or positional relationship, they are based on the orientation or positional relationship shown in the drawings, and are 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 operated in a specific orientation. Therefore, the terms describing the positional relationship in the drawings are only for illustrative purposes and cannot be understood as a limitation to the present invention. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0087] As Figure 1 shown, it is a flow chart of the present invention.

[0088] S1: Obtain the voltage-capacity data of the battery after charging at a certain current rate to the cut-off voltage under different working conditions, then continue to charge at a constant voltage until full charge, and then let it stand for a period of time;

[0089] S2: Based on the extracted voltage curve during standing, use the second-order RC equivalent circuit model of the battery to identify six parameters including the open-circuit voltage, resistance, and capacitance containing battery mechanism information as the model inputs

[0090] S3: After the static state ends, let the battery discharge at a constant current until the discharge cut-off voltage. The discharged capacity is used as the current capacity, and the ratio of it to the initial capacity is the state of health of the battery. Using this as the output, train a Gaussian process regression prediction model with multiple inputs; obtain an online state of health prediction model for the battery.

[0091] S4: Use the remaining service life of the battery during its entire life cycle before its capacity decays to 80% of the initial capacity as the output, and train a Gaussian process regression prediction model with multiple inputs; obtain an online remaining service life prediction model for the battery.

[0092] S5: Based on different states of health at the battery cycle, set dynamic thresholds, and use the battery being divided into long-life and short-life groups as the output to train a Gaussian process regression classifier with multiple inputs; obtain an online remaining life classification model for the battery.

[0093] S6: For the battery under test in different working conditions, also extract the voltage data during the static period to identify the parameters of the second-order RC equivalent circuit model, and input the identified parameters into the trained model.

[0094] S7: Achieve online prediction of the state of health, remaining service life, and long / short life classification of the battery.

[0095] I. Feature Extraction

[0096] 1.1 Machine Learning Method Driven by Physical Information

[0097] Figure 2 Build a modeling framework for online state of health and life prediction.

[0098] For the data collected on vehicles, the discharge process is usually complex and variable, and the charging process usually starts from an uncertain state of charge, making the static period after the battery is fully charged a promising stage for extracting reliable health-related features [1,2]. The voltage curve during the static period shows a regular downward trend as the battery ages. Extract the features reflecting the internal electrochemical aging process of the battery through an equivalent circuit model (ECM). While maintaining a high modeling fidelity, the parameters can be quickly identified online.

[0099] Based on the machine learning (ML) method driven by physical information, the physical features extracted from the ECM are used as inputs, and the state of health / life of the battery is used as the output. To maintain a high interpretability of the prediction model, we use lightweight machine learning methods, including Gaussian process regression (GPR) and Gaussian process classification (GPC), for state of health prediction, life prediction, and classification purposes respectively. In addition to high interpretability, lightweight ML also has a low computational burden in practice, so it can be embedded in a real BMS or effectively deployed on a cloud platform.

[0100] 1.2 Identification of Parameters of the Equivalent Circuit Model

[0101] The equivalent circuit model is as Figure 3 shown;

[0102] The equivalent circuit consists of an ohmic resistance (R0) and two RC networks (R1, R2, C1, C2) connected in series. The battery terminal voltage U t (t) is calculated as:

[0103] U t (t) = OCV - i(t)R0 - U1(t) - U2(t)

[0104] where i(t) is the load current at time t (positive for discharge and negative for charge), and U1(t) and U2(t) are the terminal voltages of the two RC networks at time t, respectively.

[0105] During the rest period after charging, when i(t) = 0, the differential form of the terminal voltage of the RC network with respect to time is:

[0106]

[0107]

[0108] Note that i(0) = i at t = 0, where i is the cut-off current during the constant voltage (CV) charging process, and by combining equations (2) and (3), U1(t) and U2(t) are derived:

[0109]

[0110]

[0111] Therefore, the battery terminal voltage is further expressed as:

[0112]

[0113] During the rest period when i(t) = 0 for t > 0, there is:

[0114]

[0115] Then, the battery voltage is fitted to the experimental data using the non-linear fitting method to obtain model parameters such as OCV, R1, R2, C1, C2 at different cycles. Let t = 0, and the parameter R0 at different cycles is calculated as follows:

[0116]

[0117] The above steps complete the identification of the six parameters of the equivalent circuit model to obtain the input features.

