AE-DRSN-ABHFA-Based Method for AUV Fault Diagnosis

Abstract

Reliable fault diagnosis of autonomous underwater vehicles (AUVs) remains challenging because fault-sensitive patterns can be obscured by measurement noise, while several fault categories may exhibit similar responses and ambiguous decision boundaries. To address these issues, this study proposes an attention-enhanced deep residual shrinkage network with adaptive boundary-aware hard feature augmentation (AE-DRSN-ABHFA). First, channel and temporal attention mechanisms are integrated into a deep residual shrinkage network (DRSN) to jointly suppress noise-related responses and emphasize fault-sensitive information across sensor and temporal dimensions. Second, the proposed adaptive boundary-aware hard feature augmentation (ABHFA) mechanism identifies ambiguous boundary samples using class prototypes and boundary-aware scores, while class-level difficulty feedback is used to determine the class-processing order during augmentation. Directional pseudo-features are then introduced during training to enrich the feature distribution near difficult decision boundaries. A joint objective incorporating pseudo-feature classification and batch-hard triplet constraints is further employed to improve intra-class compactness and inter-class separability. The main comparative, noise-robustness, and core ablation experiments were repeated using five independent random seeds. Under the current experimental setting, AE-DRSN-ABHFA achieves a mean clean-condition Accuracy of 95.27 ± 1.17% and a Macro-F1 of 0.9498 ± 0.0128. Across the four evaluated noisy conditions, the proposed model obtains an average Accuracy of 85.37 ± 2.76%, an average Macro-F1 of 0.8253 ± 0.0317, and an average worst-class Accuracy of 75.39 ± 5.98%, demonstrating favorable overall and class-level diagnostic performance under the evaluated noisy operating conditions. These results demonstrate the potential of the proposed framework for noise-robust multi-class AUV fault diagnosis, although further validation on different AUV platforms and real sea-trial data remains necessary.

1. Introduction

In the deep-sea environment, Autonomous Underwater Vehicles (AUVs) undertake tasks such as marine resource exploration, environmental monitoring, and underwater engineering operations, and have become indispensable equipment in modern ocean exploration and engineering applications [1]. Meanwhile, AUVs operate under underwater conditions characterized by high pressure, intense noise, and low visibility, which not only increase the risk of performance degradation and failure in critical components [2] but also impose higher requirements on their operational reliability and autonomous control capability. Therefore, conducting highly reliable fault diagnosis research on AUV systems in marine environments is of great theoretical significance and engineering value for ensuring their safe and stable operation and improving mission execution capability [3].

Early studies on AUV fault diagnosis mainly adopted shallow data-driven methods. For example, Li et al. [4] combined phase-space reconstruction with an extreme learning machine for AUV sensor fault diagnosis, demonstrating the feasibility of shallow learning methods in state recognition for underwater equipment. With the advancement of deep learning, fault diagnosis has shifted toward automatic feature learning. Neural architecture search and multiscale convolutional networks have been widely explored to improve feature representation capability [5,6,7]. These studies have promoted the development of AUV fault diagnosis from shallow pattern recognition to deep feature learning. However, increasing network depth and convolutional scale cannot fully address problems such as insufficient feature extraction under noisy signal conditions, unstable recognition of easily confused categories, or weak discrimination of complex class boundaries.

To further improve fault feature extraction capability, researchers have begun to introduce attention mechanisms to enhance models’ focus on critical channels and important temporal information. The attention mechanism was first applied in sequence modeling with deep neural networks. Bahdanau et al. [8] introduced the attention mechanism into deep learning-based sequence modeling for the first time, alleviating information loss caused by fixed-length encoding. As attention mechanisms were extended to visual and time-series tasks, they were gradually applied to multi-source feature fusion and fault diagnosis. Xia et al. [9] proposed hierarchical attention for multi-source feature aggregation, while Gai et al. [10] adopted a two-stage fusion strategy for thruster fault diagnosis. Xia et al. [11] further introduced adaptive fusion for multivariate time-series data, and Liu et al. [12] combined attention with mixup augmentation for cross-condition scenarios. These studies indicate that attention and fusion mechanisms can improve a model’s ability to select important information. However, existing methods mainly focus on multisource information fusion or global feature weighting. Under strong-signal noisy conditions, if random disturbances in the features are not effectively suppressed, attention weights can easily be misled by noise, making it difficult to accurately focus on weak faults.

Under noisy signal conditions, extracting weak fault features has long been an important challenge in AUV fault diagnosis. Yu et al. [13] combined optimization and signal decomposition to enhance weak fault components. Fekrmandi et al. [14] integrated neural networks with Kalman filtering for state estimation. Wu et al. [15] used GANs for anomaly detection in underwater systems. These methods improve anti-interference capability from the perspectives of front-end signal decomposition, state filtering, and anomaly detection, respectively. Although the traditional deep residual shrinkage network (DRSN) can weaken noise through adaptive soft-threshold shrinkage, the shrunk features may still contain redundant channel responses, and critical fault segments can be weakened during global aggregation. To address this limitation, this paper proposes an attention-enhanced deep residual shrinkage network (AE-DRSN), which integrates channel attention and temporal attention into the DRSN framework. The channel attention adaptively emphasizes fault-related sensor variables, while the temporal attention highlights informative time segments. By combining shrinkage-based denoising with attention-based feature refinement, AE-DRSN effectively enhances fault-sensitive representations while suppressing noise interference in a unified feature-learning framework.

In addition to signal noise interference, response similarity among multiple fault categories is also an important factor affecting diagnostic reliability. Luo et al. [16] studied the impact of difficult classes on diagnosis performance. Liu et al. [17] improved boundary adaptation in open-set settings. Guo et al. [18] used digital twins and dynamic state representation to enhance the stability of health monitoring for complex equipment. Related studies show that an improvement in overall accuracy does not necessarily mean that every fault category can be recognized stably. In particular, when responses are similar, such as in cases of mild and severe propeller damage or normal states and early degradation states, problems such as overlapping class boundaries, easy misclassification of boundary samples, and significant performance degradation for a small number of difficult classes may occur even when the sample quantities are approximately balanced. Therefore, the essence of this problem lies not only in sample quantity imbalance, but also in uneven class learning difficulty and ambiguous feature boundaries.

To optimize the fault feature space, metric learning, triplet loss, and prototype learning have gradually been introduced into fault diagnosis. Wang et al. [19] introduced triplet loss into an adversarial domain adaptation framework, enhancing cross-domain feature separability through distance constraints. Shen et al. [20] used triplet constraints to improve intra-class compactness and inter-class separation under unknown load conditions. Wang et al. [21] characterized class centers and organized the feature space through a deep prototype network. Liu et al. [22] adopted trustworthy granular feature contrastive learning to enhance discriminative feature learning under semi-supervised domain generalization. Although these studies provide effective ideas for complex boundary optimization, most existing methods employ triplet constraints, class prototypes, contrastive learning, or hard sample selection separately. A single prototype can describe the class center but cannot directly identify class boundaries; a single metric constraint can adjust distances between samples but cannot adaptively determine which classes remain difficult to learn over time; and selecting hard samples solely based on loss is susceptible to noise and incidental misclassification. Therefore, there remains a lack of a mechanism that can jointly consider class centers, boundary locations, and class learning difficulty while directionally enhancing boundary samples in difficult classes.

To address this issue, an adaptive boundary-aware hard feature augmentation mechanism, termed ABHFA, is developed. Class prototypes are first used to characterize category centers in the learned feature space. A boundary-aware score is then calculated by jointly considering classification uncertainty and the deviation of each sample from its own class prototype, allowing comparatively ambiguous samples to be prioritized for augmentation. In parallel, class-wise recognition performance evaluated on the complete training subset is used to estimate learning difficulty, and a momentum-based update is employed to reduce short-term fluctuations. The validation subset is reserved exclusively for checkpoint and hyperparameter selection, and it does not participate in the difficulty-feedback mechanism. The fault classes are processed in descending order of their smoothed difficulty scores, and within each class, samples with high boundary-aware scores are considered first for directional pseudo-feature augmentation. Pseudo-feature classification constraints are used to preserve semantic consistency, while batch-hard triplet constraints further encourage intra-class compactness and inter-class separation. Thus, the proposed method combines boundary awareness, class-difficulty feedback, targeted feature augmentation, and metric learning within a unified training framework.

From an engineering perspective, AUV fault diagnosis models must ensure not only high accuracy but also robustness and real-time performance under complex noise conditions [23]. However, most existing evaluations are conducted under limited or idealized noise settings, leaving multi-noise robustness insufficiently studied [24]. Recent studies on marine autonomous systems have further emphasized the importance of maintaining reliable operation under environmental and sensing uncertainties. Van Chuong et al. developed a fractional-integral terminal sliding-mode control strategy assisted by wavelet neural networks for unmanned surface vessels under uncertain hydrodynamics and environmental disturbances, illustrating the role of adaptive compensation in robust marine-vehicle control [25]. From the diagnostic perspective, Mai et al. proposed an AUV fault-diagnosis model based on a deep belief rule base with attribute reliability to reduce the influence of disturbed and unreliable sensor information [26]. These studies highlight two complementary aspects of autonomous marine-system reliability: robust control is required to compensate for environmental and model uncertainties, whereas reliable fault diagnosis is required to provide trustworthy health-state information when sensing or propulsion components become abnormal. Based on the above analysis, this paper proposes the AE-DRSN-ABHFA method to address problems in AUV fault diagnosis, including insufficient extraction of fault-sensitive features under noisy conditions, unstable recognition of easily confused categories, and weak discrimination of complex class boundaries. The main contributions of this paper are as follows:

(1)

An AE-DRSN method that integrates channel attention and temporal attention is proposed. Based on adaptive soft-threshold denoising, it selectively enhances fault-sensitive information from both the sensor-variable dimension and the temporal dimension, achieving collaborative modeling of noise suppression and key feature enhancement.

(2)

An ABHFA mechanism is proposed. By jointly using class prototypes, boundary-aware scores, class-difficulty-guided ordering of fault classes, and boundary-aware identification of augmentation candidates, the mechanism performs directional feature augmentation for ambiguous boundary samples according to the resulting class-processing order, thereby improving the separability of easily confused categories and the balance of class-level recognition.

(3)

A staged dynamic training strategy that integrates pseudo-feature classification constraints and batch-hard triplet constraints is constructed. Without increasing the inference burden during the testing stage, this strategy continuously optimizes intra-class compactness and inter-class separation under complex class boundaries.

The remainder of this paper is organized as follows: Section 2 introduces the proposed method; Section 3 presents the experimental design and validates the effectiveness of the method using a public laboratory dataset, along with the corresponding results and discussion; and finally, Section 4 summarizes the conclusions.

