edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 2 20 and a job report describing the same physical failure can easily be aligned if their authors refer to the same ontology concept. Finally, ontology creation and maintenance processes drive the standardization of FA terminology used in a lab ensuring that all engineers and customers understand each concept in the same way. Among other definitions, as listed above, an FA ontology can describe tasks that engineers can perform in a lab to localize the failure. This allows for the application of automated workflow systems to assist in the training of junior engineers or lead customers through the analysis specification process. For instance, the approach suggested in reference 29 relies on task definitions in the FA ontology,[27] where every task comprises two lists describing its prerequisites and consequences. In general, the workflow system creates a digital twin of the lab processes, where every sample is represented by its state and every task is associated with an internal action. To apply an action, a sample state in the workflow system must fulfill all its prerequisites and after the action is finished, the state is modified according to the consequences of the action. For instance, preliminaries of the scanning acoustic microscopy (SAM) action require a sample in the prepared state. The correct state of a sample can be obtained by executing the preparation action first. The latter, among other consequences, adds the prepared attribute to the sample state. Whenever a customer requires SAM as a part of a robustness analysis, the system can simulate the workflow of the analysis and identify that a preparation task is required and automatically suggest an extension of the customer specification. Similarly, during training, the workflow system can alert a junior engineer about an inconsistent sequence of selected tasks and suggest a correct order. A long history of symbolic AI applications to FA problems showed that these methods can reliably assist engineers in their daily work. Available knowledge representation and reasoning frameworks allow for fast development and deployment of such systems to a lab. Reasoners of these frameworks can find (optimal) solutions for encoded problems in the required time and provide additional services, such as explanations of obtained results, what-if analysis, repair of solutions to take discrepancies into account, and many more. Because experts can interpret and, whenever required, modify symbolic knowledge bases, these methods are robust and unbiased by design. That is, in case a new product is designed, ramped up, or is in production, expert knowledge can immediately be used even in the absence of any historical data. Nevertheless, these advantages come at a cost, because knowledge acquisition and management get increasingly complicated as the knowledge bases grow. This problem led to a dramatic crash of general expert systems in the past. Therefore, such systems are usually applied in situations when general domain knowledge changes rather slowly and can easily be encoded by experts, like in the case of FA ontologies, or their environments are well understood and limited, like diagnosis or workflow systems. Therefore, the next section provides an overview of machine learning methods that can help to overcome these issues. MACHINE LEARNING Efficient knowledge acquisition and maintenance techniques play a key role in AI-based applications. Machine learning methods can solve this problem by training models that capture valuable knowledge even in situations when experts are unable to formulate it in a natural language. For example, creating an expert system that can recognize a fused wire on an image would require engineers to define a set of rules discriminating between target images and others. However, in practice, the definition of this rule set is impossible due to a large number of special cases, like tool- or product-specific features of images, shape/color differences between fused and other states of wires, or expressing the level of experts’ confidence in these definitions. Machine learning methods can solve tasks that are very difficult to approach by definition of knowledge bases or programming since these algorithms can improve their performance on some task with experience.[16,30] Therefore, a classic approach to the creation of a machine learning system starts with the definition of training experience, its learning task, performance criteria, and learning algorithm, shown schematically in Fig. 5. Fig. 5 Components of a machine-learning approach.
RkJQdWJsaXNoZXIy MTMyMzg5NA==