A new energy aircraft lithium battery health condition deep learning cross-domain prediction method

Abstract

The application designs a new energy aircraft lithium battery health condition deep learning cross-domain prediction method, and belongs to the field of artificial intelligence and aviation technology; firstly, the discharge current data of the battery cycle process is collected, all SOH of the battery is extracted and arranged in time sequence to obtain the SOH time sequence of the battery; then, the source domain and the target domain of the battery are determined; a convolutional neural network is used as the framework of the prediction model, and the training is carried out, so that the differences between different battery categories can be well reflected, the differences can be made up in the optimization process, and then the prediction model meeting multiple battery health state indexes is obtained; the cross-domain health state index of the battery is predicted; the technical scheme has the advantages of low complexity, strong robustness and the like.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence and aviation technology, and in particular relates to a deep learning cross-domain prediction method for the health status of lithium batteries in new energy aircraft. Background Technology

[0002] New energy electric aircraft use lithium-ion batteries as their power source. A commonly used indicator to characterize the state of health of lithium-ion batteries is State of Health (SOH). It represents the ratio of battery capacity to its rated capacity after a period of use. The main SOH prediction methods fall into two categories: model-based and data-driven methods. Model-based prediction methods extract key features (voltage, impedance, etc.) of the centralized charging and discharging process from the data and design corresponding models to predict the SOH during battery cycling. Data-driven methods analyze historical battery health information and use statistical principles or machine learning techniques for prediction.

[0003] Deep neural networks are an effective method for training predictive models of battery state of health (SOH) using existing data samples. However, due to the differences in lithium-ion battery types and operating conditions for electric aircraft, a health prediction model trained using one type of lithium-ion battery cannot perform well on other types of batteries. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a deep learning-based cross-domain prediction method for the health status of lithium batteries in new energy aircraft.

[0005] A deep learning-based cross-domain prediction method for the health status of lithium batteries in new energy aircraft, specifically including the following steps:

[0006] Step 1: Collect discharge current data during battery cycling, extract all SOH data of the battery and arrange them in chronological order to obtain the SOH time series of the battery, which reflects the health status of the battery;

[0007] Specifically, the State of Health (SOH) defined by the battery's capacity reflects the battery's health status; it uses a percentage to represent the battery's state from the beginning to the end of its lifespan, and its expression is:

[0008]

[0009] Among them, C t Let Ct be the capacity during the t-th charge-discharge cycle, and C0 be the battery's rated capacity. The rated capacity of the battery is normalized using the State of Harm (SOH) index to avoid the potential impact of differences in rated capacity between different battery types on the prediction results. For the capacity Ct during the discharge cycle... tIt is obtained by integrating the discharge current over time; therefore, by collecting the current and time information of the battery discharge process, the SOH value of this cycle state is obtained; and by arranging all the SOH values ​​of the battery in chronological order, the SOH time series of the battery is obtained.

[0010] Step 2: Determine the source and target domains of the batteries; one type of battery containing prediction result labels is classified as the source domain data for training. The SOH data of this type of battery will be used as the input and output of the neural network architecture and as the basis for training; another type of battery without prediction result labels is used as the target domain for verification and as the basis for the neural network to learn the features of the target domain.

[0011] Step 3: Using convolutional neural networks as the prediction model framework, solve the problem of predicting health indicators across different areas of the battery. The basic framework is described below:

[0012] First, the historical time series of battery SOH extracted in step 1 is used as the input to the neural network; this time series is represented as x = [X1, X2, ..., X...]. N ], where N represents the length of the input time series; and the convolution operation is performed by a sequence of length F. L The filtering kernel w completes the process, extracting a length F from the time series x in chronological order. L The time series is represented as:

[0013]

[0014] The process of performing a convolution operation on a time series is as follows:

[0015]

[0016] Where *T represents the transpose of matrix *, and W and b represent the weight term and the bias term, respectively; As a non-linear activation function, Leaky ReLU is chosen as the activation function, expressed as:

[0017]

[0018] Compared to the ReLU function, the Leaky ReLU function solves the zero gradient problem for negative values ​​by giving a very small linear component to the negative input; typically, α is a small value.

