Method and apparatus for predicting state of health of battery by using hybrid neural network model, and model training method and apparatus

Abstract

A method for predicting the state of health of a battery by using a hybrid neural network model. The method comprises: sequentially inputting a plurality of input signals (210) at a plurality of time points comprised in a measured period data group into a hybrid neural network model (100) in terms of time, wherein the hybrid neural network model comprises a neural network sub-module (110) and a physical model sub-module (120), and each of the plurality of input signals (210) in the period data group comprises a plurality of physical quantities of a battery at one time point; during the iteration for one input signal (210) at each of the plurality of time points, the neural network sub-module (110) predicting a second value on the basis of the input signal and a first value which is output by the physical model sub-module (120) during the last iteration, and the physical model sub-module (120) calculating a first value during the current iteration on the basis of the input signal (210) and the second value until the iteration for the last input signal at the last time point; and determining the state of health of the battery on the basis of the first value.

Description

Method and device for predicting battery health status and model training using hybrid neural network model Technical Field

[0001] The present application relates to artificial intelligence technology, and more specifically, to a method and device for predicting the health status of a battery using a hybrid neural network model and a method and device for training the hybrid neural network model. Background Art

[0002] Rechargeable lithium batteries are widely used in various electric devices, such as electric vehicles. Lithium batteries generally include a negative electrode, a positive electrode, and a separator located between the negative and positive electrodes. Both electrodes contain an active material that reversibly reacts 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 so that no electrodes are electrically connected within the battery. Typically, during charging, electrons are generated at the positive electrode and an equal number of electrons are consumed at the negative electrode. During discharge, the opposite reaction occurs. Undesirable side reactions may occur during repeated charge and discharge cycles of the battery, and these undesirable side reactions can lead to a decrease in battery capacity.

[0003] Battery management systems in electric vehicles can be used to assess the battery's state of health (SOH). Existing physics-based electrochemical aging models struggle to account for complex and realistic operating conditions. Therefore, more accurate and efficient battery SOH assessments are desirable.

[0004] Summary of the Invention

[0005] The following summary is provided to introduce selected concepts in a simplified form that are further described in the detailed description below. This summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to limit the scope of the claimed subject matter.

[0006] According to one aspect of the present application, a method for predicting the health state of a battery using a hybrid neural network model is provided, comprising: temporally sequentially inputting multiple input signals at multiple time points included in a measured periodic data group into the hybrid neural network model, wherein the hybrid neural network model includes a neural network submodule and a physical model submodule, and each of the multiple input signals of the periodic data group includes multiple physical quantities of the battery at a time point; in an iteration of an input signal at each time point in the multiple time points, the neural network submodule predicts a second value based on the input signal and a first value output by the physical model submodule in a previous iteration, and the physical model submodule calculates the first value in this iteration based on the input signal and the second value, until in an iteration of a last input signal at the last time point, the neural network submodule predicts a second value based on the last input signal and the first value output by the physical model submodule in the previous iteration, and the physical model submodule calculates the first value in this iteration based on the last input signal and the second value; and determining the health state of the battery based on the first value.

[0007] According to one aspect of the present application, a training method for 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 data set for training the hybrid neural network model comprises a plurality of periodic data groups, each periodic data group comprises a plurality of input signals at a plurality of time points and a target output true value, each input signal comprises a plurality of physical quantities at a time point, wherein, in an iteration for an input signal at a 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 a previous iteration, and the physical model submodule Calculate the first value in this iteration based on the input signal and the second value; run multiple processes in parallel, and the multiple processes use the parameters of the hybrid neural network model stored in the shared memory block in parallel to train the hybrid neural network model based on the first training data set, wherein, for each of the multiple processes and each periodic data group used by the process, the process uses the parameters of the hybrid neural network model stored in the shared memory block to iteratively generate a predicted target value based on multiple input signals with a plurality of time points in a time sequence in the periodic data group, and update the parameters of the hybrid neural network model stored in the shared memory block based on the predicted target value and the target true value of the periodic data group.

[0008] According to one aspect of the present application, a device for a hybrid neural network model is provided, comprising: a memory; and one or more processing units, wherein the processing units are configured to execute the method for predicting the health status of a battery using a hybrid neural network model and the training method of the hybrid neural network model described in this document when executing program instructions.

[0009] According to one aspect of the present application, a machine-readable storage medium is provided, which stores executable instructions. When the instructions are executed, one or more processors execute the method for predicting the health status of a battery using a hybrid neural network model and the training method of the hybrid neural network model described in this article.

[0010] By using the method for predicting the health status of a battery using a hybrid neural network model according to the present application, the accuracy of predicting the health status of the battery can be effectively improved by using the neural network submodule to learn the dynamic weighting coefficients in the physical model submodule that depend on the working environment. By using the method for training a hybrid neural network model for predicting the health status of a battery according to the present application, with the help of characteristics related to the measurement data of the battery, the training time can be effectively reduced, the training efficiency can be improved, and the problem of long training time caused by the time serial processing method of the hybrid neural network model and the long time required to process the physical formulas at each time step can be alleviated, and the problem of large memory requirements can be correspondingly improved. Other advantages of the present disclosure will be explained in the detailed description below. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] A further understanding of the nature and advantages of the present disclosure may be achieved by reference to the following drawings, in which similar components or features may have the same reference numerals.

[0012] FIG1 shows a schematic diagram of a hybrid neural network model according to one embodiment.

