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

Abstract

The application discloses a lithium ion battery SOH estimation system and method based on automatic neural architecture search, through an image coding module, a neural architecture search module, a decision and model selection module, according to a Pareto optimal solution set obtained by searching, the trade-off relationship between estimation accuracy and computational complexity of different network architectures is analyzed, and a suitable network architecture is selected as a final SOH estimation model. The automatic design of the SOH estimation neural network architecture is realized, the process of manual intervention and expert trial and error is reduced, and the efficiency and accuracy of the network architecture optimization are improved. The relationship between the two can be balanced according to the actual application requirements, the model meeting different calculation resource limitations is obtained, the applicability and flexibility of the model are improved, the optimization of the calculation resources is realized, and different application scenarios and requirements can be better adapted to.

Description

Technical Field

[0001] This invention relates to battery management systems, and particularly to a lithium-ion battery SOH estimation system and method based on automatic neural architecture search. Background Technology

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

[0003] Direct measurement methods utilize testing equipment to assess the state of harmonics (SOH) value by measuring the battery's physical and electrochemical properties, such as coulomb counting and electrochemical impedance spectroscopy. While these methods are accurate, they require a significant amount of time to reach a stable battery state, especially during capacity testing, which necessitates a complete charge-discharge cycle or extensive spectral analysis, potentially proving impractical in real-world applications.

[0004] Model-based methods estimate a battery's health and degradation using physical or empirical models. For example, equivalent circuit models use circuit theory to calculate parameters such as internal resistance and open-circuit voltage, thus assessing the battery's state of health (SOH). Empirical models establish the relationship between SOH and observable variables (such as battery capacity, internal resistance, and voltage) based on extensive experimental data and empirical formulas. Kalman filtering methods use the battery's state equation and observation equation as input to a Kalman filter, employing filtering algorithms to estimate the SOH in real time. Particle filtering methods treat the SOH as a random variable, using particle filters to sample and update the battery's state, thereby obtaining a probability distribution estimate of the SOH. However, model-based methods cannot accurately capture the advanced nonlinear behavior of batteries, leading to insufficient estimation accuracy. Furthermore, these models typically rely on a precise understanding of the battery's internal physical processes; for complex battery systems, establishing accurate physical models is challenging.

[0005] Data-driven approaches utilize data from battery measurement variables to build SOH estimation models using machine learning and deep learning methods. Commonly used 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 significant 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 may also lead to suboptimal network architectures. Furthermore, traditional data-driven methods often directly input one-dimensional measurement signals into the neural network when processing battery data, failing to fully utilize the spatiotemporal correlations in the data and limiting model performance.

[0006] In summary, current methods for estimating the state of health of lithium-ion batteries have many shortcomings, which affect the ability to capture and accurately estimate the state of health of lithium-ion batteries. Summary of the Invention

[0007] This invention provides a lithium-ion battery SOH estimation system and method based on automatic neural architecture search. Through image encoding technology and a non-dominated sorting genetic algorithm-neural architecture search (NSGA-Net) algorithm, it achieves automated design and optimization of the neural network architecture for lithium-ion battery SOH estimation, solving the problem of low efficiency in network architecture optimization. It effectively utilizes the spatiotemporal correlation in the data, flexibly selects appropriate models according to different application scenarios and computational resource constraints, and improves the model's ability to capture battery degradation characteristics and its estimation accuracy.

[0008] This invention achieves the above objectives through the following technical solution: 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: The raw voltage data of lithium-ion batteries is preprocessed, and the collected battery charging voltage curves are standardized. The segmented clustering approximation (PAA) method is used to unify voltage sequences of different lengths to a specified length. Then, the voltage data is mapped to the range of [-1,1] through data scaling and image encoding is performed. The preprocessed voltage data is converted into polar coordinate representation. By calculating the sine and cosine values ​​between time series points, a Gramian Angular Field (GAF) image is constructed. 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, transforming the neural network architecture design problem based on SOH estimation into a multi-objective optimization problem, considering both 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, the Pareto optimal solution set satisfying the multi-objective optimization is searched. Finally, the network architecture is encoded and decoded, representing it as a combination of computational blocks, each block consisting of multiple nodes (such as convolution, pooling, or batch normalization operations). The connection relationships of the network architecture are encoded using binary strings, achieving the genotype representation and decoding of the network architecture. 3) Decision and Model Selection Module: First, the Pareto optimal solution set is analyzed. Based on the Pareto optimal solution set obtained by the search, the trade-off between estimation accuracy and computational complexity of different network architectures is analyzed. Then, model selection is carried out. According to the application scenario and computational resource constraints, a suitable network architecture is selected as the final SOH estimation model. Three typical solutions are provided, including the best accuracy model, the model with the lowest computational complexity, and the compromise solution model, to meet the needs of different situations.

