METHOD AND DEVICE FOR PREDICTING THE HEALTH OF A BATTERY USING A HYBRID NEUTRONAL NETWORK MODEL AND METHOD AND DEVICE FOR MODEL TRAINING

Abstract

The present disclosure provides a method for predicting the health state of a battery using a hybrid neural network model. The method comprises: temporally sequential input of multiple input signals from multiple time points, contained in a measured cycle data set, into the hybrid neural network model, wherein the hybrid neural network model comprises a neural network submodule and a physical model submodule, and each input signal of the multiple input signals of this cycle data set comprises multiple physical values ​​of the battery from a time point;In one iteration for an input signal at each of the multiple time points: prediction of a second value in the last iteration by the Neural Network submodule based on the input signal and a first value output in the previous iteration by the Physical Model submodule; calculation of a first value in the current iteration by the Physical Model submodule based on the input signal and the second value; repetition of the steps until the last iteration for the last input signal at the last time point; and determination of the battery health based on the first value.

Description

AREA OF INVENTION

[0001] The present application relates to artificial intelligence technology, in particular a method and a device for predicting the health of a battery using a hybrid neural network model, and a method and a device for training a hybrid neural network model. STATE OF THE ART

[0002] Rechargeable lithium batteries are commonly used in various electrical systems, such as electric vehicles. Lithium batteries generally consist of a negative electrode, a positive electrode, and a separator located between the two. Both electrodes contain active materials that react reversibly with lithium. For example, the negative electrode contains lithium metal, which is electrochemically dissolved and reversibly deposited. The separator contains an electrolyte with lithium cations and acts as a physical barrier between the electrodes, preventing any electrically conductive connection between them within the battery. Typically, during charging, electrons are generated at the positive electrode and consumed at the negative electrode. The reverse reaction occurs during discharging.Repeated charging and discharging cycles of the battery can cause undesirable side effects, which can lead to a decrease in battery capacity.

[0003] Battery management systems in electrical systems can be used to evaluate the state-of-health (SOH) of batteries. Existing physical electrochemical aging models struggle to account for complex and dynamic operating conditions. Therefore, the goal is to evaluate battery health more accurately and efficiently. REVELATION OF THE INVENTION

[0004] The following introduction briefly presents some selected concepts that will be described in more detail later. This introduction is not intended to highlight essential or necessary features of the claimed subject matter, nor to limit its scope.

[0005] According to one aspect of the present application, a method for predicting the health of a battery using a hybrid neural network model is provided, comprising: temporally sequential input of multiple input signals from multiple time points contained in a measured cycle data set into the hybrid neural network model, wherein the hybrid neural network model comprises a neural network submodule and a physical model submodule, and each input signal of the multiple input signals of this cycle data set comprises multiple physical values ​​of the battery from a time point;In one iteration for an input signal from each of the multiple time points: prediction of a second value by the Neural Network submodule based on the input signal and a first value output by the Physical Model submodule in the previous iteration; calculation of a first value of the current iteration by the Physical Model submodule based on the input signal and the second value; repetition of the steps until the last iteration for the last input signal at the last time point; prediction of a second value in the last iteration by the Neural Network submodule based on the last input signal and a first value output by the Physical Model submodule in the respective previous iteration; calculation of a first value in the current iteration by the Physical Model submodule based on the last input signal and the second value.Determining the battery's health status based on the first value.

[0006] According to one aspect of the present application, a method for training a hybrid neural network model is provided, comprising: storing parameters of the hybrid neural network model in a shared memory block, wherein the hybrid neural network model comprises a neural network submodule and a physical model submodule, wherein a first training dataset for training the hybrid neural network model comprises several cycle data groups, each of the cycle data groups comprising several input signals from several time points and a target output actual value, wherein each of the input signals comprises several physical values ​​from one time point, wherein in an iteration for an input signal from one time point, the neural network submodule is configured based on the input signal and a first value output by the physical model submodule in the previous iteration.a second value is predicted, and the physical model submodule calculates a first value in the current iteration based on the input signal and the second value; parallel execution of multiple processes, wherein the multiple processes in parallel, using the parameters of the hybrid neural network model stored in the shared memory block, train the hybrid neural network model based on the first training data set, wherein for each of the multiple processes and each cycle data group used by the process, the process iteratively generates a predicted target value using the parameters of the hybrid neural network model stored in the shared memory block based on the temporally ordered multiple input signals from multiple time points in the cycle data group.and, based on the predicted target value and the target-actual value of the cycle data group, updates the parameters of the hybrid neural network model stored in the shared memory block.

[0007] According to one aspect of the present application, a device for hybrid neural network models is provided, comprising: a memory; and one or more processing units configured to execute, upon execution of the program instructions, the prediction procedure of the health of a battery using a hybrid neural network model and the method for training a hybrid neural network model described herein.

[0008] According to one aspect of the present application, a machine-readable storage medium is provided on which executable instructions are stored, wherein the instructions, when executed by one or more processors, implement the battery health prediction method described herein by means of a hybrid neural network model and a method for training a hybrid neural network model.

