|
|
|
“Binary Classification of Ground Vehicles Based on the Acoustic Data Using Fuzzy Logic Rule-Based Classifiers”
by Hongwei Wu and Jerry M. Mendel
July 2002
Because the acoustic data corresponding to each run are time-variant, we segmented each run into one-second data blocks, and used the data blocks, which we called prototypes, for classification. The magnitudes of the second through 12th harmonics of each prototype were used as features. We found, by analyzing the features within each run and across runs, that the run-means and run-standard-deviations of the features vary from run to run for all kinds of vehicles. We therefore used type-2 fuzzy sets to model the uncertainties contained in these features, and then constructed type-2 fuzzy logic rule-based classifiers (FL-RBC) for three binary classification problems: tracked vs. wheeled vehicle, heavy-tracked vs. light-tracked vehicle, and heavy-wheeled vs. light-wheeled vehicle. To evaluate the performance of the type-2 FL-RBCs in a fair way, we also constructed the Bayesian classifiers and type-1 FL-RBCs and compared their performance through many experiments. The parameters of the Bayesian classifiers were estimated using the training prototypes; whereas, the parameters of both the type-1 and type-2 FL-RBCs were optimized using a steepest descent algorithm that minimized an objective function which depended upon the training prototypes. All classifiers had two working modes---non-adaptive and adaptive. When the false alarm rate (FAR) of a classifier in its non-adaptive mode is less than 0.5 then this classifier has a better performance in its adaptive mode than in its non-adaptive mode after a certain time.
We carried out the leave-one-out and leave-M-out experiments to evaluate the performance of all classifiers. In the leave-one-out experiments, only one run was used for testing, and all the other runs were used for training. In the leave-M-out experiments, one run of each kind of vehicle was used for testing, and all the other runs were used for training. Our experiments showed that for each binary classification problem, both the type-1 and type-2 FL-RBCs had significantly better performance than the Bayesian classifier, whereas the type-1 and type-2 FL-RBCs had similar performance, although most of the time, the type-2 FL-RBC had slightly better performance than the type-1 FL-RBC.
Both the type-1 and type-2 FL-RBC designs were tested for blind runs of the normal terrain. The blind test results of the type-2 FL-RBC designs were scored by our sponsor at the Army Research Laboratory. The scores were very high, which demonstrates that our type-2 FL-RBC designs for the binary classification of ground vehicles based on their acoustic emissions are successful.