Accurate and fast anomaly detection in industrial processes and IoT environments

TL;DR

This paper tackles anomaly detection in industrial IoT time series characterized by multicollinearity and unknown distributions. It introduces SAnD, a five-step semi-supervised pipeline that combines smoothing ($h$), variance inflation-factor-based multicollinearity removal, the Mahalanobis distance ($MD$), EVT-based thresholding (MVT/POT), and supervised feature-importance analysis to locate anomalies and infer potential causes. Empirical results across eight public datasets and a real case study demonstrate that SAnD outperforms nine state-of-the-art semi-supervised methods in both detection accuracy and runtime, while providing interpretable explanations of anomalies. The approach is simple, broadly applicable to various industrial domains, and comes with replicability materials to support deployment and further research.

Abstract

We present a novel, simple and widely applicable semi-supervised procedure for anomaly detection in industrial and IoT environments, SAnD (Simple Anomaly Detection). SAnD comprises 5 steps, each leveraging well-known statistical tools, namely; smoothing filters, variance inflation factors, the Mahalanobis distance, threshold selection algorithms and feature importance techniques. To our knowledge, SAnD is the first procedure that integrates these tools to identify anomalies and help decipher their putative causes. We show how each step contributes to tackling technical challenges that practitioners face when detecting anomalies in industrial contexts, where signals can be highly multicollinear, have unknown distributions, and intertwine short-lived noise with the long(er)-lived actual anomalies. The development of SAnD was motivated by a concrete case study from our industrial partner, which we use here to show its effectiveness. We also evaluate the performance of SAnD by comparing it with a selection of semi-supervised methods on public datasets from the literature on anomaly detection. We conclude that SAnD is effective, broadly applicable, and outperforms existing approaches in both anomaly detection and runtime.

Figures (16)

• Figure 1: Panels (a)-(c) show three phases of the production process of the A.Celli tissue machine. Panel (d) shows the eigenvalues of the correlation matrix of the 119 variables considered in our case study. Panel (e) shows the densities of 5 among these variables after standardization.
• Figure 2: Flow chart of the proposed procedure. The dashed blue line shows the steps used to tackle R1 , while the dashed purple line shows the step used to tackle R2.
• Figure 3: Performance of each method and threshold procedure , focusing on long-lived anomalies (y-axis) , vs runtime expressed as logarithms of seconds needed to compute anomaly scores (x-axis). Performances and runtimes are averaged across the considered datasets. Hybrid KNN uses the threshold $\Pr>0.8$.
• Figure 4: Case study results using MVT thresholding. Plot (a) shows anomalies detected in the test set by ${\mathit{SAnD}}_1$ (no smoothing , $h=1$). Plot (b) highlights such anomalies zooming in on the last 2000 observations. Plot (c) shows the time series of the 5 most important variables as identified by a Random Forest. Plot (d) is the same as Plot (b) , but for ${\mathit{SAnD}}_{10}$ (smoothing with $h=10$).
• Figure 5: Results in the case of POT. Plots (a) and (b) show the anomalies detected in the test set by ${\mathit{SAnD}}_{1}$ and ${\mathit{SAnD}}_{10}$ , respectively. Plot (c) highlights the anomalies detected by ${\mathit{SAnD}}_{10}$ by showing the last 2000 observations only. Plot (d) highlights the long-lived anomaly detected by ${\mathit{SAnD}}_{10}$.

Theorems & Definitions (2)

• Proposition 1
• Theorem 1