2. AE-DRSN-ABHFA Fault Diagnosis Method

Based on the above analysis, this paper proposes the AE-DRSN-ABHFA fault diagnosis method, as shown in Figure 1, to improve diagnostic robustness and classification stability in complex noisy environments. The proposed method consists of three core modules. First, the wide-convolution preprocessing stage in Stage 1 stabilizes the initial features of multivariate time-series signals, providing inputs with a high signal-to-noise ratio for subsequent deep robust feature learning. Second, the hierarchical multi-scale AE-RSBU module in Stage 2 incorporates residual learning, an adaptive soft-threshold shrinkage mechanism, and dual channel–temporal attention to progressively extract key fault features from local short-term disturbances, medium-scale waveform variations, and long-term dynamic evolution, while simultaneously achieving noise suppression and feature enhancement. Finally, the classification head in Stage 3 performs global decision mapping through global average pooling, fully connected layers, Dropout, and Softmax, and outputs the probability distribution over five fault states. In addition, an adaptive boundary-aware hard feature augmentation mechanism, ABHFA, is introduced to further strengthen the model’s discriminative capability for easily confused classes and samples located in boundary regions. The overall architecture enables end-to-end mapping from raw time-series inputs to fault-category outputs, fully demonstrating the effectiveness of multiscale modeling and discriminative feature enhancement under complex noisy conditions.

2.1. Robust Feature Extraction Mechanism

For AUV monitoring signals in operational environments that are simultaneously affected by propulsion disturbances, sensor measurement errors, and random signal noise, weak fault responses may still be masked by noise during feature learning if only convolution and residual connections are used. Based on this consideration, this paper further introduces adaptive soft-threshold shrinkage and attention enhancement mechanisms into the residual learning framework to strengthen signal features and construct an AE-DRSN feature purification unit. This unit operates according to the process of “feature mapping–threshold shrinkage–attention enhancement–residual fusion,” thereby achieving collaborative optimization of noise suppression and key feature enhancement. Let the input feature of the $l$-th AE-RSBU be denoted as $h^{(l-1)}\in {\mathbb{R}}^{T_l\times C_l}$, where $T_l$ denotes the temporal length and $C_l$ denotes the number of channels. The intermediate feature obtained through the residual branch is given by

$$U=F_{k_l}(h^{(l-1)};{\theta }_l)$$

(1)

Here, $F_{k_l}(·)$ denotes the feature extraction mapping, ${\theta }_l$ represents the corresponding learnable parameters, and $U$ is the intermediate feature fed into the shrinkage and enhancement stages. To suppress random noise and irrelevant disturbances in AUV monitoring signals, the AE-RSBU module introduces an adaptive soft-threshold shrinkage mechanism, which first performs global statistics on the feature magnitudes and then generates channel-wise thresholds through a gated mapping:

$$\tau =\sigma (W_2^{(l)}\delta (W_1^{(l)}GAP(\left|U\right|)))\odot GAP(\left|U\right|)$$

(2)

Here, $GAP(·)$ denotes the global average pooling operation, which is used to extract the overall response intensity of each channel. $\left|·\right|$ denotes the element-wise absolute value operation on the features; $W_1^{\left(l\right)}$ and $W_2^{\left(l\right)}$ are the parameters of the two fully connected mapping layers; $\delta (·)$ is the nonlinear activation function; and $\sigma (·)$ is the Sigmoid function. $\odot$ denotes element-wise multiplication; $\tau \in {\mathbb{R}}^C$ is the adaptive threshold vector corresponding to each channel. Unlike fixed-threshold methods, this expression allows the threshold to be jointly regulated by the current feature intensity and nonlinear inter-channel dependencies, thereby enabling adaptation to changes in feature distributions under different fault modes and noise intensities. After obtaining the adaptive thresholds, soft-threshold shrinkage is applied to the intermediate features:

$$x_{shr}^{\left(l\right)}=sign(U)\odot \mathrm{m}\mathrm{a}\mathrm{x}(\left|U\right|-{\tau }^{\left(l\right)},0)$$

(3)

Here, $sign(·)$ denotes the sign function, and $x_{shr}^{\left(l\right)}$ represents the feature output after shrinkage. Although relying solely on soft-threshold shrinkage can reduce noise interference, it remains difficult to sufficiently highlight fault-sensitive channels and key temporal segments. Therefore, this paper further introduces channel attention and temporal attention mechanisms to adaptively recalibrate the shrunk features. This mechanism is a key improvement of AE-DRSN over the traditional DRSN. First, channel attention is used to measure the importance of different sensor channels for fault discrimination, and its computation is given by

$$a_c^{\left(l\right)}=\sigma (W_{c2}^{\left(l\right)}·\delta (W_{c1}^{\left(l\right)}·GAP\left(x_{shr}^{\left(l\right)}\right)\left)\right)$$

(4)

Here, $a_c^{\left(l\right)}$ denotes the channel attention weight vector of the $l$-th module, and $W_{c1}^{\left(l\right)}$ and $W_{c2}^{\left(l\right)}$ are the fully connected parameters in the channel attention branch. The channel attention weights can adaptively enhance fault-related sensitive variables and suppress redundant or weakly related channel responses. For AUV fault diagnosis, different fault modes often exhibit different responses in variables such as propulsion control quantities, depth, attitude, acceleration, or angular velocity; therefore, channel attention helps improve the model’s utilization efficiency of key sensor information. Furthermore, to highlight the data characteristics of local temporal segments, temporal attention is constructed as follows:

$$a_t^{\left(l\right)}=\sigma \left(Conv1D_{k_t}\left([{AvgPool}_c\left(x_{shr}^{\left(l\right)}\right);{MaxPool}_c\left(x_{shr}^{\left(l\right)}\right)]\right)\right)$$

(5)

Temporal attention can assign weights along the temporal dimension, enabling the model to focus more on key moments containing fault-discriminative information rather than averaging the entire time series. Finally, channel attention and temporal attention are used to construct a joint mask, and the enhanced features are computed as follows:

$$M^{\left(l\right)}=a_c^{\left(l\right)}\otimes a_t^{\left(l\right)}$$

(6)

$$x_{att}^{\left(l\right)}=M^{\left(l\right)}\odot x_{shr}^{\left(l\right)}$$

(7)

Here, $a_c^{\left(l\right)}$ is the channel attention weight vector, $a_t^{\left(l\right)}$ is the temporal attention weight vector, $M^{\left(l\right)}$ is the joint attention mask, $\otimes$ denotes the construction of the joint attention mask after expanding the channel weights and temporal weights, and $x_{att}^{\left(l\right)}$ is the enhanced feature representation. The former helps emphasize fault-related sensitive variables, whereas the latter is more conducive to capturing key temporal patterns, such as local abrupt changes, short-term shocks, and weak degradation segments. At this stage, to avoid losing original useful information during feature selection and to maintain the training stability of the deep network, the model further fuses the enhanced features with the output of the shortcut branch to form the final output of the current module. This process corresponds to the terminal output path of each AE-RSBU module in Figure 1, and its expression can be written as

$$h^{\left(l\right)}=x_{att}^{\left(l\right)}+S^{\left(l\right)}\left(h^{\left(l-1\right)}\right)$$

(8)

Here, $h^{\left(l-1\right)}$ denotes the input feature of the $l$-th AE-RSBU module, and $x_{att}^{\left(l\right)}$ denotes the feature representation after soft-threshold shrinkage and joint channel–temporal attention enhancement. $S^{\left(l\right)}(\cdot )$ denotes the shortcut mapping, which is usually an identity mapping when the input and output dimensions are consistent and a linear projection for dimensional matching when they are inconsistent. $h^{\left(l\right)}$ is the final output feature of the $l$-th module. This fusion strategy achieves a balance between discriminative feature enhancement and the preservation of original useful information. On the one hand, soft-threshold shrinkage suppresses irrelevant noise. On the other hand, joint channel–temporal attention highlights fault-sensitive information. Meanwhile, the residual connection ensures the stable transmission of effective shallow-layer responses. As a result, AE-DRSN can extract more robust and discriminative fault features under complex noisy conditions.

2.2. Boundary-Aware Hard Feature Augmentation Mechanism

Based on the deep features extracted by AE-DRSN, this paper introduces an adaptive boundary-aware hard feature augmentation (ABHFA) mechanism, which characterizes class centers through class prototypes, identifies easily confused samples through boundary scoring, and uses class-level learning difficulty to determine the class-processing order during augmentation, as well as boundary-aware scores to identify candidate samples, thereby improving the model’s learning capability for hard classes and boundary samples.

Let the input sample be ${\mathrm{X}}_i$. The deep feature representation obtained after passing through the AE-DRSN feature extraction network is given by

$$z_i=G_{\theta }\left({\mathrm{X}}_i\right)$$

(9)

Here, $G_{\theta }(\cdot )$ denotes the feature extractor composed of the wide convolutional layer and multiscale AE-RSBU modules, $\theta$ represents its learnable parameters, and $z_i$ denotes the deep feature of the $i$-th sample, where $z_i\in {\mathbb{R}}^d$ and $d$ is the feature dimension. This feature serves as the basis for subsequent boundary awareness and hard-sample augmentation. First, to characterize the distribution center of each fault class in the feature space, this paper calculates the class prototype for the samples of each class. For the $k$-th fault class, its class prototype is defined as

$$p_k={\displaystyle \frac{1}{N_k}}\sum _{i:y_i=k}{z}_i$$

(10)

Here, $p_k$ denotes the class prototype of the $k$-th class, $N_k$ denotes the number of samples in the $k$-th class, and $y_i$ denotes the ground-truth label of the $i$-th sample. The class prototype can be understood as the central position of a certain fault class in the feature space. In the implementation, the class prototypes are recomputed from the deep features of the current training mini-batch whenever ABHFA is active. On this basis, the distance from a sample to its own class prototype and the distance from the sample to the nearest prototype of another class are calculated, respectively:

$$d_i^{+}={‖z_i-p_{y_i}‖}_2$$

(11)

$$d_i^{-}=\underset{k\ne y_i}{\min }{‖z_i-p_k‖}_2$$

(12)

Here, $d_i^{+}$ denotes the distance between the sample and the center of its own class, while $d_i^{-}$ denotes the distance between the sample and the nearest center of another class. The boundary-aware score is calculated by jointly considering classification uncertainty and the distance from the sample feature to its corresponding class prototype. Therefore, this paper defines the boundary-aware score as

$$S_i=0.45\left(1-P\left(y_i\left|z_i\right.\right)\right)+0.35{\displaystyle \frac{d_i^{+}-d_{min}^{+}}{d_{max}^{+}-d_{min}^{+}+\epsilon }}$$

(13)

Here, $P\left(y_i\left|z_i\right.\right)$ denotes the Softmax probability assigned to the ground-truth class of the ($i$)-th sample, $d_i^{+}$ denotes its distance to the corresponding own-class prototype, and $d_{min}^{+}$ and $d_{max}^{+}$ denote the minimum and maximum own-class prototype distances within the current mini-batch, respectively. $\epsilon$ is a small positive constant for numerical stability. The first term in Equation (13) characterizes classification uncertainty, whereas the second term characterizes the normalized deviation from the own-class prototype. Therefore, a larger $S_i$ indicates a comparatively more ambiguous sample and gives it a higher priority during boundary-aware augmentation. At the end of the training epoch $t$, a deterministic forward pass without gradient computation is performed over the complete training subset, and the instantaneous difficulty of the $k$-th class is defined as follows:

$$h_k^{\left(t\right)}=1-a_k^{\left(t\right)}$$

(14)

where $a_k^{\left(t\right)}$ denotes the recognition accuracy of the $k$-th class evaluated over the complete training subset at the end of epoch $t$, and $h_k^{\left(t\right)}$ denotes its instantaneous class-difficulty score. A lower class-wise training accuracy therefore corresponds to a larger difficulty score.