[0009] By applying the battery health prediction method using a hybrid neural network model according to the present application, the accuracy of the battery health prediction can be effectively improved by learning the dynamic weighting coefficients in the physical model submodule, depending on the operating environment, using the neural network submodule. Furthermore, by applying the method for training a hybrid neural network model to predict the battery health according to the present application, the training time can be effectively reduced and the training efficiency improved by leveraging the properties related to the battery measurement data.This reduces the problem of long training time, which is caused by the temporally serial processing of the hybrid neural network model and the high time expenditure for processing physical formulas at each time step; similarly, this contributes to improving the problem of high memory requirements. Further advantages of the present disclosure are explained in more detail in the following description. DESCRIPTION OF THE FIGURES

[0010] A deeper understanding of the nature and benefits of this revelation can be achieved by referring to the following figures. Similar components or features in the figures may have the same reference signs. Fig. Figure 1 shows a schematic diagram of a hybrid neural network model according to an exemplary implementation. Fig. Figure 2 shows a simplified schematic calculation diagram for training a hybrid neural network according to an exemplary implementation. Fig. Figure 3 shows a schematic diagram of a procedure for training a hybrid neural network model according to an exemplary embodiment. Fig. Figure 4 shows a schematic diagram of a procedure for training a hybrid neural network model according to an exemplary embodiment. Fig. Figure 5 shows a schematic diagram of a procedure for training a hybrid neural network model according to an exemplary embodiment. Fig. Figure 6 shows a flowchart of a prediction procedure for the health status of a battery using a hybrid neural network model according to an exemplary embodiment. Fig. Figure 7 shows a flowchart of a procedure for training a hybrid neural network model according to an exemplary implementation. Fig. Figure 8 shows a block diagram of a device for a hybrid neural network model according to an embodiment. DETAILED DESCRIPTION OF THE EXECUTION FORMS

[0011] The subject matter described herein will now be discussed with reference to exemplary embodiments. It is understood that the explanation of these embodiments serves only to improve understanding by those skilled in the art and to realize the subject matter of this document, without limiting the scope of protection, applicability, or examples in the claims. The functions and arrangement of the described elements can be modified without altering the scope of protection set forth in the claims. In each example, various processes or components can be omitted, replaced, or added as needed. For example, the described methods can be carried out in a different sequence than described, and individual steps can be added, omitted, or combined. Furthermore, features described in relation to some examples can also be combined in other examples.

[0012] The term "comprehensive" and its variants used herein can be interpreted as an open term meaning "comprehensive but not limited to." The term "based on" can mean "at least partially based on." The terms "one embodiment," "some embodiments," etc., can mean "at least one embodiment." The term "another embodiment" means "at least one other embodiment." The terms "first," "second," etc., can refer to different or the same objects. The definition of a term is consistent throughout this specification unless the context clearly indicates otherwise.

[0013] Fig. Figure 1 shows a schematic diagram of a hybrid neural network model according to an exemplary implementation.

[0014] The in Fig. The hybrid neural network model 100 shown comprises a neural network submodule 110 and a physical model submodule 120. The hybrid neural network model 100 receives a measurement signal 130 from a battery, which includes, for example, physical values ​​such as voltage, current, temperature, and state of charge (SOC) of the battery, and determines the battery's state of health (SOH) based on the received measurement signal 130. This state of charge is, for example, 30%, 60%, etc. This state of health could also be the battery's capacity or remaining capacity. Assuming that the battery's initial capacity is C0, and that the battery's capacity decreases with use and aging, the current capacity of the battery can be used to indicate the battery's state of health. For example, assuming the battery's initial capacity is 1.Then the remaining battery capacity can be represented as a fraction less than 1.

[0015] The most important aging phenomena of batteries include: growth of the solid electrolyte interface (SEI), leading to lithium loss and solvent consumption; lithium plating, which causes lithium loss; and loss of active material at the cathode and anode. In one embodiment, an aging model can be represented by a group of physical formulas: x˙SEI=(Jsr1−+Jsr2−)MLCO3ρLCO3A x˙cLi=F(2Jsr1−+2Jsr2−−Jsr1+−Jsr2+) x˙Csol1,2=(−Jsr1,2−+Jsr1,2+) / Vel x˙Li=JplF x˙AML−=FSOC−(Jsr1−+Jsr2−)(1+k1−|I(t)|) x˙AML+=FSOC(Jsr1++Jsr2+)(1+k1+|I(t)|)

[0016] Formula (1) represents the change in SEI layer thickness due to side reactions at the anode. Formula (2) represents the change in circulating lithium due to side reactions at the anode (-) and cathode (+). Formula (3) represents the consumption of the electrolyte solvent. Formula (4) represents the loss of circulating lithium due to lithium plating. Formulas (5) and (6) represent the loss of active material at the anode and cathode, respectively. The side reactions in formulas (1) to (6) can be represented by the following physical formulas: Jsr1,2−=ksr1,2−AxCsol1,21+ksr1,2−xSEI / D1,2 D1,2=D0 ​​1,2eaed1,2−(T(t)−T0)T(t)T0 ksr1,2−=ksr1,20eαesr1,2−(T(t)−T0)T(t)−T0eα1,2−FRT(t)(U−(SOC−(t))−kpol(V(t)−OCV(SOC(t)))−Usr−−I(t)xSEIkSEIA) Jsr1,2+=xCsol1,2ksr+e−αesr+(T(t)−T0)T(t)−T0eα+FRT(t)(U+(SOC(t))−(1−kpol)(V(t)−OCV(SOC(t)))−Usr+) Jpl=kple−FRT(U−(SOC−(t))−kpol(V(t)−OCV(SOC(t)))−l(t)xSEIkSEIA)

[0017] In formulas (7) and (10) Jsr1,2− and Jsr1,2+ the side reaction current at the anode or the side reaction current at the cathode in connection with the first or second electrolyte, respectively, while in formula (11) J pl Lithium plating reaction current is.

[0018] The state of health (SOH) of the battery can be defined as the remaining capacity of the battery. By solving the ordinary differential equations of formulas (1) to (6), the physical values ​​or system states x can be determined. cLi , x Li , x AML - and x AML + These factors affect the battery's capacity. The remaining capacity (SOH) C The battery can be represented as follows: SOHC=C0−(xcLi+xLi+kxxAML−+kcxAML+)

[0019] Here, C0 in formula (12) denotes the initial capacity of the battery, while k a and k c Represent coefficients (or weighting factors).