[0019] Slide the filter window from the first point to the last point of the SOH time series sample data participating in the convolution operation to obtain the feature map of the j-th filter, represented as:

[0020]

[0021] To further increase the number of layers in the neural network, the output z j It will continue to be transmitted as input to the next layer of the network until the final network output value y is obtained, which represents the SOH value of the battery in a future charge-discharge cycle test;

[0022] Step 4: Perform backpropagation calculation on the neural network and calculate its loss function. Evaluate the neural network, complete the training of the neural network framework, obtain the prediction model, and realize the prediction of cross-domain health status indicators of the battery.

[0023] For the backpropagation process, by setting a specific objective function L, the gradient values ​​obtained from the weight and bias terms in Equation 3 are used to update the weights and bias terms. To solve the prediction of the battery cross-domain SOH index, the objective function L is constructed as follows, which includes two indices L... P and L m The two are adjusted using a weight value l, as follows:

[0024] L = L p +l×L m (6)

[0025] Where L P The accuracy metric for prediction is denoted by the root mean square loss function (RMSE). Assuming the true value at the next time step is y, the output value of the trained model, i.e., the predicted value at the next time step, is... The formula for calculating RMSE loss is as follows:

[0026]

[0027] L m This represents achieving the optimization objective of reducing the prediction differences among different batteries; the constructed L m This method measures the difference between two types of batteries in the hidden layer of a neural network, achieved using the Mean-Maximum Difference Loss (MMD) approach; its calculation formula is as follows:

[0028]

[0029] Where K represents the kernel set; P and Q represent the source and target domain data in the last layer of the neural network, respectively; MMDk(P,Q) represents the MMD value between hidden layers P and Q using kernel k;

[0030] For the calculation of the MMD value, it is defined as the squared distance between the kernel embeddings of the marginal distribution in the regenerated Hilbert kernel space RKHS, expressed as:

[0031]

[0032] Where Hk Let N represent the regenerated Hilbert kernel space assigned to the feature kernel k; simultaneously, different methods of calculating the MMD loss will also have some impact; this is because different kernels embed probability distributions in different RKHS, where different RKHS emphasize sufficient statistics of different orders; for this reason, multi-kernel MMD loss is used to evaluate the differences between domains; k The Gaussian kernel function is represented as follows:

[0033]

[0034] Where k σi Indicates bandwidth σ i The Gaussian kernel below;

[0035] Using the total loss function constructed with Equation 6, the gradients of the parameters in each layer of the neural network are calculated and updated using the gradient values. This process is expressed as follows:

[0036]

[0037]

[0038] Where w and b represent the parameter values ​​in the weight matrix of each layer of the neural network, respectively; the gradient of the loss with respect to the weights is realized in Equations 11 and 12 by taking the partial derivative of the loss function with respect to the weight values; and lr represents the learning rate.

[0039] The training process of a neural network relies on multiple training rounds. When the number of training rounds has not reached the set number, after updating the parameters through backpropagation, the neural network will continue to calculate the updated value of the objective function L according to Formula 6, and continue to calculate the gradient and backpropagation process until the set number of training rounds is reached, that is, the training is completed and the final prediction model is obtained. With the help of the above process, it is possible to predict the cross-domain health status indicators of the battery.

[0040] The beneficial effects of this invention are:

[0041] The method proposed in this invention requires no prior knowledge and possesses advantages such as low complexity and strong robustness. With the continuous improvement of computing power in recent years, data-driven methods have been widely applied in practice. The optimization objective based on multi-core MMD loss can effectively reflect the differences between different battery types, and can compensate for these differences during the optimization process, thereby obtaining a predictive model that simultaneously meets multiple battery health state indicators. Attached Figure Description

[0042] Figure 1 This is a flowchart of a deep learning cross-domain prediction method for the health status of lithium batteries in new energy aircraft according to the present invention.