[0009] Furthermore, the plurality of nodes includes normalization operations on convolution, pooling, or batching.

[0010] This invention automates the design of neural network architectures for SOH estimation, reducing manual intervention and expert trial-and-error processes, and improving the efficiency and accuracy of network architecture optimization. It utilizes image encoding methods to capture the spatiotemporal correlations in battery data, enriching feature representations and contributing to improved SOH estimation accuracy. It can balance the relationship between these two factors according to practical application requirements, obtaining models that meet different computational resource constraints, improving the model's applicability and flexibility, and optimizing computational resources. It provides a variety of non-dominated solutions to choose from, covering different models from high accuracy to low complexity, offering decision-makers more choices and better adapting to different application scenarios and needs. Attached Figure Description

[0011] Figure 1 This is a basic framework diagram of the present invention; Figure 2 This is an example diagram of neural architecture search encoding in an embodiment of the present invention; Figure 3a This is a curve showing the SOH estimation results of Battery 1 using the Oxford dataset in this embodiment of the invention; Figure 3b This is a curve showing the SOH estimation results of the Oxford dataset for Battery 2 in this embodiment of the invention; Figure 3c This is a curve showing the SOH estimation results of the Oxford dataset for Battery 3 in this embodiment of the invention; Figure 3d This is a graph showing the SOH estimation results of the Oxford dataset for Battery 7 in this embodiment of the invention. Detailed Implementation

[0012] The present invention will be described in detail below with reference to embodiments. See also Figures 1 to 3d A lithium-ion battery SOH estimation system and method based on automatic neural architecture search is described, and its steps are as follows: 1) Image encoding module: The raw voltage data of lithium-ion batteries is preprocessed, and the collected battery charging voltage curves are standardized. The segmented clustering approximation (PAA) method is used to unify voltage sequences of different lengths to a specified length. Then, the voltage data is mapped to the range of [-1,1] through data scaling and image encoding is performed. The preprocessed voltage data is converted into polar coordinate representation. By calculating the sine and cosine values ​​between time series points, a Gramian Angular Field (GAF) image is constructed. 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, transforming the neural network architecture design problem based on SOH estimation into a multi-objective optimization problem, considering both 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, the Pareto optimal solution set satisfying the multi-objective optimization is searched. Finally, the network architecture is encoded and decoded, representing it as a combination of computational blocks, each block consisting of multiple nodes (such as convolution, pooling, or batch normalization operations). The connection relationships of the network architecture are encoded using binary strings, achieving the genotype representation and decoding of the network architecture. 3) Decision and Model Selection Module: First, the Pareto optimal solution set is analyzed. Based on the Pareto optimal solution set obtained through the search, the trade-off between estimation accuracy and computational complexity for different network architectures is analyzed. Then, model selection is performed. Based on the application scenario and computational resource constraints, a suitable network architecture is selected as the final SOH estimation model. Three typical solutions are provided: the model with the best accuracy, the model with the lowest computational complexity, and the compromise model, to meet the needs of different situations.

[0013] Example: The following section demonstrates a detailed SOH estimation scheme using the Oxford University publicly available lithium-ion battery dataset. It mainly consists of five parts: the overall framework of the proposed method, angle field-based image encoding, NSGA-Net for neural network architecture search, experimental setup, and experimental results.

[0014] I. Overall framework of the method proposed in this invention: To illustrate the method proposed in this invention, in Figure 1 The diagram establishes an overall framework. The proposed method follows the standard process of supervised learning. The black arrows in the diagram represent the basic steps of each stage, while the gray arrows, beginning with a dot and ending with an arrowhead, illustrate the data dependencies between different steps. The upper part of the diagram 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] The training process primarily consists of three basic sub-models: data preparation, image encoding, and neural architecture search. First, the dataset used for supervised learning is derived from aging experiments to develop basic deep learning methods for mapping measurement variables to their corresponding SOHs. In this study, voltage is used as the primary input, while the associated capacity is used to compute the SOH as the underlying truth for training. Details of the specific datasets used in this study, including both open-source and private experimental datasets, are available in Section 4.1, “Experimental Data,” of this paper. Voltage data is first encoded into images for subsequent steps. The image encoding method and corresponding analysis characteristic of this approach are presented in Section 2. Next, the neural architecture search for SOH estimation is elaborated in Section 3. For practical applications, three models have been selected as outputs for various scenarios.