[0020] The electrochemical aging model of the battery represented by formulas (1) to (12) can be used to determine the residual capacity SOH. CThe battery's aging process is calculated using physical formulas based on the measured physical values ​​of the battery, such as current I(t), voltage V(t), temperature T(t), and state of charge (SOC(t)) in formulas (7) to (11). In this document, the aging model represented by formulas (1) to (12) is used as an example to illustrate the technical solutions of this disclosure. It is understandable to a person skilled in the art that the described embodiment for modeling the electrochemical aging of the battery according to formulas (1) to (12) is merely an example to illustrate the technical solution of this disclosure. The electrochemical aging model is not limited to a specific implementation. Any suitable electrochemical aging model that complies with the laws of physics can be used.

[0021] The in Fig. The physical model submodule 120 shown comprises a physical aging model 1210. In an example, the physical model 1210 can include ordinary differential equations (ODEs), represented by the aforementioned formulas (1) to (6), and the corresponding physical formulas (7) to (11). In particular, for a measurement signal 130 at time t, the physical aging model 1210 can obtain a numerical representation of the ordinary differential equations (1) to (6) based on the measurement signal 130 by formulas (1) to (11), where the measurement signal 130 comprises the current I(t), the voltage V(t), the temperature T(t), and the state of charge (SOC) of the battery. The Physical Model submodule 120 also includes an ODE solver 1220, wherein the ODE solver 1220 solves the ordinary differential equations (1) to (6) by a forward process to determine the system states x SEI , x cLi , x Csoll,2 , xLi , x AML - and x AML + of the battery, which represents the change in SEI layer thickness due to side reactions at the anode, the change in circulating lithium due to side reactions at the anode (-) and cathode (+), the consumption of the electrolyte solvent, the loss of circulating lithium through lithium plating, the loss of active material at the anode, and the loss of active material at the cathode. It is understood that the ODE solver 1220 can be any suitable ODE solver, such as a Runge-Kutt solver, an Euler solver, etc. The physical model submodule 120 also includes a SOH determination unit 1230, where the SOH determination unit 1230 determines the residual SOH capacity by formula (12). C The battery's state of health is determined. In another embodiment, the initial capacity of the battery is represented as 1, and the remaining capacity is SOH. Ccan be expressed as a fraction less than 1. For example, the SOH determination unit 1230 can determine the remaining SOH capacity of the battery as the battery's state of health using formula (13). SOHC=1−(xcLi+xLi+kxxAML−+kcxAML+) / (C0*3600)

[0022] Here, C0 in formula (13) denotes the initial capacity of the battery, while k a and k c Represent coefficients.

[0023] In embodiments that use the physical aging model, it is in formulas (5) and (6) which are used to calculate the loss x AML - to active material at the anode and the loss x AML + an active material is used at the cathode, which is necessary to counteract the side reaction current. Jsr1,2− at the anode and the side reaction current Jsr1,2+ at the cathode in connection with the first or second electrolyte corresponding coefficients k I-|I(t)| and k I+ |I(t)| to be set. The input signal with I(t), V(t), T(t) and SOC(t) influences the loss x. AML - to active material at the anode and the loss x AML + an active material at the cathode, which are calculated in equations (5) and (6). Therefore, it is preferable to use corresponding coefficients k for different input signals. I -|(t)| and k I +|I(t)| to apply. However, since it is difficult to use exact physical formulas for representation, the coefficient values ​​k, based on empirical data, can be used. I- |(t)| and k I +|I(t)| leads to a reduction in the prediction accuracy of the physical aging model.

[0024] In the exemplary embodiment in Fig. 1 says the neural network submodule 110 based on the input signal with / (t), V(t), T(t) and SOC(t) from a time t which corresponds to the side reaction current Jsr1,2− at the anode and the side reaction current Jsr1,2+ The corresponding coefficients k at the cathode I- |(t)| and k I +|I(t)| 140 previously. In one embodiment, the neural network submodel 110 can be configured as a fully connected neural network. Accordingly, the physical model submodule 120 outputs, based on the input signal with / (t), V(t), T(t) and SOC(t) and the coefficients k output by the neural network submodule 110. I- |(t)| and k I+ |I(t)| 140 physical parameters 150 from, where the physical parameters 150 are the system states x SEI , x cLi , x Csoll,2 , x Li, x AML -, x AML + the battery and the remaining capacity SOH C the battery, including the remaining capacity SOH C The battery's state of health can be considered a factor in determining the battery's condition.

[0025] In the exemplary embodiment in Fig. 1. From a time t, the neural network sub-model 110 processes the input signal with I(t), V(t), T(t) and SOC(t) from that time t, as well as at least a part of the output signal of the physical model sub-module 120 at the previous time t-1, as input to determine the coefficients k. I- |(t)| and k I +|I(t)| 140 to obtain for time t. For example, from a time t, the neural network submodel 110 processes the input signal with / (t), V(t), T(t) and SOC(t) from that time as well as the system states x SEI , x cLi , x Csoll,2 , x Li , x AML- and x AML +, which were output by the Physical Model submodule 120 at the previous time t-1, as input to determine the coefficients k I- |I(t) and k I to obtain +|I(t)| 140 for time t.

[0026] In another embodiment, the neural network submodule 110 predicts the coefficients 140 based on the input signal I(t), V(t), T(t) and SOC(t) from a time t, where the coefficients 140 correspond to the side reaction current. Jsr1,2− at the anode and the side reaction current Jsr1,2+ The corresponding coefficients k at the cathode I- |(t)| and k I+ |(t)| and the consumption x AML active material at the anode and the consumption x AML + corresponding coefficients k on active material at the cathode a and k c This includes the Physical Model submodule 120, based on the input signal with I(t), V(t), T(t) and SOC(t) and the coefficients k output by the Neural Network submodule 110. I- |(t)|, k I+ |(t)|, k a and k c140 physical parameters 150 from, where the physical parameters 150 are the system states x SEI , x cLi , x Csoll,2 , x Li , x AML - , x AML + the battery and the remaining capacity SOH C the battery, including the remaining capacity SOH C The battery's state of health can be considered a factor in determining the battery's condition.