Directly using the class-wise accuracy from a single training stage may cause the estimated class difficulty to fluctuate with temporary changes in model performance. Therefore, a momentum-based update is introduced to obtain a more stable estimate of the class learning difficulty:

$${\overline{h}}_k^{\left(t\right)}=m{\overline{h}}_k^{(t-1)}+(1-m)h_k^{\left(t\right)}$$

(15)

where ${\overline{h}}_k^{\left(t\right)}$ denotes the smoothed difficulty score of the $k$-th class, $h_k^{\left(t\right)}$ denotes the instantaneous difficulty score, and $m$ is the momentum coefficient. In this study, the momentum coefficient is set to 0.85 to balance responsiveness to the current training-subset class-wise performance and stability against short-term fluctuations. For the first difficulty update, the smoothed score is initialized using the current instantaneous difficulty. The difficulty estimation is performed only after completion of the current training epoch. Specifically, after epoch $t$, the current model is evaluated once over the complete training subset in deterministic inference mode without gradient computation. The resulting class-wise training accuracies are used to update the difficulty scores, which are then used to determine the class-processing order for pseudo-feature generation in epoch $t+1$. Therefore, the validation subset does not participate in class-difficulty estimation or pseudo-feature generation.

The fault classes represented in the current mini-batch are then ranked in descending order according to their smoothed difficulty scores $h_k^{\left(t\right)}$. Classes with larger difficulty scores are regarded as more difficult classes. In this way, the class-difficulty feedback mechanism determines the class-processing order according to persistent class-level learning difficulty, without introducing an additional temperature-based normalization mechanism. Within each class, the samples are ranked according to their boundary-aware scores, and the top max (1, $⌊0.30\mathrm{N}\mathrm{c}⌋$) samples are considered as boundary candidates, where Nc denotes the number of samples of class c in the current mini-batch. For each boundary candidate, all other same-class samples in the current mini-batch are considered as potential neighbors. A candidate neighbor is retained only if the midpoint between it and the boundary sample remains closer to the boundary sample than to the nearest competing-class prototype; among the retained candidates, the one closest to the boundary sample is selected as the same-class neighbor. Based on the boundary-aware score and the difficulty-guided class ordering and boundary-aware sample selection described above, boundary-aware pseudo-features are generated in the deep feature space as follows:

$${\tilde{z}}_i=z_i+\alpha u_i^{+}-\beta u_i^{-}+{ϵ}_i$$

(16)

Here, ${\tilde{z}}_i$ denotes the generated pseudo-feature and $z_i$ denotes the original feature of the selected boundary candidate. $u_i^{+}$ is the normalized direction from the boundary sample toward the selected same-class neighbor, whereas $u_i^{-}$ is the normalized direction from the prototype of the ground-truth class toward the nearest competing-class prototype. The neighbor-blending coefficient is fixed at $\alpha$ = 0.25, and the competing-prototype repulsion coefficient is fixed at $\beta$ = 0.10. Therefore, the term $\alpha u_i^{+}$ moves the pseudo-feature moderately toward a same-class neighbor, whereas the term $\beta u_i^{-}$ moves it away from the nearest competing-class prototype. A small Gaussian perturbation ${ϵ}_i~\mathcal{N}(0,{0.05}^{2}I)$ is additionally introduced to provide local feature diversity. A generated pseudo-feature is retained only when it remains closer to the prototype of its original class than to the nearest competing-class prototype, thereby reducing the risk of semantic label corruption. To enable real features and pseudo-features to jointly participate in model training, this paper constructs a joint optimization objective:

$$L=L_{ce}+{\lambda }_p{L}_{pce}+{\lambda }_t{L}_{tri}$$

(17)

where $L_{\mathrm{c}\mathrm{e}}$ denotes the cross-entropy loss calculated from the original training samples, $L_{\mathrm{p}\mathrm{c}\mathrm{e}}$ denotes the pseudo-feature classification loss, and $L_{\mathrm{t}\mathrm{r}\mathrm{i}}$ denotes the batch-hard triplet loss calculated using both real and generated features. ${\lambda }_p$ and ${\lambda }_t$ denote the corresponding auxiliary loss weights and are set to 0.15 and 0.05, respectively. The triplet margin is set to 0.30. Therefore, ABHFA can effectively improve class-wise recognition balance and boundary discrimination capability in AUV multi-fault diagnosis.

2.3. Diagnostic Classification Decision

Stage 2 consists of three hierarchical AE-RSBU modules connected in series, with convolutional kernel sizes of $3\times 1$, $5\times 1$, and $7\times 1$. After progressive modeling from the first to the third module, the network finally obtains a hierarchically robust feature representation:

$$h^{\left(3\right)}={\mathcal{F}}_{{AE\text{-}RSBU}_3}\left({\mathcal{F}}_{{AE\text{-}RSBU}_2}\right({\mathcal{F}}_{{AE\text{-}RSBU}_1}\left(h^{\left(1\right)}\right)\left)\right)$$

(18)

Here, $h^{\left(1\right)}$ denotes the output feature after the Stage 1 wide-convolution preprocessing; ${\mathcal{F}}_{AE\text{-}RSB{U}_i}(\cdot )$ denotes the nonlinear mapping corresponding to the $i$-th AE-RSBU module; and $h^{\left(3\right)}$ is the final hierarchical feature output by Stage 2. On this basis, the model enters Stage 3 on the right side of Figure 1, where global average pooling is first applied to compress the hierarchical features and to aggregate discriminative information along the temporal dimension into a compact global feature vector:

$$z=GAP\left(h^{\left(3\right)}\right)$$

(19)

Here, $z$ is the global feature vector obtained after temporal aggregation. Subsequently, the model further performs feature mapping and stochastic deactivation regularization through a fully connected layer and a Dropout mechanism, and its output can be expressed as

$$o=W_{o}Dropout\left(z\right)+b_o$$

(20)

Here, $W_o$ and $b_o$ denote the weight matrix and bias vector of the classification layer, respectively; $\mathrm{D}\mathrm{r}\mathrm{o}\mathrm{p}\mathrm{o}\mathrm{u}\mathrm{t}(\cdot )$ denotes the stochastic deactivation operation; and $o$ is the output vector before being fed into the Softmax classifier. The purpose of introducing Dropout is to reduce the model’s excessive dependence on a single local feature, thereby improving its generalization ability under complex operating conditions. Finally, the posterior probability distribution of each fault class is obtained as follows:

$$p_i={\displaystyle \frac{exp\left(o_i\right)}{{\sum }_{j=1}^{M}exp\left(o_j\right)}}$$

(21)

$o_i$ denotes the output value corresponding to the $i$-th class in the vector $o$; $M$ denotes the total number of fault classes; and $p_i$ denotes the predicted probability that the sample belongs to the $i$-th fault state. Thus, the final classification result of the model can be written as

$$\widehat{y}=argmax(p_1,p_2\cdots p_i)$$

(22)

2.4. Model Training and Deployment Workflow Design

As shown in Figure 2, the training and diagnostic procedure of the proposed method can be summarized as follows. First, samples are constructed from the AUV multivariate time-series data and then partitioned into training, validation, and test sets using stratified sampling. The channel-wise min–max normalization parameters are calculated exclusively from the training set and are subsequently applied to all three subsets. Second, the training samples are fed into the AE-DRSN-ABHFA network, where deep robust features are extracted through the wide-convolution preprocessing layer and multiscale AE-RSBU modules. During the ABHFA stage, the current class-difficulty scores are used to determine the class-processing order during ABHFA and guide boundary-aware pseudo-feature generation. The real features and generated pseudo-features are then jointly optimized through the classification and metric-learning objectives. At the end of each training epoch, a deterministic forward pass without gradient computation is performed over the complete training subset to obtain class-wise training accuracies. These accuracies are converted into instantaneous class-difficulty scores and further smoothed using the momentum update. The updated difficulty scores are used to guide ABHFA pseudo-feature generation in the subsequent epoch. Importantly, the validation set does not participate in class-difficulty estimation, class ordering, or pseudo-feature generation, and is used only for checkpoint and hyperparameter selection. The test set is accessed only after the model configuration and checkpoint-selection rule have been frozen. During the testing stage, the input samples undergo only AE-DRSN-based feature extraction and classifier prediction, without pseudo-feature generation. The model finally outputs the probability distribution over the fault categories and determines the operating state or fault type of the AUV according to the maximum posterior probability. This procedure enables end-to-end diagnosis from raw multi-sensor time-series inputs to fault-category outputs.

3. Experimental Design

3.1. Data Description

To validate the effectiveness of the proposed method in AUV fault diagnosis, this study conducts experimental investigations using the publicly available dataset of the “HaiZhe” small quadrotor AUV, with fault data derived from the “HaiZhe” dataset [27]. The dataset covers five operating conditions: normal operation, added-weight fault, depth sensor fault, mild propeller damage, and severe propeller damage, thereby capturing typical abnormal patterns associated with propulsion, payload, and perception subsystems. Specifically, the added-weight fault is used to simulate payload variations caused by foreign-object attachment to the vehicle; the depth sensor fault is used to simulate bathymetric deviations induced by abnormal pressure gain; and mild and severe propeller damage correspond to different degrees of propulsion performance degradation. As illustrated in Figure 3, these three fault categories correspond to (a) added-weight fault, (b) depth sensor fault, and (c) propeller health conditions (including mild and severe damage).

For each fault type, the dataset provides raw operational records obtained through multiple repeated experiments. The state sequence from each individual trial is constructed as one sample, and the sample label is assigned based on the corresponding fault mode. The collected variables include multidimensional state quantities such as PWM control signals, depth, pressure, voltage, attitude angles, angular velocities, and accelerations, as shown in

Table 1. Experimental data acquisition variables in the “HaiZhe” dataset.

Name	Data Source	Unit
time	Recording time of the data	s
pwm1–4	High-level value of the control signal for the servo	ms
depth	Depth measured by the depth sensor	m
press	Pressure value measured by the depth sensor	Pa
voltage	Battery voltage	V
roll	Roll angle	degrees
pitch	Pitch angle	degrees
yaw	Yaw angle	degrees
a_x	Acceleration along the x-axis	m/s2
a_y	Acceleration along the y-axis	m/s2
a_z	Acceleration along the z-axis	m/s2
w_row	Angular velocity about the roll axis	degrees/s
w_pitch	Angular velocity about the pitch axis	degrees/s
w_yaw	Angular velocity about the yaw axis	degrees/s

. These variables can comprehensively characterize the coupling relationship between the dynamic response and perception information of the AUV during fault evolution.

Table 1 lists the multivariate state variables acquired from the AUV, including control signals, depth, pressure, attitude, angular velocities, and accelerations. The corresponding experimental setup and data acquisition workflow are depicted in Figure 4, which outlines the overall procedure for collecting these time-series signals under each fault condition. As shown in the figure, the data collection process involves running the AUV under controlled normal and fault-injected scenarios, with each trial recording the full dynamic response of the vehicle over a complete operational episode.

