Corrosion State Monitoring Based on Multi-Granularity Synergistic Learning of Acoustic Emission and Electrochemical Noise Signals

Abstract

Corrosion monitoring is crucial for ensuring the structural integrity of equipment. Acoustic emission (AE) and electrochemical noise (EN) have been proven to be highly effective for the detection of corrosion. Due to the complementary nature of these two techniques, previous studies have demonstrated that combining both signals can facilitate research on corrosion monitoring. However, current machine learning models have not yet been able to effectively integrate these two different modal types of signals. Therefore, a new deep learning framework, CorroNet, is designed to synergistically integrate AE and EN signals at the algorithmic level for the first time. The CorroNet leverages multimodal learning, enhances accuracy, and automates the monitoring process. During training, paired AE-EN data and unpaired EN data are used, with AE signals serving as anchors to help the model better align EN signals with the same corrosion stage. A new feature alignment loss function and a probability distribution consistency loss function are designed to facilitate more effective feature learning to improve classification performance. Experimental results demonstrate that CorroNet achieves superior accuracy in corrosion stage classification compared to other state-of-the-art models, with an overall accuracy of 97.01%. Importantly, CorroNet requires only EN signals during the testing phase, making it suitable for stable and continuous monitoring applications. This framework offers a promising solution for real-time corrosion detection and structural health monitoring.

1. Introduction

The corrosion of equipment can result in facility damage, production interruptions, environmental pollution, and even explosion. It is one of the primary factors causing the structural integrity degradation in equipment. Therefore, developing effective automated corrosion monitoring methods is crucial to ensure process safety, equipment reliability, and environment protection [1,2].

Acoustic emission (AE) and electrochemical noise (EN) technologies are two types of technologies proven to be effective for corrosion monitoring. AE involves detecting high-frequency elastic waves generated by the rapid release of energy within a material [3,4,5]. Such signals primarily reflect physical phenomena occurring during crack propagation, pitting initiation, hydrogen embrittlement, and other microscopic damage behaviors. On the other hand, EN signals primarily reflect electrochemical reactions during corrosion, with random or semi-random fluctuations in current and potential due to uneven reactions, revealing changes in the corrosion process mechanisms. [6,7,8,9]. Because EN signals are sensitive to the early stages of corrosion, they have been applied in the research of the corrosion behavior of metallic materials [10], the corrosion inhibition of steel reinforcement in concrete [11], stress corrosion cracking [12], and several other corrosion areas.

AE and EN signals for pitting corrosion are chaotic, non-linear, and complex, requiring advanced non-linear methods for processing. Deep learning algorithms effectively address this need. For example, Homborg et al. interpreted EN time–frequency spectra as an image classification problem. They explored the application of convolutional neural networks (CNNs) for the deep learning-based image classification of electrochemical noise time–frequency transient information [13]. Rubio et al. utilized a deep autoencoder model to analyze local electrochemical noise for PEM fuel cell fault diagnosis [14]. Guo et al. applied the InceptionTime model for damage classification in carbon fiber-reinforced composites using AE time-series data [15]. Liu et al. proposed a transfer learning method based on convolutional neural networks (CNN-TL) for pipeline leak detection under various working conditions [16]. Qiu et al. established a CNN-based hydrogen defect recognition (HDR) model from AE signals [17].

As a passive method, however, AE relies on damage events, lacks early warning capabilities, and struggles with noise interference and signal interpretation, while EN signals can directly represent the chemical reactions occurring during corrosion. But EN signals’ characteristics are complex, especially with the indistinct features across different corrosion stages. This complexity makes stage-specific feature analysis challenging. Due to the complementary nature of AE and EN signals in corrosion monitoring, their combined use offers significant potential for advancing corrosion research. In fact, the integrated analysis of these signals has traditionally been conducted manually [18,19,20,21,22]. But such extensive manual analyses are time consuming and labor intensive.

Recently developed multimodal techniques provide a way to integrate different types of data at the feature and/or algorithmic levels to enhance the model’s performance [23,24,25,26,27]. Inspired by these state-of-the-art (SOTA) methods, in this paper, a new network, called CorroNet, was proposed to synergistically integrate AE and EN signals at the algorithmic level for the first time. In CorroNet, AE and EN are treated as two distinct modalities with different granularities. Each modality reflects the different aspects of corrosion process while sharing semantically related features of the corrosion stages. By employing multimodal deep learning algorithms to jointly process these two modalities, more accurate corrosion state monitoring may be achieved. The contributions of this paper are summarized as follows:

(1)

A new deep learning framework, named CorroNet, for corrosion monitoring is proposed. It achieves multi-granularity synergistic learning between AE and EN signals at the model learning level;

(2)

A feature alignment method for corrosion events and a probability distribution alignment loss function are designed for CorroNet to significantly improve the performance in the corrosion monitoring task.

This paper is organized as follows: Section 2 introduces the proposed CorroNet model, the experimental setup, and the corrosion dataset. Section 3 presents an analysis of the model’s corrosion monitoring results. Section 4 provides the conclusions.

2. Methods

The goal of the acoustic-electrochemical synergistic corrosion monitoring is to achieve synergistic learning between the two modalities of acoustic emission (AE) and electrochemical noise (EN) data and to learn an accurate mapping function $F\left(x\right)$ from the training set $X_{train}$ in order to minimize the prediction error on the test set $X_{test}$.

As shown in Figure 1, CorroNet leverages both AE and EN signals to enhance corrosion monitoring. AE, as a passive method, captures signals only during active damage events, resulting in irregular and discontinuous data. In contrast, EN provides continuous-time data with uniform sampling, making it suitable for stable monitoring. CorroNet integrates AE and EN signals during training to leverage both modalities. During testing, it uses only EN signals for consistent and reliable online monitoring.

The training set consists of paired AE-EN data and additional unpaired EN samples. AE signals are selected based on significant corrosion stage features (as shown in Figure 1a), serving as anchors during the learning process. These anchors assist in aligning EN signals of the same category, enhancing the model’s ability to recognize corrosion stages and improving the classification accuracy of the EN data (as shown in Figure 1b).

As illustrated in Figure 1c, AE signals support synergistic learning with EN during training, enriching CorroNet’s understanding of corrosion mechanisms. Once fully trained, CorroNet can effectively combine AE’s intermittent insights with EN’s stable monitoring capability. Therefore, it can accurately predict corrosion stages using only EN signals. The subsequent experimental results (Section 3) provide validation and support for this.