[0027] In the exemplary embodiment in Fig. 1. From a time t, the neural network sub-model 110 processes the input signal with I(t), V(t), T(t) and SOC(t) from that time t, as well as at least a part of the output signal of the physical model sub-module 120 at the previous time t-1, as input to determine the coefficients k. I- |(t)|, k I+ |(t)|, k a and k cTo obtain 140. For example, from a time t, the neural network submodel 110 processes the input signal with I(t), V(t), T(t) and SOC(t) from that time as well as the system states x SEI , x cLi , x Csoll,2 , x Li , x AML - and x AML +, which were output by the Physical Model submodule 120 from the previous time t-1, as input to determine the coefficients k I- |I(t) , k I+ |(t)|, k a and k c to obtain 140 for time t.

[0028] The preceding statements in conjunction with equations (1) to (13) and the one in Fig. The embodiment shown in Figure 1 describes the embodiments of the various aspects of the present disclosure for predicting the battery's state of health. It is understood, however, that the present disclosure is not limited to the specific examples described. For instance, in one embodiment, the output 150 of the physical model submodule 120 may not be provided to the neural network submodule 110, but the neural network submodule 110 outputs the coefficients 140 solely based on the input signal with I(t), V(t), T(t), and SOC(t) at time t. Furthermore, the physical model submodule 120 may, for example, use any suitable physical aging model known or to those skilled in the art and is not limited to the physical aging models represented by the ODE in formulas (1) to (13).For example, the Physical Model submodule 120 can be an aging model represented by suitable partial differential equations (PDEs).

[0029] In one embodiment, to predict the battery's state of health using the hybrid neural network model, 100 data points from a single cycle are collected, comprising multiple input signals from multiple time points. For example, the cycle data might include several thousand measurement signals from several thousand time points, each of which could be, for example, the current I(t), voltage V(t), temperature T(t), and state of charge (SOC) of the battery. The data from this cycle can be referred to as the measured cycle data set. As in Fig. As shown in Figure 1, multiple input signals at multiple time points in a cycle data set are fed sequentially into the hybrid neural network model 100. In one iteration for an input signal from each of the multiple time points, a second value is predicted by the neural network submodule 110 based on the input signal and a first value output by the physical model submodule 120 in the previous iteration. A first value in the current iteration is calculated by the physical model submodule 120 based on the input signal and the second value. These steps are repeated until the last iteration for the last input signal at the last time point.In the last iteration, a second value is predicted by the Neural Network submodule 110 based on the last input signal and a first value output by the Physical Model submodule 120 in the respective previous iteration. A first value in the current iteration is calculated by the Physical Model submodule 120 based on the last input signal and the second value. Finally, the battery's state of health is determined based on the first value. In one embodiment, this first value can represent the system states x. SEI , x cLi , x Csoll,2 , x Li , x AML - and x AML + include. In one embodiment, the aforementioned second value can be the coefficients k. I- |I(t)| and k I +|I(t)| include. In one embodiment, the aforementioned second value can be the coefficients k. I- |(t)|, k I+ |I(t)|, k a and kc 140. It is understood that if different physical formulas are used to implement the physical aging model 120, the first and second values ​​mentioned may be corresponding physical values ​​and coefficients and are not limited to the specific physical values ​​and coefficients explained in the described embodiments.

[0030] As described, this must be in Fig. The hybrid neural network model 100 shown in Figure 1 processes the measurement data of a cycle in a temporally sequential or serial manner. This serial processing can lead to difficulties in the training process of the hybrid neural network model 100, one of which is the long training time.

[0031] In one example, the training dataset comprises multiple batteries, each containing multiple detection cycles, each containing multiple timestamps and the measurement signal for each timestamp. For instance, the dataset includes 33 batteries, with each battery containing 4 to 8 measurement cycles. Each cycle, in turn, contains several thousand timestamps as well as input signals such as current, voltage, temperature, and state of charge for each timestamp. Each cycle has a target-actual value (state of health), and the number of target-actual values ​​for each battery corresponds to the number of its cycles. During the training process, the hybrid neural network model 100 is used to predict the battery's state of health based on the data in the training dataset.Subsequently, the trainable parameters of the hybrid neural network model 100, such as the trainable parameters of the neural network submodule 110 and the optimizable data of the physical model submodule 120, are updated based on the predicted values ​​and the target / actual values ​​of the battery state. Since the measurement data of a cycle must be processed sequentially over time, and the physical model 120 is based on ordinary differential equations (ODEs), the joint learning of the ODEs and the neural network results in a significant training time expenditure. Simultaneously, the sequential temporal processing necessitates the storage of the parameters of each time step, leading to high memory requirements.

[0032] Fig. Figure 2 shows a simplified schematic calculation diagram for training a hybrid neural network 100 according to an exemplary implementation. The in Fig. The computation diagram shown in Figure 2 represents the operations for a training dataset of one cycle. Each node in the first row, represented as a circle, represents an input signal 210 from a specific time. Each node in the second row represents the computation 220 of the Neural Network submodule 110. Each node in the third row represents the computation 230 of the Physical Model submodule 120.

[0033] During the training of the hybrid neural network 100 with the training dataset, an input signal 210 from a specific time point is fed into the neural network submodule 110. This input signal 210 and the output data of the neural network submodule 110 are then fed into the physical model submodule 120. The physical aging model 1210 and the ODE solver 1220 of the physical model submodule 120 calculate the hidden states for the next time step, such as the system states x.SEI , x cLi , x Csoll,2 , x Li , x AML - and x AML +. Then, for each subsequent time point, or time step, the same process is performed iteratively until the last time point of the cycle. The SOH determination unit 1230 of the physical model submodule 120 calculates the remaining battery capacity as a predicted target value of 240. Based on the predicted target value of 240 and the target actual value of the cycle, a loss value can be obtained. For example, the loss value is the difference between the target value and the target actual value.