[0118] II. Machine Learning Method

[0119] 2.1 Gaussian Process Regression

[0120] In Gaussian process regression, the output y i = f(x i ) + ε i , ε i ~ N(0, σ 2 ), and it is assumed that both f(x) and ε follow a Gaussian process. And all the outputs in the training set are Y = (y1,..., y n ) T , and all the inputs are X = (x1,..., x n ). T There is F = (f(x1),..., f(x n )) ~ N(0, K), where K ij = k(x i , x j ), and k(.) is the kernel function. The kernel function selected in the present invention is the exponential covariance function:

[0121]

[0122] where l j represents the length scale of the j-th feature, and σ f is the signal standard deviation. θ = (σ, σ f , l1,..., l m ) are the hyperparameters of Gaussian process regression, and their values are obtained by minimizing the negative log marginal likelihood function NLML = -logp(Y|X, θ). The joint probability distribution of the training set and the prediction set in Gaussian process regression is:

[0123]

[0124] where I is the n-order identity matrix, and x* and y* are the input and output of the prediction set respectively. On the premise that the input and output of the training set and the input of the prediction set are known, the conditional probability distribution of the output of the prediction set is:

[0125] y*|x*, Y, X ~ N(μ, σ* 2 )

[0126] μ = K(x*, X)(K(X, X) + σ 2 I) -1 Y

[0127] σ* 2 = K(x*, x*) - K(x*, X)(K(X, X) + σ 2 I) -1 K(X, x*)

[0128] where μ is the predicted value to be obtained, and σ* 2 is the prediction variance.

[0129] To illustrate the prediction accuracy of this method, it is tested through an evaluation model. The evaluation model uses a test data set. The specific prediction model evaluation metrics are the root mean square error (RMSE) and the mean absolute percentage error (MAPE), that is, the evaluation model is:

[0130]

[0131]

[0132] where, y i refers to the actual health status or remaining useful life in the current prediction set, is the predicted value of the model, and y is the initial health status and battery life.

[0133] 2.2 Gaussian process classification

[0134] In Gaussian process classification, the outputs of all samples are assumed to follow a Bernoulli distribution, and the output has only two possibilities, +1 and -1. The probability that the output is +1 is:

[0135] p(y = 1|x) = Φ(f(x))

[0136] Here, f(.) is the latent function, and Φ(.) is the cumulative probability density function of the standard Gaussian distribution. The sigmoid function is used in this method, and its formula is:

[0137]

[0138] The purpose of this function is to convert f(x) into a value in the interval [0,1], and the conditional probability p(y|f(x)) = Φ(yf(x)). Let F = (f(x1), …, f(x n )), (F|X ∼ N(0,K)), Y = (y1, …, y n ) T , X = (x1, …, x n ) T , where x i and y i are the inputs and outputs of the training set respectively, K ij = k(x i , x j ), k(.) is the kernel function, and the same kernel function is selected for classification and prediction in this method. According to Bayes' rule:

[0139]

[0140]

[0141] p(Y|X) = ∫p(F|X)p(Y|F)df

[0142] In Gaussian process classification, the joint probability distribution of the training set and the prediction set is:

[0143]

[0144] Where x* and y* are the input and output of the prediction set respectively. From this, the conditional probability distribution of the latent function of the prediction set is obtained:

[0145] f(x*)|X,x*,F ~ N(μ,σ 2 )

[0146] μ = K(x*,X)K(X,X) -1 F

[0147] σ 2 = K(x*,x*) - K(x*,X)K(X,X) -1 K(X,x*)

[0148] Combining f(x*)|X,x*,F and p(F|X,Y) gives:

[0149] p(f(x*)|X,Y,x*) = ∫p(f(x*)|X,x*,F)p(F|X,Y)df

[0150] According to the above formula, the output probability formula of the prediction set is:

[0151] p(y*|X,Y,x*) = ∫p(y*|f(x*))p(f(x*)|X,Y,x*)df(x*)

[0152] However, this formula has no analytical solution and numerical approximation methods, such as Laplace approximation method, need to be used when solving, that is:

[0153] f(x*)|X,x*,F ~ N(μ*,σ* 2 )

[0154]

[0155] σ* 2 = K(x*,x*) - K(x*,X)(K(X,X) + W -1 ) -1 K(X,x*)

[0156]

[0157]

[0158] The output probability of the prediction set is obtained by the following formula:

[0159]

[0160] If p(y*|X,Y,x*) > 0.5, the classification result of the prediction set is +1; otherwise, it is -1.

[0161] The prediction effect of this method:

[0162] Table 1 Prediction of Battery Health State

[0163] Battery type Measurement condition RMSE(%) MAPE(%) NCM + NCA CY25 - 0.5 / 1 0.75 0.69 CY25 - 0.5 / 2 0.93 0.84 CY25 - 0.5 / 4 0.79 0.72

[0164] NCA battery: A battery with the cathode material being Li 0.86 Ni 0.86 Co 0.11 Al 0.03 O2.

[0165] NCM battery: A battery with the cathode material being Li 0.84 (Ni 0.83 Co 0.11 Mn 0.07 )O2.

