Railway Track Fault Detection Using Selective MFCC Features from Acoustic Data.

Area of Science:

• Railway Engineering
• Acoustic Signal Processing
• Machine Learning

Background:

• Railway track faults pose significant safety risks and lead to substantial financial losses.
• Manual inspection methods are labor-intensive, time-consuming, and susceptible to human error.
• Existing automatic fault detection systems face challenges like data scarcity, noise, and model inefficiency.

Purpose of the Study:

• To develop a novel and highly accurate approach for automatic railway track fault detection.
• To enhance fault detection performance by leveraging acoustic data and advanced machine learning techniques.
• To address the limitations of current methods in terms of accuracy and reliability.

Main Methods:

• Extraction of mel frequency cepstral coefficient (MFCC) features from acoustic data.
• Implementation of an ensemble machine learning model for improved classification.
• Utilization of chi-square (chi2) feature selection to identify the most relevant acoustic features.
• Experimental validation using a collected dataset to assess performance and computational complexity.

Main Results:

• The proposed approach achieved a mean accuracy score of 0.99 on the collected dataset.
• Optimal performance was observed when using a combination of 40 original features and 20 chi2-selected features (total 60 features).
• The developed method demonstrated significantly superior performance compared to existing railway fault detection techniques.

Conclusions:

• The novel acoustic-based approach using MFCC features and chi2-selected features offers a highly accurate and efficient solution for railway track fault detection.
• The ensemble model effectively integrates selected acoustic features to achieve superior detection performance.
• This method presents a promising advancement for enhancing railway safety and reducing inspection costs.