[0034] Subsequently, based on the loss value, a sequential backward propagation is performed from node 240 of the Physical Model submodule 120 and the Neural Network submodule 110 of the respective previous time step, successively continuing through the preceding time steps up to the first time step of this cycle. This updates the trainable parameters of the Physical Model submodule 120 and the Neural Network submodule 110. In one embodiment, the consumption x AML - on active material at the anode and the consumption x AML + corresponding coefficients k on active material at the cathode a and k c configured as trainable parameters so that the optimal coefficients k are determined during the training process of the hybrid neural network model. a and k cto be obtained in order to improve the prediction accuracy of the hybrid neural network model 100.

[0035] Backpropagation of ODE is complex due to the complexity of the ODE computation, so backpropagating each time step takes a relatively long time. For the example dataset, there are several thousand time steps per cycle, 4 to 8 cycles per battery, and a total of 33 batteries. The number of time steps in the entire dataset is extremely high, and each time step is very time-consuming, so the training time during the training process with serial processing, as in Fig. As shown in Figure 2, this process takes a very long time. For example, for the sample dataset, the forward processing time of a training epoch is approximately 15 minutes, and the backward processing time exceeds 2 hours. The term "epoch" refers to a complete run through all the data in the training dataset. In actual implementation, the training process of the hybrid neural network model 100 requires numerous epochs. Furthermore, real-world training datasets can be significantly larger than the sample dataset used here. Therefore, it is advantageous to shorten the training time of the hybrid neural network model 100 and thus increase the training efficiency.

[0036] Fig. Figure 3 shows a schematic diagram of a method for training a hybrid neural network model according to an exemplary implementation. The in Fig. The hybrid neural network model 300 shown in 3 can be used in Fig. The hybrid neural network model shown is 100. As in Fig. As shown in Figure 3, the hybrid neural network model 300, including parameters of the hybrid neural network model 300, is stored in a shared memory block. Several processes are executed in parallel, for example, processes P1 to P4, as shown in Figure 3. Fig. 3 shown. The four processes P1 to P4, which are in Fig. As shown in Figure 3, the same hybrid neural network model 300 is shared, and the four processes P1 to P4 are executed in parallel using the same training dataset D. This means that the four processes P1 to P4 share the parameters of the same hybrid neural network model 300 and update the parameters of the same hybrid neural network model 300 in parallel using the same dataset.

[0037] In one embodiment, the multiple processes P1 to P4 perform the training of the hybrid neural network model 300 in parallel according to different sequences of the multiple cycle data groups in the training data set D.

[0038] For example, the training dataset D is divided into four subsets D1 to D4, and processes P1 to P4 each train the hybrid neural network model 300 with different sequences of the subsets D1 to D4. For example, process P1 trains the hybrid neural network model 300 with set D in the order {D1, D2, D3, D4}. Process P2 trains the hybrid neural network model 300 with set D in the order {D2, D3, D4, D1}. Process P3 trains the hybrid neural network model 300 with set D in the order {D3, D4, D1, D2}. Process P4 trains the hybrid neural network model 300 with set D in the order {D4, D1, D2, D3}. It is understood that processes P1 to P4 can also train the model 300 in other sequences of the subsets D1 to D4.Furthermore, the training data of the training dataset D can be sorted, for example, by cycle data groups or by batteries, with processes P1 to P4 each performing the training of the hybrid neural network model 300 with the dataset D, whose training data are sorted differently.

[0039] As in Fig. As shown in Figure 3, when using the four parallel processes P1 to P4, each of them independently executes the same data set D for the same hybrid neural network model 300, as shown in Figure 3. Fig. The serial training process shown in Figure 2 is carried out. If the four processes P1 to P4 execute one epoch in parallel, this corresponds to four epochs of training for the hybrid neural network model 300 with the same dataset D, thereby increasing the efficiency of the hybrid neural network model 300.

[0040] Fig. Figure 4 shows a more detailed schematic diagram of the procedure in Fig. 3. As in Fig. As shown in Figure 4, the four processes P1 to P4 independently perform the training of the hybrid neural network model 300 with their respective cycle data groups P1-C1, P1-C2, etc., up to P4-C1, P4-C2, etc. In an embodiment as shown in Figure 4, the following applies: Fig. As shown in Figure 4, after process P2 updates the parameters of hybrid neural network model 300 with the current cycle data set P2-C1, all processes P1 to P4 perform the subsequent processing according to the updated hybrid neural network model 300. Similarly, after process P1 updates the parameters of hybrid neural network model 300 with the current cycle data set P1-C1, all processes P1 to P4 also perform the subsequent processing according to the updated hybrid neural network model 300. It should be understood that the length of each cycle of data set D is not necessarily the same.

[0041] Fig. Figure 5 shows a schematic diagram of a procedure for training a hybrid neural network model according to an exemplary embodiment.

[0042] The training dataset described here serves as an example. Since the measurement data acquired during the battery charging / discharging process often exhibit linear changes at multiple time points, this allows for the use of subsampling to train the hybrid neural network model using subsets of the training dataset. In one embodiment, several different subsampling rates 510 can be set, such as subsampling rates 1 to 8, as shown in Fig. Figure 5 shows that a subsampling rate of 1 means that all data samples in the dataset are sampled. A subsampling rate of 2 means that half of the data samples in the dataset are sampled; for example, every second data sample is sampled, and so on. A subsampling rate of 8 means that one-eighth of the data samples in the dataset are sampled; for example, every eighth data sample is sampled.

[0043] Using a training dataset D', the hybrid neural network model 100 is trained for each subsampling rate 1 to 8 using a subset of the training dataset D' obtained based on the respective subsampling rate. The performance of the trained hybrid neural network model 100 is evaluated on both the data subset and a test set, with the performance being expressed as the mean square error (MSE), which is presented in Fig. 5 is designated as 520 and 530, where 520 represents the MSE on the training subset and 530 represents the MSE on the test set. Simultaneously, the training duration can be recorded for each subsampling rate. For example, for the data subset corresponding to this subsampling rate, the time for a training epoch is recorded, which is in Fig. The 5 is designated as 540 and 550, where 540 represents the forward time and 550 the backward time of an epoch. Based on the generated performance metrics for the various subsampling rates, such as the MSE 530 test set, one or more subsampling rates are selected for training the hybrid neural network 100.