[0043] Figure 2 The neural network architecture diagram constructed for this invention;

[0044] Figure 3 This is a graph showing the prediction results of the target domain battery SOH for the technical solution of this invention;

[0045] Figure 4 This is a t-SNE diagram characterizing the domain fusion effect of the technical solution of the present invention. Detailed Implementation

[0046] To better explain and facilitate understanding of the present invention, the technical solution and effects of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] This invention uses the publicly available lithium-ion battery dataset from the University of Maryland's CALCE as an example. It contains data on two types of batteries: the CS2 prismatic lithium-ion battery and the CX2 prismatic lithium-ion battery.

[0048] Step 1: Collect discharge current data during battery cycling, extract all SOH data of the battery and arrange them in chronological order to obtain the SOH time series of the battery, which reflects the health status of the battery;

[0049] Specifically, the State of Health (SOH) defined by the battery's capacity reflects the battery's health status; it uses a percentage to represent the battery's state from the beginning to the end of its lifespan, and its expression is:

[0050]

[0051] Among them, C t Let Ct be the capacity during the t-th charge-discharge cycle, and C0 be the battery's rated capacity. The rated capacity of the battery is normalized using the State of Harm (SOH) index to avoid the potential impact of differences in rated capacity between different battery types on the prediction results. For the capacity Ct during the discharge cycle... t It is obtained by integrating the discharge current over time; therefore, by collecting the current and time information of the battery discharge process, the SOH value of this cycle state is obtained; and by arranging all the SOH values ​​of the battery in chronological order, the SOH time series of the battery is obtained.

[0052] Step 2: Determine the source and target domains of the batteries; one type of battery containing prediction result labels is classified as the source domain data for training. The SOH data of this type of battery will be used as the input and output of the neural network architecture and as the basis for training; another type of battery without prediction result labels is used as the target domain for verification and as the basis for the neural network to learn the features of the target domain.

[0053] Step 3: Using convolutional neural networks as the prediction model framework, solve the problem of predicting health indicators across different areas of the battery. The basic framework is described below:

[0054] First, the historical time series of battery SOH extracted in step 1 is used as the input to the neural network; this time series is represented as x = [X1, X2, ..., X...]. N ], where N represents the length of the input time series; and the convolution operation is performed by a sequence of length F. L The filtering kernel w completes the process, extracting a length F from the time series x in chronological order. L The time series is represented as:

[0055]

[0056] The process of performing a convolution operation on a time series is as follows:

[0057]

[0058] Where *T represents the transpose of matrix *, and W and b represent the weight term and the bias term, respectively; For non-linear activation functions, ReLU and Leaky ReLU can be chosen. This invention chooses Leaky ReLU as the activation function, expressed as:

[0059]

[0060] Compared to the ReLU function, the Leaky ReLU function solves the zero gradient problem for negative values ​​by giving a very small linear component to the negative input; usually, α is a small value; in this embodiment, the value of α is 0.01.

[0061] Slide the filtering window from the first point to the last point of the Soh time series sample data participating in the convolution operation to obtain the feature map of the j-th filter, represented as:

[0062]

[0063] To further increase the number of layers in the neural network, the output z j It will continue to be transmitted as input to the next layer of the network until the final network output value y is obtained, which represents the SOH value of the battery in a future charge-discharge cycle test;

[0064] Figure 1This paper demonstrates the workflow of the prediction model in this invention. First, the extracted lithium-ion battery health index data needs to be divided into source and target domains. For the source domain data, four CS2 battery data sets (CS2 35, CS2 36, CS2 37, and CS2 38) are selected as training objects for model training. For the target domain data, four CX2 battery data sets (CX2 34, CX2 36, CX2 37, and CX2 38) are selected as training objects for model testing.

