Machine learning in problems of automation of ultrasound diagnostics of railway tracks

TL;DR

This paper proposes a real-time system for automatic decoding of railway track ultrasound data using neural networks, enabling efficient defect detection with potential for parallel hardware implementation.

Contribution

It introduces a system architecture combining ultrasound data preprocessing, neural network classification, and decision-making for railway track defect diagnostics.

Findings

Effective preprocessing of ultrasound data for neural network input

Convolutional neural network classifier for defect identification

Potential for implementation on GPU and tensor processors

Abstract

The article presents the system architecture for automatic decoding of railway track defectograms in real time. The system includes an ultrasound data preprocessing module, a set of neutral network classifiers, a decision block. Preprocessing of data includes affine transformations of measurement information into a format suitable for the operation of a neural network, as well as a combination of information on measurement channels, depending on the type of defect being defined. The classifier is built on a convolutional neural network. The proposed solution can be effectively implemented on a modern elemental basis for performing parallel computing, including tensor processor and GPUs.