In terms of data representation, each sample can be expressed as a two-dimensional time-series matrix, $\mathrm{X}\in {\mathbb{R}}^{T\times C}$, where $T$ denotes the length of the time series and $C$ denotes the number of sensor channels. In the experiments of this study, the time-series length is fixed at 296. During data processing, the recorded time variable is removed, resulting in 16 channels. Each sample is associated with a fault label, and the label space contains five categories, denoted as Class 0 to Class 4. Different categories represent the characteristic behaviors of the system under different fault modes or operating states.

To improve the training stability of the model under limited-sample conditions, the raw time-series data are first partitioned into different subsets and subsequently normalized using statistics calculated exclusively from the training set. It should be noted that the data used in this study have been pre-constructed into independent samples according to individual experimental trials. Each sample corresponds to a complete multivariate time series, with a length of 296 and 16 channels, and is accompanied by the corresponding fault label. The dataset contains 980 independent trial-level samples in total, including 146 normal-state samples, 214 added-weight-fault samples, 213 depth-sensor-fault samples, 199 severe-propeller-fault samples, and 208 mild-propeller-fault samples. Each sample corresponds to one complete multivariate experimental trial with a temporal length of 296 and 16 input channels after removal of the recorded time variable. No overlapping sliding-window segmentation is used to artificially increase the number of samples. Therefore, this study does not adopt an overlapping sliding-window sampling method; instead, modeling and analysis are directly conducted with each sample as the basic unit. For dataset partitioning, a stratified random sampling strategy is adopted to divide the dataset into training, validation, and test sets at ratios of 70%, 15%, and 15%, respectively. This ensures consistent distributions of different fault categories across the subsets and prevents class imbalance from interfering with model training and evaluation. During data preprocessing, the minimum and maximum values of each channel are calculated based on the training set, and min–max normalization is then applied to all data. This maps features with different physical units onto a unified scale interval, thereby improving the stability and convergence speed of model training. While ensuring data independence, this sample-level partitioning strategy also provides a unified data basis for fair comparisons among different models. It avoids data leakage caused by overlapping windows and thereby ensures the objectivity and reliability of the model evaluation results. Because the complete stratified split is independently regenerated for each random seed, the identities of the samples assigned to the training, validation, and test subsets change across repeated runs, while the class-wise sample counts remain fixed under the stratified 70%/15%/15% allocation.

Table 2. Experimental dataset.

Fault Type	No. of Independent Samples	Dimension	Label
Normal state	146	296 × 16	0
Added-weight fault	214	296 × 16	1
Depth sensor fault	213	296 × 16	2
Severe propeller fault	199	296 × 16	3
Mild propeller fault	208	296 × 16	4

therefore reports the total number of independent trial-level samples for each fault category, while the common splitting protocol is described in the text.

To ensure the fairness of comparison, all models are trained and tested under a unified experimental framework. During training, the Adam optimizer and gradient clipping are adopted. The validation subset is used exclusively for checkpoint and hyperparameter selection and is not used for class-difficulty estimation, class-difficulty-guided ordering, or pseudo-feature generation. To evaluate robustness against random initialization and sampling variance, the complete experimental procedure, including data partitioning and model training, is repeated with five independent random seeds (2021, 2022, 2023, 2024, and 2025). The main comparative experiments, noise-robustness experiments, and the main-module and internal-mechanism ablation experiments are reported as the mean ± standard deviation over five runs. The hyperparameter sensitivity analysis uses three independent random seeds, as specified in Section 3.4. In this paper, DCK-CAM-TCN [28], DRSN, WDCNN-BiLSTM [29], Transformer, Informer, and PatchTST are selected as comparative models. Among them, DCK-CAM-TCN mainly employs convolutional structures to extract local temporal patterns and exhibits strong inductive capability for local features. DRSN serves as the residual shrinkage network baseline and is used to verify the effectiveness of the attention enhancement and boundary-aware feature augmentation mechanisms. WDCNN-BiLSTM models temporal dependencies through a bidirectional recurrent structure, enabling full utilization of sequential contextual information when the number of samples is limited. Transformer directly models global correlations among different temporal positions and can be used to evaluate the diagnostic performance of global temporal dependency modeling methods. The Informer model reduces the modeling cost of long sequences through a sparse attention mechanism, representing a further evolution of this route toward efficient long-sequence modeling. PatchTST combines local segment representation with long- and short-term dependency modeling through patch partitioning, representing an important model in the evolution of time-series diagnosis from point-wise modeling to segment-based temporal representation. The above models cover several fault diagnosis technical routes, including convolutional modeling, recurrent modeling, residual shrinkage feature extraction, and attention-based modeling. The evaluation metrics used in this paper include test accuracy, Macro-F1, worst-class Accuracy, and the consistency between validation-set and test-set performance. On this basis, multiple noise disturbance conditions are further constructed to systematically evaluate diagnostic accuracy, class balance, and noise robustness.

To ensure a fair comparison, the hyperparameters of WDCNN-BiLSTM, Transformer, Informer, and PatchTST were optimized using the training and validation sets only. Candidate configurations were trained on the training subset, and the final configuration was selected according to validation accuracy. The test set was used only after model selection had been completed. The selected hyperparameter settings for each baseline model are detailed in

Table 3. Validation-selected hyperparameter configurations for the baseline models.

Model	Selected Hyperparameters
WDCNN-BiLSTM	Learning rate = $5\times {10}^{-4};$ dropout = 0.2; weight decay = $1\times {10}^{-4}$
Transformer	$d_{model}=256;$ heads = 8; encoder layers = 4; dropout = 0.2; learning rate = $1\times {10}^{-4}$
Informer	$d_{model}=64;$ heads = 4; encoder layers = 3; factor = 5; dropout = 0.1; learning rate = $5\times {10}^{-4}$
PatchTST	$d_{model}=256;$ heads = 8; encoder layers = 4; patch length = 32; dropout = 0.2; learning rate = $1\times {10}^{-4}$

, including the learning rate, dropout, weight decay, and other architecture-specific parameters.

Validation-based tuning improved the test Accuracy of PatchTST from approximately 72.97% under the original configuration to 76.08 ± 2.97% under the revised configuration. Although this confirms that the original setting underestimated the achievable performance of PatchTST, a substantial performance gap remains after tuning; therefore, the remaining difference is interpreted cautiously in terms of architecture–task compatibility rather than as a result of insufficient hyperparameter optimization alone.

3.2. Comparative Model Analysis

Figure 5a presents the five-seed test Accuracy of each model under the noise-free condition. Following the revised protocol, in which class-difficulty feedback is estimated exclusively from the complete training subset, AE-DRSN-ABHFA achieves a mean Accuracy of 95.27 ± 1.17%, compared with 94.73 ± 2.05% for DRSN, 91.62 ± 1.89% for DCK-CAM-TCN, 91.22 ± 2.14% for WDCNN-BiLSTM, 84.86 ± 1.23% for Informer, 83.11 ± 3.31% for Transformer, and 76.08 ± 2.97% for PatchTST. Thus, the proposed model achieves the highest mean clean-condition Accuracy among the evaluated models. These repeated-run results indicate that AE-DRSN-ABHFA maintains favorable discriminative capability under noise-free conditions while reducing the possibility that the reported performance is attributable to a single favorable initialization or data partition.

Figure 5b further compares the accuracy of each model on the validation and test sets. The relatively small discrepancy between validation and test set performance suggests stable within-dataset generalization under the evaluated HaiZhe data partitions. DCK-CAM-TCN and DRSN also maintain high recognition accuracy, whereas PatchTST exhibits a relatively pronounced decline in test-set performance.

Figure 5c shows the five-seed worst-class Accuracy. AE-DRSN-ABHFA achieves the highest mean value of 87.89 ± 5.94%, compared with 80.23 ± 14.72% for DRSN, 76.70 ± 2.26% for WDCNN-BiLSTM, 69.20 ± 5.91% for Informer, 68.58 ± 4.11% for Transformer, 67.64 ± 5.13% for DCK-CAM-TCN, and 61.02 ± 7.89% for PatchTST. The results suggest that the proposed model raises the recognition lower bound for difficult classes on average. The relatively large SD values observed for several models also highlight the importance of repeated-run evaluation when assessing class-level stability in this limited-sample setting.

Figure 5d presents the Macro-F1 comparison across five random seeds. AE-DRSN-ABHFA obtains the highest mean Macro-F1 of 0.9498 ± 0.0128, followed by DRSN at 0.9406 ± 0.0273, DCK-CAM-TCN at 0.9065 ± 0.0181, WDCNN-BiLSTM at 0.9050 ± 0.0215, Informer at 0.8450 ± 0.0164, Transformer at 0.8265 ± 0.0349, and PatchTST at 0.7597 ± 0.0319. Since Macro-F1 assigns equal weight to each category, these results suggest that the proposed method maintains a favorable balance across fault categories rather than obtaining its overall Accuracy primarily from easily classified classes.

To further examine the stability of the observed performance differences, two-sided paired t-tests were conducted using the results obtained under the same five random seeds. Because six paired comparisons were performed across two baselines and three evaluation metrics, Holm–Bonferroni correction was applied to control for multiple comparisons, and 95% confidence intervals (CIs) of the paired mean differences and Cohen’s $d_z$ were additionally calculated.

Compared with DRSN, the mean paired differences are +0.54 percentage points for Accuracy (95% CI: −1.07 to 2.15, Cohen’s $d_z$ = 0.416, raw p = 0.4050, Holm-adjusted p = 0.6159), +0.0091 for Macro-F1 (95% CI: −0.0126 to 0.0308, $d_z$ = 0.522, raw p = 0.3079, Holm-adjusted p = 0.6159), and +7.66 percentage points for worst-class Accuracy (95% CI: −3.41 to 18.73, $d_z$ = 0.859, raw p = 0.1272, Holm-adjusted p = 0.3816). None of these differences reach the conventional significance level after multiplicity correction.

Compared with DCK-CAM-TCN, AE-DRSN-ABHFA shows a mean paired Accuracy difference of +3.65 percentage points (95% CI: 0.42 to 6.88, $d_z$ = 1.404, raw p = 0.0349, Holm-adjusted p = 0.1396) and a Macro-F1 difference of +0.0432 (95% CI: 0.0127 to 0.0738, $d_z$ = 1.758, raw p = 0.0171, Holm-adjusted p = 0.0855). Although these two comparisons exhibit positive differences and large standardized effect sizes, they do not remain statistically significant after Holm–Bonferroni correction. In contrast, the worst-class Accuracy difference is +20.25 percentage points (95% CI: 17.42 to 23.08, $d_z$ = 8.876, raw p = 3.80 × 10⁻⁵, Holm-adjusted p = 2.28 × 10⁻⁴), and remains statistically significant after correction. Given the limited number of five paired repetitions, these confidence intervals and effect sizes should be interpreted cautiously, and the results should not be regarded as definitive evidence of universal superiority.

In addition, to verify the timeliness of the proposed model under the limited computational resources of AUVs, this paper conducts a comparative analysis of different models from three aspects: the number of trainable parameters, model size, and single-sample inference time. Specifically, the number of trainable parameters reflects the structural complexity of the model; a larger number of parameters generally implies higher storage and computational resource requirements during model training and deployment. Model size reflects the storage space occupied by the model file, which is directly related to the feasibility of deploying the model on embedded or edge devices. Single-sample inference time reflects the time cost required for the model to complete one fault diagnosis and is a key indicator for evaluating the timeliness of online diagnosis. During inference-time measurement, all models were operated in evaluation mode without gradient computation on the same RTX 4060 Laptop GPU. Inference efficiency was evaluated under Batch = 1 and Batch = 32, and the corresponding per-sample latency is reported in

Table 4. Comparison of model complexity and inference efficiency.