2.1. The Overview of the CorroNet

The Transformer encoder, known for strong sequential modeling, is used as CorroNet’s backbone to extract features from EN and AE time-series data. CorroNet has two branches to process AE and EN signals separately and eventually performs synergistic learning. Different modules are connected after the Transformer encoders in each branch to accomplish various functions. The detailed architecture of CorroNet is shown in Figure 2:

(1)

Most importantly, to achieve multi-granularity synergistic learning, a feature alignment loss $L_{\mathrm{I}\mathrm{n}\mathrm{f}\mathrm{o}\mathrm{N}\mathrm{C}\mathrm{E}}$ and a probability distribution consistency loss $L_{\mathrm{K}\mathrm{L}}$ are designed. After the Transformer encoders pass $f_{\mathrm{A}}$ and $f_{\mathrm{E}}$ in the two branches, projection heads $g_{\mathrm{A}}$ and $g_{\mathrm{E}}^a$ (represented in orange and blue, respectively) are introduced to align the representations $p_{\mathrm{A}}$ and $p_{\mathrm{E}}$ of the two modalities in feature space. Additionally, supervised contrastive learning is applied to the EN data using a second projection head $g_{\mathrm{E}}^b$ (green) after the EN Transformer encoder with the loss function $L_{\mathrm{S}\mathrm{C}}$;

(2)

Secondly, since AE signals serve as anchor points to optimize the feature mapping of EN signals, a classification head $q_{\mathrm{A}}$ is connected after the Transformer encoder $f_{\mathrm{A}}$. It is used to initially train the model for accurate AE classification, for more precise anchor points in the feature space, and for the subsequent optimization of EN signals’ mapping. Similarly, a classification head $q_{\mathrm{E}}$ is connected after the EN branch’s Transformer encoder $f_{\mathrm{E}}$;

(3)

Additionally, due to the multi-granularity feature learning, the complex multi-level feature alignment and loss constraints make precise training and pre-training even more essential. Therefore, a specific training strategy is proposed for CorroNet, dividing the training process into three stages. The details can be found in Section 2.4.

Each component of the CorroNet is discussed in detail in the following section.

2.2. The Transformer Encoder

AE and EN signals are both time-series data characterized by temporal dependencies. Previous studies have demonstrated that Transformer models have achieved remarkable success in temporal modeling tasks [28,29,30,31,32]. As illustrated in Figure 2, the Transformer encoder is introduced as the backbone of CorroNet to leverage the attention mechanism for efficient feature extraction from AE and EN signals, represented as $f_{\mathrm{A}}$ and $f_{\mathrm{E}}$, respectively. Specifically, the Transformer encoder consists of three main components: the embedding block, multi-head attention, and multi-layer perceptron (MLP). The embedding block transforms the raw time-series data into multiple tokens; multi-head attention captures temporal dependencies, and the MLP performs non-linear transformations and feature fusion to enhance expressiveness. The details of the Transformer encoder are in Appendix A.1.

2.3. The AE-EN Alignment Framework

The core component of CorroNet is the multi-granularity synergistic learning between AE and EN signals. It includes three main parts, feature alignment learning, probability distribution alignment learning, and supervised contrastive learning. Each of these components addresses different aspects of the model’s learning process, and their integration enhances the overall corrosion monitoring capabilities of the model. Feature alignment captures mutual information between AE and EN signals; distribution alignment minimizes modality mismatches, and supervised contrastive learning uses labels to enhance corrosion stage feature extraction in EN. The following sections provide a detailed explanation of each module within CorroNet.

The specific structure of the multi-granularity synergistic learning consists of Transformer encoders $f_{\mathrm{A}}$ and $f_{\mathrm{E}}$ and projection heads $g_{\mathrm{A}}$, $g_{\mathrm{E}}^a$, and $g_{\mathrm{E}}^b$. Transformer encoder $f_{\mathrm{A}}$ is responsible for feature extraction from AE signals, followed by the projection head $g_{\mathrm{A}}$, which comprises an MLP to further map the extracted features. The features $z_{\mathrm{A}}$ extracted by $f_{\mathrm{A}}$ serve as the class anchors for guiding the feature learning of EN signals. Another branch extracts features from EN signals using the Transformer encoder $f_{\mathrm{E}}$. Two parallel projection heads, $g_{\mathrm{E}}^a$ and $g_{\mathrm{E}}^b$, map these features into different subspaces. Specifically, $g_{\mathrm{E}}^a$ works together with $g_{\mathrm{A}}$ to achieve a joint mapping for the AE and EN modalities. The mapping produced by $g_{\mathrm{E}}^b$ is used for supervised contrastive learning to further enhance the model’s learning of EN signals.

In the synergistic learning modules, all samples from $X_{train}$ are utilized, including the paired AE-EN dataset $X_{\mathrm{A}\mathrm{E}-\mathrm{E}\mathrm{N}}=\left\{{\left(x_{\mathrm{E}}, x_{\mathrm{A}}, y_{\mathrm{A}}\right)}_{i=1}^{N_1}\right\}$ and additional EN training data $X_{\mathrm{E}\mathrm{N}}=\left\{{\left(x, y\right)}_{i=1}^{N_2}\right\}$. As illustrated in Figure 2, AE samples $x_{\mathrm{A}}$ are fed into the Transformer encoder $f_{\mathrm{A}}$ to generate the features $z_{\mathrm{A}}$. The features $z_{\mathrm{A}}$ are then passed through the projection head $g_{\mathrm{A}}$, resulting in the representations $p_{\mathrm{A}}$. The paired EN samples $x_{\mathrm{E}}$ are fed into the Transformer encoder $f_{\mathrm{E}}$, which outputs the features $z_{\mathrm{E}}$. Subsequently, the features $z_{\mathrm{E}}$ are further mapped by the projection head $g_{\mathrm{E}}^a$ to generate the representations $p_{\mathrm{E}}$. As shown in Figure 2, the model uses feature alignment learning $L_{\mathrm{I}\mathrm{n}\mathrm{f}\mathrm{o}\mathrm{N}\mathrm{C}\mathrm{E}}$ and distribution alignment learning $L_{\mathrm{K}\mathrm{L}}$ to treat AE representations $p_{\mathrm{A}}$ as class anchors. This aligns EN representations $p_{\mathrm{E}}$ with $p_{\mathrm{A}}$, training $f_{\mathrm{E}}$ to learn corrosion stage semantics.