[0166] NCM+NCA battery: A battery whose cathode material is composed of 42% Li(NiCoMn)O2 and 58% Li(NiCoAl)O2 fused together.

[0167] Operating conditions: In CYA-B / C, A is the temperature, B is the constant current charging current rate, and C is the constant current discharging current rate.

[0168] For example, CY25-0.5 / 1 means that this group of batteries is charged at a constant current with a current magnitude of 0.5 times the nominal capacity at 25°C, then after constant voltage charging and a rest period, it is discharged at a constant current with a current magnitude of 1 times the nominal capacity.

[0169] As Figure 4 shown, the battery health state:

[0170] Root mean square error:

[0171] Mean absolute percentage error:

[0172] Table 2 Prediction of Battery Remaining Service Life

[0173] Battery type Measurement condition RMSE (number of cycles) MAPE(%) NCM + NCA CY25 - 0.5 / 1 29.72 5.88 CY25 - 0.5 / 2 28.14 5.17 CY25 - 0.5 / 4 12.30 2.19

[0174] As Figure 5As shown, the red line on the diagonal is the ideal prediction result, and the points in the figure are the actual prediction results. The closer the points are to the diagonal, the more accurate the results. The points of different colors represent different batteries under different conditions.

[0175] Q i is the current capacity of the battery, and Q init is the initial capacity of the battery.

[0176] y is the actual value, is the predicted value, y1 is the initial value of each battery, and n is the total number of cycles of all batteries in this working condition.

[0177] Table 3 Classification Results

[0178]

[0179] The selection of the threshold is related to the SOH of the battery:

[0180] The selected threshold is 450 cycles when SOH = 100%, and 0 cycles when SOH = 80%. With these two values as the limiting conditions, the threshold is set as a dynamic variable that can decrease as the battery SOH decreases. If the remaining service life of the battery is greater than or equal to the threshold, it is judged as a long-life battery; otherwise, it is judged as a short-life battery. Since the battery is generally only used until 80% SOH, the classification is generally only carried out before it reaches 80% SOH.

[0181] Accuracy calculation formula:

[0182] In the formula, n right is the number of cycles correctly classified among all batteries in this working condition, and n all is the total number of cycles among all batteries in this working condition.

[0183] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not restrictive. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the purpose and scope of the present technical solution, and they should all be covered by the scope of the claims of the present invention.

Claims

1. An online prediction method for the state of health of a battery under hybrid operating conditions, characterized in that: The method includes the following steps: S1: Obtain the voltage-capacity data of the battery after charging at a certain current rate to the cut-off voltage under different working conditions, then continue to charge at a constant voltage until full charge, and then let it stand for a period of time. S2: Based on the extracted voltage curve during the standing period, use a mechanism model to identify the characteristic parameters containing battery mechanism information as the model input. S3: After the standing period ends, let the battery discharge at a constant current to the discharge cut-off voltage, and the discharged capacity is used as the current capacity. The ratio of it to the initial capacity is the health state of the battery. Use the health state of this battery as the output to train a regression prediction model based on a machine learning method to obtain an online health state prediction model of the battery. S4: Use the remaining service life of the battery at any cycle within the full life cycle before the capacity decays to 80% of the initial capacity as the output to train a regression prediction model based on a machine learning method to obtain an online remaining service life prediction model of the battery. S5: Set classification dynamic thresholds based on the different health states of the battery at the cycle, and use the battery divided into long-life and short-life groups as the output to train a regression classifier based on a machine learning method to obtain an online remaining life classification model of the battery. S6: For the battery under test in different working conditions, also extract the voltage data during the standing period for parameter identification of the mechanism model, and input the identified parameters into the trained online health state prediction model of the battery, the online remaining service life prediction model of the battery, and the online remaining life classification model of the battery. S7: Online predict the health state, remaining service life, long-life classification of the battery, and short-life classification of the battery.

2. The online prediction method for the battery health state under a mixed operating condition according to claim 1, characterized in that: The mechanism model is a second-order RC equivalent circuit model in the battery equivalent circuit model.

3. An online prediction method for the state of health of a battery under hybrid operating conditions according to claim 2, characterized in that: In S2, for the data collected on vehicles, the standing period after the battery is fully charged becomes a promising stage for extracting reliable health-related features [1,2]; the voltage curve during the standing period shows a regular downward trend as the battery ages; extract the characteristic parameters reflecting the internal electrochemical aging process of the battery, i.e., open-circuit voltage, resistance, and capacitance, through the equivalent circuit model ECM. Based on the machine learning method, the physical features extracted from the ECM are used as the input, and the battery health state / life is used as the output; use lightweight machine learning methods, including Gaussian process regression GPR and Gaussian process classification GPC, for health state prediction, life prediction, and classification purposes respectively.

