Lithium ion battery SOH estimation system and method based on automatic neural architecture search

Abstract

The invention discloses a lithium ion battery SOH estimation system and method based on automatic neural architecture search, and the method comprises the steps: analyzing a trade-off relation between the estimation accuracy and the calculation complexity of different network architectures through an image coding module, a neural architecture search module and a decision and model selection module according to a Pareto optimal solution set obtained through search, and carrying out the estimation of the SOH of a lithium ion battery through the analysis of the trade-off relation between the estimation accuracy and the calculation complexity; and selecting a proper network architecture as a final SOH estimation model. The automatic design of the SOH estimation neural network architecture is realized, the processes of manual intervention and expert trial and error are reduced, and the efficiency and accuracy of network architecture optimization are improved. According to the method, the relationship between the two can be balanced according to actual application requirements, the model meeting different computing resource limitations is obtained, the applicability and flexibility of the model are improved, the computing resources are optimized, and the method can better adapt to different application scenes and requirements.

Description

Technical Field

[0001] The present invention relates to a battery management system, and in particular to a lithium-ion battery SOH estimation system and method based on automatic neural architecture search. Background Art

[0002] Research on lithium-ion battery state of health estimation plays a vital role in ensuring battery safety, improving battery performance, reducing maintenance costs, and promoting industrial development. Existing lithium-ion battery state of health (SOH) estimation methods include direct measurement methods, model-based methods, and data-driven methods.

[0003] Direct measurement methods use test equipment to evaluate the SOH value by measuring the physical and electrochemical characteristics of the battery, such as coulomb counting and electrochemical impedance spectroscopy. Although these methods are accurate, they require a lot of time to reach a stable battery state, especially when performing capacity testing, which requires a complete charge and discharge cycle or extensive spectrum analysis, which may not be practical in actual applications.

[0004] Model-based approaches estimate the battery's state of health and degradation using physical or empirical models of the battery. For example, equivalent circuit models use circuit theory to calculate battery parameters such as internal resistance and open-circuit voltage, thereby assessing the battery's state of health (SOH). Empirical models use extensive experimental data and empirical formulas to establish a relationship between battery SOH and observable variables (such as battery capacity, internal resistance, and voltage). Kalman filtering methods use the battery's state equation and observation equation as inputs to a Kalman filter, utilizing a filtering algorithm to estimate the battery's SOH in real time. Example filtering methods treat the battery's SOH as a random variable and use a particle filter to sample and update the battery's state, thereby deriving a probabilistic distribution estimate of the battery's SOH. Model-based approaches cannot accurately capture the battery's advanced nonlinear behavior, resulting in insufficient estimation accuracy. Furthermore, these models typically rely on a precise understanding of the battery's internal physical processes, making it challenging to establish accurate physical models for complex battery systems.

[0005] The data-driven approach uses data obtained from battery measurement variables to establish an SOH estimation model through machine learning and deep learning methods. Common machine learning methods include support vector machines (SVM), Gaussian process regression (GPR), probabilistic neural networks, and ensemble learning frameworks. Deep learning methods such as convolutional neural networks (CNN), recurrent neural networks (RNN), and their variants (such as LSTM) have achieved remarkable success in SOH estimation due to their powerful nonlinear mapping capabilities and automatic feature extraction capabilities. Although deep learning methods perform well in SOH estimation, existing neural network architecture designs often rely on expert knowledge and trial and error, which is not only time-consuming but can also lead to suboptimal network architectures. In addition, when processing battery data, traditional data-driven methods often directly input one-dimensional measurement signals into the neural network, failing to fully utilize the spatiotemporal correlations in the data, limiting the performance of the model.

[0006] In summary, there are many deficiencies in the current methods for estimating the health status of lithium-ion batteries, which affect the ability to capture the health status of lithium-ion batteries and the accuracy of estimation. Summary of the Invention

[0007] The present invention provides a lithium-ion battery SOH estimation system and method based on automatic neural architecture search. By using image coding technology and the non-dominated sorting genetic algorithm-neural architecture search (NSGA-Net) algorithm, the system realizes the automated design and optimization of the neural network architecture for lithium-ion battery SOH estimation, solves the problem of inefficient network architecture optimization, effectively utilizes the spatiotemporal correlation in data, flexibly selects the appropriate model according to different application scenarios and computing resource constraints, and improves the model's ability to capture battery degradation characteristics and estimation accuracy.

