edfas.org 17 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 2 This article provides a systematic overview of AI methods relevant to the FA domain aiming to guide researchers and practitioners in this area. Specifically, the authors follow a classic view on AI, discriminating between knowledge-based and machine-learning approaches. The former makes decisions using efficient reasoning systems with knowledge explicitly provided by experts and stored in knowledge bases. In contrast, machine learning methods acquire all knowledge required to make decisions inductively from (annotated) data. The literature survey shows that both AI approaches have their pros and cons for FA applications, focusing on the works published in the last years and covering a wide spectrum of topics including various diagnostic stages, predictive analytics, manufacturing-oriented FA tasks, or counterfeit verification. Finally, a short overview of the most representative AI methods and outline future applications of AI in the FA domain is provided. ARTIFICIAL INTELLIGENCE Artificial intelligence is a multidisciplinary science that unites computer science, mathematical logic, philosophy, probability, control and information theories, statistics, and many others. Since the early 1940s, with the development of digital electronic computers, mathematicians, engineers, and computer scientists have started to explore possibilities for the creation of intelligent machines. In 1950, Alan Turing suggested one of the first known intelligence tests known as the “Turing test,” which aims to classify machines into intelligent and not.[3] The term “artificial intelligence” was suggested by John McCarthy during a workshop held in 1965 at Dartmouth College.[4] In modern literature, for example reference 5, there are four various definitions of AI, which can be summarized as follows: Thinking humanly. Focus on activities associated with human thinking, like problem-solving, learning, or decision-making.[6] Acting humanly. Create machines performing functions that require human-level intelligence.[7] Thinking rationally. Research computations enabling a machine to perceive, reason, and act.[8] Acting rationally. Concerned with intelligent behavior in artifacts.[9] The first two definitions represent different perspectives on strong AI, also known as artificial general intelligence, capable of (nearly) human intelligence. However, after decades of research, it became clear that the initial dream of a strong AI is hard or even impossible to reach with modern technology. Therefore, researchers and practitioners are mostly focusing on the methods following the last two AI definitions, which are often referred to as weak or narrow AI. Methods of this family aim to solve a specific task, such as answering questions or recognizing objects on an image, and rely on the human ability to analyze the application domain in order to collect data, define models capturing knowledge about the domain, or determine training and reasoning procedures. There are many different classifications of weak AI methods, which differentiate between them, e.g., by applied decision-making approaches or required data. For FA applications, we classify methods by the approaches they are using to acquire and use knowledge about the domain. In particular, we differentiate between symbolic and machine-learning methods, as shown schematically in Fig. 2. The former acquire the knowledge deductively, i.e., directly from experts, who analyze the problem and use some formal language to encode their knowledge in a machine-readable form. Next, this knowledge is used by algorithms to find a solution to a problem. In turn, machine learning methods, as the name suggests, create models representing domain knowledge inductively, i.e., directly from (annotated) data. Let us exemplify these introduced notions using two FA-relevant problems, using Fig. 3 to illustrate the discussion. Symbolic approach. Consider a simple scheduling problem of an FA lab: given three sets of engineers, machines, and jobs, where every job is represented by a sequence of tasks; find an assignment of tasks to engineers and machines, such that the total processing time of all jobs is minimized. Experts can easily formulate requirements to a schedule, e.g., Fig. 2 General classification of AI methods.
RkJQdWJsaXNoZXIy MTMyMzg5NA==