Time Series Analysis Using Synthetic Data for Monitoring the Temporal Behavior of Sensor Signals
Marcos W. Rodrigues, Luis Enrique Zárate · 2019
Industrial environments demand constant monitoring in their activities, aiming to guarantee effectiveness concerning safety, production, and quality. The probability of failure in the security issue can lead to a succession of other failures, which will inevitably lead to increased risk in disaster systems, causing a high environmental, social, and economic impact. An old proverb states, “prevention is better than cure”, so we have developed a methodology for detecting and describing the patterns that lead to abrupt changes in the behavior of sensor signals in specific operating scenarios. For this, we have developed a parameterizable time series generator, which allows us to represent several types of scenarios where these sensors can operate. The generator seeks to overcome the problem of access or lack of temporal data generated by real sensors such as dams in the miner industry, for example. We approach three scenarios of time series where stationarity predominates, where variations and trends are propagated, and where anomalous or extreme events may occur. Also, we propose strategies to characterize the synthetic time series, and finally, we use unsupervised machine learning techniques to analyze the temporal evolution of the sensor system. In our analyses, we are using evaluative metrics for validating the proposed methodology, such as the stability index of the clusters along the windowing, the instability index of the sensors intracluster and intercluster, and also by the agglomerative coefficient between the clusters by temporal windowing.