Brett A. Story

Competitive probabilistic neural network

Probabilistic neural networks (PNNs) are artificial neural network algorithms widely used in pattern recognition and classification problems. In the traditional PNN algorithm, the probability density function (PDF) is approximated using the entire training dataset for each class. In some complex datasets, classmate clusters may be located far from each other and these distances between clusters may cause a reduction in the correct class’s posterior probability and lead to misclassification. This paper presents a novel PNN algorithm, the competitive probabilistic neural network (CPNN). In the CPNN, a competitive layer ranks kernels for each class and an optimum fraction of kernels are selected to estimate the class-conditional probability. Using a stratified, repeated, random subsampling cross-validation procedure and 9 benchmark classification datasets, CPNN is compared to both traditional PNN and the state of the art (e.g. enhanced probabilistic neural network, EPNN). These datasets are examined with and without noise and the algorithm is evaluated with several ratios of training to testing data. In all datasets (225 simulation categories), performance percentages of both CPNN and EPNN are greater than or equivalent to that of the traditional PNN; in 73% of simulation categories, the CPNN analyses show modest improvement in performance over the state of the art.

Bridge Impact Detection and Classification using Artificial Neural Networks

Low-clearance railroad bridges are susceptible to impacts from trucks that exceed the required clearance. These impacts can be detected by mounting instrumentation on the bridge, such as accelerometers and inclinometers, and analyzing data from these sensors. A train running over a bridge also generates a significant response, and the bridge responses produced by trains are not always dissimilar to those produced by vehicle impacts. Understanding whether an event is produced by a vehicle impact or a train is critical; a vehicle impact should be investigated, while a train crossing the bridge is part of a bridge’s regular operation. This paper presents a system of artificial neural networks capable of examining response time histories and determining if the response is a train or an impact. Event data obtained from sensor modules mounted on several diverse railroad bridges has been analyzed to identify key signal characteristics, train neural algorithms, and classify events. Signal characteristics of importance include time and frequency content and a signal center of mass metric. Combinations of these characteristics have been used to improve impact detection on specific bridges to 66%- 97% and reduce false positive rates to 0.09% - 1.82%.