Comparative analysis of fuzzy-neural network implementations on an autonomous electric vehicle / John Paolo Ramoso [and] Manuel Ramos.

Includes bibliographical references (pages 72-73).

The aim of this research is to implement an optimized hybrid fuzzy-neural (FN) algorithm for an autonomous electric vehicle’s stop-and-go decision-making and control. Four (4) different algorithms (purely fuzzy logic (FL), one (1) hidden layer (H1) FN, two (2) hidden layers (H2) FN, and purely neural network (NN)) were deployed in a buggy-type electric vehicle (EV) to compare their performances in real road conditions. The test EV was equipped with a LiDAR Lite sensor which served as the range finder to measure headway distance while an optical flow sensor and the motor’s built-in hall sensors were used to measure speed. The EV was also retrofitted with a dsPIC30F4011 microcontroller for processing and control. Both indoor and outdoor road tests were conducted to compare the difference between a controlled environment (well-lit with good road conditions) versus actual road conditions (including physical limitations), respectively. It was observed in the indoor tests that increasing the hidden layers from H1 to H2 made the algorithm more robust and decreased jerking phenomenon when the vehicle was stationary. Results from the outdoor tests also revealed that FN network with H2 (successful in eight (8) out of ten (10) runs) had better control in maintaining proper headway distance and more fluid transition in acceleration and deceleration. Hardware considerations were also outlined focusing on deploying machine learning codes and weights to a microcontroller. The ~56kB initial code size was way above the allowable 48kB program memory of the microcontroller therefore the data type of the weights were changed to shrink the code to ~38kB.