Method and apparatus for predicting battery health using a hybrid neural network model and training model

Abstract

This method uses a hybrid neural network model to predict the health of a battery. The method involves inputting multiple input signals at multiple time points included in a measured cycle dataset into a hybrid neural network model in chronological order, wherein the hybrid neural network model includes a neural network submodule and a physical model submodule, and each input signal of the multiple input signals in the cycle dataset includes multiple physical quantities of the battery at a single time point; in each iteration of one 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 the previous iteration; the physical model submodule calculates the first value in the current iteration based on the input signal and the second value, up to the iteration of the last input signal at the last time point; and identifies the health of the battery based on the first value.

Description

[Technical Field]

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

[0002] Rechargeable lithium batteries are widely used in various electrical devices, such as electric vehicles. A lithium battery typically comprises a negative electrode, a positive electrode, and a separator plate located between the two electrodes. Each of the two electrodes contains an active material that reversibly reacts with lithium; for example, the negative electrode contains lithium metal that dissolves and reversibly deposits in an electrochemical manner, while the separator plate contains an electrolyte with lithium cations and acts as a physical barrier between the electrodes, preventing them from electrically connecting within the battery. Typically, there is electron generation at the positive electrode, an equal amount of electron consumption at the negative electrode during charging, and the opposite reaction during discharge. Undesirable side reactions occur during repeated charge and discharge cycles of the battery, and these undesirable side reactions lead to a decrease in battery capacity. [Overview of the Initiative] [Problems that the invention aims to solve]

[0003] Battery management systems on electrical devices can be used to assess the state-of-health (SOH) of batteries. Existing physics-based electrochemical aging models have difficulty considering the complex and changing operating conditions of the real world. Therefore, it is desirable to be able to assess the state of health of batteries more accurately and efficiently. [Means for solving the problem]

[0004] Some concepts are described below in a simplified manner, but these concepts are further explained in the detailed description below. This description is not intended to highlight any major or necessary features of the claimed subject matter, nor is it intended to limit the scope of the claimed subject matter.

[0005] According to one aspect of the present application, a method is provided for predicting the health of a battery using a hybrid neural network model, comprising: inputting a plurality of input signals at a plurality of time points included in a measured cycle dataset into a hybrid neural network model in chronological order, wherein the hybrid neural network model comprises a neural network submodule and a physical model submodule, and each of the plurality of input signals in the cycle dataset comprises a plurality of physical quantities of the battery at a single time point; in each iteration of one input signal at each of the plurality of 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 the previous iteration; the physical model submodule calculates a first value in the current iteration based on the input signal and a second value, until it is iterated for the last input signal at the last time point, wherein the neural network submodule predicts a second value based on the last input signal and a first value output by the physical model submodule in the previous iteration; the physical model submodule calculates a first value in the current iteration based on the last input signal and a second value; and, based on the first value, identifies the health of the battery.

[0006] According to one aspect of this application, a method for training a hybrid neural network model, comprising storing the 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, and a first training dataset for training the hybrid neural network model includes a plurality of cycle datasets, each cycle dataset includes a plurality of input signals at a plurality of time points and a single target output true value, each input signal includes a plurality of physical quantities at a single time point, and in an iteration of a single input signal at a single 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 predicts a second value based on the input signal and the second value A method is provided which includes: calculating a first value in the current iteration; and running multiple processes in parallel, the multiple processes training a hybrid neural network model on a first training dataset using parameters of a hybrid neural network model stored in a shared memory block in parallel; for each process of the multiple processes and each cycle dataset used by the processes, the processes iteratively generate a single predicted target value using parameters of a hybrid neural network model stored in a shared memory block, based on the fact that multiple input signals at multiple time points have a time order within the cycle dataset; and the parameters of the hybrid neural network model stored in the shared memory block are updated based on the predicted target value and target true value of the cycle dataset.

[0007] According to one aspect of this application, there is a device for a hybrid neural network model, comprising a memory and one or more processing units, wherein the processing units are configured to, upon executing a program instruction, carry out a method for predicting the health status of a battery using a hybrid neural network model and a method for training a hybrid neural network model as described herein.

[0008] According to one aspect of this application, a machine-readable storage medium is provided which, when an instruction is executed, stores executable instructions causing one or more processors to carry out a method for predicting the health status of a battery using a hybrid neural network model, and a method for training the hybrid neural network model described herein.

[0009] By using the method for predicting the health of a battery using a hybrid neural network model according to this application, the dynamic weight coefficients in the physical model submodule, which depend on the working environment, are learned by using the neural network submodule so that the accuracy of predicting the health of the battery can be effectively improved. The method for training a hybrid neural network model to predict the health of a battery according to this application, based on characteristics related to the battery measurement data, can effectively reduce training time, improve training efficiency, reduce the time required to process physical equations at each time step, and mitigate the problem of long training times caused by serial processing methods of hybrid neural network models, thereby helping to solve the problem of large memory requirements. Other advantages of this disclosure are illustrated in the detailed description below.

