A Correlation Driven Approach with Edge Services for Predictive Industrial Maintenance

Meiling Zhu, Chen LiuSensors 2018 · DOI

TL;DR

This paper introduces a novel approach for predictive industrial maintenance by focusing on how anomalies and faults propagate through correlated sensors and devices. By modeling this propagation as an event routing problem on a Proactive Data Service Graph (PDSG) and leveraging edge computing, the system aims to predict failures earlier and more accurately than traditional methods.

What Problem Does This Paper Address?

Predictive industrial maintenance is crucial for minimizing downtime and optimizing maintenance schedules. Current methods often rely on pre-built models or sensor-level anomaly detection. However, a significant challenge is that faults don't occur in isolation; they can propagate through complex, often hidden, correlations between sensors and devices. This propagation can be uncertain, with a single anomaly potentially leading to multiple different faults, or a fault being caused by various combinations of anomalies. Existing approaches struggle to capture these dynamic, multi-faceted propagation paths, leading to delayed or missed maintenance warnings. The authors aim to address this by explicitly modeling and predicting anomaly/fault propagation.

Methods

The core of the proposed solution is a correlation-driven event routing approach. This builds upon the authors' previous work, which introduced:

  • Proactive Data Service (PDS): An abstraction layer where physical sensors are represented as software services. These "edge services" detect anomalies in sensor data and encapsulate them into "service events."
  • Service Hyperlink: A model to represent the correlations (causal relationships) between these service events.

This paper refines these concepts by:

  • Edge Computing Integration: Adapting the PDS model to run on edge nodes (e.g., industrial PCs, laptops) to handle data privacy and reduce network load, bringing computation closer to the data source.
  • Time-Constrained Frequent Co-Occurrence Pattern (TFCP) Mining: An algorithm to discover causal relationships between events, considering both their temporal proximity and conditional probability. This is crucial for identifying how anomalies propagate.
  • Proactive Data Service Graph (PDSG): A directed graph model where nodes represent edge/cloud services and edges represent service hyperlinks (correlations). This graph visually depicts anomaly/fault propagation paths.
  • Heuristic Event Routing Algorithm: An A*-based algorithm that routes service events through the PDSG, considering the probability of propagation and selecting the most probable path to a candidate destination (a service with no outgoing links, representing a potential fault). This algorithm handles the uncertainty of fault propagation.
  • Model Checking: Used to validate the completeness of the approach by ensuring that any real-world system behavior can be represented by the PDSG.

Datasets

  • Real Sensor Dataset: Sensor data from a coal power plant spanning six months (October 2014 to April 2015) was used. This dataset included data from 182 sensors across 5 interactive devices. Maintenance records from the same period were used for verification.
  • Synthetic Datasets: Five synthetic datasets were generated with varying numbers of data sources (2000 to 10,000) to test efficiency under different scales.

Key Insights and Ideas

  • Faults Propagate, They Don't Just Happen: The central idea is that understanding how faults spread through interconnected systems is key to predicting them. This moves beyond simple anomaly detection at individual sensors.
  • Uncertainty is Key: The paper acknowledges that fault propagation isn't always linear or predictable. Their approach explicitly models this uncertainty using probabilistic correlations and event routing.
  • Edge Computing for Sensitive Industries: By pushing computation to the edge, the approach becomes more practical for industries with strict data privacy requirements, like power plants.
  • Event Routing as a Predictive Tool: Mapping anomaly propagation to an event routing problem provides a structured way to predict future states and potential failures.

Key Results

  • High Performance: The approach achieved up to 100% precision and 88.89% recall in predicting maintenance warnings.
  • Early Warnings: The proposed method consistently provided earlier warnings for faults compared to range-based, outlier detection, and discord discovery approaches. For example, for the L-IPAV fault, the proposed approach warned at 70 minutes, while others took 152 minutes or couldn't warn at all.
  • Scalability: Average latency increased linearly with the number of data sources under the edge computing paradigm, while it increased exponentially under cloud computing, demonstrating the benefit of edge processing.
  • Dataset Scale Impact: Larger datasets generally led to more service hyperlinks, which in turn improved precision and recall, though the number of hyperlinks doesn't increase indefinitely.

Limitations and Future Work

  • Model-Based Approaches: Due to a lack of prior knowledge about the specific power plant, model-based predictive maintenance approaches were not directly compared. The authors plan to learn more prior knowledge to incorporate these in future work.
  • Effectiveness vs. Efficiency Trade-off: The paper notes that a large number of service hyperlinks, while improving effectiveness, increases the complexity of the event routing algorithm. Future work aims to balance effectiveness and efficiency, potentially by sacrificing some tolerable effectiveness to improve efficiency.
  • Further Improvement: The authors intend to further improve their approach based on comparative analysis and by exploring more predictive maintenance techniques designed for power plants.

Summary updated

Abstract

Predictive industrial maintenance promotes proactive scheduling of maintenance to minimize unexpected device anomalies/faults. Almost all current predictive industrial maintenance techniques construct a model based on prior knowledge or data at build-time. However, anomalies/faults will propagate among sensors and devices along correlations hidden among sensors. These correlations can facilitate maintenance. This paper makes an attempt on predicting the anomaly/fault propagation to perform predictive industrial maintenance by considering the correlations among faults. The main challenge is that an anomaly/fault may propagate in multiple ways owing to various correlations. This is called as the uncertainty of anomaly/fault propagation. This present paper proposes a correlation-based event routing approach for predictive industrial maintenance by improving our previous works. Our previous works mapped physical sensors into a soft-ware-defined abstraction, called proactive data service. In the service model, anomalies/faults are encapsulated into events. We also proposed a service hyperlink model to encapsulate the correlations among anomalies/faults. This paper maps the anomalies/faults propagation into event routing and proposes a heuristic algorithm based on service hyperlinks to route events among services. The experiment results show that, our approach can reach 100% precision and 88.89% recall at most.
BibTeX
@inproceedings{zhu2018correlation,
  author = {Meiling Zhu and Chen Liu},
  title = {{A Correlation Driven Approach with Edge Services for Predictive Industrial Maintenance}},
  booktitle = {Sensors},
  year = {2018},
  doi = {10.3390/s18061844},
}

Loading related content...

Figure