Technology Digests / Classifications of Rail Ultrasonic Signals Using Machine Learning

TD24-005

July 24, 2024
Authors
  • Anish Poudel
  • Survesh Shrestha
Share

Key Findings

  • The deep learning models trained using raw ultrasonic signals (A-scan waveforms) for rails with and without defects achieved 88 percent accuracy with balanced precision, recall, and F1-scores across the class.
  • The k-fold cross-validation procedure is commonly employed to estimate the ML skill or performance. This procedure ensures that ML models do not experience performance degradation on new inputs from the same training data distribution. Inputting 1 to 14 A-scan signals across the three-fold cross-validation yielded similar accuracy levels above 98 percent. However, a discrepancy was observed when 28 signals were inputted. Cross-validation folds 1 and 2 exhibited an accuracy of 91 percent, while cross-validation fold 3 showed a slightly higher accuracy of 93 percent. This discrepancy may be attributed to variations in the defect signal characteristics in a particular frame or other factors not captured by the model.
  • The accuracy of ML model increased as the dataset size increased.
  • The absence of a curriculum learning approach resulted in decreased accuracy and a limited learning capacity, emphasizing the importance of incorporating curriculum learning techniques in the training process.

Classifications of Rail Ultrasonic Signals Using Machine Learning

MxV Rail investigated the efficacy of artificial intelligence (AI) and machine learning (ML) algorithms for classifying ultrasonic time-series signals/data collected from rails with and without internal anomalies. By leveraging the potential of AI/ML algorithms, MxV Rail seeks to facilitate the advancement of the accuracy and efficiency of automated ultrasonic rail flaw detection systems. This work will provide both a foundation and valuable insights for future research involving automated classification and data fusion approaches for multi-dimensional time-series data involving internal rail defect detection and prediction. This project was completed in collaboration with Project JeZero, a Denver, Colorado-based firm that assisted with the design, and testing of AI algorithms. BACKGROUND In the railroad industry, rail flaw detection using ultrasonic nondestructive evaluation (NDE) is a critical task that has relied on the expertise and judgment of a trained human inspector. While automated systems such as rail detector cars provide consistent results, they often…