Using neural networks as a fault detection mechanism in MEMS devices

Micro Electro Mechanical Systems (MEMS) were first proposed about 20 years ago. Today, many different kinds have been fabricated and are used in industry, space and scientific fields. The scale and relative size of analog integrated circuits and non-electrical parts are becoming smaller. As a result, the need for automatic test of MEMS is a critical requirement in MEMS fabrication and maintenance. The rapid progress in the design of these systems has not, however, been accompanied by a similar progress in fault classification technologies. MEMS are naturally very non-linear, complex and multi-domain and systems are fabricated near to each other. A large number of faults of different types may occur. This paper presents a combination of a Competitive Neural Network (CNN) and a Robust Heteroscedastic Probabilistic Neural Network (RHPNN) for fault detection in MEMS. The RHPNN has previously been proposed for analog fault detection. Finding the optimum kernel number in the second layer is a drawback of the RHPNN method. In this paper we have used a CNN for finding the optimum kernel number automatically. In addition, as the simulation results show, the correct fault detection percentage is increased in comparison with the RHPNN alone.