edfas.org 33 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 2 UNIVERSITY HIGHLIGHT Konstantin Schekotihin, University Klagenfurt konstantin.schekotihin@aau.at ARTIFICIAL INTELLIGENCE AND CYBERSECURITY, UNIVERSITY KLAGENFURT The department of Artificial Intelligence and Cybersecurity (AI&CS), University Klagenfurt, Austria, focuses on research and application of AI methods to various practical problems. Many of our activities are conducted within research projects with industrial partners, where logic-based methods are combined with machine learning to solve complex problems related to the digitalization of business processes. In the failure analysis (FA) domain, our activities aim to develop an intelligent FA assistant that can automate many tedious and routine, but nevertheless essential activities, helping an FA engineer to localize physical failures as efficiently as possible. Therefore, the assistant’s main task is to predict the possible failure using information collected by the FA engineer so far and recommend the most probable hypothesis. This task requires an application of both symbolic and machine-learning methods. The former are used to collect and store knowledge from FA engineers about the domain in a machine-readable form, whereas the latter combine this knowledge with raw data resulting in the application of FA methods. Formalization of FA knowledge can be done in different ways. Our department studies and successfully applies methods originating in the Semantic Web domain. Developed to create machine-readable descriptions of any data artifact published on the Web, these methods provide versatile and flexible approaches to creating and maintaining sophisticated FA vocabularies describing physical failures, their electrical characterizations and possible locations, related methods and tools, as well as any other type of concepts used by FA engineers in their daily work. Every concept in these vocabularies is uniquely identified. Therefore, the terminology can be used to annotate data artifacts of different types, such as images, database entries, office documents, and others, and establish interoperability between them, thus providing a single access point for FA engineers and data collection routines. For instance, one can query for all artifacts comprising a specific physical failure at a given location identified by some method. Although data annotation can be done manually by providing FA engineers with corresponding extensions to existing software tools, the two main challenges remain unsolved: annotation of legacy data should be done manually with much additional effort, and annotation of new data is not simplified. Machine-learning methods can solve both problems by operating on the symbolic domain provided by the vocabularies. Our research in this direction focuses on applications of deep learning to natural language processing (NLP) and computer vision (CV). The most significant NLP tasks we study in the context of FA are text classification and named entity recognition. The goal of the first task is to determine a general topic of a text, like an FA report or a database entry. These topics can be any concept defined in the vocabulary, like a physical failure or its electrical characterization. The second task aims at a detailed analysis of the text, determining an exact sequence of words denoting a concept. Similarly, in the CV field, we focus on image classification, which provides general information about an image, like a degradation level of a sample after a hot-plate test, and object detection to identify interesting parts of an image, like regions depicting a void or a fused wire. Considering that the number of possible concepts in an FA vocabulary is very large, it is impossible to create a single NLP or CV model able to recognize all of them reliably. Therefore, we use the natural taxonomies of vocabulary terms to identify hierarchies of models and establish information flow between them. Finally, we look at the analysis and extraction of work- flows that FA engineers follow while localizing a failure. This problem is essential for the prediction of tasks lead- ing to failure localization and scheduling of FA jobs to ensure the most efficient use of available lab resources. Schekotihin
RkJQdWJsaXNoZXIy MTMyMzg5NA==