Analog Neural Network Implementation for a Real-Time Surface Classification Application

This paper deals with the implementation of a CMOS analog neural network (NN) that has to be integrated in a new kind of optoelectronic measurement system. The aim is to achieve real-time surface recognition using a phase-shift rangefinder and a neural network. NN architecture is a multilayer perceptron (MLP) with two analog input signals provided by the rangefinder, three processing neurons in the hidden layer, and one output neuron whose output voltage indicates the detected surface. As the complete structure is analog, no analog-to-digital conversions or signal processing between the rangefinder and the network is necessary. Furthermore, the 3.3-V voltage supply, relative to the chosen CMOS 0.35-mum technology, allows to reduce the system power consumption. This paper focuses on the implementation in an ASIC of an elementary part of the NN, called neuron, and on the achievement of the complete NN from the integration of three ASICs in a printed circuit board. Comparisons between ideal case, simulations and tests are detailed in order to validate the design and the good functioning of the complete structure.