[0044] In one embodiment, the Fig. 5. The training dataset D' described for selecting the subsampling rate is a subset of the actual training dataset D, which is in the Fig. 2 to 4 are used. In this way, the appropriate subsampling rate for the actual training process can be selected with less computational effort. Furthermore, in other embodiments, the training data set D' can be identical to the actual training data set D, or the training data set D' can intersect with the actual training data set D but be different.

[0045] In one embodiment, such as in Fig. As shown in Figure 5, subsampling rates 2 and 8 can be used for training the hybrid neural network model 100. These subsampling rates can be used in different training phases of the hybrid neural network model 100. In one embodiment, the larger subsampling rate 8 can be used for model training during the structure adaptation phase of the hybrid neural network model 100. In this embodiment, the structure of the hybrid neural network model is adapted to obtain several candidate models of the hybrid neural network with different structures. Using subsampling rate 8, these candidate models of the hybrid neural network are each trained based on the training dataset D.Based on the respective performance metrics of the trained candidate models of the hybrid neural network with different structures, one candidate model is selected. If a hybrid neural network model with an optimal structure is selected, this candidate model is trained on the training dataset D using a subsampling rate of 2.

[0046] Fig. Figure 6 shows a flowchart of a prediction procedure for the health status of a battery using a hybrid neural network model according to an exemplary embodiment.

[0047] In step 610, the multiple input signals from multiple time points contained in a measured cycle data group are entered temporally sequentially into the hybrid neural network model, wherein the hybrid neural network model comprises a neural network submodule and a physical model submodule, and each input signal of the multiple input signals of this cycle data group comprises multiple physical battery values ​​from one time point.

[0048] In step 620, in one iteration, for an input signal at each of the multiple time points, a second value is predicted by the Neural Network submodule based on the input signal and a first value output by the Physical Model submodule in the previous iteration. A first value in the current iteration is calculated by the Physical Model submodule based on the input signal and the second value. These steps are repeated until the last iteration for the last input signal at the last time point. In the last iteration, a second value is predicted by the Neural Network submodule based on the last input signal and the first value output by the Physical Model submodule in the respective previous iteration.and a first value in the current iteration is calculated by the Physical Model submodule based on the last input signal and the second value.

[0049] In step 630, the battery's state of health is determined based on the first value. For example, the battery's state of health can be expressed as the remaining capacity (SOH). C The battery, as described, can be represented. For example, the second value generated in step 620 can represent the described system states x. cLi , x Li , x AML - and x AML + include. In step 630, for example, the remaining capacity SOH is C The battery's health status, calculated based on formula (12), is considered. For example, the second value generated in step 620 can represent the described system states x cLi , x Li , x AML -, x AML + and the remaining capacity SOH Cthe battery. In step 630, for example, the remaining capacity SOH is determined. C The second value is considered the battery's health status.

[0050] In one embodiment, the physical model submodule comprises a physical formula for calculating the battery's state of health based on the measured battery values. In another embodiment, the physical formula comprises an ordinary differential equation or partial differential equation for representing the battery's state of health based on the measured battery values. In another embodiment, the neural network submodule comprises a fully connected neural network model.

[0051] In one embodiment, the multiple physical values ​​contained in each input signal at a given time include at least a current value, a voltage value, a temperature value, and a state of charge of the battery at that time. In one embodiment, the first value includes at least a change in the solid electrolyte boundary layer thickness of the battery, a change in circulating lithium due to side reactions at the anode (-) and cathode (+), consumption of electrolyte solvent, loss of circulating lithium due to lithium plating, loss of active material at the anode, and loss of active material at the cathode. In one embodiment, the second value includes a weighting factor used to calculate the battery's capacity.The second value includes, for example, weighting factors corresponding to the side reaction current at the anode and the side reaction current at the cathode of the battery. Furthermore, the second value also includes, for example, weighting factors corresponding to the loss of active material at the anode and the loss of active material at the cathode of the battery.

[0052] Fig. Figure 7 shows a flowchart of a procedure for training a hybrid neural network model according to an exemplary implementation.

[0053] In step 710, parameters of the hybrid neural network model are stored in a shared memory block, wherein the hybrid neural network model comprises a neural network submodule and a physical model submodule, wherein an initial training dataset for training the hybrid neural network model comprises multiple cycle data groups, each of which comprises multiple input signals from multiple time points and a target output actual value, wherein each of the input signals comprises multiple physical values ​​from one time point, wherein in one iteration for an input signal from one time point, the neural network submodule predicts a second value based on the input signal and a first value output by the physical model submodule in the previous iteration.and the physical model submodule calculates a first value in the current iteration based on the input signal and the second value.

[0054] In step 720, several processes are executed in parallel, whereby the several processes train the hybrid neural network model based on the first training dataset using the parameters of the hybrid neural network model stored in the shared memory block, whereby for each of the several processes and each cycle data group used by the process, the process iteratively generates a predicted target value based on the temporally ordered multiple input signals from multiple time points in the cycle data group, and based on the predicted target value and the target actual value of the cycle data group, the parameters of the hybrid neural network model stored in the shared memory block are updated.

[0055] In one embodiment, the multiple processes perform the training of the hybrid neural network model in parallel, based on the multiple cycle data groups according to different sequences of the multiple cycle data groups in the first training data set.

[0056] In one embodiment, the physical model submodule includes a physical formula for calculating the battery's state of health based on the measured battery values. In another embodiment, the neural network submodule includes a fully connected neural network model.

