Deep Learning for Railway Track Defect Detection: A Systematic Review of Validation Evidence and Implications for Kazakhstan (2020–2026)

Deep learning is widely used to detect track defects from images, laser profiles, vibration, and ultrasonic signals; however, it remains unclear how much of the reported performance carries over to lines in service. Following PRISMA 2020, 2,226 records from 2020–2026 were retrieved from OpenAlex, Crossref, and DOAJ, and 428 studies were analyzed. 63.1% of the works use only benchmark, laboratory, or simulated data, while only 23 studies have been tested on real equipment or track. The median accuracy is 96.0%. A comparison of model speeds with the required frame rates showed that 61.3% of the models are fast enough at 160 km/h. No studies based on data from Central Asian lines were found. For Kazakhstan, collecting local data, fine-tuning models, and conducting a supervised pilot in cooperation with KTZ are proposed.