[0008] The present invention achieves the above-mentioned object through the following technical solutions: A lithium-ion battery SOH estimation system and method based on automatic neural architecture search, the steps of which are as follows: 1) Image encoding module: Data preprocessing is performed on the raw voltage of lithium-ion batteries. The collected battery charging voltage curves are standardized, and the piecewise cluster approximation (PAA) method is used to unify voltage sequences of different lengths into a specified length. Data scaling is then used to map the voltage data to the range [-1, 1]. Image encoding is then performed, and the preprocessed voltage data is converted to polar coordinates. The sine and cosine values ​​between time series points are calculated to construct a Gramian Angular Field (GAF) image. The Gramian Angular Summation Field (GASF) is selected as the final image encoding method to better analyze signals with periodic characteristics. 2) Neural Architecture Search Module: First, a multi-objective optimization problem is modeled. The neural network architecture design problem for SOH estimation is modeled as a multi-objective optimization problem, considering the two objectives of estimation accuracy and computational complexity. Two optimization objective functions are defined: minimizing the mean squared error (MSE) to improve estimation accuracy and minimizing the number of floating-point operations (FLOPs) to reduce computational complexity. Then, the NSGA-Net algorithm is used for neural architecture search, encoding the neural network architecture as a genotype representation. Through crossover, mutation, and selection operations in the genetic algorithm, a Pareto optimal solution set that satisfies multi-objective optimization is searched. Finally, the network architecture is encoded and decoded, representing the neural network architecture as a combination of computational blocks, each consisting of multiple nodes (such as convolution, pooling, or batch normalization operations). The connection relationship of the network architecture is encoded by binary strings to achieve genotype representation and decoding of the network architecture. 3) Decision-making and model selection module: First, we analyze the Pareto optimal solution set. Based on the Pareto optimal solution set obtained by the search, we analyze the trade-off between estimation accuracy and computational complexity of different network architectures. Then, we perform model selection and choose the appropriate network architecture as the final SOH estimation model based on the application scenario and computing resource constraints. We provide three typical solutions, including the model with the best accuracy, the model with the lowest computational complexity, and a compromise model, to meet the needs of different situations.

[0009] Furthermore, the plurality of nodes include convolution, pooling or batch normalization operations.

[0010] The present invention realizes the automated design of the neural network architecture for SOH estimation, reduces manual intervention and expert trial and error, and improves the efficiency and accuracy of network architecture optimization. The image coding method is used to capture the spatiotemporal correlation in battery data, enriching the feature representation and helping to improve the accuracy of SOH estimation. It can balance the relationship between the two according to actual application requirements to obtain a model that meets different computing resource constraints, improve the applicability and flexibility of the model, and achieve computing resource optimization. A variety of non-dominated solutions are provided for selection, covering different models from high precision to low complexity, providing decision makers with more choices and being able to better adapt to different application scenarios and needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is the basic framework diagram of the present invention; Figure 2 is an example diagram of a neural architecture search encoding according to an embodiment of the present invention; Figure 3a is a graph showing the SOH estimation results of the Oxford data set of Battery 1 according to an embodiment of the present invention; Figure 3b is a graph showing the SOH estimation results of the Oxford data set of Battery 2 according to an embodiment of the present invention; Figure 3c is a graph showing the SOH estimation results of the Oxford data set of Battery 3 according to an embodiment of the present invention; Figure 3d 4 is a graph showing the SOH estimation results of the Oxford data set of battery 7 in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] The present invention is described in detail below with reference to the examples. Figures 1 to 3d , a lithium-ion battery SOH estimation system and method based on automatic neural architecture search, the steps are as follows: 1) Image encoding module: Data preprocessing is performed on the raw voltage of lithium-ion batteries. The collected battery charging voltage curves are standardized, and the piecewise cluster approximation (PAA) method is used to unify voltage sequences of different lengths into a specified length. Data scaling is then used to map the voltage data to the range [-1, 1]. Image encoding is then performed, and the preprocessed voltage data is converted to polar coordinates. The sine and cosine values ​​between time series points are calculated to construct a Gramian Angular Field (GAF) image. The Gramian Angular Summation Field (GASF) is selected as the final image encoding method to better analyze signals with periodic characteristics. 2) Neural Architecture Search Module: First, a multi-objective optimization problem is modeled. The neural network architecture design problem for SOH estimation is modeled as a multi-objective optimization problem, considering the two objectives of estimation accuracy and computational complexity. Two optimization objective functions are defined: minimizing the mean squared error (MSE) to improve estimation accuracy and minimizing the number of floating-point operations (FLOPs) to reduce computational complexity. Then, the NSGA-Net algorithm is used for neural architecture search, encoding the neural network architecture as a genotype representation. Through crossover, mutation, and selection operations in the genetic algorithm, a Pareto optimal solution set that satisfies multi-objective optimization is searched. Finally, the network architecture is encoded and decoded, representing the neural network architecture as a combination of computational blocks, each of which consists of multiple nodes (such as convolution, pooling, or batch normalization operations). The connection relationship of the network architecture is encoded by binary strings to achieve genotype representation and decoding of the network architecture. 3) Decision-making and model selection module: First, we analyze the Pareto optimal solution set. Based on the Pareto optimal solution set, we analyze the trade-off between estimation accuracy and computational complexity for different network architectures. Next, we conduct model selection, choosing the appropriate network architecture as the final SOH estimation model based on the application scenario and computational resource constraints. We provide three typical solutions: the model with the best accuracy, the model with the lowest computational complexity, and a compromise model to meet the needs of different scenarios.

