Estratégias não supervisionadas para identificação automática de drifts incrementais

Abstract

Industrial systems are subject to different types of changes in their operating points, arising both from deliberate actions, such as setpoint adjustments, and from faults, equipment degradation, or abnormal plant conditions. Such changes may represent new valid operating states, opportunities for optimization, or deviations that compromise process performance and safety. In this context, the adoption of automatic change detection mechanisms becomes imperative, since continuous human supervision is unfeasible in systems with a large number of variables. This limitation is even more pronounced in industrial processes, where time series are typically high-dimensional, exhibit non-stationary behavior, and are frequently affected by measurement noise and the presence of outliers. Within this framework, this work proposes two unsupervised and noise-robust methodologies for the automatic detection of incremental changes in industrial processes. The first method, termed NR-EIDDM (Noise-Resistant Equal-Intensities Drift Detection Method ), is based on an n-dimensional partitioning of the input space and on the statistical analysis of equal-density histograms, obtained through the DBSCAN algorithm, responsible for filtering noise and outliers, and a constrained K-Means algorithm, which defines partitions with equal density. Under stationary conditions, the distributions are expected to exhibit similar intensities, with significant deviations being identified through the application of the chi-square test between successive histograms. The second methodology, referred to as UDrAE (Unsupervised Drift detection via Denoising AutoEncoder reconstruction loss), employs a Denoising AutoEncoder capable of modeling the normal behavior of the system while attenuating measurement noise. Changes in the input distribution are inferred by monitoring the reconstruction loss function, whose significant increases indicate potential concept drifts. The proposed methods were evaluated using both synthetic and real datasets, including Gaussian distributions with varying noise levels and simulations of the Interactive Tank System (ITS) from the System Validation Laboratory (LVAS/PPGEE-UFMG). For validation in an industrial context, real unlabeled datasets from Petrobras systems were employed, allowing the assessment of the applicability of the proposed approaches in complex real-world scenarios. The analysis of the results, based on precision, F1-score, and average detection delay metrics, demonstrated that both approaches effectively identify the majority of changes across different datasets, while maintaining competitive and robust performance when compared to reference methods from the literature, such as QuantTree, QT-EWMA, SPLL, OCDD, D3, and IKS.