[0010] Further understanding of the nature and merits of this disclosure can be achieved by referring to the accompanying drawings below. In the drawings, similar components or features may be given the same reference numerals. [Brief explanation of the drawing]

[0011] [Figure 1] A schematic diagram of a hybrid neural network model according to one embodiment is shown. [Figure 2] A simplified schematic diagram for training a hybrid neural network, according to one embodiment, is shown. [Figure 3] This diagram shows a schematic representation of a method for training a hybrid neural network model according to one embodiment. [Figure 4] This diagram shows a schematic representation of a method for training a hybrid neural network model according to one embodiment. [Figure 5] This diagram shows a schematic representation of a method for training a hybrid neural network model according to one embodiment. [Figure 6] A flowchart illustrating a method for predicting the health status of a battery using a hybrid neural network model, according to one embodiment, is shown. [Figure 7] A flowchart illustrating a method for training a hybrid neural network model in one embodiment is shown. [Figure 8] This shows a block diagram of the equipment used in a hybrid neural network model according to one embodiment. [Modes for carrying out the invention]

[0012] In the following, the subject matter described herein will be discussed with reference to exemplary embodiments. It should be understood that the descriptions of these embodiments are not intended to limit the scope, availability, or examples of the claims, but are provided to assist those skilled in the art in better understanding and thereby in carrying out the subject matter described herein. It may be possible to modify the function and configuration of the elements described without departing from the scope of the claims. Various processes or components may be omitted, replaced, or added in various embodiments as needed. For example, the described methods may be carried out in a different order than described, and various steps may be added, omitted, or combined. Furthermore, configurations described in relation to one embodiment may also be combined in other embodiments.

[0013] As used herein, the term “comprising” and its variations are open terms, meaning “including, but not limited to.” The term “based” means “based at least in part.” The terms “one embodiment” and “one example” mean “at least one embodiment.” The term “other embodiment” means “at least one other embodiment.” Terms such as “first,” “second,” etc., may refer to different or the same subject. Unless explicitly stated in the context, the definition of a term is consistent throughout the description.

[0014] Figure 1 shows a schematic diagram of a hybrid neural network model according to one embodiment.

[0015] The hybrid neural network model 100 shown in Figure 1 includes a neural network submodule 110 and a physical model submodule 120. The hybrid neural network model 100 receives battery measurement signals 130, such as battery voltage, current, temperature, and state of charge (SOC), and determines the state of health (SOH) of the battery based on the received battery measurement signals 130. The state of charge is, for example, a charge state where the charge capacity reaches 30%, 60%, etc. For example, the state of health may be the battery capacity, or it may be referred to as the remaining capacity. Assuming the original capacity of the battery is C0, as the battery is used and ages, the battery capacity decreases, and thereafter the battery's current capacity can be used to represent the state of health of the battery. Alternatively, it can be assumed that the original capacity of the battery is 1, and therefore the remaining capacity of the battery can be expressed as a fraction less than 1.

[0016] The main aging phenomena of a battery include the growth of the solid electrolyte interface (SEI), which leads to lithium loss and solvent consumption; lithium deposition, which leads to lithium loss; and the loss of positive and negative electrode active materials. In one embodiment, the aging model may be represented by the following set of physical equations.

[0017]

number

[0018] Equation (1) represents the change in SEI film thickness due to the reaction on the negative electrode side. Equation (2) represents the change in recyclable lithium caused by the reactions on the negative (-) and positive (+) electrode sides. Equation (3) represents the consumption of the electrolyte solvent. Equation (4) represents the loss of recyclable lithium due to lithium deposition. Equations (5) and (6) represent the loss of negative electrode and positive electrode active materials, respectively. The side reactions in equations (1) to (6) can be expressed by the following physical equations.

[0019]

number

[0020] In equations (7) and (10)

number

[0021] The state of health (SOH) of a battery may also be defined as the remaining capacity of the battery, or the battery capacity or system state X. cLi , X Li , X AML- , X AML+ The physical quantities that affect this can be obtained by solving the usual difference equations (1) to (6). The remaining capacity of the battery, SOHc, can be expressed as follows:

[0022]

number

[0023] In equation (12), C0 represents the original capacity of the battery, and k a and k c This is a coefficient, or sometimes called a weighting coefficient.

[0024] The remaining capacity SOHc of the battery can be calculated by physical equations based on the measured physical quantities of the battery, for example, the current I(t), voltage V(t), temperature T(t), and state of charge SOC(t) in Equations (7) to (11) according to the electrochemical aging model of the battery represented by Equations (1) to (12). Here, the technical solution of the present disclosure is exemplified by the aging model represented by Equations (1) to (12) as an example. Those skilled in the art can understand that the above implementation of the electrochemical aging model for modeling the battery by Equations (1) to (12) is only an example for explaining the technical solution of the present disclosure. The electrochemical aging model based on physics is not limited to a specific implementation, and any suitable physical-chemical aging model can be used.