[0065] Figure 2 This demonstrates the neural network architecture built in this embodiment. First, the network input contains a time series of SOH values ​​from 30 charge-discharge cycles. The first convolutional kernel size is set to 3, and the stride is also set to 3, with no padding. After the convolution operation, the first hidden layer dimension expands from the initial 1D data to 10D, with a length of 16. To maintain the same size after the hidden layer, starting from the second hidden layer, the stride of the convolutional kernel is adjusted to 1, using a zero-padding strategy, and the kernel size is set to 3. Thus, the dimensions of the five hidden layers in the network architecture remain unchanged, and the computation continues forward through fully connected layers. Subsequently, at the 5th convolutional layer, this hidden layer is flattened to obtain a new 1D hidden layer with a size of 160. After dimensionality reduction through fully connected layers, its length is reduced to 32. The output result, the battery SOH value for the next cycle, is obtained after the final fully connected layer.

[0066] For the optimizer selection, this embodiment uses the Adam optimizer as a common approach, with a learning rate set to 0.001 and 500 training epochs. This embodiment also uses mini-batch processing to reduce the computational load per training session. In this embodiment, 128 batches of data samples are processed per training session for the network.

[0067] Step 4: Perform backpropagation calculation on the neural network and calculate its loss function. Evaluate the neural network, complete the training of the neural network framework, obtain the prediction model, and realize the prediction of cross-domain health status indicators of the battery.

[0068] For the backpropagation process, by setting a specific objective function L, the gradient values ​​obtained from the weight and bias terms in Equation 3 are used to update the weights and bias terms. To solve the prediction of the battery cross-domain SOH index, the objective function L is constructed as follows, which includes two indices L... P and L m The two are adjusted using a weight value l, as follows:

[0069] L = L p +l×L m (6)

[0070] Where L P The accuracy metric for prediction is denoted by the root mean square loss function (RMSE). Assuming the true value at the next time step is y, the output value of the trained model, i.e., the predicted value at the next time step, is... The formula for calculating RMSE loss is as follows:

[0071]

[0072] L m This represents achieving the optimization objective of reducing the prediction differences among different batteries; the constructed L m This method measures the difference between two types of batteries in the hidden layer of a neural network, achieved using the Mean-Maximum Difference Loss (MMD) approach; its calculation formula is as follows:

[0073]

[0074] Where K represents the kernel set; P and Q represent the source and target domain data in the last layer of the neural network, respectively; MMDk(P,Q) represents the MMD value between hidden layers P and Q using kernel k;

[0075] For the calculation of the MMD value, it is defined as the squared distance between the kernel embeddings of the marginal distribution in the regenerated Hilbert kernel space RKHS, expressed as:

[0076]

[0077] Where H k Let N represent the regenerated Hilbert kernel space assigned to the feature kernel k; simultaneously, different methods of calculating the MMD loss will also have some impact; this is because different kernels embed probability distributions in different RKHS, where different RKHS emphasize sufficient statistics of different orders; for this reason, multi-kernel MMD loss is used to evaluate the differences between domains; k The Gaussian kernel function is represented as follows:

[0078]

[0079] in Indicates bandwidth σ i The Gaussian kernel is used; the number of kernels selected in the implementation of this patent is 8.

[0080] Using the total loss function constructed with Equation 6, the gradients of the parameters in each layer of the neural network are calculated and updated using the gradient values. This process is expressed as follows:

[0081]

[0082]

[0083] Where w and b represent the parameter values ​​in the weight matrix of each layer of the neural network, respectively; the gradient of the loss with respect to the weights is realized by taking the partial derivative of the loss function with respect to the weight values ​​in Equations 11 and 12, respectively; lr represents the learning rate, and the learning rate value selected in this patent implementation is 0.001;

[0084] The training process of a neural network relies on multiple training rounds. When the number of training rounds has not reached the set number, after updating the parameters through backpropagation, the neural network will continue to calculate the updated value of the objective function L according to Formula 6, and continue to calculate the gradient and backpropagation process until the set number of training rounds is reached, that is, the training is completed and the final prediction model is obtained. With the help of the above process, it is possible to predict the cross-domain health status indicators of the battery.