[0057] In one embodiment, the multiple physical values ​​contained in each input signal at a given time include at least partially the current, voltage, temperature, and state of charge of the battery at that time. The first value includes at least partially a change in the solid electrolyte boundary layer thickness of the battery, a change in circulating lithium due to side reactions at the anode (-) and cathode (+), electrolyte solvent consumption, loss of circulating lithium due to lithium plating, loss of active material at the anode, and loss of active material at the cathode. The second value includes a weighting factor used to calculate the battery's capacity. In one embodiment, this second value includes weighting factors corresponding to the side reaction current at the anode and the side reaction current at the cathode of the battery.In one embodiment, this second value also includes weighting factors corresponding to the loss of active material at the anode and the loss of active material at the cathode of the battery.

[0058] In one embodiment, updating the parameters of the hybrid neural network model stored in the shared memory block based on the predicted target value and the target actual value of the cycle data group includes: updating parameters of the neural network submodule and the weighting factors of the physical model submodule corresponding to the loss of active material at the anode and the loss of active material at the cathode of the battery, based on the predicted target value and the target actual value of the cycle data group.

[0059] In one embodiment, the method further comprises: defining several different subsampling rates for each of the several different subsampling rates; training the hybrid neural network model using a training data subset obtained by the subsampling rate based on a second training dataset, wherein the second training dataset comprises several cycle data groups, each of which comprises several input signals from multiple time points and a target-output actual value, each of which comprises several physical values ​​from one time point; determining a training performance score of the trained hybrid neural network model based on the training data subset;Determination of a test performance score of the trained hybrid neural network model based on a test dataset, wherein the test dataset comprises multiple cycle data groups, each of which comprises multiple input signals from multiple time points and a target output actual value, each of which comprises multiple physical values ​​from one time point.

[0060] In one embodiment, the method further comprises: selecting one or more subsamples from the several different subsamples based on the several groups of training performance values ​​and test performance values, each corresponding to the several different subsamples; training the hybrid neural network model based on the first training dataset using the one or more selected subsamples.

[0061] In one embodiment, the selection of one or more subsampling rates from the several different subsampling rates comprises: selecting a first subsampling rate and a second subsampling rate.Training the hybrid neural network model based on the first training dataset includes: adapting the structure of the hybrid neural network model to obtain multiple candidate models of the hybrid neural network with different structures; training each of the multiple candidate models of the hybrid neural network with different structures based on the first training dataset using the first subsampling rate; selecting one candidate model of the hybrid neural network from the multiple candidate models with different structures based on the respective performance metrics of the trained candidate models; and training the selected candidate model of the hybrid neural network based on the first training dataset using the second subsampling rate.

[0062] In one embodiment, the second training dataset is a subset of the first training dataset. In another embodiment, the second training dataset is identical to the first training dataset. In yet another embodiment, the second training dataset differs from the first training dataset; for example, the second training subset intersects with the first training subset.

[0063] Fig. Figure 8 shows a block diagram of a device for a hybrid neural network model according to an embodiment.

[0064] According to one embodiment, the control system or processing system 800 can comprise one or more control units or processing units 810, wherein the control units 810 execute one or more machine-readable instructions stored in a machine-readable storage medium (i.e., a computer).

[0065] are stored or encoded in memory 820. Although in Fig. Although not shown in Figure 8, it is understandable to experts in the field that the control system 800 can contain a variety of other components, such as various communication modules, bus modules, and possibly user interface modules. In one embodiment, the processing units 810 are configured in the processing system 800, and during the execution of program instructions, the various components associated with Fig. to carry out the processes and functions described in points 1 to 7.

[0066] According to one embodiment, a program product, such as a non-volatile machine-readable medium, is provided. The non-volatile machine-readable medium can contain instructions, wherein the instructions, when executed by the processing units 810, are carried out by devices such as the processing units 810, the processing system 800, etc., the various operations and functions associated with the Fig. 1 to 7 have been described in various embodiments of the present application.

[0067] The specific embodiments described above in conjunction with the figures represent exemplary embodiments and do not constitute all embodiments that can be realized or that fall within the scope of the claims. The term "exemplary," used throughout this description, means "serving as an example or for illustration" and does not mean "preferred" or "advantageous" over other embodiments. To facilitate an understanding of the described technologies, the specific embodiments include specific details. However, these technologies can be implemented without these specific details. In some cases, known structures and devices are shown in the form of block diagrams to avoid making the concept of the described embodiments difficult to understand.

[0068] The preceding description of the present disclosure enables ordinary people skilled in the art to implement or apply the present disclosure. To ordinary people skilled in the art, the various modifications of the present disclosure are obvious, and the general principles defined herein can also be applied to other variants without altering the scope of protection of the present disclosure. Therefore, the present disclosure is not limited to the examples and embodiments described herein, but is consistent with the broadest scope of the principles and novel features disclosed herein.