[0025] The physical model sub-module 120 shown in FIG. 1 includes a physical aging model 1210. In one embodiment, the physical model 1210 may include the ordinary differential equations (ODEs) represented by the above Equations (1) to (6), and the corresponding physical equations (7) to (11). Specifically, for one measurement signal 130 at one time point t, the measurement signal 130 includes the current I(t), voltage V(t), temperature T(t), and state of charge SOC(t) of the battery, and the physical aging model 1210 can be represented by the values of the ordinary differential equations (1) to (6) obtained by Equations (1) to (11) based on the measurement signal 130. The physical model sub-module 120 also includes an ODE solver 1220, and the ODE solver 1220 solves the ordinary differential equations (1) to (6) by one forward process to obtain the system state X of the battery SEI 、X cLi 、X csoll,2 、X Li、 X AML- 、X AML+The following are obtained, each representing the change in SEI film thickness due to the negative electrode reaction, the change in recyclable lithium caused by the negative (-) and positive (+) side reactions, the consumption of the electrolyte solvent, the loss of recyclable lithium due to lithium deposition, the loss of negative electrode active material, and the loss of positive electrode active material, respectively. It will be understood that the ODE solver 1220 may be any suitable ODE solver, such as the RBMS Kutt solver or Euler. The physical model submodule 120 also comprises an SOH determination unit 1230, which determines the remaining capacity SOHc of the battery as the healthy state of the battery according to formula (12). In other embodiments, the original capacity of the battery may be represented as 1, and the remaining capacity SOHc may be represented as a fraction less than 1, for example, the SOH determination unit 1230 may determine the remaining capacity SOHc of the battery as the healthy state of the battery according to formula (13).

[0026]

number

[0027] In equation (13), C0 represents the original capacity of the battery, and k a and k c This is a coefficient.

[0028] In the embodiment using the physical aging model, the loss X of the negative electrode and positive electrode active materials AML- and X AML+ In equations (5) and (6) used to calculate the negative electrode reaction current and positive electrode reaction current related to the first electrolyte or the second electrolyte,

[0029]

number

[0030] In the embodiment shown in Figure 1, the neural network submodule 110 determines the negative electrode side reaction current and the positive electrode side reaction current based on the input signals I(t), V(t), T(t), and SOC(t) at a single time point.

[0031]

number

[0032] In the embodiment shown in Figure 1, at one time point t, the neural network submodel 110 takes as input the input signals I(t), V(t), T(t), and SOC(t) at time point t, as well as at least a portion of the output signal of the physical model submodule 120 at the previous time point t-1, and the coefficient k at time point t. I- |I(t)| and k I+ |I(t)| 140 is obtained. For example, at one time point t, the neural network submodel 110 obtains the input signals I(t), V(t), T(t), and SOC(t) at time point t, as well as the system state X output by the physical model submodule 120 at the previous time point t-1. SEI , X cLi , X Csoll,2 , X Li , X AML- , X AML+ The following is obtained as input, and the coefficient k at time t I- |I(t)| and k I+ |I(t)| Get 140.

[0033] In other embodiments, the neural network submodule 110 predicts a coefficient 140 at a single time point t based on the input signals I(t), V(t), T(t), and SOC(t), where the coefficient 140 is the negative electrode side reaction current and the positive electrode side reaction current.

[0034]

number

[0035] In the embodiment shown in Figure 1, at one time point t, the neural network submodel 110 takes as input the input signals I(t), V(t), T(t), and SOC(t) at time point t, as well as at least a portion of the output signal of the physical model submodule 120 at the previous time point t-1, and at time point t, the coefficient k I- |I(t)|, k I+ |I(t)|, k a and k c 140 is obtained. For example, at one time point t, the neural network submodel 110 takes the input signals I(t), V(t), T(t), and SOC(t) at time point t, as well as the system state X output by the physical model submodule 120 at the previous time point t-1. SEI , X cLi , X Csoll,2 , X Li , X AML- , X AML+ The coefficient k at the previous time point t is obtained as input. I- |I(t)|, k I+ |I(t)|, k a and k c Get 140.

[0036] Implementations of embodiments of this disclosure for predicting the health of a battery are described above in relation to equations (1) to (13) and the embodiments shown in Figure 1, but it should be understood that this disclosure is not limited to the embodiments described above. For example, in one embodiment, the output 150 of the physical model submodule 120 is not provided to the neural network submodule 110, rather the neural network submodule 110 outputs a coefficient 140 based only on the input signals I(t), V(t), T(t) and SOC(t) at time t. For example, the physical model submodule 120 can employ any suitable but not limited physical aging model known or to be known in the art, and is not limited to the physical aging model represented by the ODE shown in equations (1) to (13), for example the physical model submodule 120 can employ a suitable aging model represented by a partial difference equation (PDE).

