Heavy machine parts measurement through deep learning

Operational continuity of machinery involves continuously monitoring machinery parts to prevent malfunctions. Recently, it has gained popularity in the heavy industry due to its potential to ensure maintenance and address potential malfunctions before they occur. This project focuses on advancing the "Volvo undercarriage wear inspection and maintenance program." The core of this study is the wear and tear inspection process of the undercarriage parts of Volvo’s excavators and it investigates the implementation of deep learning and machine learning techniques, focusing on detecting the undercarriage part of the machine and measuring its deterioration while also aiming to minimize associated costs and labor time. The research starts with a comprehensive collection and preparation of the dataset, ensuring its validity for efficient training while addressing data quality and quantity limitations. A thorough examination and evaluation of the Mask R-CNN model for detecting and segmenting objects is conducted, followed by applying OpenCV for extracting measurements and implementing a template-matching model with a VGG16 network for image classification. The thesis concludes by training and evaluating the Mask R-CNN model three times, showcasing its promising ability to detect and segment the undercarriage part with an accuracy of up to 83.47%. The template matching approach achieved an accuracy of 16.67%, while the OpenCV method demonstrated promising capabilities with an error margin of ±0.5mm. These results indicate that inspection efficiency and accuracy could significantly increase, leading to more timely and cost-effective maintenance decisions. Finally, a validation of the approach is applied and presented in an industrial case study provided by Volvo.