Development of ai-supported software for part inspection in the automotive industry

The automotive industry is one of the most important industries where technological developments are integrated at the fastest rate and which directs the global economy. Improving quality control mechanisms in production processes is critical both in reducing costs and increasing product reliability. The new generation of technologies not only increases the complexity of vehicle components but also makes their inspection significantly more challenging. The safe and accurate operation of these systems is essential both for production processes and for end-user safety. According to a study published by the National Highway Traffic Safety Administration (NHTSA), approximately 2% of accidents are caused by vehicle components. Considering this rate of accidents worldwide, it emphasizes the importance of timely detection and elimination of production defects. In this thesis, the integration of Artificial Intelligence (AI) and Deep Learning (DL) methods into quality control and production systems in the automotive industry is analyzed and innovative approaches to improve the processes are presented. The main objective of the study is to increase quality standards, reduce effective cost, shorten production time, improve production safety and ensure sustainability in industrial processes by minimizing errors occurring in production processes. For this purpose, in this thesis, AI supported software has been developed by using the existing methods in literature and the performance of the software in real-time production environments has been analyzed in detail. In the first phase of the thesis, a real-time data augmentation over Convolutional Neural Networks (CNN) was implemented. Live (or online) data augmentation is a method of data augmentation; the main difference is that it is performed not during data pre-processing, but during the training of the model. In this thesis, a live data augmentation method, referred to as the Random Rotation Method (RRY), was integrated into the TensorFlow framework. Unlike traditional pre-augmentation methods, this method offers a data augmentation strategy that requires no additional storage, minimizes human intervention, reduces the risk of overfitting, and enables dynamic training. The performance of the proposed RRY strategy is tested in a CNN based application on data obtained from Toyota Motor Manufacturing Turkiye (TMMT) factory. The results show that the RRY can provide performance improvement over existing methods in recognizing moving parts with higher accuracy. This innovative approach is considered as an important step to provide a fast and reliable quality control mechanism, especially in production processes. In the second phase, a DL based system focusing on real-time engine part inspection was developed and implemented in TMMT factory. The system uses a hybrid structure integrating Single Shot Detector (SSD) and Faster Region-Based CNN (Faster R-CNN) algorithms for visual inspection of engine parts, implemented using a Fanuc CR-15ia collaborative robot. In the application, 2800 objects were labelled on 1060 engine part images. The training and testing phases of the algorithms used were examined in detail and the whole structure was synchronized with the existing systems. Analyses based on the data obtained from the TMMT factory over a period of four months showed that the algorithms provided 99.4% accuracy for Faster R-CNN and 95.5% accuracy for SSD with Intersection over Union (IoU) threshold of 0.5. The real-time engine part inspection system created in the hybrid framework ultimately achieved an accuracy of 99.9%. This study has made a significant contribution in terms of both hardware integration and data sharing, and has achieved significant improvements in terms of safety, ergonomics and sustainability in production processes. In the third and final stage, a versatile approach is presented by combining AI based vision systems developed by Human-Robot Collaboration (HRC) with innovative methods. In a digital environment developed using Catia V5 software, human and collaborative robots were simulated in different modes and a region-based part tracking system supported by CNN and Structural Similarity Index Measure (SSIM) was created. The system has made production processes faster and safer without the need for additional sensors or complex camera installations. Simulations and experimental studies have shown that the proposed system provides significant improvements in task completion times and accuracy rates. With the combination of collaborative robots and AI-based systems, an average accuracy rate of 99.12% was achieved in object recognition performance. In addition, a reduction of approximately 49% in task times was achieved compared to traditional methods. This system, which improves ergonomics and occupational safety standards, has made a significant contribution to making production processes smarter and more sustainable. When the studies carried out within the scope of the thesis are evaluated in general terms, it is seen that AI-supported systems make a significant contribution to reducing human errors in control processes, increasing accuracy and making processes more transparent. It has been determined that AI-supported systems can be used in a wide range of other industrial areas. It is foreseen that AI-supported systems will deliver economic and operational benefits once the necessary adjustments are made in terms of regulatory compliance and ethical guidelines. A laptop with NVIDIA RTX 3080ti graphics card, 64 GB RAM and Intel Core i9-12950HX processor was used in the applications in this thesis. Python 3.8.13, Anaconda 4.12.0 and TensorFlow 2.9.1 were used to ensure the efficient operation of the developed algorithms. In addition, the software environment was supported with comprehensive libraries and optimized packages, thus enabling data processing and model training processes to be performed more efficiently. This powerful hardware and software infrastructure provided a critical advantage, especially in processing large data sets and training complex deep learning models. This thesis aims to contribute to sectoral innovation with both technical innovations and solution-oriented approaches by using existing methods. The advanced data processing, complex decision making and forecasting capabilities provided by AI techniques enable production processes to become more effective, reliable and efficient. The findings presented in the thesis reveal the high potential of AI-based systems in industrial integration and strongly emphasize the decisive impact of these systems on sectoral transformation. The solutions developed for specific problems demonstrate how the accuracy and speed of quality control processes can be enhanced, highlighting the critical role of such systems in industrial transformation.