[0013] Example: The following section presents a detailed SOH estimation scheme using the Oxford University public lithium-ion battery dataset. It consists of five main sections: the overall framework of the proposed method, image encoding based on angle fields, the use of NSGA-Net for neural network architecture search, experimental setup, and experimental results.

[0014] 1. The overall framework of the method proposed in the present invention: In order to illustrate the method proposed by the present invention, Figure 1 The proposed method follows the standard process of supervised learning methods. The black arrows in the figure represent the basic steps in each stage, while the gray arrows, which start with a dot and end with an arrow, illustrate the data dependencies between the different steps. The upper part of the figure outlines the basic training process for the SOH estimation task. The lower part, on the other hand, provides the prediction process for new battery data.

[0015] For the training process, it mainly consists of three basic sub-models: data preparation, image encoding, and neural architecture search. First, the dataset for supervised learning is derived from the aging experiment to develop a basic deep learning method to map the measured variables to the corresponding SOH. In this study, voltage is used as the main input, and the related capacity is used to calculate the SOH as the ground truth for training. Details of the specific datasets used in this study, including open source datasets and private experimental datasets, can be obtained in Section 4.1 "Experimental Data" of this paper. The voltage data is first encoded into an image for subsequent steps. The image encoding method and corresponding analysis of the characteristics of this method are provided in Section 2. Next, the neural architecture search for SOH estimation is elaborated in detail in Section 3. For practical applications, three models have been selected as outputs for various cases.

[0016] exist Figure 1 During the prediction process depicted in the lower part of the figure, new voltage data can be obtained from the BMS. Subsequently, the decision maker can select the model that best suits the current SOH estimation. For example, if cloud-based computing is available, which may alleviate computational issues, the accurate best solution can be selected to provide more precise results. On the other hand, if the application of SOH is still limited by computing power, the Flops-Best solution can be considered the first choice. In addition to this, the output of NAS also provides a trade-off solution, which represents the best trade-off between accuracy and computational efficiency. It is worth mentioning that if there are still some labels available in the target domain, they can be used to fine-tune the existing model to further improve performance.

[0017] 2. Image coding based on angle field: In this section, we introduce the method of encoding raw measurement variables into images. We first discuss the data preprocessing steps in detail, and then introduce the encoding process in detail.

[0018] 2.1 Data preprocessing: In this study, the charging voltage curve is selected as the input for SOH estimation because it is more feasible to obtain a complete charging curve in practical applications compared to the discharge process. However, considering that various charging variables may exhibit different cycles, it is necessary to first normalize the different charging curves to a unified dimension before utilizing the data. To this end, the piecewise aggregation approximation method (PAA) is used to unify the data length of different voltage sequences. The PAA method can be regarded as a dimensionality reduction method that converts the original sequence of length m into a user-specified length n. Specifically, for the charging voltage curve v = {v1,. . . , vm}, it can be converted into a new sequence v¯ = { v ¯1 , . . . , v ¯ n}, according to the following equation: (1); Where: n < m; j is the index of the element in the corresponding interval in the original sequence used to calculate each new element.