Model	Trainable Params	Model Size (MB)	Inference Time/Sample (ms), Batch = 1	Inference Time/Sample (ms), Batch = 32
AE-DRSN-ABHFA	143,407	0.551	2.711	1.422
DCK-CAM-TCN	156,749	0.602	3.182	2.489
DRSN	141,649	0.544	2.585	1.487
WDCNN-BiLSTM	227,718	0.871	1.458	0.381
Transformer	198,597	1.978	12.953	13.360
Informer	252,358	3.404	9.806	8.444
PatchTST	257,541	2.204	9.075	9.394

.

As shown in Table 4, AE-DRSN-ABHFA has 143,407 trainable parameters and a model size of only 0.551 MB, which is only slightly higher than that of DRSN and smaller than those of the other models. This indicates that, after introducing channel attention, temporal attention, and the ABHFA mechanism, the proposed model does not significantly increase the parameter scale and still maintains a lightweight network structure. In terms of inference time, the single-sample inference time of AE-DRSN-ABHFA is 2.711 ms under Batch = 1 and further decreases to 1.422 ms under Batch = 32. Both are within the millisecond range and are markedly lower than those of temporal models such as Transformer, Informer, and PatchTST. Although WDCNN-BiLSTM exhibits faster inference speed, both its parameter count and model size are larger than those of AE-DRSN-ABHFA. Although DRSN is close to AE-DRSN-ABHFA in terms of parameter count and inference time, it lacks attention enhancement and hard feature augmentation mechanisms, resulting in relatively insufficient class recognition stability under complex noise conditions. It should be noted that ABHFA is mainly used for boundary-aware pseudo-feature augmentation during the training stage. During testing and deployment, pseudo-feature generation is no longer performed; therefore, no additional online inference burden is introduced. These results indicate computational feasibility on the tested GPU platform and suggest potential for future lightweight onboard deployment; however, actual real-time performance on embedded AUV hardware remains to be validated.

3.3. Performance Analysis Under Noisy Conditions

To systematically evaluate the diagnostic robustness of the models under complex interference conditions, this study considers four noise-contaminated scenarios: Gaussian noise at 20 dB, Gaussian noise at 10 dB, impulse noise parameterized using a 20 dB reference SNR, and mixed Gaussian–sinusoidal interference parameterized using a 20 dB reference SNR. The proposed model is compared with WDCNN-BiLSTM, DRSN, DCK-CAM-TCN, Transformer, Informer, and PatchTST.

Artificial disturbances were introduced after data normalization. For each normalized training, validation, or test subset $X$, the mean signal power was calculated as follows:

$$P_s=\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{n}\left(X^2\right)$$

(23)

and the reference noise power corresponding to a specified SNR is defined as

$$P_n={\displaystyle \frac{P_s}{{10}^{{\mathrm{S}\mathrm{N}\mathrm{R}}_{dB}/10}}}$$

(24)

For Gaussian noise, independent zero-mean Gaussian disturbances were sampled from

$$n_G\sim \mathcal{N}\left(0,P_n\right)$$

(25)

For the impulse-noise condition, a random corruption mask with a probability of 0.03 was generated independently over the signal elements. The disturbance was zero at unselected positions, whereas the selected positions were perturbed using zero-mean Gaussian impulses with a standard deviation of $3\sqrt{P_n}$, thereby representing sparse transient high-amplitude disturbances. For the mixed-noise condition, Gaussian noise was combined with periodic sinusoidal interference. The sinusoidal component was defined as

$$n_s\left(t\right)=0.5\mathrm{s}\mathrm{i}\mathrm{n}\left(2\pi \cdot 15t\right)\sqrt{P_n}$$

(26)

where $t$ was uniformly sampled over the normalized temporal interval $\left[0,1\right]$. The mixed disturbance was therefore defined as

$$n_{\mathrm{m}\mathrm{i}\mathrm{x}}=n_G+n_s$$

(27)

For each investigated noisy condition, the corresponding disturbance was independently introduced into the normalized training, validation, and test subsets. Each model was therefore retrained, validation-selected, and tested under the same specified noisy condition rather than being trained only on clean data. Within each random seed and noise condition, all compared models used the same data split and the same artificially generated noise realization. Accordingly, the present experiments evaluate diagnostic performance under controlled matched noisy operating conditions rather than the out-of-distribution robustness of a clean-trained model to previously unseen test disturbances. For AE-DRSN-ABHFA, under each noisy condition, class-difficulty scores were still calculated exclusively from the corresponding complete training subset, while the validation subset was used only for checkpoint selection and did not participate in class-difficulty-guided ordering or pseudo-feature generation.

In the noise experiments, the models are trained using the Adam optimizer, together with a cosine annealing learning rate scheduling strategy. The main training parameters of AE-DRSN-ABHFA are set as follows: the learning rate is 0.0003, the batch size is 32, the number of training epochs is 200, and the weight decay coefficient is 0.0001. ABHFA is activated after 50 warm-up epochs, and the maximum number of accepted pseudo-features is set to four per mini-batch. The actual number of retained pseudo-features may be smaller than four depending on the available boundary candidates, valid same-class neighbors, and the semantic-validity check; the weight of the pseudo-feature classification loss is 0.15; the weight of the batch-hard triplet loss is 0.05; the triplet margin is 0.3; the boundary sample selection ratio is 0.30; and the momentum coefficient for difficult classes is 0.85. The experimental hardware platform is a Lenovo Legion R7000 laptop (manufacturer: Lenovo (Beijing) Limited, Beijing, China), equipped with an AMD Ryzen 7 8745H CPU, an NVIDIA GeForce RTX 4060 Laptop GPU with 8 GB of video memory, and 16 GB of DDR5 memory. The experimental programs were implemented using Python 3.8 and PyTorch 2.13.0, with key dependencies including NumPy 1.24.3, scikit-learn 1.2.2, Matplotlib 3.7.1, and Seaborn 0.12.2. The code is configured to prioritize the CUDA GPU for training and testing; when the GPU is unavailable, the program automatically switches to CPU execution. Each noise condition is evaluated over the same five independent random seeds, and

Table 5. Model performance analysis under noisy conditions.

Model	Clean Acc. (%)	Gaussian 20 dB Acc. (%)	Gaussian 10 dB Acc. (%)	Impulse 20 dB Acc. (%)	Mixed 20 dB Acc. (%)	Avg. Noisy Acc. (%)	Avg. Noisy Macro-F1	Avg. Noisy Worst-Class Acc. (%)
AE-DRSN-ABHFA	95.27 ± 1.17	89.43 ± 0.91	70.69 ± 6.44	92.03 ± 3.49	89.32 ± 2.21	85.37 ± 2.76	0.8253 ± 0.0317	75.39 ± 5.98
DRSN	94.73 ± 2.05	86.76 ± 3.50	63.11 ± 8.20	89.32 ± 3.04	86.62 ± 3.62	81.45 ± 3.31	0.8073 ± 0.0344	62.76 ± 6.94
DCK-CAM-TCN	91.62 ± 1.89	88.78 ± 3.33	65.14 ± 3.90	90.54 ± 1.07	86.89 ± 2.73	82.84 ± 2.34	0.8150 ± 0.0260	58.21 ± 10.77
WDCNN-BiLSTM	91.22 ± 2.14	81.08 ± 5.73	62.57 ± 1.13	86.76 ± 1.02	82.70 ± 1.40	78.28 ± 1.27	0.7751 ± 0.0128	59.38 ± 5.60
Transformer	83.11 ± 3.31	80.41 ± 1.85	64.46 ± 4.32	82.43 ± 1.91	80.41 ± 2.74	76.93 ± 1.92	0.7591 ± 0.0236	51.70 ± 5.79
Informer	84.86 ± 1.23	77.30 ± 3.53	56.35 ± 2.60	83.65 ± 1.30	77.57 ± 2.36	73.72 ± 1.47	0.7282 ± 0.0171	50.51 ± 2.03
PatchTST	76.08 ± 2.97	63.11 ± 2.51	42.70 ± 3.32	68.65 ± 5.46	62.84 ± 1.26	59.32 ± 2.60	0.5833 ± 0.0273	35.28 ± 5.67

reports mean ± SD. Figure 6, Figure 7, Figure 8 and Figure 9 are retained as representative single-run confusion-matrix and t-SNE visualizations for qualitative interpretation; the quantitative conclusions are based on the five-seed aggregate results in Table 5.

As shown in Table 5, AE-DRSN-ABHFA achieves the highest mean clean-condition Accuracy of 95.27 ± 1.17% among the evaluated models. Across the four noisy operating conditions, the proposed model achieves an average Accuracy of 85.37 ± 2.76%, an average Macro-F1 of 0.8253 ± 0.0317, and an average worst-class Accuracy of 75.39 ± 5.98%. These are the highest mean values among the compared models, suggesting that AE-DRSN-ABHFA maintains a favorable balance between overall diagnostic accuracy and class-level recognition performance under the evaluated noisy conditions.

Among the investigated disturbances, Gaussian noise at 10 dB produces the most pronounced performance degradation for all models. Under this severe condition, AE-DRSN-ABHFA achieves a mean Accuracy of 70.69 ± 6.44%, compared with 65.14 ± 3.90% for DCK-CAM-TCN, 64.46 ± 4.32% for Transformer, 63.11 ± 8.20% for DRSN, 62.57 ± 1.13% for WDCNN-BiLSTM, 56.35 ± 2.60% for Informer, and 42.70 ± 3.32% for PatchTST. Thus, the proposed model retains the highest mean Accuracy under the strongest Gaussian interference considered in this study, although the relatively large standard deviation of 6.44% indicates noticeable run-to-run variability.

Under Gaussian noise at 20 dB, AE-DRSN-ABHFA achieves a mean Accuracy of 89.43 ± 0.91%, exceeding DCK-CAM-TCN by 0.65 percentage points (88.78 ± 3.33%). Under impulse noise at 20 dB, the proposed model achieves 92.03 ± 3.49%, compared with 90.54 ± 1.07% for DCK-CAM-TCN, corresponding to a numerical advantage of 1.49 percentage points. Under mixed Gaussian–sinusoidal noise at 20 dB, AE-DRSN-ABHFA obtains 89.32 ± 2.21%, which is 2.43 percentage points higher than DCK-CAM-TCN at 86.89 ± 2.73%. Among the architecture-level models compared in Table 5, the proposed method achieves the highest mean Accuracy under all four evaluated noisy conditions, although the magnitude of the advantage varies across disturbance types.

When the four noisy conditions are aggregated, AE-DRSN-ABHFA achieves an average Accuracy of 85.37 ± 2.76%, compared with 82.84 ± 2.34% for DCK-CAM-TCN and 81.45 ± 3.31% for DRSN. The corresponding numerical improvements are 2.53 and 3.92 percentage points, respectively. Among the architecture-level models listed in Table 5, AE-DRSN-ABHFA also achieves the highest average noisy-condition Macro-F1 of 0.8253 ± 0.0317, compared with 0.8150 ± 0.0260 for DCK-CAM-TCN and 0.8073 ± 0.0344 for DRSN.