The feature alignment learning $L_{\mathrm{I}\mathrm{n}\mathrm{f}\mathrm{o}\mathrm{N}\mathrm{C}\mathrm{E}}$ is based on the InfoNCE loss [33], which enhances the similarity between paired representations $(p_{\mathrm{A}}, p_{\mathrm{E}})$ while pushing apart unpaired samples, as formulated in Equation (1):

$$L_{\mathrm{I}\mathrm{n}\mathrm{f}\mathrm{o}\mathrm{N}\mathrm{C}\mathrm{E}}=-\sum _{i=1}^{N_1}\log \left({\displaystyle \frac{\exp \left(p_{\mathrm{E}}^i·p_{\mathrm{A}}^i/\tau \right)}{\exp \left(p_{\mathrm{E}}^i·p_{\mathrm{A}}^i/\tau \right)+\sum _{j\ne i}^{N_1}\exp \left(p_{\mathrm{E}}^i·p_{\mathrm{A}}^j/\tau \right)}}\right),$$

(1)

where $\tau$ is the temperature parameter [34], used to control the smoothness during training.

Distribution alignment learning $L_{\mathrm{K}\mathrm{L}}$ is designed based on the Kullback–Leibler divergence. $L_{\mathrm{K}\mathrm{L}}$ minimizes the distribution mismatch between AE and EN modalities. Specifically, it first computes the individual distribution probability scores $s_{\mathrm{A}\mathrm{E}} \mathrm{a}\mathrm{n}\mathrm{d} s_{\mathrm{E}\mathrm{N}}$ for AE and EN samples. Then, an exponential kernel is applied to fit the distribution probability scores of the two modalities to compute their similarity. Taking AE as an example, the distribution probability score $s_{\mathrm{A}\mathrm{E}}^{i,j}$ for the $\left(i,j\right)$-th pair of AE samples is calculated using Equation (2):

$$s_{\mathrm{A}\mathrm{E}}^{i,j}={\displaystyle \frac{\mathrm{e}\mathrm{x}\mathrm{p}(p_{\mathrm{A}}^i·p_{\mathrm{A}}^j/\tau )}{\sum _{k=1}^{N_1}\mathrm{e}\mathrm{x}\mathrm{p}(p_{\mathrm{A}}^i·p_{\mathrm{A}}^k/\tau )}}$$

(2)

The probability score of EN is obtained in the same way and is denoted as $s_{\mathrm{E}\mathrm{N}}^{i,j}$. The probability distribution similarity alignment loss $L_{\mathrm{K}\mathrm{L}}$ is given by Equation (3):

$$L_{\mathrm{K}\mathrm{L}}=\sum _{i=1}^{N_1}\sum _{j=1}^{N_1}{s}_{\mathrm{E}\mathrm{N}}^{i,j}·\log \left({\displaystyle \frac{s_{\mathrm{E}\mathrm{N}}^{i,j}}{s_{\mathrm{A}\mathrm{E}}^{i,j}}}\right).$$

(3)

Unpaired EN samples $x$ are processed by $f_{\mathrm{E}}$, producing features $z$, which are mapped by $g_{\mathrm{E}}^b$ into representations $p$. The model is further trained using supervised contrastive learning to enhance the model’s ability to learn discriminative features for different stages of EN signals. Supervised contrastive learning $L_{SC}$ encourages the model to pull closer the feature representations of samples belonging to the same class, while pushing apart the feature representations of samples from different classes, as described in Equation (4):

$$L_{\mathrm{S}\mathrm{C}}=\sum _{i\in I}{\displaystyle \frac{-1}{\left|P\left(i\right)\right|}}\sum _{j\in P\left(i\right)}\log \left({\displaystyle \frac{\exp \left(z_i·z_j/\tau \right)}{\sum _{aϵA\left(i\right)}\exp \left(z_i·z_a/\tau \right)}}\right),$$

(4)

where $I$ denotes the set of indices for all samples in a batch, and $P\left(i\right)$ represents the set of positive samples that share the same label as sample $i$. $A\left(i\right)$ is the set of all other samples in the training batch except for the $i$-th sample itself. $\tau$ is the temperature parameter that controls the smoothness of contrastive learning. In summary, the training loss function for multi-granularity synergistic learning consists of three components: $L_{\mathrm{I}\mathrm{n}\mathrm{f}\mathrm{o}\mathrm{N}\mathrm{C}\mathrm{E}}$, $L_{\mathrm{K}\mathrm{L}}$, and $L_{\mathrm{S}\mathrm{C}}$. The detailed optimization process for the parameters of each module in the model is provided in Section 2.4.

Additionally, since AE signals serve as anchors to assist in EN signal learning, it is essential to first obtain a feature extractor which is capable of extracting effective corrosion stage discriminative features (denoted as $z_{\mathrm{A}}$) from AE signals. Therefore, before the AE-EN synergistic learning, the Transformer encoder $f_{\mathrm{A}}$ is pre-trained with a supervised classification task. After encoding AE signals with $f_{\mathrm{A}}$, a classification head $q_{\mathrm{A}}$ composed of an MLP is added to classify the corrosion stages of AE signals. This outputs the predicted stage label $\widehat{y_{\mathrm{A}}}=q_{\mathrm{A}}\left(f_{\mathrm{A}}\right(x_{\mathrm{A}}\left)\right)$. The AE portion $\left\{{\left(x_{\mathrm{A}},y_{\mathrm{A}}\right)}_{i=1}^{N_1}\right\}$ of the $X_{\mathrm{A}\mathrm{E}-\mathrm{E}\mathrm{N}}$ is used as the training data. The AE signals $x_{\mathrm{A}}$ are input into the classifier, and the predicted outputs $\widehat{y_{\mathrm{A}}}$ are compared with the true labels $y_{\mathrm{A}}$ to compute the cross-entropy loss $L_{\mathrm{C}\mathrm{N}}^{\mathrm{A}}$, as shown in Equation (5):

$$L_{\mathrm{C}\mathrm{N}}^{\mathrm{A}}=-\sum _{i=1}^{N_1}{y_{\mathrm{A}}}^i\mathrm{l}\mathrm{o}\mathrm{g}\left({\widehat{y_{\mathrm{A}}}}^i\right),$$

(5)

where $i$ represents the sample index, and $N_1$ is the number of AE samples.

