Surface Defect Recognition (SDR) is crucial in the manufacturing industry. Recent advancements in deep learning and computer vision have significantly improved the precision and efficiency of SDR. However, the scarcity of defect samples presents a challenge for training deep learning models, making few-shot learning necessary for the SDR task. For this, this paper explores innovative contributions of Large Generative Models (LGM) to few-shot SDR by integrating Large Language Models (LLM) and Multimodal Generative Models (MGM). We present an LGM-based Training-Free Data Augmentation (LTDA) method to efficiently expand few-shot datasets without additional training. LTDA employs a conditional generation framework that leverages text instructions provided by LLM as conditions to guide the generation process of MGM. Additionally, multiple prompting templates have been designed for LLM to provide more precise instructions, thereby better guiding the conditional generation process. Finally, we employ two strategies for utilizing generated data to accomplish enhanced few-shot SDR by LTDA. The effectiveness of these approaches is demonstrated through two cases of defect recognition on steel surfaces, where the results demonstrate that the proposed method achieves an average accuracy improvement of 29% and 13% over the baseline method. The code will be released at https://github.com/jackiddd/LTDA.

Deep neural networks (DNNs) have achieved satisfactory performance in multiple fields. However, recent studies have shown that DNNs can be easily fooled by adversarial examples. To mitigate the threats caused by adversarial attacks, a highly effective strategy is to design detectors to reject adversarial examples. This article proposes an unsupervised class- and classifier-free adversarial detection method. It only takes unlabeled clean data for training to discriminate illegal samples, and does not require any knowledge about the adversarial examples, sample classes, and the original classifier. More specifically, motivated by the idea that adversarial examples may differ significantly from benign data in terms of sample structural information, we develop an adversarial detector that can simultaneously capture the residual information and the variable-wise structural relationships of data. After that, we design an attribute called data identity (ID) that combines the extracted residual and structural information of data to identify adversarial examples. We validate the superiority of the proposed method through detecting adversarial attacks on CIFAR-10 and ImageNet datasets, and the experimental results demonstrate that the performance of our model is the best among various state-of-the-art adversarial detectors. Besides, we also conduct visualization experiments to illustrate the role of structural information in detecting adversarial examples.

Probabilistic latent variable models (PLVMs), such as probabilistic principal component analysis (PPCA), are widely employed in process monitoring and fault detection of industrial processes. This article proposes a novel deep PPCA (DePPCA) model, which has the advantages of both probabilistic modeling and deep learning. The construction of DePPCA includes a greedy layer-wise pretraining phase and a unified end-to-end fine-tuning phase. The former establishes a hierarchical deep structure based on cascading multiple layers of the PPCA module to extract high-level features. The latter builds an end-to-end connection between the raw inputs and the final outputs to further improve the representation of the model to high-level features. After constructing the model structure of DePPCA, we first present the detailed training processes of the pretraining and fine-tuning stages, then clarify the theoretical merits of the proposed model from the perspective of variational inference. For process monitoring purposes, we develop two statistics based on the established DePPCA. The monitoring performance of these two statistics can remain superior even if the features extracted by DePPCA are significantly compressed to univariate. This makes the feature extraction process and online monitoring procedure of DePPCA quite fast. In other words, the proposed DePPCA can achieve accurate and efficient process monitoring by only extracting one feature for each sample. Finally, the effectiveness of DePPCA is evaluated on the Tennessee Eastman (TE) process and the multiphase flow (MPF) facility.

In recent years, deep neural networks (DNNs) have been widely applied in fault classification tasks. Their adversarial security has received attention, but little consideration has been given to the robustness of adversarial attacks against imperfect DNNs. Owing to the data scarcity and quality deficiencies prevalent in industrial data, the performance of DNNs may be severely constrained. In addition, black-box attacks against industrial fault classification models have difficulty in obtaining sufficient and comprehensive data for constructing surrogate models with perfect decision boundaries. To address this gap, this article analyzes the outcomes of adversarial attacks on imperfect DNNs and categorizes their decision scenarios. Subsequently, building on this analysis, we propose a robust adversarial attack strategy that transforms traditional adversarial attacks into an iterative targeted attack (ITA). The ITA framework begins with an evaluation of DNNs, during which a classification confidence score (CCS) is designed. Using the CCS and the prediction probability of the data, the labels and sequences for targeted attacks are defined. The adversarial attacks are then carried out by iteratively selecting attack targets and using gradient optimization. Experimental results on both a benchmark dataset and an industrial case demonstrate the superiority of the proposed method.

Due to the complex reaction mechanisms of industrial process units, causality and correlations exist between industrial process variables. Causal discovery algorithms have been utilized to discover the knowledge on variable relationships and guide process modeling and control optimization. However, most of them are limited by strict assumptions, such as linear relationships, additive noise, steady-state process, etc. Therefore, these methods cannot gain good performance for most practical industrial processes. To solve these problems, a novel weight comparison causal mining (WCCM) algorithm is proposed in this article for industrial causal graph discovery. It first trains a group of hidden layer neural networks with process data, then mines an undirected skeleton of the process variables according to the comparison of the network weights, and further determines the causal directions of the undirected edges in the skeleton to get a directed causal graph. The effectiveness of WCCM is verified on a benchmark and a practical industrial case from the urea synthesis process. The undirected and direct edges mined by WCCM show high consistency with the ground truths. Moreover, the causal discovery results of WCCM are utilized to guide the feature selection of soft sensor modeling, resulting in improved prediction accuracy and enhanced model interpretability.

