The design and training of an intelligent sensor implementation

The design and VLSI implementation of an intelligent sensor that can be used for process control applications requiring image capture or non-contact measurement is presented in this paper. The sensor architecture is based on an analog VLSI realization of an artificial neural network with an integrated photosensitive array. A 10×10 array of photosensitive cells has been designed as the input nodes to a two-layer feedforward neural network. Four groups each with 25 photosensors are fully connected to 4 neurons respectively in the hidden layer. The 16 hidden neurons are fully connected to 5 neurons in the output layer. A development program has been written that determines the optimal set of weights using the modified backpropagation algorithm. The neural network has been trained to recognize 18 different types of patterns. The final implementation of this intelligent sensor has the dimensions of 4500×4820 design scale microns and contains approximately 1630 analog devices fabricated using a 3μ single-polysilicon, double-metal, p-well CMOS process