Ain’t got time for this? Reducing manual evaluation effort with Machine Learning based Grouping of Analog Waveform Test Data

Abstract:
Design of integrated circuits (ICs) requires sophisticated evaluation of their functionality following defined characterisation and test procedures. The functional requirements are typically examined using automated setups by means of rule-based evaluation on measured transient waveforms. Verification of the IC includes characterisation of the behaviour beyond specified operating conditions. This is done to identify where and how the ICs start to fail by showing a deviation compared to the behavior in specified operating conditions. Each of these fails requires individual manual and hence time consuming and expensive analysis of its corresponding waveforms and classification of the fail reason. To drastically reduce the effort of this process, we propose to employ unsupervised machine learning algorithms for automated grouping of similar fail scenarios. For this process, we extract features from the recorded waveforms and use them for further manual analysis or input for cluster algorithms. We evaluate our proposed method by a case study using an industrial test data set. Our results show the automatically generated groups to efficiently summarise the behaviour in case of potential fail scenarios and to provide users with representative examples and data visualisation. In the evaluated case study, our approach shows the potential to cut down manual effort by reducing the number of scenarios to be evaluated by a factor of 14.