After multi-granularity AE-EN synergistic training, the Transformer encoder $f_{\mathrm{E}}$ gains the ability to effectively extract corrosion-related semantic information from EN signals. Finally, $f_{\mathrm{E}}$ is fine-tuned to perform the final corrosion stage classification task. As shown in Figure 2, a classification head $q_{\mathrm{E}}$, composed of an MLP, is attached to $f_{\mathrm{E}}$. During the fine-tuning stage, only EN data are used to train the model. The samples $x$ from $X_{\mathrm{E}\mathrm{N}}$ are fed into $f_{\mathrm{E}}$ and $q_{\mathrm{E}}$, which output the predicted labels $\widehat{y}$. Similar to Equation (5), the cross-entropy loss $L_{\mathrm{C}\mathrm{N}}^{\mathrm{E}}$ is computed between $\widehat{y}$ and the true labels $y$ and used to fine-tune the model, further improving its classification performance.

$$L_{\mathrm{C}\mathrm{N}}^{\mathrm{E}}=-\sum _{i=1}^{N_2}{y}^i\mathrm{l}\mathrm{o}\mathrm{g}\left({\widehat{y}}^i\right).$$

(6)

2.4. Model Training and Optimization

This section describes the training strategy proposed for CorroNet. The proposed CorroNet model undergoes a three-stage training process to maximize its ability to extract semantic corrosion stage features from EN signals, thereby improving its performance in corrosion stage monitoring. Stage 1 aims to pre-train the transformer encoder of the AE branch to extract highly discriminative representations of corrosion stages from AE signals to provide anchor points for feature mapping of EN signals. Stage 2 focuses on achieving synergistic learning between AE and EN signals to enhance the model’s ability to recognize features of different corrosion stages through multi-granularity synergistic learning. Stage 3 further utilizes corrosion label information to enable the model to fully learn the corrosion stage features from EN signals. The parameters of the CorroNet model are optimized using gradient backpropagation, with the specific process detailed in Equations (7)–(16). The parameters for each component $f_{\mathrm{A}}$, $f_{\mathrm{E}}$, $q_{\mathrm{A}}$, $q_{\mathrm{E}}$, $g_{\mathrm{A}}$, $g_{\mathrm{E}}^a$, and $g_{\mathrm{E}}^b$ are denoted as $W_{f_{\mathrm{A}}}$, $W_{f_{\mathrm{E}}}$, $W_{q_{\mathrm{A}}}$, $W_{q_{\mathrm{E}}}$, $W_{g_{\mathrm{A}}}$, $W_{g_{\mathrm{E}}^a}$, and $W_{g_{\mathrm{E}}^b}$, respectively.

(1)

Stage 1 Parameter Optimization:

AE signals serve as anchors to optimize the feature mapping of EN signals, so the AE branch’s Transformer encoder $f_{\mathrm{A}}$ is first trained to extract corrosion stage features. Therefore, only the parameters of $f_{\mathrm{A}}$ and $q_{\mathrm{A}}$ are updated in Stage 1, while other modules are frozen.

$$W_{f_{\mathrm{A}}}\leftarrow W_{f_{\mathrm{A}}}-{\eta }_1\left({\displaystyle \frac{\mathsf{\partial }{L}_{\mathrm{C}\mathrm{N}}^{\mathrm{A}}}{\mathsf{\partial }{W}_{f_{\mathrm{A}}}}}\right)$$

(7)

$$W_{q_A}\leftarrow W_{q_A}-{\eta }_1\left({\displaystyle \frac{\mathsf{\partial }{L}_{CN}^A}{\mathsf{\partial }{W}_{q_A}}}\right)$$

(8)

where ${\eta }_1$ represents the learning rate for Stage 1.

(2)

Stage 2 Parameter Optimization:

In the Stage 2, multi-granularity synergistic learning between AE and EN, as well as supervised contrastive learning for EN signals, is performed. Thus, the parameters of $f_{\mathrm{A}}$, $g_{\mathrm{A}}$, $f_{\mathrm{E}}$, $g_{\mathrm{E}}^a$, and $g_{\mathrm{E}}^b$ are updated, while other modules remain frozen.

$$W_{f_{\mathrm{A}}}\leftarrow W_{f_{\mathrm{A}}}-{\eta }_2\left(\lambda {\displaystyle \frac{\mathsf{\partial }{L}_{\mathrm{K}\mathrm{L}}}{\mathsf{\partial }{W}_{f_{\mathrm{A}}}}}+\left(1-\lambda \right){\displaystyle \frac{\mathsf{\partial }{L}_{\mathrm{I}\mathrm{n}\mathrm{f}\mathrm{o}\mathrm{N}\mathrm{C}\mathrm{E}}}{\mathsf{\partial }{W}_{f_{\mathrm{A}}}}}\right)$$

(9)

$$W_{g_{\mathrm{A}}}\leftarrow W_{g_{\mathrm{A}}}-{\eta }_2\left(\lambda {\displaystyle \frac{\mathsf{\partial }{L}_{\mathrm{K}\mathrm{L}}}{\mathsf{\partial }{W}_{f_{\mathrm{A}}}}}+\left(1-\lambda \right){\displaystyle \frac{\mathsf{\partial }{L}_{\mathrm{I}\mathrm{n}\mathrm{f}\mathrm{o}\mathrm{N}\mathrm{C}\mathrm{E}}}{\mathsf{\partial }{W}_{f_{\mathrm{A}}}}}\right)$$

(10)

$$W_{f_{\mathrm{E}}}\leftarrow W_{f_{\mathrm{E}}}-{\eta }_2\left(\lambda {\displaystyle \frac{\mathsf{\partial }{L}_{\mathrm{K}\mathrm{L}}}{\mathsf{\partial }{W}_{f_{\mathrm{E}}}}}+\left(1-\lambda \right){\displaystyle \frac{\mathsf{\partial }{L}_{\mathrm{I}\mathrm{n}\mathrm{f}\mathrm{o}\mathrm{N}\mathrm{C}\mathrm{E}}}{\mathsf{\partial }{W}_{f_{\mathrm{E}}}}}\right)$$

(11)

$$W_{g_{\mathrm{E}}^a}\leftarrow W_{g_{\mathrm{E}}^a}-{\eta }_2\left(\lambda {\displaystyle \frac{\mathsf{\partial }{L}_{\mathrm{K}\mathrm{L}}}{\mathsf{\partial }{W}_{g_{\mathrm{E}}^a}}}+\left(1-\lambda \right){\displaystyle \frac{\mathsf{\partial }{L}_{\mathrm{I}\mathrm{n}\mathrm{f}\mathrm{o}\mathrm{N}\mathrm{C}\mathrm{E}}}{\mathsf{\partial }{W}_{g_{\mathrm{E}}^a}}}\right)$$

(12)

$$W_{f_{\mathrm{E}}}\leftarrow W_{f_{\mathrm{E}}}-{\eta }_2\left({\displaystyle \frac{\mathsf{\partial }{L}_{\mathrm{S}\mathrm{C}}}{\mathsf{\partial }{W}_{f_{\mathrm{E}}}}}\right)$$

(13)

$$W_{g_{\mathrm{E}}^b}\leftarrow W_{g_{\mathrm{E}}^b}-{\eta }_2\left({\displaystyle \frac{\mathsf{\partial }{L}_{\mathrm{S}\mathrm{C}}}{\mathsf{\partial }{W}_{f_{\mathrm{E}}}}}\right)$$

(14)

where ${\eta }_2$ represents the learning rate for Stage 2, and $\lambda$ is the hyperparameter that balances the loss function components.

