Long Term Prediction of Hydraulic Oil Properties with Sample Convolution and Interaction Networks

Hydraulic oil is an important transmission medium in hydraulic systems. With the operation of hydraulic systems, the hydraulic oil may get contaminated, and its quality may become worse. The degradation process of hydraulic oil can be observed by monitoring its physical properties, such as temperature and viscosity. As a result, in order to get the degradation process of hydraulic oil in advance, it is important to predict the oil properties in a long term. In this study, we utilized modified sample convolution and interaction networks (SCINet) to predict the density, temperature, dynamic viscosity and kinematic viscosity of hydraulic oil in an EPB shield machine for a period of 10 minutes using the data collected in the past 30 minutes. In the revised model, to alleviate the distribution shift problem in the dataset, a simple but effective re-normalization and denormalization method is used. The results show that compared to the existing machine learning and deep learning-based methods, SCINet can better extract useful information from historical data, and is more accurate in long term forecasting of hydraulic oil properties. By using the proposed method, it becomes possible to detect the change in oil property in advance, facilitating timely inspect and repair of the hydraulic system and preventing possible system failures.