Yue Zhuo

This paper proposes ABIGX (Adversarial fault reconstruction-Based Integrated Gradient eXplanation), a unified framework for explainable fault detection and classification (FDC). ABIGX builds on the foundational principles of established fault diagnosis methods, including contribution plots (CP) and reconstruction-based contribution (RBC), while extending their applicability to general FDC models and improving fault explanation results. Central to ABIGX is the Adversarial Fault Reconstruction (AFR) method, which rethinks fault reconstruction from the perspective of adversarial attacks, introducing a novel fault index applicable to both fault detection and classification tasks. In fault detection, we theoretically bridge ABIGX with conventional fault diagnosis methods by proving that CP and RBC are the linear specifications of ABIGX. For fault classification, we address the challenge of fault class smearing, an inherent issue that can obscure accurate explanations. We demonstrate that ABIGX effectively mitigates this issue, outperforming current gradient-based explanation methods. The experiments evaluate the explanations of FDC by quantitative metrics and intuitive illustrations. The results validate the generality and accuracy of AFR, and show that ABIGX provides more comprehensive and precise explanations across various FDC models, offering a significant improvement over existing methods.

In the field of intelligent manufacturing and industrial big data, data-driven Industrial Intelligence Models (IIMs) based on machine learning have become indispensable for modern industrial systems. While IIMs are renowned for their effective learning capabilities, a critical challenge persists: Their performance is severely compromised by the substantial quality gap between raw industrial data and model-ready data. To address this core issue, Industrial Data Augmentation (IDA) has emerged as a transformative solution, yet existing research lacks systematic frameworks and implementation guidelines. This paper presents the first comprehensive survey establishing IDA as an independent research domain. We propose a novel taxonomy categorizing IDA methods by transformation-based, interpolation-based, and distribution estimation-based approaches. Beyond methodology analysis, we conduct a systematic review of frontier IDA applications spanning key performance indicator prediction, anomaly monitoring, fault diagnosis, and defect detection. Significantly, an open-source IDA toolbox implementing twenty IDA algorithms is introduced to facilitate ongoing development and application, available at https://github.com/3uchen/IdaLy. Finally, this paper highlights current challenges and future prospects for the IDA, seeking to motivate and steer further research in this area.

Numerous methodologies have been proposed to tackle the pervasive issue of catastrophic forgetting. However, most current methods tend to segregate the historical and current information during the training process. In contrast, we acknowledge the significance of global properties throughout the entire incremental learning process to consolidate memory. We propose a feature-level bi-update mechanism that facilitates concurrent learning and revision by fostering mutual awareness between learning and revising. To learn new knowledge, we design a cross-session contrastive learning function. This function benefits from contrastive learning, facilitating the acquisition of robust representations and inherently avoiding collapse. Notably, it accomplishes this without necessitating asymmetric network architectures or relying on data augmentation techniques. To revise prior knowledge, we propose a prototype update approach. This method involves the dynamic interaction between previous semantic representations and the ongoing learning process, thus aligning decision boundaries and promoting coherence. In addition, a novel classification method compatible with the aforementioned objectives is introduced. Extensive experiments have been conducted to validate the proposed method, and the results show that the proposed method reaches new state-of-the-art levels.

Feature attribution explains Artificial Intelligence (AI) at the instance level by providing importance scores of input features’ contributions to model prediction. Integrated Gradients (IG) is a prominent path attribution method for deep neural networks, involving the integration of gradients along a path from the explained input (explicand) to a counterfactual instance (baseline). Current IG variants primarily focus on the gradient of explicand's output. However, our research indicates that the gradient of the counterfactual output significantly affects feature attribution as well. To achieve this, we propose I terative G radient path I ntegrated G radients (IG ² ), considering both gradients. IG ² incorporates the counterfactual gradient iteratively into the integration path, generating a novel path ( GradPath ) and a novel baseline ( GradCF ). These two novel IG components effectively address the issues of attribution noise and arbitrary baseline choice in earlier IG methods. IG ² , as a path method, satisfies many desirable axioms, which are theoretically justified in the paper. Experimental results on XAI benchmark, ImageNet, MNIST, TREC questions answering, wafer-map failure patterns, and CelebA face attributes validate that IG ² delivers superior feature attributions compared to the state-of-the-art techniques.

Feature attribution explains Artificial Intelligence (AI) at the instance level by providing importance scores of input features’ contributions to model prediction. Integrated Gradients (IG) is the prevalent path attribution method for deep neural networks, involving the integration of gradients along a path from the explained input (explicand) to a counterfactual instance (baseline). Nonetheless, current IG variants exclusively consider the gradient of explicand’s output, but we have discovered that the gradient of the counterfactual output also exerts a substantial influence on feature attribution. To achieve this, we propose Iterative Gradient path Integrated Gradients (IG ² ), considering both two gradients. IG ² incorporates the counterfactual gradient iteratively into the integration path, consequently obtaining a novel path (GradPath) and a novel baseline (GradCF). These two novel IG components effectively address the issues of attribution noise and arbitrary baseline choice in earlier IG methods. IG ² , as a path method, stratifies many desirable axioms, which are theoretically justified in the paper. Experimental results on XAI benchmark, ImageNet, MNIST, TREC questions answering, wafer-map failure patterns, and CelebA face attributes validate that IG ² delivers superior feature attributions compared to the state-of-the-art techniques.