The advantage is more pronounced in worst-class recognition. AE-DRSN-ABHFA achieves an average noisy-condition worst-class Accuracy of 75.39 ± 5.98%, compared with 62.76 ± 6.94% for DRSN and 58.21 ± 10.77% for DCK-CAM-TCN, corresponding to numerical improvements of 12.63 and 17.18 percentage points, respectively. This result is consistent with the design objective of ABHFA, which uses class-difficulty information to determine the class-processing order and boundary-aware scores to identify ambiguous augmentation candidates during training.

Nevertheless, the repeated-run results also show that the numerical advantage of AE-DRSN-ABHFA is not equally pronounced under every disturbance type. In particular, the difference relative to DCK-CAM-TCN is only 0.65 percentage points under Gaussian noise at 20 dB, and the standard deviation reaches 6.44% under Gaussian noise at 10 dB. Therefore, the present results are interpreted as evidence of favorable average robustness and class-level recognition balance under the evaluated conditions rather than deterministic superiority for every individual data partition or noise realization.

The remaining temporal models exhibit comparatively lower average performance under the present limited-sample and noisy diagnostic setting. Transformer and Informer achieve average noisy-condition Accuracies of 76.93 ± 1.92% and 73.72 ± 1.47%, respectively, while PatchTST obtains 59.32 ± 2.60%. Its average worst-class Accuracy decreases to 35.28 ± 5.67%, indicating substantial difficulty in maintaining recognition of the hardest class under noisy conditions. Overall, the repeated-run results suggest that the combination of adaptive soft-threshold shrinkage, channel–temporal attention, and boundary-aware hard feature augmentation provides a favorable balance among overall Accuracy, Macro-F1, and worst-class recognition under the evaluated noisy operating conditions.

To further investigate the classification behavior and learned feature distributions of the proposed model under different noise conditions, representative confusion matrices and t-SNE visualizations are shown in Figure 6, Figure 7, Figure 8 and Figure 9. The confusion matrices illustrate class-specific recognition patterns, whereas the t-SNE visualizations provide a qualitative view of intra-class compactness and inter-class separability in the learned feature space. It should be emphasized that these visualizations correspond to representative individual runs, whereas the quantitative conclusions in Table 5 are based on the mean ± standard deviation across five independent random seeds.

Under the noise-free condition, as shown in Figure 6, the representative confusion matrix shows that most predictions are concentrated along the main diagonal, indicating high recognition accuracy across the different operating conditions. The corresponding t-SNE visualization shows that samples from different classes form relatively compact and distinguishable clusters, with comparatively clear inter-class boundaries. This qualitative observation is consistent with the five-seed quantitative result of 95.27 ± 1.17% reported in Table 5 and indicates that the proposed model maintains strong discriminative performance under the noise-free condition.

Under impulse noise at 20 dB, as shown in Figure 7, the representative confusion matrix still exhibits a strong diagonal distribution, while the t-SNE visualization indicates that most class clusters retain useful intra-class compactness and inter-class separation despite the presence of transient disturbances. Across five independent runs, AE-DRSN-ABHFA achieves a mean Accuracy of 92.03 ± 3.49% under this condition, representing a moderate decrease relative to the clean-condition Accuracy of 95.27 ± 1.17%. The result suggests that the proposed architecture retains favorable diagnostic performance in the presence of transient high-amplitude disturbances.

Under the strong Gaussian noise condition of 10 dB, the classification difficulty of both models increases substantially. In the representative run shown in Figure 8, DCK-CAM-TCN achieves an overall Accuracy of 68.24%. In particular, only 14 of the 32 Class 4 samples are correctly recognized, corresponding to a Class 4 accuracy of 43.75%. The remaining Class 4 samples are mainly misclassified as Class 3, with additional confusion involving Classes 1 and 2. The corresponding t-SNE visualization also shows substantial overlap among several fault categories, indicating reduced feature separability under strong Gaussian interference.

In contrast, the representative AE-DRSN-ABHFA result shown in Figure 9 achieves an overall Accuracy of 70.27%. Although noticeable overlap remains among some classes in the t-SNE space under severe Gaussian interference, the corresponding confusion matrix illustrates that useful class-specific discriminative information is still retained in this representative run. Figure 9 is used only for qualitative interpretation and should not be interpreted as the aggregate quantitative results.

The five-seed results in Table 5 show that AE-DRSN-ABHFA achieves a mean Accuracy of 70.69 ± 6.44% under Gaussian noise at 10 dB, compared with 65.14 ± 3.90% for DCK-CAM-TCN. Across the four evaluated noisy conditions, the proposed model achieves an average worst-class Accuracy of 75.39 ± 5.98%. These aggregate results indicate favorable class-level recognition under the evaluated disturbances, although the relatively large variability under severe Gaussian noise also highlights the sensitivity of the model to random initialization and data partitioning in particularly challenging conditions.

Overall, the noise-condition experiments indicate that AE-DRSN-ABHFA maintains favorable diagnostic performance across Gaussian, impulse, and mixed disturbances. AE-DRSN combines adaptive soft-threshold shrinkage with channel and temporal attention to enhance fault-sensitive feature extraction, while ABHFA directs feature augmentation toward ambiguous boundary samples through class prototypes, boundary-aware scoring, and training-derived class-difficulty-guided ordering. Among the architecture-level models in Table 5, the five-seed aggregate results show that the proposed model achieves the highest mean Accuracy under all four evaluated noisy conditions, together with the highest average noisy-condition Macro-F1 and worst-class Accuracy. Nevertheless, the relatively large variability under severe Gaussian noise indicates that the method does not guarantee deterministic superiority for every individual run.

3.4. Ablation and Mechanism Analyses

To systematically evaluate the effectiveness of each core module in the proposed AE-DRSN-ABHFA method and to further verify that the performance improvement of ABHFA does not originate from simple random perturbations, two sets of ablation experiments are designed in this section: a main-module ablation experiment, which analyzes the contributions of channel attention (CA), temporal attention (TA), and ABHFA to the overall performance; and an internal mechanism ablation experiment of ABHFA, which evaluates the roles of prototype guidance, boundary score, difficulty-guided ordering, pseudo-feature cross-entropy (CE) constraint, and triplet loss under the given experimental sample conditions.

The experiments adopt Accuracy, Macro-F1, and worst-class Accuracy as evaluation metrics. Among them, Accuracy reflects the overall classification performance of the model, Macro-F1 reflects the balance of multi-class recognition, and Worst-class Accuracy is used to measure the recognition lower bound for the most difficult class. Since unstable recognition of certain classes is likely to occur in AUV multi-fault diagnosis, comparing only the overall accuracy is insufficient; Macro-F1 and Worst-class Accuracy should also be considered in the analysis.

As shown in

Table 6. Main module ablation experiment.

Model	CA	TA	ABHFA	Acc. (%)	Macro-F1	Worst-Class Acc. (%)	ΔAcc. vs. DRSN (%)
DRSN	No	No	No	94.73 ± 2.05	0.9406 ± 0.0273	80.23 ± 14.72	0.00
DRSN + CA	Yes	No	No	94.73 ± 2.50	0.9447 ± 0.0246	85.28 ± 5.86	0.00
DRSN + TA	No	Yes	No	94.59 ± 1.26	0.9429 ± 0.0133	87.33 ± 5.02	−0.14
AE-DRSN	Yes	Yes	No	94.73 ± 1.11	0.9441 ± 0.0127	85.50 ± 5.80	0.00
DRSN + ABHFA	No	No	Yes	94.46 ± 1.21	0.9401 ± 0.0141	84.66 ± 7.58	−0.27
AE-DRSN-ABHFA	Yes	Yes	Yes	95.27 ± 1.17	0.9498 ± 0.0128	87.89 ± 5.94	0.54

, the DRSN baseline achieves an Accuracy of 94.73 ± 2.05%, a Macro-F1 of 0.9406 ± 0.0273, and a worst-class Accuracy of 80.23 ± 14.72%. Adding channel attention alone maintains the same mean Accuracy of 94.73 ± 2.50%, while increasing the Macro-F1 to 0.9447 ± 0.0246 and the worst-class Accuracy to 85.28 ± 5.86%. Temporal attention alone achieves an Accuracy of 94.59 ± 1.26%, corresponding to a difference of −0.14 percentage points relative to DRSN, while increasing the worst-class Accuracy to 87.33 ± 5.02%. Combining both attention mechanisms in AE-DRSN yields an Accuracy of 94.73 ± 1.11%, a Macro-F1 of 0.9441 ± 0.0127, and a worst-class Accuracy of 85.50 ± 5.80%. These results suggest that the attention mechanisms primarily improve class-level recognition balance rather than producing a large increase in overall Accuracy.

When ABHFA is directly incorporated into the DRSN backbone without the attention-enhanced feature extractor, DRSN + ABHFA achieves an Accuracy of 94.46 ± 1.21%, which is 0.27 percentage points lower than that of the DRSN baseline. Its Macro-F1 is 0.9401 ± 0.0141, while its worst-class Accuracy reaches 84.66 ± 7.58%, which is numerically 4.43 percentage points higher than that of DRSN. This result suggests that ABHFA alone provides limited improvement in overall Accuracy, while its effect is more evident in difficult-class recognition. In contrast, the complete AE-DRSN-ABHFA model achieves the highest mean Accuracy of 95.27 ± 1.17%, a Macro-F1 of 0.9498 ± 0.0128, and a worst-class Accuracy of 87.89 ± 5.94%. Relative to DRSN, the corresponding mean differences are +0.54 percentage points in Accuracy, +0.0092 in Macro-F1, and +7.66 percentage points in worst-class Accuracy. Compared with AE-DRSN, introducing ABHFA further increases the mean Accuracy by 0.54 percentage points, the Macro-F1 by 0.0057, and the worst-class Accuracy by 2.39 percentage points. Overall, these results support a complementary relationship between the attention-enhanced feature representation and boundary-aware hard-feature augmentation.

To further investigate the contributions of the individual mechanisms within ABHFA and their interactions, a systematic internal ablation study was conducted under the same five-seed data-partitioning and evaluation protocol. AE-DRSN without ABHFA was adopted as the reference backbone. Starting from this backbone, prototype-guided pseudo-feature augmentation, boundary-aware sample selection, class-difficulty-guided ordering, and the batch-hard triplet constraint were progressively incorporated. The pseudo-feature classification constraint was introduced together with prototype-guided augmentation so that the generated features could participate in supervised optimization. All variants used the same AE-DRSN backbone, data-splitting protocol, optimizer settings, and training schedule. Because these mechanisms jointly affect augmentation-target selection and feature-space optimization, the results are interpreted in terms of their complementary interactions rather than assuming that each individual addition must produce a monotonic improvement in all evaluation metrics.

The five-seed internal ablation results in

Table 7. Internal ablation experiments on AE-DRSN-ABHFA.

