Incremental learning based fault diagnosis is effective to learn from continuous industrial data on an ongoing basis. However, in the case of non-stationary industrial processes, new data distribution often gradually shifts away from that of historical data, due to equipment aging and manufacturing strategies, etc. Thus, conventional incremental methods with the identical independent distribution (i.i.d) assumption may no longer promise satisfied diagnosis performance. This paper concerns the conditional drift phenomenon, a relaxation of the i.i.d assumption, in which the conditional distribution of industrial data changes within different time. From a mathematical point of view, we firstly give the problem formulation of conditional drift and introduce a target mapping strategy for drift adaptation, under the minimum risk criteria. Then, following this strategy, an incremental diagnosis model with adaptation ability is designed. Particularly, a transformation matrix keeps matching the distributions of historical and new data. Thus, our method can quickly adapt to the conditional drift and be more robust against the evolving environment. The proposed method is applied for diagnosing faults in two industrial processes to demonstrate its effectiveness.

This study investigates the sample value imbalance problem of process monitoring. A fault detection approach based on variable selection and support vector data description (SVDD) is developed for efficient process monitoring. First, Kullback–Leibler divergence serves as the variable selection algorithm, which highlights the most beneficial information about the concerned faults. The attained variables are segmented by block division to avoid faults information being covered in single space monitoring, so that the relevant variables and the most beneficial information are concentrated in the same block. Then, Kernel principal component analysis is applied in each block to address the challenge that variables may still be high-dimensional and nonlinear. After that, the monitoring result is given based on the proposed SVDD with a restructured radius index, which is more sensitive to the fault. As demonstrated from experimental results on the Tennessee Eastman process, this method is effective and outperforms counterparts with higher mean fault detection rate. Note to Practitioners—Recently, multivariate statistical process monitoring (MSPM) has attracted much attention. In general, MSPM incorporates all variables for the large-scale process. However, only a small number of variables are fault-dependent. Namely, the sample value imbalance problem is encountered in application. In this scenario, the monitoring performance degrades and the online computational complexity increases. To this end, a SVDD-based fault detection method, which considers the fault-related variables, is proposed for process monitoring. The proposed method is verified by the Tennessee Eastman process and it is more sensitive to the concerned fault.

Zero-sample fault diagnosis (ZSFD) achieves remarkable success and has attracted considerable attention. However, existing methods suffer from the limitations in fault attribute labeling and feature extraction, leading to poor generalization and robustness. To be specific, fault attribute labeling that depends on expert knowledge is time-consuming and laborious. Fault feature extraction is carried out in one projecting space, and the useful knowledge of target data is ignored, which results in nonoptimal ZSFD performance. To tackle the above problems, a novel ZSFD based on transfer learning is proposed. First, a shared knowledge dictionary that automatically learns from source data with labels is transferred into the target data, which reduces the dependence of ZSFD on fault description. Second, a novel multiclass space projection model is designed to obtain the discriminative fault features. Finally, the pseudolabel mechanism is introduced to excavate the interclass and intraclass information in the target domain. The experiment results on the Tennessee–Eastman process and a real hot roll of steel process show the effectiveness of our method as well as its superiority.

Evolving monitoring requirements may necessitate the addition of new sensors or the exclusion of old ones. Unfortunately, traditional data-driven fault diagnosis methods usually hold the assumption that the number of sensors remains constant throughout the monitoring process, so they need to be retrained with intractable computation to account for the varying sensor behaviors. This article designs a fault diagnosis method that deals with varying sensor behaviors in an online fashion. First, we list potential sensor varying behaviors by providing definitions of sensor states and sensor state transitions. Then, this article proposes the incremental varying sensory data-driven fault diagnosis model (IVSM). IVSM is able to update in an incremental manner under varying sensory data, with a theoretical performance guarantee. The primary objective of IVSM is to continuously map the heterogeneous sensory data within different time into a unified subspace, thereby enabling the direct measurement of heterogeneous and varying sensory data. Subsequently, it constructs a fault identification classifier within this unified subspace to determine the presence of faulty conditions in the systems. Its effectiveness and efficiency are verified by experimental results obtained from two public industrial systems and one practical industrial plant.

Many real-world streaming data are sequentially collected over time and with skew-distributed classes. In this situation, online learning models may tend to favor samples from majority classes, making the wrong decisions for those from minority classes. Previous methods try to balance the instance number of different classes or assign asymmetric cost values. They usually require data-buffers to store streaming data or pre-defined cost parameters. This study alternatively shows that the imbalance of instances can be implied by the imbalance of gradients. Then, we propose the Online Harmonizing Gradient Descent (OHGD) for one-pass online classification. By harmonizing the gradient magnitude occurred by different classes, the method avoids the bias of the proposed method in favor of the majority class. Specifically, OHGD requires no data-buffer, extra parameters, or prior knowledge. It also handles imbalanced data streams the same way that it would handle balanced data streams, which facilitates its easy implementation. On top of a few common and mild assumptions, the theoretical analysis proves that OHGD enjoys a satisfying sub-linear regret bound. Extensive experimental results demonstrate the high efficiency and effectiveness in handling imbalanced data streams.