Modern industrial processes are characterized by highly interconnected and interdependent units, where disturbances and faults propagate dynamically across variables, equipment, and subsystems. This complex interaction makes root cause identification particularly challenging in real-world applications, such as semiconductor manufacturing. Although numerous existing methods are capable of performing root cause analysis, most fall short in effectively integrating temporal dynamics with spatial causal dependencies, thereby constraining their overall model performance. In this study, a novel network called graph autoencoder with causal relationship inference (GACRI) is proposed to tackle this challenge. Firstly, a dual-view decoder graph autoencoder (i.e., feature-level and causal-level decoders) is proposed to consider temporal features and spatial causal relationships concurrently. Secondly, a global index based on the learned features and the reconstructed residual space is developed for fault detection. The variable contribution degree is analyzed to isolate the fault variables via the reconstruction loss. Finally, a causal discovery network is designed to predict the causal relationship among fault variables for root-cause identification. The testing results on four processes (i.e., a numerical process, the Tennessee Eastman process, a semiconductor process and a hydraulic system) demonstrate the superior performance of GACRI in process monitoring.

Vibration sensing and infrared thermal image technology have been widely used in the health monitoring of machines. Multimodal fault diagnosis combining vibration and infrared thermal data has shown considerable potential. However, due to the heterogeneity of multimodal data, the performance of some models will significantly decrease under small samples. It is challenging to implement fault diagnosis based on multimodal data under small samples. Thus, a new DNN, i.e., adaptive multimodal feature interactive network (ADMMFI), is proposed in this study, where a multimodal feature interactive (MMFI) module is proposed to perform adaptive feature fusion under small samples. Firstly, a dimension reshaping module is proposed in ADMMFI to preserve discriminative thermal features during modality transformation. Secondly, a multimodal feature interaction module (MMFI) dynamically separates features into private and shared components, enabling adaptive fusion across modalities. Finally, the experimental results on a rolling bearing test bench and a rotor system test bench show that ADMMFI has a good performance on multimodal data fusion and feature extraction for machinery fault diagnosis under small samples. The fault recognition accuracy of ADMMFI on the two multimodal datasets was 79.17% and 95.15%, respectively. It demonstrates the effectiveness of ADMMFI compared with other DNNs.

Metal defect semantic segmentation is a crucial process for classifying and locating defects during the industrial production process, which holds paramount importance in elevating the quality of metal products. Recently, deep learning has exhibited impressive capabilities in identifying and segmenting defects on metal surfaces . However, the prevalent use of fully supervised segmentation techniques demands a substantial amount of annotated data for effective model training, which is hard to obtain in real scenarios. Additionally, most defects of metal products exhibit indistinct edge details, which hinders precise defect localization. In this study, a Crop-Paste and diffusion-based semi-supervised segmentation network (CPDNet) is proposed to identify pixel-level defects on metal surfaces by utilizing data that are both labeled and unlabeled. Firstly, a semi-supervised training method Crop-Paste is proposed to facilitate the learning of comprehensive semantic features from an extensive of unlabeled images and a restricted set of labeled images. Secondly, a frequency-directed diffusion model is proposed to recover high frequency features of defects to generate more accurate segmentation results. Lastly, an edge aware module is proposed in Sobel mean-teacher (M-T) UNet to improve the boundary information representation associated with defects. The experimental results on four datasets related to metal surface defects and a multimodal dataset show that CPDNet achieves a better performance in comparison with those state-of-the-art methods.

Valve-regulated lead–acid (VRLA) battery module is one of the important components of the uninterruptible power supply system in a large Internet data center. Battery health monitoring and anomaly diagnosis in time is significant to ensure the safe operation of a data center. In this article, a multitask joint network (MTJNet) is proposed to perform anomaly diagnosis and inconsistency identification simultaneously. First, an unsupervised learning-based encoder-decoder structure is proposed in MTJNet for battery inconsistency identification, where only health data are required for training. Second, the other branch with a classifier is further constructed to recognize the anomaly of the battery module. Third, a multitask joint training method is used to update the parameters of the two task models in MTJNet. The reconstruction error obtained by inconsistency identification task is fed into the anomaly diagnosis task. The predicted pseudo labels by the anomaly diagnosis task are feedback to inconsistency identification task to guide the encoder-decoder to provide the discriminate features. Finally, the effectiveness of MTJNet is verified on VRLA battery modules in a large data center. The experimental results illustrate that MTJNet is a good tool for anomaly diagnosis and inconsistency identification of VRLA battery modules.

Many machines generally operate under time-varying speed conditions. The significant and frequent changes of speeds often lead to missed defect detections and false alarms, which hinders the further applications of health monitoring technology. In this article, a novel self-supervised learning motivated network that is composed of extractor, projector, spyder and detector, i.e., EPSDNet, is proposed for self-adaptive health monitoring of machines under time-varying speed conditions. First, contrastive learning is used in EPSDNet to implement self-extraction of trend manifold over speeds, which adjusts the representational features and reduces their entropy. Second, a nonlinear regression network, i.e., Spyder, is proposed to identify the relationship between vibration signals and corresponding rotational speeds, which facilitates the adaptive feature alignment for each transient speed condition. Finally, a refined skip connection-based feature fusion is proposed to balance the calculation cost and signal reconstruction among encoder and decoder of detector. The real-time health indicator can be obtained based on the reconstruction error of detector for machinery. The effectiveness of EPSDNet is validated on two bearing test rigs with time-varying rotational speeds. The results show that EPSDNet outperforms other state-of-the-art methods for machinery health monitoring, which provides a novel approach to address the challenging issue of time-varying speed conditions.