4. The online prediction method for the battery health state under a mixed working condition according to claim 3, wherein: The machine learning method is a machine learning ML method driven by physical information.

5. The online prediction method for the battery health state under a mixed operating condition according to claim 4, characterized in that: In the above S2, the equivalent circuit is composed of an ohmic resistor R0 and two RC networks R1, R2, C1 and C2 connected in series; the battery terminal voltage U t (t) is calculated as follows: U t U(t) = OCV - i(t)R0 - U1(t) - U2(t) Where i(t) is the load current at time t, positive for discharge and negative for charge, and U1(t) and U2(t) are the terminal voltages of the two RC networks at time t respectively. During the standing process after charging, when i(t) = 0, the differential form of the terminal voltage of the RC network with respect to time is: i(0) = i at t = 0, where i is the cut-off current during the constant voltage CV charging process, and U1(t) and U2(t) are derived. The battery terminal voltage is expressed as: During the standing process, when i(t) = 0 and t > 0, there is: Use the nonlinear fitting method to fit the experimental data with the equivalent circuit model, and identify the model parameters of OCV, R1, R2, C1, and C2 at different cycles; let t = 0, and the parameter R0 at different cycles is calculated as follows: Complete the identification of the six parameters of the equivalent circuit model to obtain the input features.

6. The online prediction method for the battery health state under a mixed operating condition according to claim 5, wherein: In S6, Gaussian process regression outputs y i = f(x i ) + ε i , where ε i ~ N(0, σ 2 ), and it is assumed that both f(x) and ε follow Gaussian process distributions; and all the outputs in the training set are Y = (y1,..., y n ), T all the inputs are X = (x1,..., x n ), T and F = (f(x1),..., f(x n )) ~ N(0, K), where K ij = k(x i , x j ), and k(.) is the kernel function; the kernel function is the exponential covariance function: where l j represents the length scale of the j-th feature, and σ f is the signal standard deviation; θ = (σ, σ f , l1, …, l m ) are the hyperparameters of Gaussian process regression, and their values are obtained by minimizing the likelihood function NLML = -log p(Y|X, θ); in Gaussian process regression, the joint probability distribution of the training set and the prediction set is: Where I is the n-order identity matrix, and x* and y* are the input and output of the prediction set respectively; on the premise that the input and output of the training set and the input of the prediction set are known, the conditional probability of the output of the prediction set is: y*|x*,Y,X~N(μ,σ* 2 ) μ = K(x*, X)(K(X, X)+σ 2 I) -1 Y σ* 2 = K(x*, x*) - K(x*, X)(K(X, X)+σ 2 I) -1 K(X, x*) where μ is the predicted value to be found, and σ* 2 is the prediction variance; In Gaussian process classification, the outputs of all samples are assumed to follow a Bernoulli distribution, and the output has only two possibilities, +1 and -1, where the probability of the output being +1 is: p(y = 1|x) = Φ(f(x)) Here f(.) is the latent function, and Φ(.) is the cumulative probability density function of the standard Gaussian distribution. The sigmoid function is used, and its formula is: The purpose of this function is to convert f(x) into values in the interval [0, 1], with the conditional probability p(y|f(x)) = Φ(yf(x)); let F = (f(x1), …, f(x n ), (F|X ~ N(0, K)), Y = (y1, …, y n ) T , X = (x1, …, x n ) T , where x i and y i are the inputs and outputs of the training set respectively, K ij = k(x i , x j ), k(.) is the kernel function, and the same kernel function is used in classification and prediction; according to Bayes' rule: p(Y|X) = ∫p(F|X)p(Y|F)df In Gaussian process classification, the joint probability distribution of the training set and the prediction set is: Where x* and y* are the input and output of the prediction set respectively; find the conditional probability distribution of the latent function of the prediction set: f(x*)|X,x*,F~N(μ,σ 2 ) μ = K(x*, X)K(X, X) -1 F σ 2 = K(x*, x*) - K(x*, X)K(X, X) -1 K(X, x*) Combined with f(x*)|X, x*, F and p(F|X, Y), we get: p(f(x*)|X, Y, x*) = ∫p(f(x*)|X, x*, F)p(F|X, Y)df According to the above formula, the output probability formula of the prediction set is: p(y*|X, Y, x*) = ∫p(y*|f(x*))p(f(x*)|X, Y, x*)df(x*) The Laplace numerical approximation method is used in the solution, that is: f(x*)|X,x*,F~N(μ*,σ* 2 ) σ* 2 = K(x*, x*) - K(x*, X)(K(X, X) + W -1 ) -1 K(X, x*) Furthermore, the output probability of the prediction set is obtained by the following formula: If p(y*|X, Y, x*) > 0.5, the classification result of the prediction set is +1, otherwise it is -1.

Images