Damage Classification of Oil & Gas Equipment Components Using Siamese Networks

Abstract Visual inspections are non-destructive testing methods that involve the assessment of a material, component, or product to determine its current condition and maintenance schedule. This is done visually, either unaided or with the assistance of a tool. It serves as the initial check for both new and in-service equipment. The process involves scrutinizing the quality of surfaces to uncover any superficial flaws that may have occurred during production or might develop during operation. Although visual inspection is an integral part of any integrity asset management strategy, current visual inspection workflows are marked by experience-driven rules, based on which one can assign damage severity levels to any type of equipment or component. The evolution of the latest technologies for capturing images (e.g., drones, 360° cameras) has reduced the requirements of physical inspections in the field. However, the number of images captured by these devices in a single iteration is huge (e.g., 100,000+). Manually inspecting these images is a persistent challenge and requires a significant amount of human effort, time, cost, and consistency across the entire process. Hence, the need for an automated visual inspection workflow is paramount. The paper introduces a Siamese network-based damage classification application that enables faster execution of the visual inspection workflow. Siamese networks work by learning the features associated with a class based on differences and similarities between given triplets of images. The methodology has an added advantage over conventional classification approaches in terms of the data required to train the model for different equipment. The embeddings for a pre-defined set of images are generated and stored. Comparison of any unseen image embedding is performed against this set of predefined embeddings. Based on the similarity of the unseen embedding with the existing set, a damage class is assigned to the given image. By reducing the amount of data needed, the methodology enables quick extension of the existing workflow to previously unseen equipment.