LSTM and TCN application for airport surface distress detection

This paper evaluates the feasibility of using smartphone accelerometers to identify and categorize airport pavement distresses. Using experimental measurements taken on a test road section, we tested the smartphone accelerometers to recognize and categorize selected distress patterns. In our work, we investigated the capacity of neural networks with a Long Short-Term Memory (LSTM) layer with normalized data weighted and not weighted, bidirectional LSTM, and Temporal Convolutional Networks (TCNs). The networks tested for sequence data classification displayed considerably high accuracy. However, many associations of sequence versus distress were needed to adequately complete the training process. In contrast, the TCNs we used for sequence-to-sequence classification showed lower accuracy. However, a much smaller data set was sufficient to complete the training. As a consequence, we chose the TCN to implement the practical application. Both algorithms demonstrated high accuracy on training and validation data and performed well on other independent test samples. • Detection of pavement distress is introduced using smartphone accelerometers and a standard vehicle. • It can classify current pavement distresses, with vertical changes tailored to airport requirements. • Dynamic parameters are proposed using LSTM, bidirectional LSTM, and TCN. • TCN can be effectively trained with far fewer samples compared to LSTM and bidirectional LSTM.