[0037] In one embodiment, to predict the health of a battery using a hybrid neural network model 100, data from a single cycle containing multiple input signals at multiple time points is acquired. For example, the cycle data contains thousands of measurement signals at thousands of time points, each measurement signal containing, for example, the battery current I(t), voltage V(t), temperature T(t), and charge state (SOC(t)). The cycle data may also be referred to as a measured cycle dataset. As shown in Figure 1, the multiple input signals at multiple time points within the cycle dataset are input to the hybrid neural network model 100 in chronological order. At each iteration of a single input signal at multiple time points, the neural network submodule 110 predicts a second value based on the input signal and a first value output by the physical model submodule 120 in the previous iteration, and the physical model submodule 120 calculates a first value in the current iteration based on the input signal and the second value. At the last iteration of the final input signal at the last time point, the neural network submodule 110 predicts a second value based on the last input signal and a first value output by the physical model submodule 120 in the previous iteration, and the physical model submodule 120 calculates a first value in the current iteration based on the last input signal and the second value. Finally, the health status of the battery is determined based on the first value. In one embodiment, the first value is the system state X. SEI , X cLi , X Csoll,2 , X Li , X AML- , X AML+ It may include the coefficient k. In one embodiment, the second value above is the coefficient k I- |I(t)| and k I+ It may include |I(t)|. In one embodiment, the second value above is the coefficient k I- |I(t)|, k I+ |I(t)|, k a and k cIt may include 140. If the physical aging model 120 is achieved using different physical equations, it will be understood that the first and second values ​​mentioned above may be the respective physical quantities and coefficients, not limited to the specific physical quantities and coefficients described in the above embodiment.

[0038] As mentioned above, for example, when the hybrid neural network model 100 shown in Figure 1 processes measurement data for one cycle, the multiple measurement data for the cycle must be processed either chronologically or sequentially. Such a sequential processing method can cause difficulties in the training process of the hybrid neural network model 100. One difficulty is the problem of long training times.

[0039] In one embodiment, the training dataset includes multiple batteries, each battery includes multiple detection cycles, each cycle includes multiple timestamps and the measurement signal for each timestamp. For example, the dataset includes 33 batteries, each battery includes 4 to 8 measurement cycles, each cycle includes thousands of timestamps and input signals such as current, voltage, temperature, and charge state for each timestamp. For each cycle, there is one target true value (healthy state), and the number of target true values ​​for each battery is equal to the number of cycles. In the training process, the hybrid neural network model 100 is used to predict the healthy state of the batteries based on the data in the training dataset, and to update 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 based on the predicted and true values ​​of the battery state. However, since multiple measurement data must be processed continuously over time, and the nature of the physical model 120 is that of a normal difference equation (ODE), the main problem faced by joint learning of the ODE and the neural network is the long training time. At the same time, for continuous processing over time, it is necessary to save the parameters for each time step, and therefore a large amount of memory is required.

[0040] Figure 2 shows a simplified schematic diagram for training the hybrid neural network 100 according to an embodiment. The computational graph shown in Figure 2 represents the operation of one cycle of the training dataset. The nodes indicated by each circle in the first row represent the input signal 210 at one point in time. Each node in the second row represents the calculation 220 of the neural network sub-module 110. Each node in the third row represents the calculation 230 of the physical model sub-module 120.

[0041] In the process of training the hybrid neural network 100 using the training dataset, the input signal 210 at one point in time is input into the neural network sub-module 110. The input signal 210 and the output data from the neural network sub-module 110 are input into the physical model sub-module 120. The physical aging model 1210 and the ODE solver 1220 of the physical model sub-module 120 calculate the hidden state of the next time step, such as the system state X SEI , X cLi , X Csoll,2 , X Li , X AML- , X AML+ , etc. Then, the same operation is repeatedly performed for each next point in time (or referred to as a time step) until the last point in time of this cycle. The SOH determination unit 1230 of the physical model sub-module 120 calculates the remaining capacity of the battery as the prediction target value 240. The loss value may be obtained based on the prediction target value 240 and the target true value of this cycle. For example, the loss value is the difference between the target value and the target true value.

[0042] Next, based on the loss value, backpropagation is sequentially performed from node 240 to the physical model submodule 120 and the neural network submodule 110 of the previous time step. Backpropagation is sequentially performed to the following previous time steps until the first time step of this cycle, thereby updating the trainable parameters of the physical model submodule 120 and the neural network submodule 110. In one example, the losses X AML- and X AML+ corresponding coefficients k a and k c are configured as trainable parameters, thereby obtaining the optimal coefficients k a and k c in the training process of the hybrid neural network model, thereby improving the prediction accuracy of the hybrid neural network model 100.

[0043] Due to the complexity of the ODE calculation, the backpropagation of the ODE becomes more complex, so the backpropagation for each time step takes a relatively long time. For the entire illustrated dataset and for a total of 33 batteries, there are thousands of time steps per cycle and 4 to 8 cycles per battery. Since the number of time steps in the entire dataset is very large and each time step takes a long time, the training of the continuous processing shown in FIG. 2 takes a very long training time. For example, for the illustrated dataset, one training epoch has a forward processing time of about 15 minutes and a backward processing time of more than 2 hours. Here, the term "epoch" refers to running all the data in the training dataset once. In a practical implementation, if the training process of the hybrid neural network model 100 requires training for several epochs, the training dataset may be much larger than the illustrated training dataset, and it would be advantageous if the training time of the hybrid neural network model 100 could be shortened, thereby improving its training efficiency.

