FPGA Platform for Prototyping and Evaluation of Neural Network Automotive Applications

In this paper we present an FPGA based reconfigurable computing platform for prototyping and evaluation of advanced neural network based applications for control and diagnostics i n an automotive sub-systems. Recurrent networks with internal or extemal feedback have proven to be very useful as models or controllcrs for dynamical systems. Our system implements a recurrent multi-layer perceptron (RMLP), which may be regarded as a combined generalization of a feedforward MLP and a one-layer fully recurrent network. The neuroprocessor design was implemented using a single Xilinx Virtex FPGA and made extensive use of a variety of techniques; including intra-layer parallelism, inter-layer multiplexing, and fixed-point bit-serial based computational techniques to achieve extreme design compactness. Flexibility of the design was achieved by allowing the architecture to be on-the-fly programmable from RAM. The intra-layer architecture of the neuroprocessor was organized in a SIMD configuration. While the overall goal is to develop an inexpensive and commercially viable ASIC based on the FPGA IP core, we describe here our FPGA design implementation and the two challenging applications, misfire detection and engine idle speed control, that had served a s the focus for this effort. 1