[0013] FIG2 shows a simplified schematic computational diagram for training a hybrid neural network according to one embodiment.

[0014] FIG3 shows a schematic diagram of a method for training a hybrid neural network model according to one embodiment.

[0015] FIG4 shows a schematic diagram of a method for training a hybrid neural network model according to one embodiment.

[0016] FIG5 shows a schematic diagram of a method for training a hybrid neural network model according to one embodiment.

[0017] FIG6 shows a flowchart of a method for predicting the health status of a battery using a hybrid neural network model according to one embodiment.

[0018] FIG7 shows a flowchart of a method for training a hybrid neural network model according to one embodiment.

[0019] FIG8 shows a block diagram of an apparatus for a hybrid neural network model according to one embodiment. DETAILED DESCRIPTION

[0020] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that discussing these embodiments is intended only to enable those skilled in the art to better understand and implement the subject matter described herein, and is not intended to limit the scope of protection, applicability, or examples set forth in the claims. The functions and arrangements of the elements discussed may be changed without departing from the scope of protection of this disclosure. Various examples may omit, replace, or add various processes or components as needed. For example, the described method may be performed in an order different from the described order, and various steps may be added, omitted, or combined. In addition, features described relative to some examples may also be combined in other examples.

[0021] As used herein, the term "including" and its variations are open-ended terms meaning "including but not limited to." The term "based on" means "based at least in part on." The terms "one embodiment" and "an embodiment" mean "at least one embodiment." The term "another embodiment" means "at least one other embodiment." The terms "first," "second," etc. may refer to different or the same objects. Unless the context clearly indicates otherwise, the definition of a term is consistent throughout the specification.

[0022] FIG1 shows a schematic diagram of a hybrid neural network model according to one embodiment.

[0023] The hybrid neural network model 100 shown in FIG1 includes a neural network submodule 110 and a physical model submodule 120. The hybrid neural network model 100 receives a battery measurement signal 130, such as the battery's voltage, current, temperature, state of charge (SOC) and other physical quantities, and determines the battery's state of health (SOH) based on the received battery measurement signal 130. The charge state is, for example, a charge state of 30%, 60%, and so on. For example, the health state can be the capacity of the battery or the remaining capacity. Assuming that the original capacity of the battery is C0, the capacity of the battery will decrease as the battery is used and ages, then the current capacity of the battery can be used to represent the battery's health state. It can also be assumed that the original capacity of the battery is 1, then the remaining capacity of the battery can be represented by a fraction less than 1.

[0024] [Corrected 25.11.2024 according to Rule 26] The main aging phenomena of batteries include: Solid Electrolyte Interface (SEI) growth, which leads to lithium loss and solvent consumption; lithium deposition, which leads to lithium loss; and loss of active materials at the cathode and anode. In one embodiment, the aging model can be expressed by a set of physical formulas:

[0025] [Corrected 25.11.2024 according to Rule 26] Formula (1) represents the change in SEI film thickness due to the anode side reaction. Formula (2) represents the change in recyclable lithium due to the anode (-) and cathode (+) side reactions. Formula (3) represents the consumption of electrolyte solvent. Formula (4) represents the loss of recyclable lithium due to lithium deposition. Formulas (5) and (6) represent the loss of active materials at the anode and cathode, respectively. The side reactions in formulas (1) to (6) can be represented by the following physical formulas:

[0026] In formulas (7) and (10), and are the anode side reaction current and cathode side reaction current associated with the first or second electrolyte, respectively. J in formula (11) pl is the lithium deposition reaction current.

[0027] [Corrected 25.11.2024 according to Regulation 26] The state of health (SOH) of a battery can be defined as the remaining capacity of the battery. By solving the ordinary differential equations (1) to (6), the physical quantity or system state x that affects the battery capacity can be obtained. cLi 、x Li 、x AML -、x AML +. The remaining capacity of the battery SOH C It can be expressed as SOH C =C0-(x cLi +x Li +k a x AML- +k c x AML+ ) (12)

[0028] C0 in formula (12) represents the original capacity of the battery, k a and k c It is a coefficient or called a weighting factor.

[0029] The electrochemical aging model of the battery represented by formulas (1) to (12) can be used to calculate the remaining charge SOH of the battery based on the measured physical quantities of the battery, such as the current I(t), voltage V(t), temperature T(t) and state of charge SOC(t) in formulas (7) to (11). C In this article, the aging model represented by formulas (1) to (12) is used as an example to illustrate the technical solution of the present disclosure. Those skilled in the art will understand that the above-mentioned implementation method of modeling the electrochemical aging model of the battery using formulas (1) to (12) is only an example for illustrating the technical solution of the present disclosure. The physics-based electrochemical aging model is not limited to a specific implementation method, and any suitable electrochemical aging model that conforms to the laws of physics can be used.