(3)

Stage 3 Parameter Optimization:

In the final stage, fine-tuning is conducted to achieve the final corrosion stage classification. In this stage, only the parameters of $f_{\mathrm{E}}$ and $q_{\mathrm{E}}$ are updated, and all other modules are frozen.

$$W_{f_{\mathrm{E}}}\leftarrow W_{f_{\mathrm{E}}}-{\eta }_3\left({\displaystyle \frac{\mathsf{\partial }{L}_{\mathrm{C}\mathrm{N}}^{\mathrm{E}}}{\mathsf{\partial }{W}_{f_{\mathrm{E}}}}}\right),$$

(15)

$$W_{q_{\mathrm{E}}}\leftarrow W_{q_{\mathrm{E}}}-{\eta }_3\left({\displaystyle \frac{\mathsf{\partial }{L}_{\mathrm{C}\mathrm{N}}^{\mathrm{E}}}{\mathsf{\partial }{W}_{q_{\mathrm{E}}}}}\right),$$

(16)

where ${\eta }_3$ represents the learning rate for Stage 1.

Based on the above formulas and descriptions, the pseudocode for the training and inference process can be found in Algorithm 1.

Algorithm 1: Training and Inference Procedure of CorroNet

Input: Training sample sets $X_{train}$ consist of the paired AE-EN dataset $X_{\mathrm{A}\mathrm{E}-\mathrm{E}\mathrm{N}}=\left\{{\left(x_{\mathrm{E}},x_{\mathrm{A}},y_{\mathrm{A}}\right)}_{i=1}^{N_1}\right\}$ and the unpaired EN dataset $X_{\mathrm{E}\mathrm{N}}=\left\{{\left(x,y\right)}_{i=1}^{N_2}\right\}$. Testing sample sets $X_{test}=\left\{{\left(x_t,y_t\right)}_{i=1}^{N_3}\right\}$ consist of the EN data. Output: The predicted corrosion stage label $\widehat{y_t}$ for the sample $x_t$ from the testing set.

Initialize randomly the network weight $W_{f_{\mathrm{A}}}$$, W_{f_{\mathrm{E}}}$$, W_{q_{\mathrm{A}}}$$, W_{q_{\mathrm{E}}}$$, W_{g_{\mathrm{A}}}$$, W_{g_{\mathrm{E}}^a}$, and $W_{g_{\mathrm{E}}^b}$. Set the number of iterations for each stage as $n_1$, $n_2$, and $n_3$. Training:   for $(i=1 to n_1)$ do:  # Stage 1 training process       Frozen $W_{f_{\mathrm{E}}}$$, W_{q_{\mathrm{E}}}$$, W_{g_{\mathrm{A}}}$$, W_{g_{\mathrm{E}}^a}$$, \mathrm{and} W_{g_{\mathrm{E}}^b}$.       obtain AE data and labels $\left\{\left(x_{\mathrm{A}},y_{\mathrm{A}}\right)\right\}$$\mathrm{from} X_{\mathrm{A}\mathrm{E}-\mathrm{E}\mathrm{N}}$.       Calculate $L_{\mathrm{C}\mathrm{N}}^{\mathrm{A}}$ by solving Equation (5).       Update $W_{f_{\mathrm{A}}}$$\mathrm{and} W_{q_{\mathrm{A}}}$ by solving Equations (7) and (8).   end for.   for $(i=1 to n_2)$ do:  # Stage 2 training process       for obtain a batch of paired AE-EN data $\left\{\left(x_{\mathrm{E}},x_{\mathrm{A}}\right)\right\}$ from $Dataloader\left(X_{\mathrm{A}\mathrm{E}-\mathrm{E}\mathrm{N}}\right)$   do:        Frozen $W_{q_{\mathrm{A}}}$$, W_{q_{\mathrm{E}}}$$, \mathrm{and} W_{g_{\mathrm{E}}^b}$.        Calculate $L_{\mathrm{I}\mathrm{n}\mathrm{f}\mathrm{o}\mathrm{N}\mathrm{C}\mathrm{E}}$$\mathrm{and} L_{\mathrm{K}\mathrm{L}}$ by solving Equations (1) and (3).        Update $W_{f_{\mathrm{A}}}$$, W_{g_{\mathrm{A}}}$$, W_{f_{\mathrm{E}}}$$, \mathrm{and} W_{g_{\mathrm{E}}^a}$ by solving Equations (9)–(12).       end for.       for obtain a batch of EN data $\left\{x\right\}$$\mathrm{from} Dataloader\left(X_{EN}\right)$ do:        Frozen $W_{f_{\mathrm{A}}}$$, W_{q_{\mathrm{A}}}$$, W_{q_{\mathrm{E}}}$$, W_{g_{\mathrm{A}}},$$\mathrm{and} W_{g_{\mathrm{E}}^a}$.        Calculate $L_{\mathrm{S}\mathrm{C}}$ by solving Equation (4).        Update $W_{f_{\mathrm{E}}}$$\mathrm{and} W_{g_{\mathrm{E}}^b}$ by solving Equations (13) and (14).       end for.   for $(i=1 to n_3)$ do:  # Stage 3 training process       for obtain a batch of EN data $\left\{x\right\}$$\mathrm{from} Dataloader\left(X_{EN}\right)$ do:        Frozen $W_{f_{\mathrm{A}}}$$, W_{q_{\mathrm{A}}}$$, W_{g_{\mathrm{A}}}$$, W_{g_{\mathrm{E}}^a}$$, \mathrm{and} W_{g_{\mathrm{E}}^b}$.        Calculate $L_{\mathrm{C}\mathrm{N}}^{\mathrm{E}}$ by solving Equation (6).        Update $W_{f_{\mathrm{E}}}$$\mathrm{and} W_{q_{\mathrm{E}}}$ by solving Equations (15) and (16).       end for. Inference:   Obtain a test sample $x_t$$\mathrm{from} X_{test}.$   Calculate the predicted label $\widehat{y_t}=q_{\mathrm{E}}\left(f_{\mathrm{E}}\right(x_t\left)\right)$.   Provide the model’s predicted result for the corresponding corrosion stage of the sample.

