Knowledge Discovery and Anomaly Identification for Low Correlation Industry Data

Abstract

With the development of information technology, industry data is increasingly generated during the manufacturing process. Companies often want to utilize the data they collected for more than the initial purposes. In this paper, we report a case study with an industrial equipment manufacturer to analyze the operation data and the failure records of the equipment. We first tried to map the working condition of the equipment according to the daily recorded sensor data. However, we found the collected sensor data is not strongly correlated with the failure data to capture the phenomenon of the recorded failure categories. Thus, we proposed a data driven-based method for anomaly identification of such low correlation data. Our idea is to apply a deep neural network to learn the behavior of collected records to calculate the severity degree of each record. The severity degree of each record indicates the difference of performance between each record and all other records. Based on the value of severity degree, we identified a few anomalous records, which have very different sensor data with other records. By analyzing the sensor data of the anomalous records, we observed some unique combinations of sensor values that can potentially be used as indicators for failure prediction. From the observations, we derived hypotheses for future validation.