Graph Neural Networks for Anomaly Detection in Multivariate time-series datasets

异常现象几乎是每个系统不可或缺的一部分，包括制造过程。它会导致系统故障，延误制造过程，浪费资源和时间。异常检测的关键挑战之一是定义正常/异常数据点之间的精确边界，多变量时间序列数据。最先进的方法不显式地学习现有关系的结构，这限制了其在许多现实世界应用中的适用性。近年来，基于深度学习的图神经网络作为一种成功的方法出现在图结构数据中的复杂模式建模中。本研究的目的是设计和开发用于制造过程异常检测的图神经网络。当查看晶圆时间线的图形视图时，开发的模型将能够识别制造数据中的潜在异常。博士学位期间将实现以下目标1。提出了一种新的用于检测制造过程多变量时间序列数据异常的图神经网络。在所提供的商业图形数据集上对所提出的方法进行评估，并与现有最先进的方法进行比较。以人类可理解的形式可视化输出。开发的方法将被应用于纳米制造领域，以基于知识图来识别制造过程中的潜在问题和缺陷。它将考虑生成的事件流，以捕获在半导体晶片制造过程中创建的过程和事件。在生产中使用这些技术将减少工艺工程师为晶圆潜在问题收集和分析数据所花费的时间，并能够为制造工艺中的异常情况提供高级通知，这些异常情况不容易被查看单个晶片时间表或工艺步骤的工艺工程师发现。图神经网络由于其固有的与数据的结构关系关联的性质，有望在解决异常检测问题上表现得更好。这类研究的成果可应用于：(A)粮食供应链，以发现粮食安全方面的异常情况；(B)交通运输；(C)商业/住宅建筑的能源消费模式，以减少能源浪费，并有助于支持气候变化的使命。

Anomalies are integral part of almost every system including manufacturing process. It can cause system fault, delay the manufacturing process, waste of resources and time. One of the key challenges in anomaly detection include defining the precise boundaries between normal/abnormal data points multivariate time series data. The state-of-the-art methods do not explicitly learn the structure of existing relationships, which limits the applicability in many real-world applications. In recent years, deep learning-based graph neural networks have emerged as a successful approach for modelling complex patterns in graph-structured data. The goal of this research is to design and developed the graph neural networks for anomalies detection in manufacturing process. The developed model will be able to identify potential anomalies within the manufacturing data when looking at the graph view of the wafer timelines. The following objectives are set to be achieved during the Ph.D. duration.1. A novel graph neural networks to detect anomalies for multivariate time-series data of manufacturing process.2. Evaluate the proposed approach on the provided commercial graph datasets and comparison with the existing state-of-the-art methods.3. Visualization of the output in a human understandable format.The developed approach will be applied to the area of nano-manufacturing to identify potential issues and defects within the manufacturing process based on a knowledge-graph. It will consider the generated event stream to capture the processes and events created during the construction of a semiconductor wafer. Using these techniques in production will reduce the amount of time that process engineers spend collecting and analysing data for potential issues with wafers as well as being able to provide advanced notification to anomalies within the manufacturing processes which are not easily identified by process engineers looking at single wafer timelines or process steps. The graph neural networks are expected to outperform to solve anomaly detection problem because of its inherent nature of structural relationship association with the data. The outcome of such research can be applied into (a) food-supply chain to detect the anomalies for food security, (b) transportation, and (c) energy consumption patterns in commercial/residential buildings to reduce the energy wastage and contribute to support the mission of climate change.