Variant	Prototype Guidance	Boundary Score	Difficulty-Guided Ordering	Pseudo-Feature CE	Triplet Loss	Acc. (%)	Macro-F1	Worst-class Acc. (%)	ΔAcc. vs. AE-DRSN (%)
AE-DRSN	No	No	No	No	No	94.73 ± 1.11	0.9441 ± 0.0127	85.50 ± 5.80	0.00
AE-DRSN +Prototype Aug	Yes	No	No	Yes	No	94.73 ± 0.88	0.9443 ± 0.0113	83.81 ± 8.31	−0.00 ± 0.68
AE-DRSN +Boundary Score	Yes	Yes	No	Yes	No	95.00 ± 0.37	0.9478 ± 0.0050	85.23 ± 1.18	0.27 ± 0.60
AE-DRSN + Difficulty-Guided Ordering	Yes	Yes	Yes	Yes	No	94.86 ± 1.40	0.9463 ± 0.0144	86.59 ± 5.59	0.13 ± 0.79
Full ABHFA	Yes	Yes	Yes	Yes	Yes	95.27 ± 1.17	0.9498 ± 0.0128	87.89 ± 5.94	0.54 ± 0.96

show that the contributions of the ABHFA components are complementary but not strictly monotonic. Starting from the AE-DRSN baseline, which achieves an Accuracy of 94.73 ± 1.11%, a Macro-F1 of 0.9441 ± 0.0127, and a worst-class Accuracy of 85.50 ± 5.80%, introducing prototype-guided augmentation alone maintains essentially the same overall Accuracy at 94.73 ± 0.88% and Macro-F1 at 0.9443 ± 0.0113, while the worst-class Accuracy decreases to 83.81 ± 8.31%. This indicates that prototype-guided feature generation by itself does not consistently improve difficult-class recognition.

After boundary-aware sample selection is incorporated, the mean Accuracy increases to 95.00 ± 0.37%, and the Macro-F1 increases to 0.9478 ± 0.0050, while the worst-class Accuracy recovers to 85.23 ± 1.18%. Further introducing class-difficulty-guided ordering results in an Accuracy of 94.86 ± 1.40%, a Macro-F1 of 0.9463 ± 0.0144, and a worst-class Accuracy of 86.59 ± 5.59%. Compared with the boundary-score variant, this configuration does not further increase the overall Accuracy or Macro-F1, but it raises the mean worst-class Accuracy by 1.36 percentage points. This suggests that the class-difficulty mechanism is more closely associated with improving recognition of difficult classes than with uniformly increasing overall classification performance.

The complete ABHFA configuration, which additionally incorporates the batch-hard triplet constraint, achieves the best overall balance among the internal variants, with an Accuracy of 95.27 ± 1.17%, a Macro-F1 of 0.9498 ± 0.0128, and a worst-class Accuracy of 87.89 ± 5.94%. Relative to the AE-DRSN baseline, the corresponding mean changes are +0.54 percentage points in Accuracy, +0.0057 in Macro-F1, and +2.39 percentage points in worst-class Accuracy. These results indicate that the individual ABHFA components should not be interpreted as providing strictly additive gains. Instead, prototype guidance, boundary-aware selection, class-difficulty feedback, and metric learning interact to produce a more balanced feature-augmentation mechanism, with the complete configuration providing the most favorable combined performance under the present experimental setting.

To further examine whether the additional feature-space optimization introduced by ABHFA provides advantages over simpler and widely used training strategies, AE-DRSN was additionally combined with Focal Loss, Hard Negative Mining (HNM), and SMOTE. All methods used the same AE-DRSN backbone, data-partitioning protocol, optimizer settings, and training schedule, and were evaluated using the same five random seeds (2021–2025). The results are reported as mean ± standard deviation in

Table 8. (a) Comparison with conventional training strategies under the clean condition. (b) Focused comparison between AE-DRSN + SMOTE and AE-DRSN-ABHFA under clean and noisy conditions.

(a)
Method			Accuracy (%)		Macro-F1		Worst-Class Accuracy (%)
AE-DRSN (Baseline)			94.73 ± 1.11		0.9441 ± 0.0127		85.50 ± 5.80
AE-DRSN + Focal Loss			95.10 ± 0.89		0.9468 ± 0.0118		84.51 ± 6.76
AE-DRSN + Hard Negative Mining			95.37 ± 2.01		0.9483 ± 0.0237		83.70 ± 9.66
AE-DRSN + SMOTE			95.92 ± 1.67		0.9553 ± 0.0191		85.70 ± 6.44
AE-DRSN-ABHFA (Proposed)			95.27 ± 1.17		0.9498 ± 0.0128		87.89 ± 5.94
(b)
Method	Clean Acc. (%)	Gaussian 20 dB	Gaussian 10 dB	Impulse 20 dB	Mixed 20 dB	Avg. Noisy Acc. (%)	Avg. Noisy Macro-F1	Avg. Noisy Worst-Class Acc. (%)
AE-DRSN + SMOTE	95.92 ± 1.67	89.52 ± 2.70	63.54 ± 2.83	91.84 ± 2.26	90.48 ± 3.47	83.84 ± 1.39	0.8290 ± 0.0222	66.23 ± 9.02
AE-DRSN-ABHFA	95.27 ± 1.17	89.43 ± 0.91	70.69 ± 6.44	92.03 ± 3.49	89.32 ± 2.21	85.37 ± 2.76	0.8253 ± 0.0317	75.39 ± 5.98
Δ (ABHFA − SMOTE)	−0.65	−0.09	+7.15	+0.19	−1.16	+1.53	−0.0037	+9.16

.

As shown in Table 8, under the noise-free condition, AE-DRSN + SMOTE achieves an Accuracy of 95.92 ± 1.67% and a Macro-F1 of 0.9553 ± 0.0191, slightly exceeding the 95.27 ± 1.17% Accuracy and 0.9498 ± 0.0128 Macro-F1 obtained by AE-DRSN-ABHFA. In contrast, the proposed method achieves the highest worst-class Accuracy of 87.89 ± 5.94%, compared with 85.70 ± 6.44% for SMOTE. Therefore, the clean-condition comparison does not indicate the universal superiority of ABHFA over generic feature interpolation. Rather, the results suggest that SMOTE is effective at increasing feature diversity in the present limited-sample setting, whereas the effect of ABHFA is more closely associated with maintaining the recognition performance of difficult classes.

To further clarify this distinction and directly evaluate the reviewer-raised noisy-condition scenario, AE-DRSN + SMOTE was additionally evaluated using the same four controlled noise conditions and the same five random seeds as AE-DRSN-ABHFA. The focused comparison is shown in Table 8. Across the four noisy conditions, AE-DRSN-ABHFA achieves an average Accuracy of 85.37 ± 2.76%, compared with 83.84 ± 1.39% for SMOTE, corresponding to a numerical improvement of 1.53 percentage points. The difference becomes substantially more pronounced for worst-class recognition. AE-DRSN-ABHFA obtains an average noisy-condition worst-class Accuracy of 75.39 ± 5.98%, whereas SMOTE obtains 66.23 ± 9.02%, corresponding to a 9.16-percentage-point improvement. In contrast, their average Macro-F1 values remain close, with 0.8253 ± 0.0317 for ABHFA and 0.8290 ± 0.0222 for SMOTE.

The largest overall-Accuracy difference between the two approaches is observed under Gaussian noise at 10 dB, which represents the strongest Gaussian interference considered in this study. AE-DRSN-ABHFA achieves an Accuracy of 70.69 ± 6.44% under this condition, whereas AE-DRSN + SMOTE achieves 63.54 ± 2.83%, corresponding to a numerical improvement of 7.15 percentage points. This result suggests that targeted boundary-aware augmentation becomes more beneficial when severe interference substantially degrades the separability between fault categories.

The advantage is not uniform across all disturbance types. Under Gaussian noise at 20 dB, the two methods achieve almost identical mean Accuracy, with 89.43% for ABHFA and 89.52% for SMOTE. A similarly small difference is observed under impulse noise at 20 dB, where ABHFA and SMOTE achieve 92.03% and 91.84%, respectively. Under mixed Gaussian–sinusoidal noise at 20 dB, SMOTE achieves a slightly higher mean Accuracy of 90.48%, compared with 89.32% for ABHFA. Therefore, these results should not be interpreted as evidence that ABHFA deterministically or uniformly outperforms SMOTE under every noisy operating condition.

Instead, the comparison reveals a more specific advantage that is consistent with the motivation of ABHFA. SMOTE primarily increases feature diversity through interpolation between same-class samples, whereas ABHFA uses class prototypes, boundary-aware scoring, and training-derived class-difficulty-guided ordering to preferentially augment ambiguous samples associated with difficult decision boundaries. As a result, the benefit of ABHFA is more evident under severe boundary-degrading interference and in the recognition lower bound of the most difficult class. In particular, although SMOTE remains highly competitive in terms of average Macro-F1, ABHFA increases average noisy-condition worst-class Accuracy by 9.16 percentage points while simultaneously improving average noisy-condition Accuracy by 1.53 percentage points.

Overall, the additional comparison indicates that the principal advantage of ABHFA over SMOTE does not lie in universally maximizing clean-condition overall Accuracy. Rather, ABHFA provides a more favorable balance between overall diagnostic robustness and difficult-class recognition when the feature space is affected by substantial interference. This finding is consistent with the specific design objective of the proposed boundary-aware and class-difficulty-guided augmentation mechanism.

Table 9. Sensitivity analysis of the main ABHFA hyperparameters.

Parameter	Evaluated Values	Accuracy (%)	Macro-F1	Worst-Class Accuracy (%)
Pseudo-feature loss weight	0, 0.05, 0.10, 0.15, 0.20	94.56–95.92	0.9443–0.9583	84.70–91.22
Triplet loss weight	0, 0.05, 0.10, 0.15	94.78–96.60	0.9466–0.9641	86.77–91.33
Boundary sample ratio	0.10, 0.20, 0.30, 0.40, 0.50	94.78–95.92	0.9466–0.9575	86.77–90.25
Neighbor-blending coefficient $\alpha$	0.10, 0.20, 0.25, 0.35, 0.50	94.78–96.83	0.9466–0.9678	86.77–92.40
Competing-prototype repulsion coefficient $\beta$	0.05, 0.10, 0.20, 0.30, 0.40	94.78–96.60	0.9466–0.9645	85.84–90.25
Difficulty momentum	0.70, 0.80, 0.85, 0.90, 0.95	94.78–96.60	0.9466–0.9652	86.77–91.41
Triplet margin	0.10, 0.20, 0.30, 0.40, 0.50	94.78–96.15	0.9466–0.9596	86.77–91.18

shows that the performance of AE-DRSN-ABHFA remains relatively stable across reasonable ranges of the investigated hyperparameters. Overall Accuracy and Macro-F1 vary within comparatively limited ranges for most settings, whereas worst-class Accuracy exhibits greater sensitivity to several parameters directly related to boundary-sample selection and pseudo-feature displacement. This observation is consistent with the design of ABHFA, because these parameters primarily determine which difficult samples are augmented and how far the generated features are displaced within the local boundary region.

