Patch-Wise Neighborhood Feature Distribution Modeling for Anomaly Detection in Locomotive Roof Inspections

Abstract

Electric locomotives depend on their roof power supply systems, which are threatened by unknown foreign objects due to the complex outdoor environment. Timely and accurate detection of these objects is vital for train safety. Existing methods are limited by the diversity and unpredictability of foreign objects, difficulty in collecting abnormal samples, and poor image consistency. To address these challenges, this paper proposes a Patch-Wise Neighborhood Feature Distribution Modeling (PNFDM) algorithm for anomaly detection. It uses multiple multivariate Gaussian distributions to model normal image features at the patch level and computes Mahalanobis distances to generate an anomaly heatmap. A patch-wise neighborhood feature sharing strategy improves the signal-to-noise ratio of the anomaly heatmap. To reduce storage requirements, a low-rank approximation using SVD decomposition reduces model size by nearly 110 times and speeds up loading by 103 times. A chi-square distribution-based normalization remaps the Mahalanobis distance to a 0-1 range, facilitating threshold selection in anomaly localization. Experiments show that this method significantly outperforms others on a locomotive roof anomaly dataset and generalizes well to the public MVTec AD dataset. The approach allows for rapid model updates without needing abnormal samples, training on 100 samples in just 39.76 seconds. Dataset and code are available at: https://github.com/zzfu-buaa/PNFDM.

Keywords

• Anomaly detection
• line array image alignment
• train fault detection
• multivariate Gaussian distribution
• chi-square normalization