[0085] To demonstrate the advantages of the method proposed in this invention, this embodiment sets the weight parameter l in Equation 6; experiments are conducted using six values: 0, 1, 5, 10, 20, and 40. When the weight parameter is set to 0, L... m The term "not functioning" represents the traditional deep learning prediction method. As the value of the weight parameter l increases, the difference between the source and target domain batteries gradually decreases with the increase of the number of learning epochs. Figure 3 The prediction results under these six weighting indices are presented. The method proposed in this invention uses the RMSE loss defined in Formula 7 to quantitatively demonstrate the prediction effect on the four cells in the target domain.

[0086] Because the source and target domains differ significantly in the later stages, the health indicators of the battery's cycle tests will be divided into three segments—short-term, medium-term, and long-term—when evaluating the predictive performance of the proposed method. For CX2 34, CX2 37, and CX2 38, the short-term segment is defined as having 500 or fewer cycles, the medium-term segment as having more than 500 but less than 1000 cycles, and the long-term segment as having more than 1000 but less than 1500 cycles. For the CX236 battery, due to its shorter time series, the evaluation range is defined as follows: short-term segment as having 500 or fewer cycles, medium-term segment as having more than 500 but less than 1000 cycles, and long-term segment as having more than 1000 but less than 1500 cycles.

[0087] The table below lists the prediction results of the proposed method for four target domain batteries within the defined range.

[0088]

[0089] The results show that as the value of l in Equation 6 increases, the RMSE error of the model prediction gradually decreases. This demonstrates that the method proposed in this invention significantly improves the accuracy of long-term prediction indicators. In the short and medium term, the prediction effects under different weights l are not significantly different, which also reflects that the machine learning method proposed in this invention can effectively improve the prediction accuracy of battery health indicators in the target domain in the medium and long term.

[0090] To visualize the prediction effect of the method proposed in this invention, Figure 4 The t-SNE method is used to describe the difference between the source and target domains. This method can reduce high-dimensional data to a two-dimensional space; the degree of overlap between the source and target domains can qualitatively reflect the difference between them.

[0091] Figure 4 This demonstrates the degree of fusion between the two domains after 500 rounds of training in transfer learning. Figure 4 In the diagram, a black "1" represents the source domain, and a gray "2" represents the target domain. When the weight value 1 is low, the overlap between the source and target domains is small, indicating a significant difference between the two domains. However, when the weight value is set to 10, partial overlap begins to appear between the two domains. As the degree of transfer learning continues to increase, the overlap between the two domains gradually increases until they almost completely overlap. This demonstrates that the method proposed in this invention can compensate for the differences between the two types of batteries.

Claims

1. A deep learning cross-domain prediction method for the health status of lithium batteries in new energy aircraft, characterized in that, Specifically, the following steps are included: Step 1: Collect discharge current data during battery cycling, extract all SOH data of the battery and arrange them in chronological order to obtain the SOH time series of the battery, which reflects the health status of the battery; Step 2: Determine the source and target domains of the batteries; one type of battery containing prediction result labels is classified as the source domain data for training. The SOH data of this type of battery will be used as the input and output of the neural network architecture and as the basis for training; another type of battery without prediction result labels is used as the target domain for verification and as the basis for the neural network to learn the features of the target domain. Step 3: Using convolutional neural networks as the framework for prediction models, solve the problem of predicting health indicators across different areas of the battery. Step 4: Perform backpropagation calculation on the neural network and calculate its loss function. Evaluate the neural network, complete the training of the neural network framework, obtain the prediction model, and realize the prediction of cross-domain health status indicators of the battery. The framework described in step 3 is as follows: First, a historical time series of battery SOH extracted in step 1 is used as the input to the neural network; this time series is represented as x = [ , , ..., ], where N represents the length of the input time series; and the convolution operation is performed by a sequence of lengths of . The filtering kernel w is completed, and a length of [value] is extracted from the time series x in chronological order. The time series is represented as: （2） The process of performing a convolution operation on a time series is as follows: （3） Where *T represents the transpose of matrix *, W and b represent the weight and bias terms, respectively; φ is the non-linear activation function, Leaky ReLU is chosen as the activation function, expressed as: （4） Compared to the ReLU function, the Leaky ReLU function solves the zero gradient problem for negative values ​​by giving a very small linear component to the negative input; typically, α is a small value. Slide the filter window from the first point to the last point of the SOH time series sample data participating in the convolution operation to obtain the feature map of the j-th filter, represented as: （5） To further deepen the neural network layers, the output... It will continue to be transmitted as input to the next layer of the network until the final network output value y is obtained, which represents the SOH value of the battery in a future charge-discharge cycle test.