Based on the validation-based model-development procedure and the consideration that the auxiliary objectives should not dominate the primary classification objective, the final configuration was fixed at a pseudo-feature classification loss weight of 0.15, a triplet-loss weight of 0.05, a triplet margin of 0.30, a boundary-sample ratio of 0.30, a neighbor-blending coefficient $\alpha$ of 0.25, a competing-prototype repulsion coefficient $\beta$ of 0.10, and a difficulty-momentum coefficient of 0.85. Importantly, these values were frozen before the final five-seed test evaluation. Therefore, the sensitivity analysis was used to characterize the local stability of the proposed mechanism rather than to select hyperparameters retrospectively from the test results.

As shown in

Table 10. Independent ablation of Stage 1 wide-convolution preprocessing.

Variant	Stage 1	Accuracy (%)	Macro-F1	Worst-Class Accuracy (%)	Parameters
AE-DRSN-ABHFA	Wide Conv, (k = 7)	95.27 ± 1.17	0.9498 ± 0.0128	87.89 ± 5.94	143,407
AE-DRSN-ABHFA w/o Wide Conv	Point-wise Conv, (k = 1)	93.38 ± 1.68	0.9304 ± 0.0197	81.88 ± 4.29	137,263
Δ (Full − Ablation)	—	+1.89	+0.0194	+6.01	+6144

, replacing the Stage 1 wide convolution ($k=7$) with a pointwise projection ($k=1$) decreases the mean Accuracy from 95.27 ± 1.17% to 93.38 ± 1.68%, corresponding to a mean paired improvement of 1.89 percentage points. Macro-F1 decreases from 0.9498 ± 0.0128 to 0.9304 ± 0.0197, while worst-class Accuracy decreases from 87.89 ± 5.94% to 81.88 ± 4.29%. The full model achieves higher values than the pointwise variant in four out of five matched runs for all three metrics. These results suggest that initial local temporal aggregation provides a consistent practical benefit to the subsequent robust feature-learning process, particularly for difficult fault categories.

The two variants are not exactly parameter-matched. The Stage 1 wide-convolution configuration contains 7296 trainable parameters, whereas the point-wise projection contains 1152, resulting in 6144 additional parameters. Correspondingly, the complete models contain 143,407 and 137,263 trainable parameters, respectively. Therefore, the observed performance difference should not be attributed exclusively to the enlarged temporal receptive field. Paired comparisons across the five matched random seeds show mean Full-minus-Ablation differences of 1.89 percentage points for Accuracy (95% CI: −1.21 to 5.00, $p=0.1660$, Cohen’s $d_z=0.76$), 0.0194 for Macro-F1 (95% CI: −0.0155 to 0.0542, $p=0.1973$, $d_z=0.69$), and 6.01 percentage points for worst-class Accuracy (95% CI: −5.40 to 17.42, $p=0.2175$, $d_z=0.65$). After Holm correction across the three outcomes, the adjusted $p$-value is 0.4980 for each comparison. Thus, although the full configuration exhibits moderate paired effect sizes and wins in four of five matched runs, the differences do not reach conventional statistical significance with only five paired repetitions. The Stage 1 ablation is therefore interpreted as evidence of a consistent performance trend rather than definitive statistical proof of superiority.

3.5. Robustness to Sensing Abnormalities

In addition to additive noise, practical AUV sensing systems may experience persistent sensor degradation or intermittent measurement loss. To evaluate robustness to such abnormalities, two additional test-stage experiments were conducted using a model trained only on the original clean training data: single-sensor gain degradation and random missing measurements. No degraded or incomplete test samples were used for model retraining or parameter adjustment. Therefore, these experiments evaluate the direct tolerance of a clean-trained model to previously unseen sensing abnormalities.

After removing the time variable, the model input contains 16 channels. Because PWM1–PWM4 are actuator/control commands rather than measurement variables, sensor degradation was evaluated only on the remaining twelve measurement channels. For each experiment, one measurement channel was degraded at a time while all remaining channels were kept unchanged. Degradation levels of 10%, 30%, and 50% were considered. The resulting data were normalized using the channel-wise minimum and maximum values calculated exclusively from the training set. For the sensing-abnormality experiments, the clean-trained checkpoint obtained using random seed 2024 was fixed for all degradation and missing-data evaluations. The quantitative results for single-sensor gain degradation at these three levels are summarized in

Table 11. Robustness under single-sensor gain degradation.

Degradation	Accuracy (%)	Macro-F1	Worst-Class Accuracy (%)	Accuracy Median [IQR] (%)
10% degradation	84.92 ± 24.83	0.8345 ± 0.2806	73.09 ± 32.64	96.60 [92.18, 96.60]
30% degradation	78.85 ± 29.65	0.7635 ± 0.3425	63.90 ± 36.36	94.90 [81.12, 95.92]
50% degradation	73.87 ± 30.55	0.7067 ± 0.3547	53.75 ± 40.04	91.84 [61.05, 95.41]

.

Mean ± SD represents variability across the 12 independently degraded measurement channels rather than variability across random model initializations. The results exhibit pronounced channel-dependent robustness. Although the mean Accuracy decreases with increasing degradation, the corresponding median remains substantially higher, reaching 91.84% even under 50% single-channel degradation. This difference indicates that the low mean values are driven mainly by a small number of highly informative channels rather than by uniform sensitivity across all sensors. In particular, pressure and voltage are the most sensitive channels: under 50% degradation, the associated Accuracies decrease to approximately 21.77% and 14.97%, respectively, whereas several attitude and angular-velocity channels retain approximately 95% Accuracy. Therefore, the proposed model should not be interpreted as universally robust to persistent failure of every sensor; its tolerance depends strongly on the affected measurement channel.

Random measurement loss was further introduced into the same 12 measurement channels at missing ratios of 10%, 20%, and 30%, whereas PWM1–PWM4 remained complete. For each missing ratio, ten independent missing-pattern realizations were generated. Missing values within each sensor channel were reconstructed using temporal linear interpolation; missing values at sequence boundaries were filled using the nearest valid observation. When an entire channel within one sample contained no valid observation, the corresponding training-set channel mean was used as a fallback. The reconstructed samples were subsequently normalized using the training-set Min-Max statistics. The diagnostic performance under random missing measurements with these three missing ratios is summarized in

Table 12. Robustness under random missing measurements.

Missing Ratio	Accuracy (%)	Macro-F1	Worst-Class Accuracy (%)
10%	92.86 ± 0.92	0.9244 ± 0.0095	76.77 ± 2.54
20%	89.73 ± 1.13	0.8917 ± 0.0120	71.94 ± 2.66
30%	85.24 ± 1.02	0.8434 ± 0.0111	62.90 ± 3.80

.

Values are reported as mean ± SD across ten independently generated missing-pattern realizations. The model retains an Accuracy of 92.86 ± 0.92% when 10% of the measurements are randomly missing. As the missing ratio increases to 20% and 30%, the Accuracy decreases to 89.73 ± 1.13% and 85.24 ± 1.02%, respectively. The relatively small variation across the ten independently generated missing patterns indicates a stable degradation trend rather than sensitivity to one particular missing-data realization. Together with the single-sensor degradation results, these experiments reveal two different robustness characteristics: the model exhibits useful tolerance to spatially distributed random measurement loss after temporal reconstruction, whereas persistent degradation of a small number of highly informative sensors can cause substantially larger performance deterioration. Accordingly, sensor-health monitoring, redundant sensing, abnormal-channel detection, or dedicated measurement-reconstruction mechanisms remain necessary for safety-critical deployment.

4. Discussion and Conclusions

This study addresses the problems of signal noise suppression, difficult-class recognition, and complex decision-boundary optimization in AUV fault diagnosis, and presents three main innovative contributions. First, an attention-enhanced deep residual shrinkage network, AE-DRSN, is proposed. By introducing channel attention and temporal attention mechanisms into the deep residual shrinkage network, the proposed model preserves the denoising advantage of soft-thresholding while improving the extraction capability of fault-sensitive features under noisy operating conditions. Second, an adaptive boundary-aware hard feature augmentation mechanism, ABHFA, is developed. Through class prototype construction, boundary-aware scoring, and class-difficulty-guided ordering, this mechanism performs directional augmentation of ambiguous boundary samples according to the resulting class-processing order. Third, a joint optimization strategy integrating pseudo-feature classification constraints and batch-hard triplet metric constraints is proposed, together with a staged dynamic augmentation mechanism. Repeated five-seed experiments show that AE-DRSN-ABHFA achieves a clean-condition mean Accuracy of 95.27 ± 1.17% and a Macro-F1 of 0.9498 ± 0.0128. Across the four evaluated noisy operating conditions, the proposed model achieves an average Accuracy of 85.37 ± 2.76%, an average Macro-F1 of 0.8253 ± 0.0317, and an average worst-class Accuracy of 75.39 ± 5.98%. Among the architecture-level models in the main comparison, it obtains the highest mean Accuracy under each of the four evaluated noise conditions. The additional comparison with SMOTE further indicates that the principal advantage of ABHFA lies in difficult-class recognition and average noisy-condition robustness rather than universal superiority across every metric and disturbance type. The improvement is particularly pronounced in average worst-class recognition, which is consistent with the use of class-difficulty-guided ordering together with boundary-aware augmentation. Nevertheless, the relatively large standard deviation observed under Gaussian noise at 10 dB indicates non-negligible run-to-run variability under severe random interference. Therefore, the present results support favorable average diagnostic performance under the evaluated conditions but do not establish deterministic superiority for every data partition or noise realization.

Several limitations should be considered when interpreting the present results. First, all principal experiments were conducted on a single public laboratory dataset. Although repeated stratified splits improve the statistical reliability of the evaluation within the HaiZhe dataset, they do not establish cross-platform or cross-environment generalization. Second, the Gaussian, impulse, and mixed-noise experiments represent controlled noisy operating conditions in which the corresponding artificial disturbance is introduced into the training, validation, and test subsets. These experiments therefore evaluate learning and diagnosis under specified noisy conditions rather than the out-of-distribution robustness of a clean-trained model to previously unseen disturbances. Real underwater measurements may exhibit more complex nonstationary and temporally correlated disturbances, long-term sensor drift, biofouling, communication loss, and platform-dependent interference. Third, the reported inference time is measured on an RTX 4060 Laptop GPU and should not be interpreted as validated onboard AUV latency. Fourth, the current dataset contains only five predefined and approximately balanced operating/fault categories; consequently, the present results do not establish robustness to unknown fault modes, severe class-frequency imbalance, or extremely few-shot minority classes. Finally, the additional sensing-abnormality experiments indicate that persistent degradation of highly informative sensors can still cause substantial performance deterioration. Therefore, the present results support the potential of AE-DRSN-ABHFA under the evaluated conditions but do not establish universal superiority or engineering-ready deployment across arbitrary AUV systems.

Future work can be extended in the following directions. First, the experiments in this study are mainly conducted on a public dataset; therefore, cross-condition, cross-task, or cross-platform data can be introduced in subsequent studies to verify the generalization capability of the model under different sea conditions and across different AUV platforms. In addition, online continuous diagnosis and early warning mechanisms can be further investigated to make the model more suitable for real-time health monitoring during actual AUV operation. Considering the limited computational resources of AUV platforms, future research should also focus on model compression and lightweight deployment, edge deployment, and real-time inference efficiency, so as to enhance the applicability and practical value of the proposed method in real-world engineering scenarios.