A Lightweight Anomaly Detection Model in Aero Turbine Borescope Using Unsupervised Deep Learning

Visual inspection is a critical task typically performed by humans, but human decisions can be inconsistent due to various factors. While many studies have introduced artificial intelligence (AI) to assist aircraft technicians, supervised approaches often encounter limitations, such as the need for human intervention and a lack of sufficient data. By understanding these challenges, researchers developed a lightweight anomaly detection (AD) model in aero turbine borescope inspection using unsupervised deep learning. Considering the utilization of the borescope AD model, two usability perspectives are considered, 1) lightweight operation and 2) input image resolution. The selected AD algorithm is Patch Distribution Model (PaDiM). This algorithm is trained with normal data and tested with normal and abnormal data. One of the key benefits of the PaDiM model is that it reduces the need for human intervention during turbine borescope inspections. It requires only normal data for training, provides anomaly scores, and can localize potential defect areas when abnormal data is input To operate this model for light operations, MobilNetV3 was selected for the backbone, with layers selectively adjusted based on different settings and each model's performance using matrices such as AUROC, F1 score, training time, and prediction time. When compared to the original PaDiM Models, this research found the performance of the new model to be relatively consistent across different settings while requiring less time to train and predict, and that higher resolution images can better detect small defects than lower resolution images. The proposed unsupervised deep learning based AD models can aid aircraft technicians in improving the performance of borescope inspection.