[0016] exist Figure 1 The lower part of the description depicts the prediction process, where new voltage data can be obtained from the BMS. Decision-makers can then choose the model best suited for the current SOH estimate. For example, if cloud-based computing is available, which may alleviate computational issues, the most accurate solution can be chosen to provide more precise results. On the other hand, if the SOH application remains computationally limited, the Flops-Best solution can be considered the preferred choice. In addition, NAS outputs offer trade-off solutions, representing the optimal balance between accuracy and computational efficiency. It's worth noting that if some labels are still available in the target domain, they can be used to fine-tune existing models to further improve performance.

[0017] II. Image Coding Based on Angle Field: This section introduces methods for encoding raw measurement variables into images. First, the data preprocessing steps are discussed in detail, followed by a detailed explanation of the encoding process.

[0018] 2.1 Data Preprocessing: In this study, the charging voltage curve was chosen as the input for SOH estimation because obtaining a complete charging curve is more feasible in practical applications compared to the discharge process. However, considering that various charging variables may exhibit different periods, it is necessary to standardize the different charging curves to a uniform dimension before utilizing the data. To this end, the Piecewise Aggregate Approximation (PAA) method was used to unify the data length of different voltage sequences. The PAA method can be viewed as a dimensionality reduction method that can convert an 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 elements in the corresponding interval in the original sequence when calculating each new element.

[0019] Equation (1) clearly shows 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 with reduced size is retained. Using this method, various cycles are standardized to each charging curve with an aligned length of n.

[0020] The next step for the data is to extend the data to a consistent range to mitigate the impact of the data amplitude. Assuming a unified voltage of v¯ = { v ¯1 , . . . , v ¯ n}, the corresponding scaled data within the range [-1, 1] can be obtained through the following formula: (2) 2.2 Encoding images through coordinate transformation: Next, the preprocessed voltage data is reclassified into a polar representation: (3) In the formula: and r correspond to the radial direction and the radius respectively. For each timestamp t i belonging to the natural number N, W is a constant factor used to adjust the span of the polar coordinate system. Subsequently, the sine and cosine values between time series points are calculated and integrated into the constructed Gramian angle field (GAF) matrix. When comparing two typical forms of the submitted GAF - Gramian angle difference (GADF) and Gramian Angular summation (GASF), it is found that GASF is more suitable for analyzing signals with periodic characteristics. Therefore, GASF is selected as the final encoding method for obtaining the encoded image. More specifically, the GASF operator can be defined as follows: (4) And its corresponding matrix form can be expressed as follows: (5) Finally, the obtained encoded image is regarded as the input for the following architecture selection.

[0021] III. NSGA - Net for neural network architecture search: So far, images have been encoded, and it is known that encoded images may possess richer features compared to the original time-series data. Next, we will explore the subsequent steps for SOH estimation using encoded images. Following the basic process of data-driven methods, the next 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, feature extraction and learning can be combined within a single framework, utilizing different layers to handle both tasks simultaneously. Therefore, the core challenge shifts from designing feature extraction methods and estimators to designing the architecture of task-specific deep learning networks. To achieve ideal estimation results, different kinds 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 architecture design of deep neural networks will be a key task in subsequent steps.

[0022] The architecture of deep neural networks is crucial for effectiveness in SOH estimation. It is anticipated that appropriate network architectures can be automatically constructed for the available data. Therefore, in this work, NSGA-Net is used to provide neural network architectures for the SOH estimation task. Using this approach, the architecture search problem is treated as a multi-objective optimization problem, rather than a dominant ranking genetic algorithm used to determine the Pareto set for this search problem.

[0023] In NSGA-Net, the architecture search problem is first transformed into a multi-objective problem, where two objectives are considered simultaneously. First, naturally, estimation accuracy is considered as one objective. Given a search space A, a battery dataset D, and a training method P, the first objective is to find the architecture A∈A that achieves the minimum estimation loss.

[0024] This objective can be expressed mathematically as follows: (6) MSE(a) The test loss representing the estimated SOH and its ground truth is defined as: (7) Where D is the entire dataset, and Dt represents the test dataset. is the output value predicted by the neural network built by the architecture solution a for Dt, Iv is the encoded image data, and SOHi represents the i-th true value of the test dataset.

