Foreign object detection algorithm for railway freight cars based on a novel adversarial autoencoder
DING Fengxia
Abstract：The Trouble of Moving Freight car Detection System(TFDS)utilizes manual methods to detect faults in captured images of key sections of railway freight cars,which is not only inefficient but also susceptible to false alarms.Non-manual fault detection methods typically employ traditional im-age processing techniques and deep learning-based target detection networks,which are limited by the available image data.To overcome the challenges associated with collecting and annotating fault im-ages,addressing the most frequently occurring fault in railway freight cars,that is,foreign objects fault,this study proposes a novel vehicle foreign object detection algorithm.The algorithm is based on an innovative adversarial autoencoder,and it uses an unlabeled training dataset of normal images.To address small foreign object targets,the attention mechanism is incorporated into the adversarial auto-encoder structure,and the effectiveness of various attention mechanisms in the target scenario is com-pared to select the most suitable configuration.Additionally,feature matching loss is employed to opti-mize the loss function and improve the stability of adversarial training.Furthermore,a feature vector outlier scoring mechanism is introduced to assess overall abnormal performance,considering both the application context and the characteristics of the generation model.The experiment results show that the proposed vehicle foreign object detection algorithm is effective in two scenarios:the bottom and the side of the bogie,achieving Area Under Curve(AUC)indicators of 96.9%and 99.3%,respectively.
Keywords：railway freight carvehicle foreign object detectionadversarial autoencoderfault d-etectionattention mechanism
Publication Date：2023-10-28
Online Publishing Date：2025-08-15(First online date of this platform, not the publication date of the document)
Pages：9( 25-33 )