[0030] [Corrected 25.11.2024 according to Rule 26] The physical model submodule 120 shown in FIG1 includes a physical aging model 1210. In one example, the physical model 1210 may include ordinary differential equations (ODEs) represented by the above formulas (1) to (6) and corresponding physical formulas (7) to (11). Specifically, for a measurement signal 130 at a time point t, where the measurement signal 130 includes the current I(t), voltage V(t), temperature T(t) and state of charge SOC(t) of the battery, the physical aging model 1210 may obtain numerical representations of the ordinary differential equations (1) to (6) based on the measurement signal 130 through formulas (1) to (11). The physical model submodule 120 also includes an ODE solver 1220, which solves the ordinary differential equations (1) to (6) through a forward process to obtain the system state x of the battery. SEI 、x cLi 、x Csol1,2 、x Li 、x AML -、x AML +, which respectively represent the change in SEI film thickness caused by the anode side reaction, the change in cyclable lithium due to the anode (-) and cathode (+) side reactions, the consumption of electrolyte solvent, the loss of cyclable lithium due to lithium deposition, the loss of active material at the anode, and the loss of active material at the cathode. It can be understood that the ODE solver 1220 can be any suitable ODE solver, such as Runge Kutt solver, Euler, etc. The physical model submodule 120 also includes a SOH determination unit 1230, which determines the remaining capacity SOH of the battery according to formula (12) C As the health status of the battery. In another embodiment, the original capacity of the battery is represented as 1, and the remaining capacity SOH is represented as CIt can be expressed as a fraction less than 1. For example, the SOH determination unit 1230 can determine the remaining capacity SOH of the battery according to formula (13): C As the battery's state of health SOH C =1-(x CLi +x Li +k a x AML- +k c x AML+ ) / (C0*3600) (13)

[0031] C0 in formula (13) represents the original capacity of the battery, k a and k c is the coefficient.

[0032] In the embodiment using the physical aging model, the loss of active materials x of the anode and cathode is calculated. AML - and x AML In formulas (5) and (6) of +, it is necessary to set the anode side reaction current and cathode side reaction current related to the first or second electrolyte and The corresponding coefficient k I -|I(t)| and k I+ |I(t)|. The input signals I(t), V(t), T(t) and SOC(t) affect the loss of active materials x at the anode and cathode calculated in equations (5) and (6). AML - and x AML +, so it is preferred to adopt an appropriate coefficient k for different input signals I -|I(t)| and k I+ |I(t)|, however, since it is difficult to express it using an exact physical formula, the coefficient value k is based on experience I -|I(t)| and k I+ |I(t)| may lead to a decrease in the prediction accuracy of the physical aging model.

[0033] In the embodiment shown in FIG1 , the neural network submodule 110 predicts the anode side reaction current and the cathode side reaction current based on the input signals I(t), V(t), T(t) and SOC(t) at a time point t. and The corresponding coefficient k I -|I(t)| and k I+ |I(t)| 140. In one embodiment, the neural network sub-model 110 may be a fully connected neural network. Accordingly, the physical model sub-module 120 is based on the input signals I(t), V(t), T(t) and SOC(t) and the coefficient k output by the neural network sub-module 110.I -|I(t)| and k I+ |I(t)| 140 outputs a physical parameter 150 , which includes the system state x of the battery. SEI 、x cLi 、x Csol1，2 、x Li 、x AML -、x AML + and the remaining capacity of the battery SOH C , where the remaining capacity of the battery SOH can be C As the health status of the battery.

[0034] In the embodiment shown in FIG1 , at a time point t, the neural network sub-model 110 takes as input the input signals I(t), V(t), T(t), and SOC(t) at the time point t and at least a portion of the output signal of the physical model sub-module 120 at the previous time point t-1, and obtains the coefficient k at the time point t. I -|I(t)| and k I+ |I(t)| 140. For example, at a time point t, the neural network sub-model 110 combines the input signals I(t), V(t), T(t) and SOC(t) at that time point t with the system state x output by the physical model sub-module 120 at the previous time point t-1. SEI 、x cLi 、x Csol1，2 、 xLi 、x AML -、x AML + as input, get the coefficient k at the time point t I -|I(t)| and k I+ |I(t)|140.

[0035] In another embodiment, the neural network submodule 110 predicts the coefficient 140 based on the input signals I(t), V(t), T(t) and SOC(t) at a time point t, wherein the coefficient 140 includes the coefficients related to the anode side reaction current and the cathode side reaction current. and The corresponding coefficient k I -|I(t)| and k I+ |I(t)| and the loss of active materials at the anode and cathode x AML - and x AML + corresponding coefficient k a and k c Accordingly, the physical model submodule 120 is based on the input signals I(t), V(t), T(t) and SOC(t) and the coefficient k output by the neural network submodule 110. I -|I(t)|, k I+|I(t)|, k a and k c 140 outputs physical parameters 150, which include the system status x of the battery SEI 、x cLi 、x Csol1，2 、x Li 、x AML -、x AML + and the remaining capacity of the battery SOH C , where the remaining capacity of the battery SOH C As the health status of the battery.

[0036] In the embodiment shown in FIG1 , at a time point t, the neural network sub-model 110 takes as input the input signals I(t), V(t), T(t), and SOC(t) at the time point t and at least a portion of the output signal of the physical model sub-module 120 at the previous time point t-1, and obtains the coefficient k at the time point t. I -|I(t)|, k I+ |I(t)|, k a and k c 140. For example, at a time point t, the neural network sub-model 110 combines the input signals I(t), V(t), T(t) and SOC(t) at the time point t with the system state x output by the physical model sub-module 120 at the previous time point t-1. SEI 、x cLi 、x Csol1，2 、x Li 、x AML -、x AML + as input, get the coefficient k at the time point t I -|I(t)|, k I+ |I(t)|, k a and k c 140.