2.5. Evaluation Metrics

To evaluate the effectiveness of the model in corrosion monitoring, Accuracy and F1 score are chosen as evaluation metrics. Accuracy measures the proportion of correctly classified samples out of the total number of samples. It provides a straightforward measure of the model’s overall performance by indicating the percentage of correct predictions. The F1 score is the harmonic mean of precision and recall, balancing the trade-off between these two metrics. Precision measures the proportion of true positives out of all predicted positives, and recall measures the proportion of true positives out of all actual positives. Specifically, in addition to calculating the Accuracy (Acc.) and F1 score for each class individually, the Average Accuracy (Ave. Acc.) and the average Macro-F1 score on the test set were also calculated.

For each class $c$, Accuracy and F1 score are calculated using Equations (17) and (18):

$${Acc}_c={\displaystyle \frac{{TP}_c+{TN}_c}{{TP}_c+{TN}_c+{FP}_c+{FN}_c}}$$

(17)

$${F1}_c=2\times {\displaystyle \frac{{Precision}_c\times {Recall}_c}{{Precision}_c+{Recall}_c}}$$

(18)

$${Precision}_c={\displaystyle \frac{{TP}_c}{{TP}_c+{FP}_c}}$$

(19)

$${Recall}_c={\displaystyle \frac{T{P}_c}{{TP}_c+{FN}_c}}$$

(20)

${TP}_c$ is the true positives for class $c$; ${TN}_c$ is true negatives for class $c$; ${FP}_c$ is false positives for class $c$; ${FN}_c$ is false negatives for class $c$. The Average Accuracy (Ave. Acc) across all classes is the mean of the per-class Accuracies, and the Macro-F1 score is the unweighted average of the per-class F1 scores.

2.6. Experimental Setup and Data Description

Pitting corrosion damage is a major issue affecting material strength as it can lead to catastrophic failure of metallic systems and structures, which is often difficult to predict. Therefore, this paper uses pitting corrosion as an example to evaluate the performance of CorroNet and other comparative methods. To obtain AE and EN data for model training and testing, pitting corrosion experiments were conducted on 304 stainless steels. The experimental setup is shown in Figure 3, with Figure 3a depicting a schematic diagram and Figure 3b showing the physical experimental setup. The experiments were conducted at room temperature, using 3.5% NaCl (Sinopharm Chemical Reagent Co. Ltd., Shanghai, China) as the corrosive medium. The number of 304 stainless steel specimens (Shen Xin Ke Ji Co. Ltd., Tengzhou, Shandong, China) used in this study was 20. The working surface of the 304 stainless steel specimens (25 × 50 × 3 mm) was ground with 2000-grit silicon carbide paper, cleaned with ethanol, and then air-dried. A circular area of 1 cm² was exposed for corrosion, with the rest sealed using D-31 epoxy/polyurethane primer. Except for the NaCl solution, all other materials were supplied by Shen Xin Ke Ji Co. Ltd. (Tengzhou, Shandong, China). Sampling details are in Appendix A.2.

The EN and AE data were labeled into three categories according to the corrosion states: no corrosion, metastable corrosion, and stable corrosion, denoted as C1, C2, and C3, respectively. In the no corrosion stage, the metal remains in a healthy state with no obvious signs of corrosion. In the metastable corrosion stage, corrosion has initiated but remains unstable, where corrosion reactions may be localized or intermittent. In the stable corrosion stage, the corrosion process enters a stable phase, with a steady rate and intensified damage [22,35].

Similar to previous studies [36], the synergistic AE-EN data pairs were generated by capturing both signals occurring simultaneously at the same time step. Each AE event comprises 1024 sampling points, and EN signals within 5 min before and after the event were paired as synergistic data. Given the small sample size of AE data, the goal was to achieve effective learning with a limited dataset. As a result, approximately 90 AE-EN aligned data segments were sampled.

In addition, we conducted multiple corrosion experiments with the same experimental setup as described above and collected additional unpaired EN data to expand the dataset. Each experiment lasted approximately 100 h in total, metastable corrosion began after about 10 h and stable corrosion after about 70 h. At the end of the experiment, the long-sequence data collected throughout the entire experiment were segmented. Each segment consists of 3600 sampling points, corresponding to a 1 h duration. In total, the EN dataset comprised 324 samples, with 108 samples for each corrosion stage. To ensure a fair comparison between models and consistency with previous studies, the EN data were randomly split into a train set and a test set in an 8:2 ratio. Using a sliding window approach (window length: 600, stride: 30), the original 3600-time-step sample segments were further divided into smaller segments. The details of the EN dataset are shown in

Table 1. Composition of electrochemical noise dataset.

	Training Set		Testing Set
	Original Count	Count After Sliding Window	Original Count	Count After Sliding Window
C1	86	8686	22	2222
C2	86	8686	22	2222
C3	86	8686	22	2222

.

The proposed CorroNet method was implemented in PyTorch 12.1, which was built using Anaconda (Anaconda Inc., Austin, TX, USA), and the code editor used was Visual Studio Code (Microsoft Corporation, Redmond, WA, USA). The Transformer encoders, $f_{\mathrm{A}}$ and $f_{\mathrm{E}}$, were both configured with two layers ($L=2$) and the multi-head attention mechanism included two heads. The optimizer was configured as AdamW with a learning rate of 0.003, the temperature parameter $\tau$ of 0.07, and the weight $\lambda$ set at 0.5. All tasks used the same preprocessing method, and all models were trained and tested using the same workstation (Windows, Intel Xeon Gold 6248R, NVIDIA GeForce RTX 4090, and 256 GB RAM).

3. Results and Discussion

3.1. Comparison of Corrosion Monitoring Performance

To evaluate the corrosion monitoring performance of CorroNet, it was compared against commonly used deep learning models, i.e., CNN and LSTM. The dataset described in Section 2.6 was used to train and test all models. CNN is a representative model widely utilized in corrosion analysis [16,17,37]. LSTM is a deep learning model extensively employed for analyzing and modeling time-series data [38]. Previous studies have demonstrated their effectiveness in corrosion classification tasks.