[0019] Equation (1) clearly illustrates that the basic idea of ​​this dimensionality reduction method is to first divide the original curve into segments of size n and then sample the data in each frame. Subsequently, the most representative data of the reduced size are retained. This method is used to normalize each charging curve of length n with various cyclic variations aligned.

[0020] The next step in data advancement is to scale the data to a uniform range to mitigate the impact of data amplitude. Assuming a uniform voltage of v = { v ¯1 , . . . , v ¯ n}, the corresponding scaling data in the range of [-1, 1] can be obtained by the following formula: (2);

[0021] 2.2 Encoding images by coordinate transformation: Next, the pre-processed voltage data is reclassified into a polar representation: (3); Where: and r correspond to the radial direction and radius respectively. i N is a natural number, and W is a constant factor used to adjust the span of the polar coordinate system. Subsequently, the sine and cosine values ​​between the time series points are calculated and integrated to construct the Gramian Angular Field (GAF) matrix. When comparing two typical forms of the proposed GAF—Gramian Angular Difference (GADF) and Gramian Angular Summation (GASF)—GASF was found to be more suitable for analyzing signals with periodic characteristics. Therefore, GASF was selected as the final encoding method to obtain the encoded image. More specifically, the GASF operator can be defined as follows: (4); And its corresponding matrix table can be expressed as follows: (5); Finally, the obtained encoded image is considered as input for the following architecture selection.

[0022] 3. NSGA-Net for neural network architecture search: By now, we have encoded images, and we know that they may contain richer features than the original time series data. Next, we will explore how to utilize encoded images for subsequent steps in SOH estimation. Following the basic process of data-driven methods, the following steps in SOH estimation are feature extraction and estimator training. Given that deep learning methods can automatically extract features from both shallow and deep layers, it is possible to combine feature extraction and learning in a single framework, utilizing its different layers to simultaneously handle both tasks. Therefore, the core challenge shifts from designing feature extraction methods and estimators to designing the architecture of the task-specific deep learning network. To achieve the desired estimation results, different types of deep neural networks are typically handcrafted through trial and error. Therefore, based on the theoretical mechanisms of deep neural networks and previous empirical works, it can be concluded that the architectural design of deep neural networks will be a key task in the subsequent steps.

[0023] The structure of deep neural networks is crucial for their effectiveness in SOH estimation. It is desirable to automatically construct an appropriate network architecture for the available data. To this end, in this work, NSGA-Net is used to propose a neural network architecture for the SOH estimation task. Using this approach, the architecture search problem is formulated as a multi-objective optimization problem, and a non-dominant sorting genetic algorithm is used to determine the Pareto set for this search problem.

[0024] In NSGA-Net, we first transform the architecture search problem into a multi-objective problem, where two objectives are considered simultaneously. First, it is natural to consider estimation accuracy as a single objective. Given a search space A, a battery dataset D, and a training method P, the first goal is to find the architecture A∈A that achieves the lowest estimation loss.

[0025] This goal can be expressed mathematically as follows: (6); MSE(a) represents the test loss between the estimated SOH and its ground truth, which is defined as: (7); Among them, D is the entire data set, Dt represents the test data set, is the output value of Dt predicted by the neural network constructed by architecture solution a, Iv is the encoded image data, and SOHi represents the i-th true value of the test dataset.

[0026] In addition to estimation accuracy, computational complexity must be added back as a separate objective because the implementation of SOH estimation methods is still limited by computational power. It is not feasible to only emphasize estimation accuracy while ignoring computational cost. In NSGA-Net, inference time and the number of floating-point operations (FLOPS) are considered as proxies for evaluating the computational complexity of a particular network. Compared to other metrics, FLOPS directly quantifies the number of floating-point operations required by the neural network, providing a concrete measure of the computational work required for execution. Therefore, the second goal of minimizing computational complexity can be expressed as follows: (8); The architecture search problem has been transformed into a multi-objective optimization problem and awaits a solution. Before addressing this problem with multi-objective optimization algorithms, it is necessary to consider encoding methods that define the search space for all potential solutions to the neural architecture. The NSGA-Net used in this work considers CNN architectures and their variations as a combination of computational blocks that specify layer computations. The basic computational unit in NSGA-Net is defined as a node, which may be a single operation such as convolution, pooling, or batch processing. Several nodes form a computational block, called a stage. Ultimately, designing a neural network architecture involves organizing several stages and establishing connections between different nodes. To incorporate this approach into an evolutionary search framework, it is necessary to establish a mapping between phenotypes, which represent neural network genotypes, and encoding representations of solutions in the context of architecture search.