[0037] The above describes the implementation of various aspects of the present disclosure for predicting the health status of a battery in combination with formulas (1) to (13) and the embodiment shown in Figure 1. However, it should be understood that the present disclosure is not limited to the specific examples described above. For example, 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 coefficient 140 based only on the input signals I(t), V(t), T(t) and SOC(t) at time point t. For another example, the physical model submodule 120 may adopt any appropriate physical aging model known in the art or known in the future, and is not limited to the physical aging model represented by the ODE shown in formulas (1) to (13). For example, the physical model submodule 120 may adopt an aging model represented by an appropriate partial differential equation (PDE).

[0038] In one embodiment, to predict the health status of a battery using the hybrid neural network model 100, data is collected for a cycle of multiple input signals at multiple time points. For example, the data for this cycle includes thousands of measurement signals at thousands of time points, each measurement signal including, for example, the battery's current I(t), voltage V(t), temperature T(t), and state of charge (SOC(t). This cycle of data can be referred to as a measured cycle data set. As shown in FIG1 , the multiple input signals at multiple time points in a cycle data set are sequentially input into the hybrid neural network model 100 in time. In an iteration for an input signal at each of the multiple time points, the neural network submodule 110 predicts a second value based on the input signal and the first value output by the physical model submodule 120 in the previous iteration. The physical model submodule 120 calculates the first value in the current iteration based on the input signal and the second value. This continues until, in an iteration for the last input signal at the last time point, the neural network submodule 110 predicts a second value based on the last input signal and the first value output by the physical model submodule 120 in the previous iteration. The physical model submodule 120 calculates the first value in the current iteration based on the last input signal and the second value. Finally, the battery health status is determined based on the first value. In one embodiment, the first value may include the system status x. SEI 、x cLi 、x Csol1，2 、x Li 、x AML -、x AML In one embodiment, the second value may include a coefficient k I -|I(t)| and k I+ |I(t)|. In one embodiment, the second value may include the coefficient k I-|I(t)|, k I+ |I(t)|, ka and kc 140. It can be understood that when different physical formulas are used to implement the physical aging model 120, the first value and the second value can be corresponding physical quantities and coefficients, and are not limited to the specific physical quantities and coefficients described in the above embodiments.

[0039] As mentioned above, for example, when the hybrid neural network model 100 shown in Figure 1 processes the measurement data of a cycle, it is necessary to process multiple measurement data of the cycle sequentially or serially in time. Such serial processing method will bring difficulties to the training process of the hybrid neural network model 100, one of which is the problem of long training time.

[0040] In one example, a training dataset includes multiple batteries, each battery comprising multiple measurement cycles, each cycle comprising multiple timestamps and measurement signals for each timestamp. For example, the dataset includes 33 batteries, each battery comprising 4 to 8 measurement cycles, each cycle comprising thousands of timestamps and input signals such as current, voltage, temperature, and state of charge for each timestamp. For each cycle, there is a target true value (state of health), and the number of target true values ​​for each battery is equal to the number of cycles. During training, a hybrid neural network model 100 is used to predict the battery state of health based on the data in the training dataset. The trainable parameters of the hybrid neural network model 100, such as the trainable parameters of the neural network submodule 110 and the trainable data in the physical model submodule 120, are updated based on the predicted and true values ​​of the battery state. However, due to the need to process multiple measurement data for a cycle serially in time and the nature of the physical model 120 being an ordinary differential equation (ODE), the main challenge faced by the joint learning of ODEs and neural networks is the long training time. Furthermore, due to the serial processing in time, the parameters need to be stored at each time step, which requires a large memory capacity.

[0041] FIG2 shows a simplified schematic computational graph for training the hybrid neural network 100 according to one embodiment. The computational graph shown in FIG2 represents operations for a training data set for one cycle, where each circled node in the first row represents an input signal 210 at a time point, each node in the second row represents a computation 220 of the neural network submodule 110, and each node in the third row represents a computation 230 of the physical model submodule 120.

[0042] In the process of training the hybrid neural network 100 using the training data set, an input signal 210 at a time point is input to the neural network submodule 110, and the input signal 210 and the output data of the neural network submodule 110 are input to the physical model submodule 120. The physical aging model 1210 and the ODE solver 1220 of the physical model submodule 120 calculate the hidden state of the next time step, such as the system state x SEI 、x cLi 、x Csol1，2 、x Li 、x AML -、x AML Then, the same operation is iteratively performed for each subsequent time point, or time step, until the last time point of the cycle, at which point the SOH determination unit 1230 of the physical model submodule 120 calculates the remaining capacity of the battery as the predicted target value 240. A loss value can be obtained based on the predicted target value 240 and the target true value of the cycle. For example, the loss value is the difference between the target value and the target true value.

[0043] Next, based on the loss value, backpropagation is performed from node 240 to the physical model submodule 120 and the neural network submodule 110 of the previous time step, and then to the next previous time step, until the first time step of the cycle, thereby updating the trainable parameters of the physical model submodule 120 and the neural network submodule 110. In one embodiment, the loss x of the active material at the anode and cathode is AML - and x AML + corresponding coefficient k a and k c Configured as a trainable parameter to obtain the optimal coefficient k during the training process of the hybrid neural network model a and k c , thereby improving the prediction accuracy of the hybrid neural network model 100.