[0044] Figure 3 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 Figure 3 may be the same as the hybrid neural network model 100 shown in Figure 1. As shown in Figure 3, the parameters of the hybrid neural network model 300, for example, are stored in a shared memory block. Multiple processes (for example, processes P1 to P4 shown in Figure 3) are executed in parallel. The four processes P1 to P4 shown in Figure 3 share the same hybrid neural network model 300, 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 each updates the parameters of the same hybrid neural network model 300 in parallel with the same dataset.

[0045] In one embodiment, multiple processes P1 to P4 train a hybrid neural network model 300 in parallel according to different orders of multiple cycle datasets within the training dataset D. For example, the training dataset D is divided into four subsets D1 to D4, and each of processes P1 to P4 trains the hybrid neural network model 300 according to D1 to D4 in a different order. For example, process P1 trains the hybrid neural network model 300 in set D in the order {D1, D2, D3, D4}, process P2 trains the hybrid neural network model 300 in set D in the order {D2, D3, D4, D1}, process P3 trains the hybrid neural network model 300 in set D in the order {D3, D4, D1, D2}, and process P4 trains the hybrid neural network model 300 in set D in the order {D4, D1, D2, D3}. It will be understood that processes P1 to P4 may also train model 300 in a different order of subsets D1 to D4. In other embodiments, the training data of training dataset D is ordered in terms of a cycle dataset or battery units, and processes P1 to P4 each train the hybrid neural network model 300 according to dataset D ordered in a different order.

[0046] As shown in Figure 3, by using four parallel processes P1 to P4, each process independently executes the continuous training process on the same hybrid neural network model 300 using the same dataset D, as described above in Figure 2. When the four processes P1 to P4 execute one epoch in parallel, it is equivalent to executing four epochs on the hybrid neural network model 300 using the same dataset D, and therefore the efficiency of the hybrid neural network model 300 can be improved.

[0047] Figure 4 shows a more detailed schematic diagram of the method shown in Figure 3. As shown in Figure 4, the four processes P1 to P4 independently train the hybrid neural network model 300 according to their respective cycle datasets P1-C1, P1-C2, ..., P4-C1, P4-C2, etc. In one embodiment, as shown in Figure 4, for example, after process P2 updates the parameters of the hybrid neural network model 300 with the current cycle dataset 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 with the current cycle dataset 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 in dataset D is not necessarily the same.

[0048] Figure 5 shows a schematic diagram of a method for training a hybrid neural network model according to one embodiment.

[0049] Still, taking the training dataset described above as an example, the measurement data at multiple time points acquired during the battery charging and discharging process often fits linear changes, so it is possible to train a hybrid neural network model using a subset of the training dataset by downsampling. In one embodiment, several different downsampling rates 510 can be set, for example, downsampling rates 1 to 8 shown in Figure 5. Downsampling rate 1 represents all data samples in the sampling dataset, downsampling rate 2 represents half of the data samples in the sampling dataset, for example, one sample is taken for every two data samples, and similarly, downsampling rate 8 represents one-eighth of the data samples in the sampling dataset, for example, one sample is taken for every eight data samples.

[0050] In a single training dataset D', for each downsampling rate 1 to 8, the hybrid neural network model 100 is trained on a subset of the training dataset D' acquired based on the downsampling rate. The performance of the trained hybrid neural network model 100 is evaluated on the respective data subset and test set, and the performance can be expressed as the mean squared error (MSE), as shown by 520 and 530 in Figure 5, representing MSE 520 on the training subset and MSE 530 on the test set, respectively. Simultaneously, the training process for each downsampling rate may be time-set, for example, as shown by 540 and 550 in Figure 5, representing forward time 540 and backward time 550 of the same epoch, respectively, with respect to the data subset corresponding to the downsampling rate. One or more downsampling rates are selected to train the hybrid neural network 100 based on performance indicators generated for each downsampling rate, such as the test set MSE 530.

[0051] In one embodiment, the training dataset D' for selecting the downsampling rate, as described in relation to Figure 5, is a subset of the actual training dataset D used in Figures 2 to 4, thereby allowing the downsampling rate of the actual training process to be selected with less computation. Furthermore, in other embodiments, the training dataset D' may be identical to the actual training dataset D, or it may be different from the actual training dataset D, intersecting with it.

[0052] In one embodiment, for example, as shown in Figure 5, downsampling rates 2 and 8 may be employed to train the hybrid neural network model 100, and downsampling rates 2 and 8 may be used in different training stages of the hybrid neural network model 100. In one embodiment, a larger downsampling rate 8 may be used to train the model during the structural tuning stage of the hybrid neural network model 100. In this embodiment, the structure of the hybrid neural network model is tuned to obtain multiple candidate hybrid neural network models with different structures, downsampling rate 8 is used to train multiple candidate hybrid neural network models with different structures based on the training dataset D, and one candidate hybrid neural network model is selected from the multiple candidate hybrid neural network models with different structures based on the performance metrics of each of the multiple candidate hybrid neural network models with different structures. Once one hybrid neural network model 100 with the optimal structure is selected, the candidate hybrid neural network model is trained on the training dataset D using downsampling rate 2.