[0027] For a neural network containing the np stage, its compact genotype can express the mutation operation in the evolutionary method as x= [x 1 , x 2 , . . . , x n p ] Each stage encodes a directed acyclic structure represented by a binary string, which includes no operation nodes, where bits, encoding connections. Assume that nodes and their connections can be represented as G=(V,E) ,in V = {1, 2, . . . , n o } is a node. If the first bit is one, it indicates that there is an edge connection between the first node and the second node, that is, E ={(1, 2)} The next two bits indicate the connection E ={(i,3)|i∈{1,2}} This process continues until the final Bits are used to indicate connections E ={(i, no)|i∈{1,...,n o }} .

[0028] An example is Figure 2 To illustrate the encoding process, the left side of the image represents a portion of the neural architecture that can accommodate up to six nodes, denoted as V={1, 2,...,6}. The right half of the image represents the corresponding genotype encoded by this architecture. Decoding the genotype into a matching architecture requires five steps. The first bit is 1, indicating that the first node must be connected to the second node, that is, E={(1, 2)}. The next two bits are 00, indicating that there is no node linked to the third node. The subsequent code 010 indicates that the fourth node is only connected to the second node, thus establishing E ={(1,2),(2,4)}. This process continues until the last five bits are used to represent the connection from the sixth node to the remaining nodes. The ultimate connection between these six nodes can be decoded as E = {(1 , 2) , (2 , 4) , (1 , 5) , (4 , 6) , (5 , 6)}.

[0029] From the above description and examples, it is clear that although the search space mainly refers to the connections between various operators, the selection of operators is also considered. By using this encoding strategy, some operators represented as nodes can be disabled if other nodes cannot be connected to the current node. For example, since there are no other connected nodes, Figure 2 The third node in will be discarded. Essentially, the constructed search space consists of several operator choices and potential connection configurations.

[0030] 4. Experimental setup: 4.1 Experimental data: Oxford Dataset: Four Kokam Pouch lithium-ion cells from the University of Oxford were selected for validation. The cells' cathode materials consisted of a mixture of 11a cobalt carbon oxide and nickel carbon cobalt oxide, while the anode material was graphite. Initially, the cells underwent a characteristic test phase, in which all cells were charged and discharged at 1C. Subsequently, during the aging experiment phase, when the voltage reached 4.2V, the cells were charged at 2C using a constant current and constant voltage (CC-CV) model. During the discharge phase, the process was conducted using a specific profile to simulate realistic operating conditions, ending when the voltage dropped to 2.7V.

[0031] 4.2 Peer method: An overview of peer methods and evolutionary models is given below. We strive to expand the diversity of selected methods to provide more comprehensive experimental results.

[0032] Resnet: This model can be considered the most complex for SOH estimation. The original Resnet has 34 layers and implements a large number of shortcut connections to avoid the vanishing / exploding gradient problem. Unlike VGG, the Resnet used here was trained from scratch using the training dataset.

[0033] In addition to the typical models used for output models, three different types of models are intentionally selected from non-dominant solutions for better comparison. These models cover requirements in different fields and their overview is as follows: Best solution for accuracy: This solution (specifically for sampling with the best accuracy) still maintains a non-dominant relationship with the other solutions. It achieves the best accuracy among the other architectures, making it very valuable to observe its specific structure and try to summarize the design principles of SOH estimation.

[0034] FLOPS Best Solution: This is also a non-dominant solution but has the lowest computational complexity. This solution can be considered as the simplest architecture that retains the essential elements for efficient estimation.

[0035] Trade-off solutions: We specifically sample a solution that strikes a balance between the best accuracy and the optimal solution. In the present work, the knee-point solution on the Pareto front is considered to be the optimal solution that represents the best trade-off between accuracy and computational complexity.