[0044] Because the computation of the ODE is complex, the backward propagation of the ODE is even more complex, so the backward propagation of each time step takes a relatively long time. For the entire dataset in the example, each cycle has thousands of time steps, each battery has 4 to 8 cycles, and there are 33 batteries in total. The number of time steps in the entire dataset is very large, and each time step takes a long time. Therefore, in the serial processing training process shown in Figure 2, the training time will be very long. For example, for the dataset in the example, the forward processing time of one training epoch is approximately 15 minutes, and the backward processing time is over 2 hours. Here, the term "epoch" refers to running all the data in the training dataset once. In actual implementation, the training process of the hybrid neural network model 100 requires many epochs, and the training dataset may be much larger than the training dataset in the example. It would be beneficial to shorten the training time of the hybrid neural network model 100 and thus improve its training efficiency.

[0045] FIG3 shows a schematic diagram of a method for training a hybrid neural network model according to one embodiment. The hybrid neural network model 300 shown in FIG3 may be the hybrid neural network model 100 shown in FIG1 . As shown in FIG3 , the hybrid neural network model 300, for example, the parameters of the hybrid neural network model 300, are stored in a shared memory block. A plurality of processes, for example, processes P1 to P4 shown in FIG3 , are run in parallel. The four processes P1 to P4 shown in FIG3 share the same hybrid neural network model 300, and the four processes P1 to P4 are run in parallel using the same training data set 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 using the same data set in parallel.

[0046] In one embodiment, multiple processes P1 to P4 train the hybrid neural network model 300 in parallel according to different orders of multiple periodic data groups in the training dataset D. For example, the training dataset D is divided into four subsets D1 to D4, and processes P1 to P4 respectively train the hybrid neural network model 300 according to different orders of D1 to D4. For example, process P1 trains the hybrid neural network model 300 with set D in the order of {D1, D2, D3, D4}, process P2 trains the hybrid neural network model 300 with set D in the order of {D2, D3, D4, D1}, process P3 trains the hybrid neural network model 300 with set D in the order of {D3, D4, D1, D2}, and process P4 trains the hybrid neural network model 300 with set D in the order of {D4, D1, D2, D3}. It is understood that processes P1 to P4 may also train the model 300 according to other orders of subsets D1 to D4. For another example, the training data in the training data set D is sorted in units of cycle data groups or batteries, and processes P1 to P4 respectively train the hybrid neural network model 300 according to the differently sorted data sets D.

[0047] As shown in FIG3 , by adopting four parallel processes P1 to P4, each process independently executes the serial training process as described above with reference to FIG2 for the same hybrid neural network model 300 using the same data set D. When the four processes P1 to P4 run one epoch in parallel, it is equivalent to training the hybrid neural network model 300 with the same data set D for four epochs, thereby improving the efficiency of the hybrid neural network model 300.

[0048] FIG4 shows a more detailed schematic diagram of the method shown in FIG3 . As shown in FIG4 , four processes P1 to P4 independently perform training of the hybrid neural network model 300 according to their respective periodic data groups P1-C1, P1-C2, etc., P4-C1, P4-C2, etc. In one embodiment, as shown in FIG4 , for example, after process P2 updates the parameters of the hybrid neural network model 300 using the current periodic data group P2-C1, all processes P1 to P4 perform subsequent processing according to the updated hybrid neural network model 300. Similarly, after process P1 updates the parameters of the hybrid neural network model 300 using the current periodic data group P1-C1, all processes P1 to P4 perform subsequent processing according to the updated hybrid neural network model 300. It will be understood that the length of each cycle of the data set D is not necessarily the same.

[0049] FIG5 shows a schematic diagram of a method for training a hybrid neural network model according to one embodiment.

[0050] Still taking the above training data set as an example, since the measurement data collected at multiple time points during the battery's charge and discharge process often conforms to linear changes, it provides the possibility of using a data subset of the training data set to train the hybrid neural network model using a downsampling method. In one embodiment, multiple different downsampling rates 510 can be set, such as downsampling rates 1 to 8 shown in Figure 5. Downsampling rate 1 represents all data samples in the sampled data set, downsampling rate 2 represents half of the data samples in the sampled data set, for example, one sample is taken for every two data samples, and so on. Downsampling rate 8 represents one-eighth of the data samples in the sampled data set, for example, one sample is taken for every eight data samples.

[0051] Using a training dataset D', for each downsampling rate 1 to 8, a data subset of the training dataset D' obtained based on the downsampling rate is used to train the hybrid neural network model 100, and the performance of the trained hybrid neural network model 100 is evaluated on the data subset and on the test set, respectively. The performance can be expressed as a mean square error (MSE), as shown in 520 and 530 of FIG5, which represent the MSE 520 on the training subset and the MSE 530 on the test set, respectively. At the same time, the training process for each downsampling rate can be timed, for example, the time taken for a training epoch for the data subset corresponding to the downsampling rate is timed, as shown in 540 and 550 of FIG5, which represent the forward time 540 and backward time 550 of an epoch, respectively. Based on the performance indicators generated for each downsampling rate, such as the test set MSE 530, one or more downsampling rates are selected for training the hybrid neural network 100.

[0052] In one embodiment, the training dataset D' used to select the downsampling rate described in conjunction with FIG5 is a subset of the actual training dataset D used in FIG2 to 4. In this way, a downsampling rate that can be used in the actual training process can be selected with less computational effort. In addition, in other embodiments, the training dataset D' can be the same as the actual training dataset D, or the training dataset D' can intersect with the actual training dataset D but be different.