Deep learning has made remarkable achievements in various applications in recent years. With the increasing computing power and the "black box" problem of neural networks, however, the development of deep neural networks (DNNs) has entered a bottleneck period. This article proposes a novel deep belief network (DBN) based on knowledge transfer and optimization of the network structure. First, a neural-symbolic model is proposed to extract rules to describe the dynamic operation mechanism of the deep network. Second, knowledge fusion is proposed based on the merge and deletion of the extracted rules from the DBN model. Finally, a new DNN, knowledge transfer-based sparse DBN (KT-SDBN) is constructed to generate a sparse network without excessive information loss. In comparison with DBN, KT-SDBN has a more sparse network structure and better learning performance on the existing knowledge and data. The experimental results in the benchmark data indicate that KT-SDBN not only has effective feature learning performance with 30% of the original network parameters but also shows a large compression rate that is far larger than other structure optimization algorithms.

Although deep neural networks (DNNs) have been widely applied in machinery fault diagnosis, the key problems of impulsive component extraction and noise filtering in the learning procedure are not addressed very well. Thus, a sparse representation network (SRNet) is developed to extract impulses from collected signals and then used for machinery fault recognition. For the purpose of improving the feature extraction capacity of SRNet, a convolutional sparse graph is developed in a sparse representation layer to suppress noise and reserve impulsive characteristics of signals. A selective residual learning is developed to effectively optimize gradient propagation and further enhance the feature learning performance of SRNet. Finally, the feature learning and fault classification capacity of SRNet is evaluated on two gearbox cases. The recognition accuracies of SRNet are 98.72% and 99.68% in two cases, respectively, demonstrating the effectiveness of SRNet compared with other DNNs.

Wafer map defect recognition (WMDR) is an important part of the integrated circuit manufacturing system. Accurate recognition of wafer map defects can help operators troubleshoot root causes of the abnormal process, and then accelerate the process adjustment. Although deep neural networks (DNNs) have been applied successfully in WMDR, class imbalance and lack of data with class labels affect their performance significantly. In view of these issues in semiconductor manufacturing processes, a new generative adversarial network (GAN), multigranularity GAN (MGGAN), is proposed for wafer map augmentation and enhancement. To alleviate instability and mode collapse of traditional GANs, the lightweight convolution and a two-way information interaction of three subnetworks are considered. MGGAN consists of an auxiliary feature extractor (AFE), a generator (G) and a discriminator (D) for wafer map generation and WMDR. First, a pretrained deep convolutional neural network (CNN), ResNet101, is employed as AFE to extract multigranularity features from wafer maps, which is used to guide the generator to reconstruct the real images. Second, in order to improve effectiveness of the adversarial training, a feature matching term is considered in the objective function of the feature generator to minimize the statistical difference between the real images and the generated images. Finally, MGGAN has been successfully applied to WMDR. The experiment results on an industrial dataset WM-811K demonstrate that MGGAN outperforms other typical GANs that aim to solve class imbalance problems and gains better recognition performance than those state-of-the-art DNNs on WMDR. This indicates effectiveness of MGGAN in image enhancement and generation. MGGAN obtains an accuracy of 88.02% on the original data and the pre-trained ResNet101 obtains 93.43% on the data enhanced by MGGAN.

Deep neural networks (DNNs), e.g., convolutional neural network (CNN), are able to learn effective features from wafer maps for dimensional reduction and feature extraction. However, very large image data are needed to train DNNs to obtain high generalization performance. It is still a difficult task due to the lack of sufficient labeled images with various defects. This article proposes a semisupervised deep transfer learning algorithm called joint feature and label adversarial network (JFLAN). JFLAN uses CNNs to extract transferable features of wafer maps and then introduces a multilayer domain adaptation and pseudolabel learning block based on the generative adversarial network (GAN). This effectively reduces the distribution discrepancy and the among-class distance of the transferable features. Finally, JFLAN transfers knowledge from wafer image source data collected offline and then achieves the goal of significantly improved accuracy of wafer defect recognition and realizes online adaptive defect recognition.Note to Practitioners—The defect recognition on wafer maps plays a key role to recognize fault sources of the semiconductor manufacturing processes. Transfer learning is able to use the existing labeled data to assist in the classification of unlabeled data and is very effective to solve the problem of small samples and nonstationary generalization errors. In particular, the infusion of adversarial learning in transfer learning will provide a new idea for deep feature learning. This article provides a novel method based on transfer learning to implement wafer map defect recognition (WMDR) to quickly identify defect root causes for yield enhancement. This article provides a novel way for quality control of semiconductor manufacturing processes based on transfer and adversarial learning.

Vibration signals are generally utilized for machinery fault diagnosis to perform timely maintenance and then reduce losses. Thus, the feature extraction on one-dimensional vibration signals often determines accuracy of those fault diagnosis models. These typical deep neural networks (DNNs), e.g., convolutional neural networks (CNNs), perform well in feature learning and have been applied in machine fault diagnosis. However, the supervised learning of CNN often requires a large amount of labeled images and thus limits its wide applications. In this article, a new DNN, one-dimensional residual convolutional autoencoder (1-DRCAE), is proposed for learning features from vibration signals directly in an unsupervised-learning way. First, 1-D convolutional autoencoder is proposed in 1-DRCAE for feature extraction. Second, a deconvolution operation is developed as decoder of 1-DRCAE to reconstruct the filtered signals. Third, residual learning is employed in 1-DRCAE to perform feature learning on 1-D vibration signals. The results show that 1-DRCAE has good signal denoising and feature extraction performance on vibration signals. It performs better on feature extraction than the typical DNNs, e.g., deep belief network, stacked autoencoders, and 1-D CNN.