Determining Asset Health Using Multi-Parametric Data and Machine Learning

IMC-2018 Learning Zone 42:35

by Abhinav Khushraj, Petasense

No doctor depends solely on an X-ray to determine a patient's health; likewise, a single vibration alarm is not sufficient for understanding an asset's health. Due to the variability of asset types and running conditions, alarm levels can be triggered even while a machine is healthy, or bypassed even while a machine is unhealthy. Since sensor readings can fluctuate depending on plant operating conditions, it is good practice to assess asset health based upon multiple sensor parameters. Analyzing multiple sensor values for a particular machine, or multiparametric analysis, will prevent false positives and false negatives. When the selection of sensor parameters is based on deep domain knowledge, reliability professionals are enabled to make accurate assessments about the health of their assets. In order to scale a condition-based maintenance program, multi-parametric data can also be integrated with a machine learning system. When the data from these multiple sensor parameters trigger an anomaly, an analyst can observe the relationship between the parameters, and decide if this is normal operation. An analyst can then provide feedback by indicating if a sensor measurement is indicating a machine defect or simply a different operating condition. Not only will this prevent the machine learning algorithm from triggering false alarms, but it also trains the system to identify anomalies even during variable operating conditions. This presentation will cover practical examples of how domain knowledge, in conjunction with multi-parametric data analysis and machine learning, can vastly improve a condition monitoring program.