Application of Autoencoder for Control Valve Predictive Analytics

In this paper, we investigated the application of an autoencoder neural network for predictive analytics of control valves, which are crucial components in industrial processes with significant consequences in case of failure. The autoencoder was created using Python and Keras deep learning framework, comprising encoding and decoding sections. By comparing the difference between the input sensor data and its reconstructed output, referred to as the reconstruction error, we were able to identify anomalies. The result which is based on the data of an actual asset was compared with the random forest regressor to ensure the effectiveness of the approach. We have also proposed a practical approach to generate alerts when the magnitude exceeds a predefined threshold, thereby enabling proactive maintenance and avoiding unplanned shutdowns. We emphasized the diagnostic capability of the autoencoder in identifying anomalous sensors, which is not present in traditional regression approaches. Furthermore, we argued that this capability could be more valuable for a complex equipment with many input sensors. The proposed approach can be further improved to provide prognostic capability by forecasting the trend of the reconstruction error.