Sensor Failure Identification in Industrial Big Data

With the rise of information and sensor technologies, sensors play an increasingly significant role in modern production systems. The reliability, safety, and productivity of a production system may largely depend on sensor performance. However, there has been a lack of unsupervised methods for sensor anomaly identification under the environment of industrial big data. This paper proposed an approach to detect sensor failure for industrial big data in an unsupervised manner with the help of random forest and long short-term memory neural networks. The data used in this research are time-series data collected from a gas turbine with 107 sensors. The dataset includes sensor data with 700,000 timestamps in recent years. In this research, random forest regression was first applied to identify the relationship among those sensor values. Afterward, a long short-term memory network is established to predict the values of the target sensor at the current time step. Then, sensor failures can be identified according to the difference between the predicted and actual sensor values. The conducted experiments show promising results that the approach successfully identifies the sensor failure in a completely unsupervised manner.