An improved short-wave rail irregularity detection method based on frequency-related Recurrence Plot and Convolutional Neural Network

Short-wave rail irregularities pose significant risks to high-speed rail safety, as their high-frequency shock features can be effectively captured by axle-box acceleration (ABA) signals. However, visualizing features embedded these signals remains challenging. This paper proposes an improved method for intelligently identifying short-wave rail diseases by combining Adaptive Chirp Mode Decomposition (ACMD), Recurrence Plot (RP), and Convolutional Neural Network (CNN). The method was tested on ABA signals contaminated with rail corrugation and rail impact. In the proposed approach, an ABA signal is first decomposed through ACMD to eliminate noise. Then, the signal is converted to a two-dimensional phase space trajectory image using a recurrence plot. The experimental results demonstrate that RPs of the same rail disease exhibit significant visual similarities, while those of different diseases show distinct patterns. Specifically, RPs of normal signals appear uniform and stochastic with low complexity, RPs of rail corrugation display a periodic checkerboard pattern, and RPs of rail impact feature a prominent cross-shaped void. Finally, a two-dimensional CNN model was developed to adaptively extract irregularity features. The CNN model achieved an accuracy of 96.3%, thus validating its effectiveness in accurately identifying rail impact and rail corrugation, two critical types of short-wave rail irregularities.