In prior research, CNN and LSTM have predominantly been applied to single-modal signal corrosion classification tasks. For a fair comparison, experiments using CNN and LSTM within the proposed CorroNet multimodal framework were also conducted. Specifically, the Transformer backbone in CorroNet was replaced with CNN and LSTM, respectively. This created two variants: Corro-CNN and Corro-LSTM. This setup was intended to demonstrate the robustness and effectiveness of the proposed acoustic-electrochemical alignment framework in improving corrosion classification performance across different backbones. The comparison results are presented in

Table 2. Results for corrosion monitoring. (The bolded values in the table indicate the optimal values).

	Ave. Acc.	Macro-F1	C1		C2		C3
			Acc.	F1	Acc.	F1	Acc.	F1
CNN	0.9310	0.9307	0.8573	0.9221	0.9356	0.9037	1.0000	0.9663
Corro-CNN	0.9548	0.9547	0.8816	0.9351	0.9847	0.9362	0.9982	0.9928
LSTM	0.9389	0.9385	0.8807	0.9275	0.9361	0.9271	1.0000	0.9609
Corro-LSTM	0.9602	0.9601	0.9761	0.9445	0.9046	0.9466	1.0000	0.9893
CorroNet	0.9701	0.9701	0.9203	0.9549	0.9928	0.9600	0.9973	0.9953

.

As shown in Table 2, CorroNet achieved the highest Average Accuracy (Ave. Acc.) of 0.9701 and Macro-F1 score of 0.9701 among all models. For specific classes, C1 samples were the most challenging to classify for all models. This may be attributed to the less distinctive features and higher noise interference in the “no corrosion” stage. Notably, CorroNet achieved an Accuracy of 0.9203 and an F1 score of 0.9549 for C1, which are 11.28% higher in Accuracy and 4.8% higher in F1 score than the CNN and 8.94% higher in Accuracy and 4.26% higher in F1 score than the LSTM. Additionally, CorroNet maintained remarkable Accuracies and F1 scores for C2 and C3 with minimal variance while CNN and LSTM exhibited comparatively larger performance drops. This superior performance is attributed to CorroNet’s Transformer backbone for effective feature extraction. Its synergistic learning framework that aligns AE and EN signals enables more precise corrosion stage classification compared to the single-modal CNN and LSTM models.

The comparison between the regular models (CNN and LSTM) and their variants (Corro-CNN and Corro-LSTM) also demonstrated the effectiveness of synergistic learning. The regular CNN achieved the lowest performance with an Ave. Acc. of 0.9310 and a Macro-F1 of 0.9307. This suggested that CNN may have encountered difficulties when handling time-series data with complex temporal dependencies. LSTM performed slightly better than CNN (Ave. Acc. 0.9389, Macro-F1 0.9385). This result was expected, as LSTMs were generally more capable of handling time-series information compared to CNNs. However, LSTM still fell short compared to CorroNet, particularly in C1 (Accuracy 0.8807), where it struggled to capture the necessary features for accurate classification.

Integrating CNN into the CorroNet framework (Corro-CNN) significantly improved its performance, with a 2.38% increase in Ave. Acc. And a 2.4% increase in Macro-F1 score. The enhanced performance, especially in C1 (Accuracy +2.43%) and C2 (Accuracy +4.91%), indicated that the CorroNet framework helped CNN handle complex temporal tasks. Similarly, Corro-LSTM outperformed the LSTM, with a 2.13% increase in Ave. Acc. And a 2.16% increase in Macro-F1 score. The most significant improvement was in C1 (Accuracy +9.54%), where Corro-LSTM achieved the highest Accuracy (0.9761) among all the models. These results suggested that the CorroNet framework enhanced the ability to capture and align features from dual-modal data to further improve the overall performance.

To further analyze each model’s corrosion monitoring performance and ability to learn corrosion-related semantic features, Figure 4 presents the confusion matrixes of the predicted results on the test set. The horizontal axis represents the predicted labels, and the vertical axis represents the true labels.

The analysis of the confusion matrices for all five models revealed a general pattern, samples from adjacent corrosion stages exhibited more similar features, resulting in a relatively higher misclassification rate. The non-corroded stage (C1) and the metastable corrosion stage (C2) were misclassified by most of the models. For instance, in CNN model (Figure 4a), 13.5% of C1 samples were misclassified as C2, while in LSTM (Figure 4c), 8.33% of C1 samples were also misclassified as C2. Similarly, 6.21% of C2 samples in the CNN model were also misclassified as C3, and 4.55% in the LSTM model. This indicated that most models exhibited weaker recognition capabilities for early-stage corrosion data (C1 and C2). Despite these challenges, CorroNet (Figure 4e) maintained strong corrosion monitoring performance, with only 0.72% of C2 samples misclassified as C1.

Both Corro-CNN and Corro-LSTM significantly reduced misclassification compared to regular CNN and LSTM models. Corro-CNN (Figure 4b) reduced the misclassification rate of C1 as C2 by 1.8% and the misclassification rate of C2 as C3 by 5.08%. Corro-LSTM (Figure 4d) reduced the misclassification rate of C1 as C2 by 7.65%. The improvement in the performance of CNN and LSTM after incorporating the CorroNet framework further demonstrated the effectiveness of the synergistic learning between AE and EN data. In contrast, all models performed exceptionally well on the stable corrosion stage (C3), nearly achieving 100% Accuracy. This indicated that C3 features are more distinct and easier to classify with minimal errors.

In addition to the 8:2 split ratio, we analyzed how different data split ratios affected the model’s robustness in corrosion detection. The results and analysis are in Appendix A.3.

3.2. Ablation Experiments

To evaluate the key components in CorroNet, ablation experiments were conducted. CorroNet consists of the following critical components: (1) Using AE signals as anchors to enhance the learning of corrosion-related semantic features in EN signals. This is achieved through the multi-granularity synergistic learning of AE and EN, including feature alignment learning and probability distribution consistency learning; (2) introducing supervised contrastive learning to capture class-specific features of EN signals; (3) pre-training AE on the corrosion stage classification task to provide precise anchor points for optimizing feature mapping during training.