[0053] Figure 6 shows a flowchart illustrating a method for predicting the health status of a battery using a hybrid neural network model, according to one embodiment.

[0054] In step 610, multiple input signals at multiple time points included in the measured cycle dataset are input to a hybrid neural network model in chronological order, the hybrid neural network model including a neural network submodule and a physical model submodule, and each input signal of the multiple input signals in the cycle dataset includes multiple physical quantities of the battery at a single time point.

[0055] In step 620, at each iteration of one input signal at 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 the previous iteration, and the physical model submodule calculates the first value in the current iteration based on the input signal and the second value, until 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 at the last time point for the last input signal, the physical model submodule calculates the first value in the current iteration based on the last input signal and the second value.

[0056] In step 630, the health status of the battery is determined based on the first value. For example, the health status of the battery may be represented by the remaining capacity SOHc of the battery, as detailed above. For example, the second value generated in step 620 is the system state X of the system, as detailed above. cLi , X Li , X AML- , X AML+ It may also include the remaining capacity of the battery, SOHc, which is calculated in step 630 and functions as the healthy state of the battery, for example, based on equation (12). In other embodiments, the second value generated in step 620 is the system state X cLi、 XLi , X AML- , X AML+ The value may also include the remaining capacity of the battery, SOHc, and the remaining capacity of the battery in the second value above is obtained as the healthy state of the battery in step 630.

[0057] In one embodiment, the physical model submodule includes physical equations for calculating the health state of the battery based on battery measurements. In one embodiment, the physical equations include ordinary difference equations or partial difference equations for representing the health state of the battery based on battery measurements. In one embodiment, the neural network submodule includes a fully connected neural network model.

[0058] In one embodiment, the multiple physical quantities at a single point in time included in each input signal include at least a portion of the battery's current, voltage, temperature, and charge state at that point in time. In one embodiment, the first value includes at least a portion of the change in the solid electrolyte interface film thickness of the battery, the change in recyclable lithium caused by the negative (-) and positive (+) side reactions, the consumption of the electrolyte solvent, the loss of recyclable lithium due to lithium deposition, the loss of negative electrode active material, and the loss of positive electrode active material. In one embodiment, the second value includes weighting coefficients used to calculate the battery's capacity. For example, the second value includes weighting coefficients corresponding to the reaction currents on the negative and positive electrode sides of the battery. In other embodiments, the second value also includes weighting coefficients corresponding to the loss of negative and positive electrode active material of the battery.

[0059] Figure 7 shows a flowchart illustrating a method for training a hybrid neural network model according to one embodiment.

[0060] In step 710, the parameters of the hybrid neural network model are stored in a shared memory block, the hybrid neural network model includes a neural network submodule and a physical model submodule, the first training dataset for training the hybrid neural network model includes multiple cycle datasets, each cycle dataset includes multiple input signals at multiple time points and one target output true value, each input signal includes multiple physical quantities at one time point, in an iteration of one input signal at 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 the first value in the current iteration based on the input signal and the second value.

[0061] In step 720, multiple processes are executed in parallel, and the multiple processes train the hybrid neural network model on a first training dataset using the parameters of the hybrid neural network model stored in a shared memory block in parallel. For each of the multiple processes and each cycle dataset used by the processes, the processes iteratively generate a single predicted target value using the parameters of the hybrid neural network model stored in the shared memory block, based on the fact that multiple input signals at multiple time points have a temporal order within the cycle dataset. The parameters of the hybrid neural network model stored in the shared memory block are updated based on the predicted target value and target true value of the cycle dataset.

[0062] In one embodiment, multiple processes train a hybrid neural network model in parallel based on multiple cycle datasets, according to a different order of multiple cycle datasets within a first training dataset.

[0063] In one embodiment, the physical model submodule includes physical equations for calculating the health state of the battery based on battery measurements. In one embodiment, the neural network submodule includes a fully connected neural network model.

[0064] In one embodiment, a plurality of physical quantities at a single time point included in each input signal include at least a portion of the battery's current value, voltage value, temperature value, and charge state at that point in time, the first value including at least a portion of the change in the solid electrolyte interface film thickness of the battery, the change in recyclable lithium caused by the negative and positive electrode reactions, the consumption of the electrolyte solvent, the loss of recyclable lithium caused by lithium deposition, the loss of negative electrode active material, and the loss of positive electrode active material, the second value including weighting coefficients for calculating the battery capacity. In one embodiment, the second value includes weighting coefficients corresponding to the negative and positive electrode reaction currents of the battery. In one embodiment, the second value also includes weighting coefficients corresponding to the loss of negative electrode and positive electrode active material of the battery.

[0065] In one embodiment, updating the parameters of a hybrid neural network model stored in a shared memory block based on the predicted target values ​​and target true values ​​of a cycle dataset includes updating the parameters of a neural network submodule, as well as updating the weight coefficients of a physical model submodule corresponding to the negative and positive electrode active material losses of the battery based on the predicted target values ​​and target true values ​​of the cycle dataset.