[0053] In one embodiment, for example, as shown in FIG5 , downsampling rates of 2 and 8 can be used for training the hybrid neural network model 100, wherein the downsampling rates of 2 and 8 can be used in different training stages of the hybrid neural network model 100. In one embodiment, during the structural adjustment stage of the hybrid neural network model 100, a larger downsampling rate of 8 can be used to train the model. In this embodiment, the structure of the hybrid neural network model is adjusted to obtain a plurality of candidate hybrid neural network models with different structures, and the plurality of candidate hybrid neural network models with different structures are trained based on the training data set D using the downsampling rate of 8. Based on the performance indicators of the trained plurality of candidate hybrid neural network models with different structures, a candidate hybrid neural network model is selected from the plurality of candidate hybrid neural network models with different structures. After a hybrid neural network model 100 with an optimal structure is selected, the candidate hybrid neural network model is trained based on the training data set D using the downsampling rate of 2.

[0054] FIG6 shows a flowchart of a method for predicting the health status of a battery using a hybrid neural network model according to one embodiment.

[0055] In step 610, multiple input signals at multiple time points included in the measured periodic data group are temporally sequentially input into a hybrid neural network model, wherein the hybrid neural network model includes a neural network submodule and a physical model submodule, and each of the multiple input signals of the periodic data group includes multiple physical quantities of the battery at a time point.

[0056] In step 620, in an iteration of an input signal for each of the multiple time points, the neural network submodule predicts a second value based on the input signal and the first value output by the physical model submodule in the previous iteration, and the physical model submodule calculates the first value in this iteration based on the input signal and the second value, until in an iteration of the last input signal for the last time point, the neural network submodule predicts a second value based on the last input signal and the first value output by the physical model submodule in the previous iteration, and the physical model submodule calculates the first value in this iteration based on the last input signal and the second value.

[0057] At step 630, the state of health of the battery is determined based on the first value. For example, the state of health of the battery can be determined by the remaining capacity SOH of the battery as described in detail above. C For example, the second value generated in step 620 may include the system state x described in detail above. cLi 、x Li 、x AML -、x AML+, in step 630, the remaining capacity SOH of the battery is calculated based on formula (12) C As the health status of the battery. For another example, the second value generated in step 620 may include the system status x described in detail above. cLi 、x Li 、x AML -、x AML + and the remaining capacity of the battery SOH C In step 630, the remaining battery capacity SOH in the second value is taken. C As the health status of the battery.

[0058] In one embodiment, the physical model submodule includes a physical formula for calculating the battery's state of health based on the battery's measured values. In one embodiment, the physical formula includes an ordinary differential equation or a partial differential equation for representing the battery's state of health based on the battery's measured values. In one embodiment, the neural network submodule includes a fully connected neural network model.

[0059] In one embodiment, the multiple physical quantities included in each input signal at one time point include at least a portion of the current value, voltage value, temperature value and charge state of the battery at that time point. In one embodiment, the first value includes at least a portion of the change in the thickness of the solid electrolyte interface film of the battery, the change in the circulatory lithium due to the anode (-) and cathode (+) side reactions, the consumption of the electrolyte solvent, the loss of circulatory lithium due to lithium deposition, the loss of active material at the anode and the loss of active material at the cathode. In one embodiment, the second value includes a weighting factor for calculating the capacity of the battery. For example, the second value includes a weighting factor corresponding to the anode and cathode side reaction currents of the battery. For another example, the second value also includes a weighting factor corresponding to the loss of active material at the anode and cathode of the battery.

[0060] FIG7 shows a flowchart of a method for training a hybrid neural network model according to one embodiment.

[0061] In step 710, the parameters of the hybrid neural network model are stored in a shared memory block, wherein the hybrid neural network model includes a neural network submodule and a physical model submodule, wherein a first training data set for training the hybrid neural network model includes multiple periodic data groups, each periodic data group includes multiple input signals at multiple time points and a target output true value, each input signal includes multiple physical quantities at a time point, wherein, in an iteration of an input signal at a 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 the first value in this iteration based on the input signal and the second value.

[0062] In step 720, a plurality of processes are run in parallel, and the plurality of processes utilize the parameters of the hybrid neural network model stored in the shared memory block in parallel to train the hybrid neural network model based on the first training data set, wherein, for each of the plurality of processes and each periodic data group used by the process, the process utilizes the parameters of the hybrid neural network model stored in the shared memory block to iteratively generate a predicted target value based on a plurality of input signals having a plurality of time points in a time sequence in the periodic data group, 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 true value of the periodic data group.

[0063] In one embodiment, the multiple processes train the hybrid neural network model based on the multiple periodic data groups in parallel according to different orders of the multiple periodic data groups in the first training data set.

[0064] In one embodiment, the physical model submodule includes a physical formula for calculating the health state of the battery based on the measured values ​​of the battery. In one embodiment, the neural network submodule includes a fully connected neural network model.

[0065] In one embodiment, the multiple physical quantities included in each input signal at a time point include at least a portion of the current value, voltage value, temperature value and charge state of the battery at that time point, the first value includes at least a portion of the change in the thickness of the solid electrolyte interface film of the battery, the change in cyclable lithium due to anode and cathode side reactions, the consumption of electrolyte solvent, the loss of cyclable lithium due to lithium deposition, the loss of active material at the anode and the loss of active material at the cathode, and the second value includes a weighting factor for calculating the capacity of the battery. In one embodiment, the second value includes a weighting factor corresponding to the anode and cathode side reaction currents of the battery. In one embodiment, the second value also includes a weighting factor corresponding to the loss of active material at the anode and cathode of the battery.