[0025] Besides estimation accuracy, computational complexity must be reinstated as a separate objective because the implementation of SOH estimation methods remains limited by computational power. It is not feasible to emphasize estimation accuracy while ignoring computational cost. In NSGA-Net, inference time and the number of floating-point operations (FLOPS) are considered surrogate indicators for evaluating the computational complexity of a particular network. Compared to other metrics, FLOPS directly quantifies the number of floating-point operations required for a neural network, thus providing a concrete measure of the computational work required to execute it. Therefore, the second objective of minimizing computational complexity can be expressed as follows: (8) The architecture search problem has been transformed into a multi-objective optimization problem, and a solution awaits. Before leveraging multi-objective optimization algorithms to solve this problem, it is necessary to consider an encoding method for defining the search space for all potential solutions to neural architectures. The NSGA-Net used in this work considers CNN architectures and their variations as combinations of computational blocks computed to specified layers. The basic computational unit in NSGA-NET is defined as a node, which may be a single operation such as convolution, merging, or batching. Several nodes form a computational block called a stage. Ultimately, designing a neural network architecture involves organizing several stages and establishing connections with different nodes. To incorporate this approach into the evolutionary search framework, it is necessary to establish a mapping between phenotypes that represent neural networks as genotypes, which represent the encoded representation of the solution within the context of architecture search.

[0026] For a neural network containing np stages, its compact genotype can be represented by the mutation operation in the evolutionary method as follows: x = [x 1 , x 2 , . . . , x n p ] Each stage is encoded as a directed acyclic structure represented by a binary string, which includes nodes without operations. 1 + 2 + · · · + ( ) = 1 n o ( ) Bits, encoding connections. Assume a node and its connections can be represented as... G=(V,E) ,in V = {1, 2, ..., n} o } It is a node. If the first bit is a single character, it indicates that there is an edge connection between the first node and the second node, i.e. E ={(1, 2)} The next two bits indicate the connection. E = {(i,3)|i∈{1,2}}This process continues until the final... The bit is used to indicate the connection. E ={(i, no)|i∈{1,...,n o }} .

[0027] An example is as follows Figure 2 The diagram illustrates the encoding process. The left side of the image represents a portion of a neural architecture, accommodating up to six nodes, denoted as V={1, 2,...,6}. The right half represents the corresponding genotype encoded by this architecture. Decoding the genotype into a matching architecture requires five steps. The starting bit is 1, indicating that the first node must be connected to the second node, i.e., 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 connections from the sixth node to the remaining nodes. The final connection between these six nodes can be decoded as... E = {(1 , 2) , (2 , 4) , (1 , 5) , (4 , 6) , (5 , 6)}.

[0028] From the description and examples above, it is clear that while the search space primarily refers to connections between various operators, operator selection is also considered. By using this encoding strategy, certain operators represented as nodes can be disabled if other nodes cannot connect to the current node. For example, due to the absence of other connected nodes, Figure 2 The third node in the search space will be discarded. Essentially, the constructed search space consists of several operator choices and potential connection configurations.

[0029] IV. Experiment Setup: 4.1 Experimental Data: Oxford Dataset: Four Kokam Pouch lithium-ion batteries from Oxford University were selected for validation. The cathode material of the cells consisted of a mixture of 11a cobalt carbon oxide and nickel cobalt carbon oxide, while the anode material was graphite. Initially, the batteries underwent a characteristic testing phase in which all cells were charged and discharged at 1C. Subsequently, in an aging test phase, the cells were charged at 2C using a constant current and constant voltage (CC-CV) model when the voltage reached 4.2V. In the discharge phase, the process was performed through a specific profile to simulate realistic operating conditions, ending when the voltage dropped to 2.7V.

[0030] 4.2 Peer method: An overview of peer methods and evolutionary models is presented below. Efforts have been made to expand the diversity of the selected methods to provide more comprehensive experimental results.

[0031] ResNet: This model can be considered the most complex for SOH estimation. The original reconnection system has 34 layers and implements a large number of fast connections to avoid the vanishing / exploding gradient problem. Unlike VGG, the reconnections used here are trained from scratch using the training dataset.