[0066] In one embodiment, the method further comprises setting a plurality of different downsampling rates, training a hybrid neural network model for each of the plurality of different downsampling rates using a subset of training data obtained from the downsampling rate based on a second training dataset, wherein the second training dataset comprises a plurality of cycle datasets, each cycle dataset comprising a plurality of input signals and a target output true value at a plurality of time points, and each input signal comprising a plurality of physical quantities at a single time point, and determining the training performance value for the trained hybrid neural network model based on the subset of training data, and determining the test performance value for the trained hybrid neural network model based on a test dataset, wherein the test dataset comprises a plurality of cycle datasets, each cycle dataset comprising a plurality of input signals and a single target output true value at a plurality of time points, and each input signal comprising a plurality of physical quantities at a single time point.

[0067] In one embodiment, the method further includes selecting one or more downsampling rates from among several different downsampling rates based on several sets of training performance values ​​and test performance values ​​corresponding to several different downsampling rates, and training a hybrid neural network model based on a first training dataset based on one or more downsampling rates.

[0068] In one embodiment, selecting one or more downsampling rates from a plurality of different downsampling rates includes selecting a first downsampling rate and a second downsampling rate. Training a hybrid neural network model on a first training dataset includes tuning the structure of the hybrid neural network model to obtain a plurality of candidate hybrid neural network models with different structures; training each of the plurality of candidate hybrid neural network models with different structures on the first training dataset using the downsampling rates; selecting one candidate hybrid neural network model from the plurality of candidate hybrid neural network models with different structures based on the performance indicators of each of the plurality of candidate hybrid neural network models with different structures; and training the selected candidate hybrid neural network model on the first training dataset using the second downsampling rate.

[0069] 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 is different from the first training dataset; for example, the second training subset intersects with the first training subset.

[0070] Figure 8 shows a block diagram of the equipment used in a hybrid neural network model according to one embodiment.

[0071] In one embodiment, the control system, or processing system 800, may comprise 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 Figure 8, those skilled in the art will understand that the control system 800 may comprise various other components, such as various communication modules, e.g., bus modules, and possible user interface modules. In one embodiment, the processing unit 810 in the processing system 800 is configured to execute various operations and functions described above in relation to Figures 1 to 7 when it executes a program instruction.

[0072] According to one embodiment, a program product such as a non-temporary machine-readable medium is provided. When executed by a machine such as a processing unit 810, the non-temporary machine-readable medium may contain instructions that cause devices such as the processing unit 810 and the processing system 800 to perform the various operations and functions described above in relation to Figures 1 to 7 in various embodiments of this application.

[0073] Exemplary embodiments are described above with reference to specific examples shown in the accompanying drawings, but do not represent all embodiments that may be practiced or included within the scope of the claims. Throughout this specification, the term “exemplary” means “serves as an example or illustration,” and does not mean “preferred” or “advantageous” over other examples. Specific embodiments include specific details to facilitate understanding of the described technology. However, these technologies may be practiced without these specific details. In some cases, known structures and apparatus are shown in block diagram form to avoid causing difficulty in understanding the concepts of the described embodiments.

[0074] The foregoing descriptions in this disclosure are provided to enable those skilled in the art to implement or use the disclosure. Various modifications of this disclosure will be obvious to those skilled in the art, and the general principles defined in this disclosure may be applied to other modifications without departing from the scope of protection of this disclosure. Accordingly, this disclosure is not limited to the exemplary embodiments and designs described herein, but corresponds to the broadest scope defined by the principles and novel features disclosed herein.

Claims

1. A method for predicting the health of a battery using a hybrid neural network model, The method involves inputting multiple input signals at multiple time points included in a measured cycle dataset into the hybrid neural network model in chronological order, wherein the hybrid neural network model includes a neural network submodule and a physical model submodule, and each of the multiple input signals in the cycle dataset includes multiple physical quantities of the battery at a single time point. In each of the multiple time points, in the iteration of one input signal at each 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; the physical model submodule calculates the first value in the current iteration based on the input signal and the second value until it is repeated for the last input signal at the last time point, wherein 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 a first value in the current iteration based on the last input signal and the second value. Based on the aforementioned first value, the healthy state of the battery is identified, A method that includes this.

2. The method according to claim 1, wherein the physical model submodule includes physical equations for calculating the healthy state of the battery based on measurements of the battery.

3. The method according to claim 2, wherein the physical equation includes a normal difference equation or a partial difference equation for representing the healthy state of the battery based on the measured values ​​of the battery.

4. The method according to claim 2, wherein the neural network submodule includes a fully connected neural network model.

5. The method according to claim 2, wherein the plurality of physical quantities at a single point in time included in each input signal include at least a portion of the current value, voltage value, temperature value and the charge state of the battery at that point in time, the first value includes at least a portion of the change in the film thickness of the solid electrolyte interface of the battery, the change in recyclable lithium caused by the negative electrode and positive electrode reactions, the consumption of the electrolyte solvent, the loss of recyclable lithium due to lithium deposition, and the loss of negative electrode active material and the loss of positive electrode active material, and the second value includes a weighting coefficient for calculating the capacity of the battery.

