A Deep Learning First Approach to Remaining Useful Lifetime Prediction of Filtration System With Improved Response to Changing Operational Parameters Using Parameterized Fully-connected Layer

For the remaining useful lifetime prediction, apart from the normal sensor data which is updated regularly, there are also operational parameters, which do not change during a cycle of operation. Different sets of parameters result in essentially different, but relevant systems and thus require the adaptation from the statistical model for better prediction. We noticed that neural networks could easily overfit into one set of operational parameters and demonstrate constant bias in the prediction for other sets (underfitting). An aspect of major contribution from our work is the use of Parameterized Fully-Connected Layer (PFL). The PFL builds the parameter dependency right into each layer, in this way the parameters act as ”meta-inputs” which adapt the model of neural network models to the different operating conditions. In another aspect of contribution, our work demonstrated that, instead of using feature engineering, convolutional layers could be used to automatically learn the features which are relevant for the prediction. In this way, the deep learning architecture could be reused for different problems or systems. We conduct experiments on the filtration system datasets provided by the Data Challenge 2020 and received results that compare favorably to the prize winners.