The salient progress of deep learning is accompanied by nonnegligible deficiencies, such as: 1) interpretability problem; 2) requirement for large data amounts; 3) hard to design and tune parameters; and 4) heavy computation complexity. Despite the remarkable achievements of neural networks-based deep models in many fields, the practical applications of deep learning are still limited by these shortcomings. This article proposes a new concept called the lightweight deep model (LDM). LDM absorbs the useful ideas of deep learning and overcomes their shortcomings to a certain extent. We explore the idea of LDM from the perspective of partial least squares (PLS) by constructing a deep PLS (DPLS) model. The feasibility and merits of DPLS are proved theoretically, after that, DPLS is further generalized to a more common form (GDPLS) by adding a nonlinear mapping layer between two cascaded PLS layers in the model structure. The superiority of DPLS and GDPLS is demonstrated through four practical cases involving two regression problems and two classification tasks, in which our model not only achieves competitive performance compared with existing neural networks-based deep models but also is proven to be a more interpretable and efficient method, and we know exactly how it improves performance, how it gives correct results. Note that our proposed model can only be regarded as an alternative to fully connected neural networks at present and cannot completely replace the mature deep vision or language models.

Adversarial attack can fabricate imperceptible fake samples to cheat a well-trained artificial intelligence (AI) model, and it has shown strong destructive power in many classification tasks. In real-world AI applications, there is another popular kind of machine learning paradigm—regression. The threats of adversarial attack may also exist in the regression scenario, however, the research on the adversarial vulnerability of the regression model has been basically neglected. This article first systematically explores the adversarial attack on regression problems. Starting from analyzing the difference between the attacking classification models and regression systems, we show the existing attack framework of classification problems is unsuitable for attacking regression systems. Then, we discuss the essence of regression tasks and design an appropriate attack objective for regression problems. After that, we propose two algorithms with different properties based on gradient optimization to achieve the attack objective. The proposed attack methods are evaluated on three real-world regression cases, and the results show that our attacks can successfully make the prediction deviate a lot from its original value by only exerting a tiny perturbation on the inputs. Finally, we conduct further experiments and analyses to discuss the effectiveness and characteristics of the proposed methods from various perspectives.

Neural-network-based soft sensors are widely employed in the industrial process. Such models have great significance to smart manufacturing. Considering the strict requirements of industrial production, it is vital to ensure the safety and robustness of these models in their actual deployment. However, recent research has shown that neural networks are quite vulnerable to adversarial attacks. By imposing tiny perturbation to the original sample, the fabricated adversarial sample can cheat these models to make wrong decisions. Such a phenomenon may bring serious trouble to the practical application of soft sensors. This article focuses on the adversarial attacks on industrial soft sensors. For the first time, we verify and analyze the effectiveness and deficiencies of the existing attack methods in the industrial soft sensor scenario. Based on solving these defects, this article proposes a novel perspective for attacking soft sensors. We analyze the optimization mechanism behind this new idea and then design two algorithms to perform attacks. The proposed methods more conform to the actual situation. Besides, compared with the existing approaches, the proposed methods have potentials to cause severer damages since their attacks are not only more concealed but also more likely to cheat the technicians to execute wrong operations. The research and analyses of the proposed methods lay a solid foundation for more thorough defenses against various attacks, which is quite necessary for making the deployed soft sensors more robust and secure.

Data-driven process monitoring based on latent variable models are widely employed in industry. This article proposes a novel monitoring framework for latent variable models using hierarchical feature extraction, Bayesian inference, and weighting strategy. We first establish a deep structure to implement hierarchical latent variables extraction, the extracted features are used to construct diverse monitoring statistics. Then, we utilize Bayesian inference and proper weighting strategy to fuse various useful information. In line with the different characteristics of principal component analysis (PCA) and independent component analysis (ICA), we construct a deep PCA-ICA model for process monitoring according to the proposed framework. The deep PCA-ICA model performs hierarchical feature extraction, which can simultaneously extract deep Gaussian information and deep non-Gaussian information. The features extracted by different layers are then transformed to posterior probabilities through Bayesian inference. After that, different posterior probabilities are combined through appropriate weighting strategy to build new probabilistic statistic, which can give more synthetic monitoring results. Moreover, the Bayesian inference and weighting strategy are further used to integrate the advantages of different models by transforming various probabilistic statistics into an overall monitoring index, which can comprehensively indicate the process status. The Tennessee Eastman process is used to validate the superiority of the proposed model over the existing methods. Besides, the extracted features are further analyzed to show the effectiveness and benefits of the deep hierarchical feature extraction structure.