CAFE & Test: Causal feature selection for machine learning test of analog, mixed-signal and RF systems


The integration capabilities offered by current nanoscale CMOS technologies enable the fabrication of com- plete and very complex mixed-signal systems on a single die. However, manufacturing processes are prone to imperfections and defects that may degrade –sometimes catastrophically– the intended functionality of the fabricated circuits. Extensive production tests are then needed in order to separate these defective or unreliable parts from functionally correct devices. Unfortunately, the co-integration of blocks of very distinct nature (analog, mixed-signal, digital, RF, MEMS, etc.) as well as the limited access to internal nodes in an integrated system make the test of these devices a very challenging and costly task.

Standard test methods for analog, mixed-signal and RF (AMS-RF) circuits are based on the direct measurement of complex circuit specifications, relying on the use of expensive dedicated test equipment. The test of AMS-RF blocks embedded in a complex systems has become a challenging, costly and time consuming task that has been identified as one of the main bottlenecks in the production of current and future integrated systems.

Positioning of the project

Machine learning-based test, also known as indirect test or alternate test, is a promising strategy for overcoming these issues. Indirect test reduces the complexity and cost of production tests by replacing conventional functional tests at the production line for a set of low-cost indirect observations, often called signatures. Test results are then inferred by post-processing these signatures by building non-linear multi-dimensional regression models. The underlying idea is that signatures are easier to measure than specifications and can be extracted using low-cost equipment, or, even more advantageously, by simple on-chip test instruments that can be integrated together with the Device Under Test (DUT). That is, indirect test naturally enable efficient and practical Built-In Self-Test (BIST) for AMS-RF circuits. The combination of machine learning-based test and built-in test instruments has the potential to overcome many of the issues in current AMS-RF test.

Moreover, it is also worth to remark that AMS-RF BIST techniques, if made feasible, not only reduce test complexity and cost at the production line, but they enable interesting features such as in-the-field test, on-line test, diagnosis capabilities, adaptive self-calibration, and self-healing that can have a huge socio-economic impact in many different applications. Indeed, these features are key attributes for circuits deployed in applications that require high reliability (e. g. harsh environments, systems with limited or impossible access), or that are safety critical (e. g. automotive, avionics, space, healthcare).


Within the framework of the Advanced Data Mining research axe in the Persyval-Lab Labex initiative, the proposed exploratory project brings together the expertise of microelectronic designers, test engineers and data mining mathematicians with the goal of exploring, identifying and developing systematic methodologies for reliable and accurate built-in Indirect Test strategies for AMS-RF complex systems. For this purpose, the project is built around the collaboration of three research institutions: TIMA Laboratory, Grenoble, France (Unite ́ Mixte de Recherche 5159), IMSE-CNM, Seville, Spain (Centro Mixto Universidad de Sevilla – CSIC), and LIRIS, Lyon, France (UMR 5205). 

Activities and news

3-4-2016: Seminar Bayesian networks & Causal inference by Prof. Alexander Aussem, LIRIS, Lyon

Abstract: The central aim of many studies in the social and medical sciences is the elucidation of cause-effect relationships among variables or events. While, the appropriate methodology for extracting such causal relationships from data is still an open question (and fiercly debated), graphical models provide a simple and convenient way of communicating causal claims. The causal diagram represents the investigator's understanding of the causal influences among measurable - but partially observed - quantities in the domain. In this talk, I review  the basics  of causal inference in graphical model and demonstrate, using simple examples, how non-trivial causal phenomena, paradoxes and controversies in causal analysis can be understood, exemplified and analyzed using the do-calculus developped by Judea Pearl. I also show how selection bias and missing data mechanisms can be represented in the graph and discuss the conditions under which consistent causal (or probabilistic) inferences can be made from such corrupted data sets.


- G. Leger and M. J. Barragan, "Questioning the reliability of Monte Carlo simulation for machine learning test validation," 2016 21th IEEE European Test Symposium (ETS), Amsterdam, 2016, pp. 1-6.

- G. Leger and M. J. Barragan, "Brownian distance correlation-directed search: Afast feature selection technique for alternate test," Integration, the VLSI Journal, vol. 55, September 2016, pp- 401-414

- M. J. Barragan, G. Leger, A. Gines, E. Peralias, A. Rueda, "On the limits of machine learning-based test: a calibrated mixed-signal system case study," Design Automation and Test in Europe 2017 (DATE'17), accepted for publication.