[0066] 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 true value of the periodic data group includes: updating the parameters of the neural network sub-module and the weighting factors of the physical model sub-module corresponding to the anode and cathode active material losses of the battery based on the predicted target value and the target true value of the periodic data group.

[0067] In one embodiment, the method further includes: setting a plurality of different downsampling rates; for each of the plurality of different downsampling rates: using a training data subset obtained by the downsampling rate based on a second training data set to train the hybrid neural network model, wherein the second training data set includes a plurality of periodic data groups, each periodic data group includes a plurality of input signals at a plurality of time points and a target output true value, and each input signal includes a plurality of physical quantities at a time point; determining a training performance value of the trained hybrid neural network model based on the training data subset; determining a test performance value of the trained hybrid neural network model based on a test data set, wherein the test data set includes a plurality of periodic data groups, each periodic data group includes a plurality of input signals at a plurality of time points and a target output true value, and each input signal includes a plurality of physical quantities at a time point.

[0068] In one embodiment, the method further includes: selecting one or more downsampling rates from the multiple different downsampling rates based on multiple groups of the training performance values ​​and the test performance values ​​corresponding to the multiple different downsampling rates; and training the hybrid neural network model based on the first training data set based on the one or more downsampling rates.

[0069] In one embodiment, selecting one or more downsampling rates from the plurality of different downsampling rates includes: selecting a first downsampling rate and a second downsampling rate. Training the hybrid neural network model based on the first training data set includes: adjusting the structure of the hybrid neural network model to obtain a plurality of candidate hybrid neural network models with different structures; using the first downsampling rate to train the plurality of candidate hybrid neural network models with different structures based on the first training data set; selecting a candidate hybrid neural network model from the plurality of candidate hybrid neural network models with different structures based on respective performance indicators of the trained plurality of candidate hybrid neural network models with different structures; and using the second downsampling rate to train the selected candidate hybrid neural network model based on the first training data set.

[0070] In one embodiment, the second training dataset is a subset of the first training dataset. In one embodiment, the second training dataset is the same as the first training dataset. In one embodiment, the second training dataset is different from the first training dataset, for example, the second training subset intersects with the first training subset.

[0071] FIG8 shows a block diagram of an apparatus for a hybrid neural network model according to one embodiment.

[0072] According to one embodiment, the control system or processing system 800 may include one or more control units or processing units 810, which execute one or more machine-readable instructions stored or encoded in a machine-readable storage medium (i.e., memory 820). Although not shown in FIG8, those skilled in the art will appreciate that the control system 800 may include various other components, such as various communication modules, bus modules, and possibly user interface modules. In one embodiment, the processing units 810 in the processing system 800 are configured to perform the various operations and functions described above in conjunction with FIG1-7 when executing program instructions.

[0073] According to one embodiment, a program product, such as a non-transitory machine-readable medium, is provided. The non-transitory machine-readable medium may have instructions that, when executed by a machine, such as processing unit 810, cause a device, such as processing unit 810 or processing system 800, to perform the various operations and functions described above in conjunction with Figures 1-7 in various embodiments of the present application.

[0074] The specific embodiments described above in conjunction with the accompanying drawings describe exemplary embodiments, but do not represent all embodiments that can be implemented or fall within the scope of protection of the claims. The term "example" used throughout this specification may mean "used as an example, instance or illustration" and does not mean "preferred" or "having advantages" over other embodiments. For the purpose of providing an understanding of the described technology, the specific embodiments include specific details. However, these technologies can be implemented without these specific details. In some instances, in order to avoid obscuring the concepts of the described embodiments, well-known structures and devices are shown in block diagram form.

[0075] The foregoing description of the present disclosure is provided to enable any person skilled in the art to implement or use the present disclosure. Various modifications to the present disclosure will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the scope of the present disclosure. Therefore, the present disclosure is not limited to the examples and designs described herein, but is intended to be consistent with the widest range of principles and novel features disclosed herein.

Claims

1. A method for predicting the health status of a battery using a hybrid neural network model, comprising: Inputting a plurality of input signals at a plurality of time points included in the measured periodic data group into the hybrid neural network model sequentially in time, wherein the hybrid neural network model comprises a neural network submodule and a physical model submodule, and each of the plurality of input signals of the periodic data group comprises a plurality of physical quantities of the battery at a time point; In an iteration of an input signal at each of the multiple time points, the neural network submodule predicts a second value based on the input signal and a first value output by the physical model submodule in a previous iteration, and the physical model submodule calculates the first value in this iteration based on the input signal and the second value, until in an iteration of a last input signal at a last time point, the neural network submodule predicts a second value based on the last input signal and the first value output by the physical model submodule in the previous iteration, and the physical model submodule calculates the first value in this iteration based on the last input signal and the second value; and A state of health of the battery is determined based on the first value.

2. The method of claim 1, wherein: The physical model submodule includes physical formulas for calculating the state of health of the battery based on the measured values ​​of the battery.

3. The method of claim 2, wherein: The physical formula includes an ordinary differential equation or a partial differential equation for representing the state of health of the battery based on the measured values ​​of the battery.

