METHOD FOR SEQUENTIALLY AGGREGATED ASSESSMENT OF STRUCTURAL CHANGES IN DATA STREAMS FOR DETECTING DEGRADATION OF TECHNICAL SYSTEMS

Abstract

This paper proposes a method for sequentially aggregated assessment of structural changes in data streams aimed at the early detection of technical system degradation. The relevance of the study is determined by the need to identify initial changes in technical condition that may be weak and unstable and may not be accompanied by the exceedance of established diagnostic thresholds. Current approaches to condition monitoring, anomaly detection, and degradation prediction are analyzed, revealing that the problem of interpretable aggregation of a sequence of local structural-probabilistic assessments while preserving information on the pattern of their evolution remains insufficiently investigated. The proposed method sequentially transforms local assessments of structural atypicality obtained from short data-stream fragments into a multicomponent aggregated representation. Five complementary components are introduced to characterize the dynamics of structural changes: level, trend, accumulation, duration, and variability. Their joint analysis makes it possible to distinguish isolated local deviations, directional growth of structural atypicality, and an established persistent degradation state. Rules for interpreting sequentially aggregated assessments are defined, and their responses to characteristic scenarios of condition change are investigated. The experimental simulation results demonstrate that an isolated deviation is primarily reflected by increased variability without substantial accumulation or duration, whereas degradation development is characterized by a pronounced positive trend followed by increases in the accumulation and duration components. It is shown that, once a persistent degradation state has been established, the trend may decrease while the level, accumulation, and duration components remain high, which substantiates the need for their joint interpretation. The practical applicability of the method is investigated using rolling-bearing vibration monitoring as a case study, based on the operating conditions and structure of the open XJTU-SY dataset. The simulation results confirm the possibility of detecting the directional development of structural changes before a persistent degradation state is established. The proposed method does not require the training of a complex parametric model on large datasets, provides an interpretable representation of the dynamics of the monitored process, and can be applied to streaming condition monitoring of technical systems as well as to the generation of additional features for diagnostic and prognostic tasks.

Keywords: data stream analysis, structural changes, sequentially aggregated assessment, time series analysis, technical system degradation, condition monitoring, condition prediction, machine learning