Enabling advanced analytics for industry applications using machine learning and graph mining

Abstract

Industrial applications require robust analytical solutions, yet designing these solutions is fraught with many challenges. The vast and diverse datasets, along with the demand for sophisticated algorithms, present considerable difficulties. Effective modeling, domain expertise and explainable results are crucial for success. Mastering these aspects can unlock the potential of advanced analytics, driving industrial progress. This thesis delves into the integration of graph-based techniques and Artificial Intelligence (AI) methods in blockchain networks and semiconductor engineering, focusing on network analysis, device performance optimization and manufacturing process anomaly detection to advance the efficiency of industrial operations. The Ethereum blockchain network stands as the largest public blockchain supporting smart contracts. Recent works have modeled Ethereum's transactions, tokens and other interactions as static graphs, providing insights through graph analysis. This research investigates the evolutionary nature of the large-scale Ethereum interaction networks from a temporal graph perspective, examining growth rates, active lifespans and update rates of high-degree vertices, detecting anomalies based on temporal changes in global network properties as well as designing forecasting models for community survival using graph mining techniques and machine learning models. In semiconductor product engineering for the design and production of GPUs, continuous performance enhancement is critical in the development of high-performance hardware and software. However, for device manufacture, manually inspecting extensive codes and applications to tweak a GPU's performance is impractical, especially when more and more new applications are introduced and tuning needs to be done on top of the current optimized devices. This work requires extensive expertise, and it is made more difficult by the fact that the source codes of the applications are not transparent to the GPU designers. ShaderAnalyzer, a novel model integrating graph mining and machine learning to investigate low-level shader codes executed on GPUs to enhance performance tuning efficiency has been proposed. ShaderAnalyzer can identify and optimize frequently occurring sub-structures in the shader codes. This framework aids engineers by clustering sub-structures and predicting their efficiency. Thus, ShaderAnalyzer can guide tuning engineers to focus on critical algorithmic structures to maximize performance enhancement and expedite the entire tuning process. It also offers valuable insights for future hardware architecture design. As can be seen in the above-mentioned use case on enhancing shader codes to enhance GPU performance, the semiconductor industry is keeping in tandem with the advancement of data engineering and AI to improve productivity and efficiency. Chip fabrication is a complicated process involving numerous steps and each of them requires precise control on quality. Among them, inline wafer map monitoring is crucial for quality assurance as it helps identify abnormal wafer maps among a large number of normal wafers. Hence, it will help to flag production issues or excursions in the early stage. However, identifying abnormal wafer maps is challenging due to the large difference in volume between normal and abnormal wafer maps. Recent research on supervised methods requires considerable time and resources for data labeling. Furthermore, creating a balanced dataset for every category in real-world manufacturing is not practical. In this study, an end-to-end anomaly detection solution is proposed that uses only normal wafer maps for training to identify abnormal wafer maps. Moreover, the proposed model can accurately localize defective areas in the wafer maps. The proposed model outperforms other unsupervised methods and is on par with existing supervised learning approaches even though it does not require the annotation of the datasets by domain experts to train the model. The research demonstrates the potential of the integration of AI and graph techniques in complex distributed systems and advanced engineering and contributes to a broader understanding of how advanced analytics can be exploited to address complex industrial issues. The methodologies and findings in this thesis not only tackle the challenges in the three use cases but also lay the foundation for future research using similar data-driven techniques in various industrial contexts. The combination of specialized knowledge and sophisticated analytical techniques is becoming increasingly important to enable advanced analytics for industrial applications. This thesis exemplifies the potential of these methods to drive significant advancements in industrial efficiency, predictive analytics and process optimization.