ABSTRACT
Lake Okeechobee data are routinely collected by sensors and transcribed from field forms to an electronic format. Occasionally, the effects of extraneous objects, device failure, and human errors may distort these field data. At present, the data are inspected manually by qualified engineers/scientists/technicians to identify abnormalities. This manual process is slow, costly, and sometimes inconsistent among inspectors. This paper presents a neural network approach that can efficiently, effectively, and consistently mimic human expertise in identification of days that contain abnormal water temperature. The approach begins with training 12 backpropagation neural networks (one per month) using the normal and abnormal daily patterns selected by the domain expert. The trained networks can then be used for identification in lieu of human experts. A total of 623 days of data from Lake Okeechobee were used: 241 days for training and 382 days for testing the networks. The results for the test data indicate that this approach achieves 98% accuracy in rejecting abnormal data and 95% accuracy in accepting normal data.