2. The deep learning cross-domain prediction method for the health status of lithium batteries in new energy aircraft according to claim 1, characterized in that, Step 1 is as follows: The State of Health (SOH) defined by the battery's capacity reflects the battery's health status; it expresses the battery's state from the beginning to the end of its lifespan as a percentage, and its expression is: （1） in, Let be the capacity during the t-th charge-discharge cycle. The rated capacity of the battery is used; the rated capacity of the battery is normalized using the SOH index to avoid the potential impact of differences in the rated capacity of different types of batteries on the prediction results; the capacity for the t-th cycle of discharge is... It is obtained by integrating the discharge current over time; therefore, by collecting the current and time information of the battery discharge process, the SOH value of this cycle state is obtained; and by arranging all the SOH values ​​of the battery in chronological order, the SOH time series of the battery is obtained.

3. The deep learning cross-domain prediction method for the health status of lithium batteries in new energy aircraft according to claim 1, characterized in that, Step 4 specifically involves: For the backpropagation process, by setting a specific objective function L, the gradient values ​​obtained from the weight and bias terms in formula (3) are used to update the weights and bias terms; in order to solve the prediction of the battery cross-domain SOH index, the objective function L is constructed as follows, which includes two indices respectively and The two are adjusted using a weight value l, as follows: （6） in The accuracy metric for prediction is denoted by the root mean square loss function (RMSE). Assuming the true value at the next time step is y, the output value of the trained model, i.e., the predicted value at the next time step, is... Then, the formula for calculating RMSE loss is expressed as: （7） This represents achieving the optimization objective of reducing the prediction differences among different batteries; the constructed This method measures the difference between two types of batteries in the hidden layer of a neural network, achieved using the Mean-Maximum Difference Loss (MMD) approach; its calculation formula is as follows: （8） Where K represents the kernel set; P and Q represent the source and target domain data in the last layer of the neural network, respectively; MMDk(P,Q) represents the MMD value between hidden layers P and Q using kernel k; For the calculation of the MMD value, it is defined as the squared distance between the kernel embeddings of the marginal distribution in the regenerated Hilbert kernel space RKHS, expressed as: MMDk(P,Q) = (9) in Let k represent the regenerated Hilbert kernel space assigned to the feature kernel k; at the same time, different methods of calculating MMD loss will also have some impact; this is because different kernels embed probability distributions in different RKHS, where different RKHS emphasize sufficient statistics of different orders; for this reason, multi-kernel MMD loss is used to evaluate the differences between domains. The Gaussian kernel function is represented as follows: （10） in Indicates bandwidth The Gaussian kernel below; Using the total loss function constructed with formula (6), the gradients of the parameters in each layer of the neural network are calculated and updated using the gradient values. This process is expressed as follows: （11） （12） Where w and b represent the parameter values ​​in the weight matrix of each layer of the neural network, respectively; the gradient of the loss with respect to the weights is realized in Equations 11 and 12 by taking the partial derivative of the loss function with respect to the weight values; and lr represents the learning rate. The training process of the neural network relies on multiple training rounds. When the number of training rounds has not reached the set number of training rounds, after backpropagation to update the parameters, the neural network will continue to calculate the updated value of the objective function L according to formula (6), and continue to calculate the gradient and backpropagation process until the set number of training rounds is reached, that is, the training is completed and the final prediction model is obtained. With the help of the above process, the prediction of the cross-domain health status index of the battery can be realized.

Images