Data-driven sparse coding for onboard condition monitoring of railway tracks

Continuous monitoring of the rail condition plays an important role in railway maintenance and the planning of noise- and vibration-reducing measures. Rail monitoring can be carried out efficiently using vibro-acoustic measurements with onboard sensors. However, this approach generates large amounts of acoustic and vibration data, which makes real-time transmission, processing and storage a challenge. This paper presents a sparse coding framework applied in the time–frequency domain that aims to overcome these challenges by significantly reducing the amount of data while preserving important information for rail defect detection and diagnosis. The Short-Time Fourier Transform is used as a preprocessing step to transform raw signals into a time–frequency representation, capturing the non-stationary characteristics of the signals. The spectrum at each time window is then represented by a sparse linear combination of basis spectra, which form a dictionary. Online sparse dictionary learning is used to create a data-driven, adaptive representation tailored to the frequency characteristics of vibro-acoustic signals related to rail defects. Experimental data acquired with a microphone and an accelerometer mounted on the wheelset of a tram are used to evaluate the framework. The experimental results show that the framework is able to achieve high compression rates and reduce noise. A reduction in data size of 98% was obtained without loss of relevant information. The proposed approach offers significant advantages for modern railway condition monitoring systems. It is scalable for large amounts of data, energy efficient and suitable for real-time implementation. By reducing data bottlenecks, it enables efficient track monitoring with on-board sensors. This work thus contributes to the development of intelligent and cost-effective solutions for infrastructure management.