Claims

[1] Method for predicting the health status of a battery using a hybrid neural network model, comprising: Temporally sequential input of multiple input signals from multiple time points, contained in a measured cycle data set, into the hybrid neural network model, wherein the hybrid neural network model comprises a neural network submodule and a physical model submodule, and each input signal of the multiple input signals of this cycle data set comprises multiple physical values ​​of the battery from one time point; In one iteration for an input signal from each of the multiple time points: prediction of a second value by the Neural Network submodule based on the input signal and a first value output in the previous iteration by the Physical Model submodule, calculation of a first value in the current iteration by the Physical Model submodule based on the input signal and the second value, repetition of the steps until the last iteration for the last input signal at the last time point, prediction of a second value in the last iteration by the Neural Network submodule based on the last input signal and a first value that was output in the respective previous iteration by the Physical Model submodule, calculation of a first value in the current iteration by the Physical Model submodule based on the last input signal and the second value; and Determining the battery's health status based on the first value. [2] Method according to claim 1, wherein the physical model submodule comprises a physical formula for calculating the health status of the battery based on the measured values ​​of the battery. [3] Method according to claim 2, wherein the physical formula comprises an ordinary differential equation or partial differential equation for representing the health status of the battery based on the measured values ​​of the battery. [4] Method according to claim 2, wherein the neural network submodule comprises a fully connected neural network model. [5] Method according to claim 2, wherein the multiple physical values ​​contained in each input signal from one time point at least partially comprise a current value, a voltage value, a temperature value and a state of charge of the battery from that time point, the first value comprising at least partially a change in the solid electrolyte boundary layer of the battery, a change in the circulating lithium due to side reactions at the anode and cathode, a consumption of electrolyte solvent, a loss of circulating lithium due to lithium plating, a loss of active material at the anode and a loss of active material at the cathode, and the second value comprising a weighting factor used to calculate the capacity of the battery. [6] Method according to claim 5, wherein the second value comprises weighting factors corresponding to the side reaction current at the anode and the side reaction current at the cathode of the battery. [7] Method according to claim 6, wherein the second value also includes weighting factors corresponding to the loss of active material at the anode and the loss of active material at the cathode of the battery. [8] Methods for training a hybrid neural network model, comprising: Storing parameters of the hybrid neural network model in a shared memory block, wherein the hybrid neural network model comprises a neural network submodule and a physical model submodule, wherein an initial training dataset for training the hybrid neural network model comprises multiple cycle data groups, each of which comprises multiple input signals from multiple time points and a target output actual value, wherein each of the input signals comprises multiple physical values ​​from one time point, wherein in an iteration for an input signal from one time point, the neural network submodule predicts a second value based on the input signal and a first value output by the physical model submodule in the previous iteration.and the physical model submodule calculates a first value in the current iteration based on the input signal and the second value; and, Parallel execution of multiple processes, wherein the multiple processes in parallel, using the parameters of the hybrid neural network model stored in the shared memory block, train the hybrid neural network model based on the first training data set, wherein for each process of the multiple processes and each cycle data group used by the process, the process iteratively generates a predicted target value based on the temporally ordered multiple input signals from multiple time points in the cycle data group, using the parameters of the hybrid neural network model stored in the shared memory block, and updates the parameters of the hybrid neural network model stored in the shared memory block based on the predicted target value and the target actual value of the cycle data group. [9] Method according to claim 8, wherein the multiple processes perform the training of the hybrid neural network model in parallel based on the multiple cycle data groups according to different sequences of the multiple cycle data groups in the first training data set. [10] Method according to claim 9, wherein the physical model submodule comprises a physical formula for calculating the health status of the battery based on the measured values ​​of the battery, wherein the neural network submodule comprises a fully connected neural network model. [11] Method according to claim 10, wherein the multiple physical values ​​contained in each input signal from one time point at least partially comprise a current value, a voltage value, a temperature value and a state of charge of the battery from that time point, the first value comprising at least partially a change in the solid electrolyte boundary layer of the battery, a change in the circulating lithium due to side reactions at the anode and cathode, a consumption of electrolyte solvent, a loss of circulating lithium due to lithium plating, a loss of active material at the anode and a loss of active material at the cathode, and the second value comprising a weighting factor used to calculate the capacity of the battery. [12] Method according to claim 11, wherein the second value comprises weighting factors corresponding to the side reaction current at the anode and the side reaction current at the cathode of the battery. [13] Method according to claim 12, wherein the second value also includes weighting factors corresponding to the loss of active material at the anode and the loss of active material at the cathode of the battery. [14] Method according to claim 12, wherein the updating of the parameters of the hybrid neural network model stored in the shared memory block based on the predicted target value and the target actual value of the cycle data group comprises: updating parameters of the neural network submodule and the weighting factors of the physical model submodule corresponding to the loss of active material at the anode and the loss of active material at the cathode of the battery based on the predicted target value and the target actual value of the cycle data group. [15] Method according to any one of claims 8 to 14, further comprising: Setting several different subsampling rates for each of the several different subsampling rates: Training of the hybrid neural network model using a training data subset obtained by subsampling based on a second training dataset, wherein the second training dataset comprises multiple cycle data groups, each of which comprises multiple input signals from multiple time points and a target-output actual value, with each input signal comprising multiple physical values ​​at a time point; Determination of a training performance score of the trained hybrid neural network model based on the training data subset; Determination of a test performance score of the trained hybrid neural network model based on a test dataset, wherein the test dataset comprises multiple cycle data groups, each cycle data group comprising multiple input signals from multiple time points and a target output actual value, each input signal comprising multiple physical values ​​from one time point. [16] The method of claim 15, further comprising: Selection of one or more subsamples from the multiple different subsamples based on the multiple groups of training performance values ​​and test performance values, each corresponding to the multiple different subsamples; Training of the hybrid neural network model based on the first training dataset using one or more selected subsampling rates. [17] Method according to claim 16, wherein, The selection of one or more subsampling rates from the several different subsampling rates includes the following: selection of a first subsampling rate and a second subsampling rate; The training of the hybrid neural network model based on the first training dataset includes the following: Adapting the structure of the hybrid neural network model to obtain multiple candidate models of the hybrid neural network with different structures; Training of the multiple candidate models of the hybrid neural network with different structures, each based on the first training dataset using the first subsampling rate; Selection of a candidate model of the hybrid neural network from the multiple candidate models of the hybrid neural network with different structures based on the respective performance metrics of the trained candidate models of the hybrid neural network; Training of the selected candidate model of the hybrid neural network based on the first training dataset using the second subsampling rate. [18] Method according to claim 16, wherein the second training data set is a subset of the first training data set, or the second training data set is identical to the first training data set, or the second training data set differs from the first training data set. [19] Device for a hybrid neural network model, comprising: A storage facility; and one or more processing units that are configured to perform tasks such as... Program instructions to execute a method according to any one of claims 1 to 18. [20] Machine-readable storage medium on which executable instructions are stored, wherein, in the execution of the instructions by one or more processors, a method according to one of claims 1 to 18 is implemented.

Images