A novel method for detecting anomaly in Pump Vibration data in a multi-regime setting using a combination of changepoint detection, clustering and density based anomaly detecting algorithms to improve operational efficiency in Nuclear Power Plant
In industrial systems, vibration signals are the most important measurements indicating asset health. Based on these measurements, an engineer with expert knowledge on the assets, industrial process and vibration monitoring can perform spectral analysis to identify failure modes. When measurements are performed continuously, it becomes impossible to act in real time on this data. Therefore, the objective of this paper is to use analytics to perform vibration analysis in near real time to monitor health of Boiler feed pumps and detect anomaly.
The Nuclear power plants have Main Boiler feed pumps. Vibration Data is generated from the Boiler feed pumps and it is analysed to detect anomaly in the time domain. We employ a combination of statistical methods of changepoint detection algorithms like Pelt with cost function being L1, L2, Gaussian or Mahalanobis, Kmeans clustering and local outlying factor analysis (LOF) to detect the anomaly at specific point in time based on the Regime of operation. Based on the number of data anomaly we can infer on the health condition of the Boiler feed pump. The cost of unplanned equipment downtime comes in thousands of dollars. This paper illustrates the strength of combining expert engineering knowledge with advanced data analytics techniques to improve asset performance.
Outline/Structure of the Talk
1. Understanding the Business problem
2. Understand the IoT data
3. Proposed algorithms and solution
Time breakup of the presentation :
1. Anomaly detection in IoT
2. Operational analytics
Business stakeholders and data scientists