[0032] In addition to the typical model used for the output model, three different types of models were deliberately derived from non-dominant solutions for better comparison. These models cover the requirements of different domains, and their overview is as follows: Optimal Accuracy Solution: This solution (specifically designed for sampling with optimal accuracy) remains non-dominant compared to other solutions. It achieves optimal accuracy across other architectures, making it valuable to examine its specific structure and attempt to summarize the design principles behind SOH estimation.

[0033] The best FLOPS solution: This is also a non-dominant solution, but with the lowest computational complexity. This solution can be viewed as the simplest architecture, which preserves the essential elements for efficient estimation.

[0034] Trade-off Solution: A solution that balances optimal accuracy with optimal computational complexity was specifically sampled. In the current work, the knee-point solution on the Pareto front is considered the optimal solution representing the best trade-off between accuracy and computational complexity.

[0035] It is worth noting that the hyperparameters of each competitor have been carefully tuned to ensure that each competitor represents its best performance, thus ensuring a fair comparison.

[0036] V. Experimental Results: Each validation experiment employs residual cross-validation to evaluate the performance of the proposed method and its competitors. Data from one cell is used as the test dataset, while data from the remaining cells is used for training and optimization.

[0037] Figure 3 a- Figure 3d Estimation results for batteries 1, 3, 5, and 7 on the Oxford dataset are presented separately. For better clarity, a portion of the results has been intentionally magnified. The thick black line in the figure represents the true SOH value, serving as a reference for evaluation. The distance of the predicted SOH value from this baseline indicates the accuracy of the estimate relative to the actual SOH. Intuitively, it is clear that the most accurate optimal solution achieves the best performance, as it closely matches the reference. Conversely, the ResNet method exhibits the worst performance in the current case, suggesting that the residual structure of RESNET may be unfavorable for predicting degradation trajectories that steadily decline. Beyond the best and worst methods, it is noteworthy that the best method also achieves satisfactory results. In most test cases, the performance difference between the optimal and tradeoff solutions is small. Based on these observations and considering the characteristics of the current dataset, an overall conclusion is drawn that for SOH estimation, there is a consistent downward trend, which is a problem with models using simpler architectures.

[0038] Table 1 lists the specific evaluation results using different metrics, providing a detailed comparison of the various methods. Clearly, the most accurate solution outperforms its peers across all test units. However, it is evident that the performance gap between the most accurate and the trade-off solutions is small, suggesting that the relationship between the computational resources allocated to improving prediction accuracy is not linear, and that increasing complexity may be necessary in later stages to further improve accuracy.

[0039] 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 is that while LSTM is capable of managing long-term data, it may exhibit lower performance in extracting complex patterns and functions from the data compared to other models. The optimal solution outperforms other classical models. Therefore, to improve accuracy, it is recommended to design a specific architecture suitable for a given SOH estimation task.

[0040] Table 1. RMSE and MAE of trajectory predictions from the Oxford dataset. ; *Best viewed in bold with a gray background.

Claims

1. A lithium-ion battery SOH estimation method based on automatic neural architecture search, characterized by the following steps: 1) Image encoding module: The raw voltage data of lithium-ion batteries is preprocessed, and the collected battery charging voltage curves are standardized. A segmented clustering approximation method is used to unify voltage sequences of different lengths to a specified length. Then, the voltage data is mapped to the range of [-1,1] through data scaling and image encoding is performed. The preprocessed voltage data is converted into polar coordinate representation. By calculating the sine and cosine values ​​between time series points, a Gram angle field image is constructed. The Gram angle 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, transforming the neural network architecture design problem based on SOH estimation into a multi-objective optimization problem, considering both estimation accuracy and computational complexity. Two optimization objective functions are defined: minimizing the mean squared error to improve estimation accuracy and minimizing the number of floating-point operations 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, the Pareto optimal solution set satisfying the multi-objective optimization is searched. Finally, the network architecture is encoded and decoded, representing it as a combination of computational blocks, each block consisting of multiple nodes. The connection relationships of the network architecture are encoded using binary strings, realizing the genotype representation and decoding of the network architecture. 3) Decision and Model Selection Module: First, the Pareto optimal solution set is analyzed. Based on the Pareto optimal solution set obtained by the search, the trade-off between estimation accuracy and computational complexity of different network architectures is analyzed. Then, model selection is carried out. According to the application scenario and computational resource constraints, a suitable network architecture is selected as the final SOH estimation model. Three typical solutions are provided, including the best accuracy model, the model with the lowest computational complexity, and the compromise solution model, to meet the needs of different situations.

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

Images