4. The method of claim 2, wherein: The neural network submodule includes a fully connected neural network model.

5. The method of claim 2, wherein: The multiple physical quantities at the one time point included in each input signal include at least a part of the current value, voltage value, temperature value and charging state of the battery at the time point, the first value includes at least a part of the change in the thickness of the solid electrolyte interface film of the battery, the change in the circulatory lithium due to anode and cathode side reactions, the consumption of electrolyte solvent, the loss of circulatory lithium due to lithium deposition, the loss of active substances at the anode and the loss of active substances at the cathode, and the second value includes a weighting factor for calculating the capacity of the battery.

6. The method of claim 5, wherein: The second value includes weighting factors corresponding to anode and cathode side reaction currents of the battery.

7. The method of claim 6, wherein: The second value also includes weighting factors corresponding to the anode and cathode active material losses of the battery.

8. A method 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 includes a neural network submodule and a physical model submodule, wherein a first training data set for training the hybrid neural network model includes a plurality of periodic data groups, each periodic data group includes a plurality of input signals at a plurality of time points and a target output true value, each input signal includes a plurality of physical quantities at a time point, wherein in an iteration for an input signal at a 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 a previous iteration, and the physical model submodule calculates a first value in this iteration based on the input signal and the second value; and A plurality of processes are run in parallel, and the plurality of processes utilize the parameters of the hybrid neural network model stored in the shared memory block in parallel to train the hybrid neural network model based on the first training data set, wherein, for each of the plurality of processes and each periodic data group used by the process, the process utilizes the parameters of the hybrid neural network model stored in the shared memory block to iteratively generate a predicted target value based on a plurality of input signals having a plurality of time points in a time sequence in the periodic data group, 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 true value of the periodic data group.

9. The method of claim 8, wherein: The multiple processes perform training of the hybrid neural network model in parallel based on the multiple periodic data groups in different orders of the multiple periodic data groups in the first training data set.

10. The method of claim 9, wherein: The physical model submodule includes a physical formula for calculating the health state of the battery based on the measured values ​​of the battery, wherein the neural network submodule includes a fully connected neural network model.

11. The method of claim 10, wherein: The multiple physical quantities at the one time point included in each input signal include at least a part of the current value, voltage value, temperature value and charging state of the battery at the time point, the first value includes at least a part of the change in the thickness of the solid electrolyte interface film of the battery, the change in the circulatory lithium due to anode and cathode side reactions, the consumption of electrolyte solvent, the loss of circulatory lithium due to lithium deposition, the loss of active substances at the anode and the loss of active substances at the cathode, and the second value includes a weighting factor for calculating the capacity of the battery.

12. The method of claim 11, wherein: The second value includes weighting factors corresponding to anode and cathode side reaction currents of the battery.

13. The method of claim 12, wherein: The second value also includes weighting factors corresponding to the anode and cathode active material losses of the battery.

14. The method of claim 12, wherein: The method of updating the parameters of the hybrid neural network model stored in the shared memory block based on the predicted target value and the target true value of the periodic data group includes: updating the parameters of the neural network sub-module and the weighting factors of the physical model sub-module corresponding to the anode and cathode active material losses of the battery based on the predicted target value and the target true value of the periodic data group.

15. The method of one of claims 8 to 14, further comprising: Set multiple different downsampling rates; For each of the multiple different downsampling rates: The hybrid neural network model is trained using a training data subset obtained by the downsampling rate based on a second training data set, wherein the second training data set includes a plurality of periodic data groups, each periodic data group includes a plurality of input signals at a plurality of time points and a target output true value, and each input signal includes a plurality of physical quantities at a time point; Determining a training performance value of the trained hybrid neural network model based on the training data subset; A test performance value of the trained hybrid neural network model is determined based on a test data set, wherein the test data set includes multiple periodic data groups, each periodic data group includes multiple input signals at multiple time points and a target output true value, and each input signal includes multiple physical quantities at a time point.

16. The method of claim 15, further comprising: Selecting one or more downsampling rates from among the multiple different downsampling rates based on multiple groups of the training performance values ​​and the test performance values ​​corresponding to the multiple different downsampling rates respectively; Based on the one or more downsampling rates, the hybrid neural network model is trained based on the first training data set.

17. The method of claim 16, wherein: The selecting one or more downsampling rates among the plurality of different downsampling rates comprises: selecting a first downsampling rate and a second downsampling rate; The training of the hybrid neural network model based on the first training data set comprises: Adjusting the structure of the hybrid neural network model to obtain a plurality of candidate hybrid neural network models with different structures; Using the first downsampling rate, respectively train the plurality of candidate hybrid neural network models with different structures based on the first training data set; Selecting a candidate hybrid neural network model from the plurality of candidate hybrid neural network models with different structures based on respective performance indicators of the trained plurality of candidate hybrid neural network models with different structures; The selected candidate hybrid neural network model is trained based on the first training data set using the second downsampling rate.

18. The method of claim 16, wherein: The second training data set is a subset of the first training data set, or the second training data set is the same as the first training data set, or the second training data set is different from the first training data set.

19. An apparatus for a hybrid neural network model, comprising: Memory; as well as One or more processing units, which, when executing program instructions, are configured to perform the method of one of claims 1 to 18.

20. A machine-readable storage medium storing executable instructions which, when executed, cause one or more processors to perform the method of any one of claims 1 to 18.

Images