Therefore, the following seven ablation experiments were conducted:

(1) w/o FAL: the feature alignment learning (FAL) loss function $L_{\mathrm{I}\mathrm{n}\mathrm{f}\mathrm{o}\mathrm{N}\mathrm{C}\mathrm{E}}$ was removed; (2) w/o DAL: the probability distribution alignment learning (DAL) loss function $L_{\mathrm{K}\mathrm{L}}$ was removed; (3) w/o SCL: the supervised contrastive learning (SCL) $L_{\mathrm{S}\mathrm{C}}$ of EN signals was removed; (4) w/o AE pre-training: the pre-training of the AE feature extractor in Stage 1 was removed, and the subsequent learning was conducted directly; (5) w/o AE: the AE signals were entirely excluded from the learning model, which means that only the supervised contrastive learning and the fine-tuning of EN signals were performed; (6) w/o pre-train: both pre-training stages (Stages 1 and 2) were excluded, and only the classification fine-tuning of EN signals was performed directly; (7) w/o EN: The EN signals was removed during model training, and only AE signals were used for model training and prediction. The full model is represented by CorroNet. The results are shown in

Table 3. Results for ablation experiments. (The bolded values in the table indicate the optimal values).

	Ave. Acc.	Macro-F1	C1		C2		C3
			Acc.	F1	Acc.	F1	Acc.	F1
w/o FAL	0.9440	0.9441	0.8794	0.9320	0.9667	0.9225	0.9860	0.9777
w/o FAL	0.9554	0.9552	0.8722	0.9315	0.9977	0.9398	0.9964	0.9944
w/o SCL	0.9470	0.9465	0.8479	0.9152	0.9946	0.9315	0.9986	0.9928
w/o AE pre-training	0.9586	0.9584	0.8870	0.9346	0.9887	0.9429	1.0000	0.9978
w/o AE	0.9373	0.9371	0.8636	0.9268	0.9482	0.9103	1.0000	0.9741
w/o Pre-train	0.9295	0.9287	0.8177	0.8872	0.9743	0.9104	0.9964	0.9884
w/o EN	0.8334	0.8311	0.7083	0.7907	0.9167	0.8627	0.8750	0.8400
CorroNet	0.9701	0.9701	0.9203	0.9549	0.9928	0.9600	0.9973	0.9953

.

w/o FAL (without feature alignment learning): Removing the feature alignment learning $L_{\mathrm{I}\mathrm{n}\mathrm{f}\mathrm{o}\mathrm{N}\mathrm{C}\mathrm{E}}$ caused a 2.61% drop in Ave. Acc., with C1 Accuracy decreasing by 4.09%. This showed that feature alignment learning is essential for learning the corrosion features across stages.

w/o DAL (without distribution alignment learning): distribution alignment learning $L_{\mathrm{K}\mathrm{L}}$ was particularly important for C1 classification, as its absence caused a noticeable 4.81% decline in C1 Accuracy and a 2.34% drop in F1 score.

w/o SCL (without supervised contrastive learning): Removing supervised contrastive learning resulted in a 7.24% drop in C1 Accuracy and a 3.97% decrease in F1 score. These showed that SCL is essential for learning class-specific features, especially for early-stage corrosion (C1).

w/o AE pre-training (without AE pre-training): Removing the pre-training of AE signals reduced Ave. Acc. by 1.05% and Macro-F1 score decreased by 1.17%. Additionally, the C1 Accuracy significantly dropped by 3.33%, highlighting the importance of AE pre-training for distinguishing early corrosion stages (C1).

w/o AE (without AE signals): Removing AE signals resulted in a significant performance degradation, especially in C1 (Accuracy drops by 5.67% and F1 score dropped by 2.81%) and C2 (Accuracy dropped by 4.46% and F1 score dropped by 4.97%). Additionally, the Ave. Acc. decreased by 3.28%, and the Macro-F1 score dropped by 3.3%. These highlight the importance of using AE signals as anchors to improve corrosion feature learning.

w/o pre-train (without pre-training in Stages 1 and 2): Removing both pre-training stages led to the worst performance, with Ave. Acc. decreasing by 4.06% and Macro-F1 by 4.14%. C1 Accuracy significantly dropped by 10.26%, with the F1 score decreasing by 6.77%. The F1 score for C2 also decreased by 4.96%, demonstrating that pre-training is critical for learning robust features. The sharp decline in C1 and C2 performance further underscored the importance of synergistic learning of AE and EN signals and the supervised contrastive learning for effective corrosion feature extraction.

w/o EN (without EN signals): Removing EN signals resulted in a significant performance drop, with the Ave. Acc. decreasing by 13.67% and the Macro-F1 score dropping by 13.9%. This highlights the importance of combining EN and AE signals for model learning.

Comparing the seven ablation experiments, it was found that removing EN signals caused the largest performance drop, which confirms their crucial role. Excluding AE signals or pre-training also led to considerable performance degradation, highlighting the importance of AE-EN multimodality and the pre-training stages. Overall, the ablation experiments strongly supported the design choices made in CorroNet, as each component contributed significantly to the model’s ability to accurately monitor corrosion stages, especially in the early stages where feature differences are more subtle.

4. Conclusions

A novel deep learning framework, named CorroNet, is proposed in this paper, which provides a new approach to automated corrosion monitoring. The proposed CorroNet has the potential to detect corrosion processes on the surface of stainless steels in chloride solutions. The main conclusions are summarized as follows:

(1)

CorroNet pioneers the integration of multi-scale learning from both AE and EN signals, effectively capturing the intricate characteristics of corrosion evolution at different stages. By combining the complementary strengths of these two signals, CorroNet offers a comprehensive approach to corrosion monitoring;

(2)

To further improve the model’s performance, CorroNet incorporates an adaptive learning mechanism that dynamically adjusts to the complexities of corrosion data. This mechanism not only optimizes feature extraction but also allows for more reliable detection of early-stage corrosion, which is often difficult to identify with traditional monitoring methods;

(3)

A comprehensive metal corrosion monitoring task was conducted to validate CorroNet’s real-world applicability. The results highlight its superior performance in predicting various corrosion stages. Ablation studies further underscore the significance of the key model components in achieving this performance.

Despite the promising capabilities of CorroNet, there are opportunities for further improvement. The current research centers on pitting corrosion, one of the most prevalent corrosion forms. However, distinct types of corrosion have different mechanisms. As different corrosion mechanisms may require tailored approaches, future work should explore diverse corrosion types and extend CorroNet’s capabilities to handle multiple corrosion scenarios. Additionally, the development of a user-friendly graphical interface (GUI) is planned to facilitate better model interaction and accessibility for users in industrial settings.