For explainable fault detection and classification (FDC), this paper proposes a unified framework, ABIGX (Adversarial fault reconstruction-Based Integrated Gradient eXplanation). ABIGX is derived from the essentials of previous successful fault diagnosis methods, contribution plots (CP) and reconstruction-based contribution (RBC). It is the first explanation framework that provides variable contributions for the general FDC models. The core part of ABIGX is the adversarial fault reconstruction (AFR) method, which rethinks the FR from the perspective of adversarial attack and generalizes to fault classification models with a new fault index. For fault classification, we put forward a new problem of fault class smearing, which intrinsically hinders the correct explanation. We prove that ABIGX effectively mitigates this problem and outperforms the existing gradient-based explanation methods. For fault detection, we theoretically bridge ABIGX with conventional fault diagnosis methods by proving that CP and RBC are the linear specifications of ABIGX. The experiments evaluate the explanations of FDC by quantitative metrics and intuitive illustrations, the results of which show the general superiority of ABIGX to other advanced explanation methods.

The visual anomaly diagnosis can automatically analyze the defective products, which has been widely applied in industrial quality inspection. The anomaly classification can classify the defective products into different categories. However, the anomaly samples are hard to access in practice, which impedes the training of canonical machine learning models. This paper studies a practical issue that anomaly samples for training are extremely scarce, i.e., few-shot learning (FSL). Utilizing the sufficient normal samples, we propose PatchProto networks for few-shot anomaly classification. Different from classical FSL methods, PatchProto networks only extract CNN features of defective regions of interest, which serves as the prototypes for few-shot learning. Compared with basic few-shot classifier, the experiment results on MVTec-AD dataset show PatchProto networks significantly improve the few-shot anomaly classification accuracy.

While the data-driven fault classification systems have achieved great success and been widely deployed, machine-learning-based models have recently been shown to be unsafe and vulnerable to tiny perturbations, i.e., adversarial attack. For the safety-critical industrial scenarios, the adversarial security (i.e., adversarial robustness) of the fault system should be taken into serious consideration. However, security and accuracy are intrinsically conflicting, which is a trade-off issue. In this article, we first study this new trade-off issue in the design of fault classification models and solve it from a brand new view, hyperparameter optimization (HPO). Meanwhile, to reduce the computational expense of HPO, we propose a new multiobjective (MO), multifidelity (MF) Bayesian optimization (BO) algorithm, MMTPE. The proposed algorithm is evaluated on safety-critical industrial datasets with the mainstream machine learning (ML) models. The results show that the following hold: 1) MMTPE is superior to other advanced optimization algorithms in both efficiency and performance and 2) fault classification models with optimized hyperparameters are competitive with advanced adversarially defensive methods. Moreover, insights into the model security are given, including the model intrinsic security properties and the correlations between hyperparameters and security.

Data-driven fault detection and classification (FDC) systems play an important role in ensuring the stability and security of modern industry. However, the security issue of the data-driven FDC itself poses new challenges, where the model prediction can be seriously damaged by maliciously manipulated imperceptible perturbations, known as the adversarial attack. Since the adversarial attacks may threaten the FDC models and even the whole safety-critical industrial systems, there is an urgent need for the guarantee of data-driven model security. In this article, a scheme is presented for formally and completely verifying the security properties by treating the target problem as a convex mathematical programming. The major contribution on methodology is the verification under multiple norms via multiobjective optimization with a novel Pareto front approximation algorithm. Moreover, this work studies security verification for both supervised (fault classification) and unsupervised (fault detection) models under all mainstream norms simultaneously. For four basic data-driven models on two industrial datasets, our exclusive verification scheme provides deep and novel security insight into FDC systems. Moreover, we compare with related works to validate the algorithm performances of verification and Pareto front approximation.

Adversarial attack is to craft tiny perturbations on inputs, causing neural networks to give incorrect outputs with high confidence, while adversarial training is the de facto most successful method to obtain robust neural networks. Nevertheless, adversarial training suffers from robust overfitting: the increasing robustness gap between train and test datasets. This paper aims at reducing this robust overfitting (i.e., improving adversarially robust generalization / generalized adversarial robustness). Firstly, we study the robust bias-variance trade-off and find that the robust overfitting even happens on the simple dataset (100MNIST), which is considered free from that in previous research. Next, based on 100MNIST and Gaussian data, the correlation between feature augmentation and robust overfitting is analyzed on a theoretical basis. A novel insight is obtained: the augmentation on robust features can improve generalized adversarial robustness, through alleviating the over-optimization on non-robust features. Then, we propose a regularization term, Feature Alignment (FA), to guide synthesized sample directions. Adversarial FA aided Generative Adversarial Network, AFAGAN, generates samples aligning robust features, which can train more adversarially robust generalized models. Experiments on CIFAR10 show that augmented data of AFAGAN significantly improves the generalized adversarial robustness.