In the integrated circuit (IC) packaging, the surface defect detection of flexible printed circuit boards (FPCBs) is important to control the quality of IC. Although various computer vision (CV)-based object detection frameworks have been widely used in industrial surface defect detection scenarios, FPCB surface defect detection is still challenging due to non-salient defects and the similarities between diverse defects on FPCBs. To solve this problem, a decoupled two-stage object detection framework based on convolutional neural networks (CNNs) is proposed, wherein the localization task and the classification task are decoupled through two specific modules. Specifically, to effectively locate non-salient defects, a multi-hierarchical aggregation (MHA) block is proposed as a location feature (LF) enhancement module in the defect localization task. Meanwhile, to accurately classify similar defects, a locally non-local (LNL) block is presented as a SEF enhancement module in the defect classification task. What is more, an FPCB surface defect detection dataset (FPCB-DET) is built with corresponding defect category and defect location annotations. Evaluated on the FPCB-DET, the proposed framework achieves state-of-the-art (SOTA) accuracy to 94.15% mean average precision (mAP) compared with the existing surface defect detection networks. Soon, source code and dataset will be available at https://github.com/SCUTyzy/decoupled-two-stage-framework.

Industrial process monitoring is a significant task and has started to get better solved by deep learning. However, ensuring that the learned deep features from process data are effective and interpretable for the monitoring task remains a challenge. In this article, a slow feature analysis-aided autoencoder (SFA-AE) is proposed for interpretable process monitoring. The SFA-AE, which combines the advantages of SFA and an autoencoder (AE), enables the learning of deep slow variation patterns from the high-level features extracted by the AE. Particularly, the AE additionally incorporates in convolutional long short-term memories for representing the spatiotemporal relationships of multivariate time-series process data. These slow patterns that are able to heavily highlight abnormal states are used to construct monitoring statistics and then improve the reliability of fault detection. Most importantly, variable relative attention and its control limit based on self-attention mechanisms are developed for fault diagnosis. The proposed self-attention emphasizes the different contributions of process variables to fault detection, making it possible to interpret fault variables in an end-to-end way. In the experiments on the Tennessee Eastman process, the proposed process monitoring method obtains superior fault detection and diagnosis performance than state-of-the-art ones.

Deep neural networks (DNNs) have achieved superior performance in diagnosing process faults, but they lack robustness when encountering novel fault types absent from training sets. Such unknown faults commonly appear in industrial settings, and conventional DNNs often misclassify them as one of the known fault types. To address this limitation, we formulate the concept of open-set fault diagnosis (OSFD), which seeks to distinguish unknown faults from known ones while correctly classifying the known faults. The primary challenge in OSFD lies in minimizing both the empirical classification risk associated with known faults and the open space risk without access to training data for unknown faults. In order to mitigate these risks, we introduce a novel approach called hypersphere-guided reciprocal point learning (SRPL). Specifically, SRPL preserves a DNN for feature extraction while constraining features to lie on a unit hypersphere. To reduce empirical classification risk, it applies an angular-margin penalty that explicitly increases intra-class compactness and inter-class separation for known faults on the hypersphere, thereby improving discriminability among known faults. Additionally, SRPL introduces reciprocal points on the hypersphere, with each point acting as a classifier by occupying the extra-class region associated with a particular known fault. The interactions among multiple reciprocal points, together with the deliberate synthesis of unknown fault features on the hypersphere, serve to lower open-space risk: the reciprocal-point interactions provide an indirect estimate of unknowns, and the synthesized unknowns provide a direct estimate, both of which enhance distinguishability between known and unknown faults. Extensive experimental results on the Tennessee Eastman process confirm the superiority of the proposed method compared to state-of-the-art OSR algorithms, e.g., an 82.32% AUROC score and a 71.50% OSFDR score.