[0036] Notably, the hyperparameters of each competitor are carefully tuned, ensuring that each competitor represents its best performance, thus ensuring a fair comparison.

[0037] 5. Experimental results Each validation experiment adopts the residual cross-validation method to evaluate the performance of the proposed method and its competitors. The data from one cell is used as the testing dataset, while the data from the remaining cells are used for training and optimization.

[0038] Figure 3a- Figure 3dEstimation results for Cells 1, 3, 5, and 7, demonstrated on the Oxford dataset, are presented separately. Portions of the results are intentionally exaggerated for better clarity. The thick black line in the figure represents the true SOH value, which serves as a reference for evaluation. The distance of the predicted SOH value from this baseline indicates the accuracy of the estimated value relative to the true SOH. Intuitively, it is clear that the accurate best solution achieves the best performance, as it closely matches the reference. In contrast, the ResNet method exhibits the worst performance in the present case, suggesting that the residual structure of ResNet may not be conducive to predicting degradation trajectories that steadily decrease. Beyond the best and worst-case scenarios, it is noteworthy that the best approach also achieves satisfactory results. In most test cases, the performance difference between the best and trade-off solutions is minimal. Based on these observations and considering the characteristics of the present dataset, an overall conclusion is that for SOH estimation, there is a consistent degradation trend, favoring models with simpler architectures.

[0039] Table 1 lists the specific evaluation results using different metrics, providing a detailed comparison of the various methods. Clearly, the best-accuracy solution outperforms its peers across all tested units. However, it is also clear that the performance gap between the best-accuracy and trade-off solutions is small, suggesting that the relationship between computational resources allocated to improving prediction accuracy is not linear, and that increasing complexity may be necessary at later stages to further improve accuracy.

[0040] Furthermore, for specific cells (e.g., Cell-5 and Cell-7), the LSTM method achieved the worst performance compared to other methods. One potential reason for this is that, despite its ability to manage long-term data, the LSTM may be less effective than other models at extracting complex patterns and features from the data. The optimal solution outperformed other classical models. Therefore, to improve accuracy, it is recommended to design a specific architecture suitable for a given SOH estimation task.

[0041] ; *Best results are highlighted in bold on a grey background.

Claims

1. A lithium-ion battery SOH estimation system and method based on automatic neural architecture search is characterized by the following steps: 1) Image encoding module: Data preprocessing is performed on the raw voltage of lithium-ion batteries. The collected battery charging voltage curves are standardized, and a segmented clustering approximation method is used to unify voltage sequences of different lengths into a specified length. Data scaling is then used to map the voltage data to the range [-1, 1] for image encoding. The preprocessed voltage data is converted to polar coordinates. The sine and cosine values ​​between time series points are calculated to construct a Gram angle field image. The Gram angle sum field is selected as the final image encoding method to better analyze signals with periodic characteristics. 2) Neural Architecture Search Module: First, a multi-objective optimization problem is modeled. The neural network architecture design problem for SOH estimation is modeled as a multi-objective optimization problem, considering the two objectives of estimation accuracy and computational complexity. Two optimization objective functions are defined, namely minimizing the mean square error to improve estimation accuracy and minimizing the number of floating-point operations to reduce computational complexity. Then, the NSGA-Net algorithm is used to search for neural architectures. The neural network architecture is encoded as a genotype representation. Through the crossover, mutation, and selection operations in the genetic algorithm, the Pareto optimal solution set that satisfies multi-objective optimization is searched. Finally, the network architecture is encoded and decoded. The neural network architecture is represented as a combination of computational blocks, each of which consists of multiple nodes. The connection relationship of the network architecture is encoded by binary strings to achieve genotype representation and decoding of the network architecture. 3) Decision-making and model selection module: First, we analyze the Pareto optimal solution set. Based on the Pareto optimal solution set obtained by the search, we analyze the trade-off between estimation accuracy and computational complexity of different network architectures. Then, we perform model selection and choose the appropriate network architecture as the final SOH estimation model based on the application scenario and computing resource constraints. We provide three typical solutions, including the model with the best accuracy, the model with the lowest computational complexity, and a compromise model, to meet the needs of different situations.

2. The lithium-ion battery SOH estimation system and method based on automatic neural architecture search according to claim 1 is characterized in that the multiple nodes include normalization operations on convolution, pooling or batch.

Images