Anomaly detection in a fleet of industrial assets with hierarchical statistical modeling

Maharshi Dhada, Mark Girolami, Ajith Kumar ParlikadData-Centric 2020 · DOI

Introduction

Reliable operation of industrial assets depends heavily on accurate anomaly detection. However, statistical classifiers often struggle with "early-life" assets—those newly commissioned or recently serviced—due to a lack of historical "normal" data. This "cold start" problem leads to unstable models and high false alarm rates. This paper proposes a hierarchical statistical framework that enables assets to collaboratively learn from similar peers in a fleet, significantly improving detection accuracy during early operations.

The Cold Start Problem in Industrial IoT

In predictive maintenance, anomaly detection typically involves learning the probability distribution of sensor data during normal operation. If new data deviates significantly from this learned "normal," it is flagged. For a reliable model, substantial training data is required. When data is scarce, independent models for each asset are prone to error. Conversely, a single global model for an entire fleet often fails to capture the unique operational nuances of individual machines.

Plate notation for an independent asset model.

Hierarchical Modeling Framework

The researchers address this by implementing a two-level hierarchical model:

  1. Asset Level: Individual asset sensor data is modeled using a Multivariate Gaussian (MVG) distribution, capturing the mean ($\mu_i$) and covariance ($C_i$) of sensor relationships.
  2. Fleet Level: To allow for collaborative learning, these asset-level parameters are treated as samples from a higher-level "parent" distribution—specifically a Normal-Inverse Wishart (NIW) distribution.

To handle diverse fleets, the model incorporates clustering. An Expectation-Maximization (EM) algorithm iteratively assigns assets to clusters ($K$) and estimates the parent parameters for each group. This structure allows a new asset to "borrow" experience from its cluster until it has generated enough of its own data.

The full hierarchical model with clustering.

Experimental Setup and Results

The framework was tested on a synthetic dataset of 800 assets divided into four clusters based on model types and operating conditions. Performance was measured using the Area Under the ROC Curve (AUC).

Key findings include:

  • Early-Life Advantage: For assets with only 5 training points, the hierarchical model achieved AUCs >0.8, while independent models performed no better than random chance (~0.5).
  • Stability Threshold: The hierarchical model consistently outperformed independent modeling until assets accumulated approximately 13 to 21 data points.
  • Clustering Efficacy: Learning from a small group of similar assets proved more effective than using a single model for the entire fleet, which often introduced excessive variance.

Conclusion

Hierarchical modeling provides a mathematically rigorous path for automated transfer learning in industrial fleets. By leveraging the "wisdom of the crowd" through structured priors, the system ensures robust anomaly detection even for assets with minimal operational history. Future work will focus on transitioning from batch processing to online updates and incorporating categorical metadata to improve clustering accuracy in extremely low-data scenarios.

Summary updated

Abstract

Abstract Anomaly detection in asset condition data is critical for reliable industrial asset operations. But statistical anomaly classifiers require certain amount of normal operations training data before acceptable accuracy can be achieved. The necessary training data are often not available in the early periods of assets operations. This problem is addressed in this paper using a hierarchical model for the asset fleet that systematically identifies similar assets, and enables collaborative learning within the clusters of similar assets. The general behavior of the similar assets are represented using higher level models, from which the parameters are sampled describing the individual asset operations. Hierarchical models enable the individuals from a population, comprising of statistically coherent subpopulations, to collaboratively learn from one another. Results obtained with the hierarchical model show a marked improvement in anomaly detection for assets having low amount of data, compared to independent modeling or having a model common to the entire fleet.
BibTeX
@inproceedings{dhada2020anomaly,
  author = {Maharshi Dhada and Mark Girolami and Ajith Kumar Parlikad},
  title = {{Anomaly detection in a fleet of industrial assets with hierarchical statistical modeling}},
  booktitle = {Data-Centric},
  year = {2020},
  doi = {10.1017/dce.2020.19},
}
Figure