A Temporal Pyramid Pooling-Based Convolutional Neural Network for Remaining Useful Life Prediction

Remaining Useful Life (RUL) prediction is a key issue in Prognostics and Health Management (PHM). Accurate RUL assessments are crucial for predictive maintenance planning. Deep neural networks such as Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) have been widely applied in RUL prediction due to their powerful feature learning capabilities in dealing with high-dimensional sensor data. The sliding time window method with a predefined window size is typically employed to generate data samples to train such deep neural networks. However, the disadvantage of using a fixed-size time window is that we might not be able to apply the resulting predictive model to predict new sensor data whose length is shorter than the predetermined time window size. Besides, as the length of sensor data varies, the traditional unchanged and subjectively set time window size may be inappropriate and impair the prediction model’s performance. Therefore, we propose a Temporal Pyramid Pooling-Based Convolutional Neural Network (TPP-CNN) to increase model practicability and prediction accuracy. With the temporal pyramid pooling module, we can generate data samples of arbitrary time window sizes and use them as inputs of CNN. In the training phase, CNN can learn to capture temporal dependencies of different lengths since we feed in samples with different time window sizes. In this novel manner, the learned model can be used to test data with arbitrary sizes, and its predictive ability is also improved. The proposed TPP-CNN model is validated on the C-MPASS turbofan engine dataset, and the experiments have demonstrated its effectiveness.