Multivariate Data-Based Anomaly Characterisation of a Potential Pallet-Associated Event in a Manufacturing Conveyor System

Abstract

Manufacturing conveyor disturbances may not be adequately characterised by single-sensor thresholds because operating changes can appear across multiple sensor channels. This study examined whether a recorded M4 pallet-static event contained multivariate sensor patterns that differed from a recorded M4 stall condition. The analysis used temperature (TempC), raw sensor output (Raw), three directional components (x, y and z), and a resultant magnitude derived from the directional components. Following timestamp reconstruction and data screening, observations were aggregated into one-minute medians. Descriptive statistics, Pearson correlation, principal component analysis (PCA), Isolation Forest and two-sided Mann–Whitney U tests with rank-biserial effect sizes were used for exploratory analysis. The pallet-static record yielded 317 one-minute windows, of which 16 (5.05%) were flagged as candidate anomalies under a 5% Isolation Forest contamination setting. Temperature and Raw output were strongly inversely correlated (r = −0.9991), while comparisons with the M4 stall record indicated distributional differences in y, z and resultant magnitude but not in temperature or Raw output, according to the reported unadjusted tests. PCA showed partial separation between the event groups. These findings suggest that directional channels may add descriptive information beyond temperature and Raw output alone. However, the source records do not independently verify a pallet-caused mechanical fault, and the statistical comparisons require confirmation in light of temporal dependence and multiple testing. The framework should therefore be treated as exploratory and validated against independently labelled events before operational deployment.