6. The method according to claim 5, wherein the second value includes weighting coefficients corresponding to the reaction currents on the negative and positive sides of the battery.

7. The method according to claim 6, wherein the second value further includes weighting coefficients corresponding to the losses of the negative electrode and positive electrode active materials of the battery.

8. A method for training a hybrid neural network model, 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, and a first training dataset for training the hybrid neural network model includes a plurality of cycle datasets, each cycle dataset includes a plurality of input signals at a plurality of time points and a single target output true value, each input signal includes a plurality of physical quantities at a single time point, and in an iteration of a single input signal at a single 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. The method involves executing multiple processes in parallel, wherein each of the multiple processes trains the hybrid neural network model based on a first training dataset, using the parameters of the hybrid neural network model stored in the shared memory block in parallel; for each of the multiple processes and each cycle dataset used by the processes, the process iteratively generates a single predicted target value based on the fact that the multiple input signals at multiple time points have a time order within the cycle dataset, using the parameters of the hybrid neural network model stored in the shared memory block; and the parameters of the hybrid neural network model stored in the shared memory block are updated based on the predicted target value and target true value of the cycle dataset. A method that includes this.

9. The method according to claim 8, wherein the plurality of processes train the hybrid neural network model in parallel based on the plurality of cycle datasets according to a different order of the plurality of cycle datasets in the first training dataset.

10. The method according to claim 9, wherein the physical model submodule includes physical equations for calculating the health state of the battery based on measurements of the battery, and the neural network submodule includes a fully connected neural network model.

11. The method according to claim 10, wherein the plurality of physical quantities at a single point in time included in each input signal include at least a portion of the current value, voltage value, temperature value and the charge state of the battery at that point in time, the first value includes at least a portion of the change in the film thickness of the solid electrolyte interface of the battery, the change in recyclable lithium caused by the negative electrode and positive electrode reactions, the consumption of the electrolyte solvent, the loss of recyclable lithium due to lithium deposition, the loss of negative electrode active material and the loss of positive electrode active material, and the second value includes a weighting coefficient for calculating the capacity of the battery.

12. The method according to claim 11, wherein the second value includes weighting coefficients corresponding to the reaction currents on the negative and positive sides of the battery.

13. The method according to claim 12, wherein the second value further includes weighting coefficients corresponding to the losses of the negative electrode and positive electrode active materials of the battery.

14. The method according to claim 12, wherein updating the parameters of the hybrid neural network model stored in the shared memory block based on the predicted target values ​​and target true values ​​of the cycle dataset includes updating the parameters of the neural network submodule and the weight coefficients of the physical model submodule corresponding to the losses of the negative electrode and positive electrode active materials of the battery, based on the predicted target values ​​and target true values ​​of the cycle dataset.

15. Setting multiple different downsampling rates, The process involves training the hybrid neural network model for each of the multiple different downsampling rates based on a second training dataset, using a subset of training data obtained from the downsampling rates, wherein the second training dataset comprises multiple cycle datasets, each cycle dataset comprising multiple input signals and target output true values ​​at multiple time points, and each input signal comprising multiple physical quantities at a single time point. Based on the aforementioned training data subset, the training performance value for the trained hybrid neural network model is determined. Determining the test performance value of the trained hybrid neural network model based on a test dataset, wherein the test dataset comprises multiple cycle datasets, each cycle dataset comprising multiple input signals and one target output true value at multiple time points, and each input signal comprising multiple physical quantities at one time point. The method according to any one of claims 8 to 14, further comprising:

16. Selecting one or more downsampling rates from the multiple different downsampling rates based on multiple sets of training performance values ​​and test performance values ​​corresponding to each of the multiple different downsampling rates, Training the hybrid neural network model based on the first training dataset, based on the one or more downsampling rates, The method according to claim 15, further comprising:

17. Selecting one or more downsampling rates from the aforementioned multiple 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 dataset is: The structure of the aforementioned hybrid neural network model is adjusted to obtain several candidate hybrid neural network models with different structures. Using the first downsampling rate, the candidate hybrid neural network models of the multiple different structures are trained based on each of the first training datasets. Based on the performance indicators of each of the trained candidate hybrid neural network models of multiple different structures, one candidate hybrid neural network model is selected from the multiple candidate hybrid neural network models of multiple different structures. Training the selected candidate hybrid neural network model based on the first training dataset using the second downsampling rate, The method according to claim 16, including the method described in claim 16.

18. The method according to claim 16, wherein the second training dataset is a subset of the first training dataset, or the second training dataset is identical to the first training dataset, or the second training dataset is different from the first training dataset.

19. A device for hybrid neural network models, Memory and One or more processing units, each configured to perform the method described in any one of claims 1 to 18 when it executes a program instruction, A device equipped with the following features.

20. A machine-readable storage medium, A machine-readable storage medium that, when executed, stores executable instructions for causing one or more processors to carry out the method according to any one of claims 1 to 18.

Images