The goal of this model is to form a better predictive model to classify wire breaks in prestressed concrete cylinder pipes.
The following model utilizes a dataset of 2000 sections of PCCP including 10 features such as length, class, wire diameter, wire spacing, # of wire turns, thickness of cylinder, soil resistivity, potential value, line current discharge, and land use. A machine learning ensemble of four state-of-the-art algorithms was created to achieve best of breed results. These models were: a) two-class averaged perceptron, b) two-class bayes point machines, c) two class decision jungle, and d) two-class locally deep support vector machines. All performed similarly. Their accuracies were 0.9664, 0.9674, 0.9677, and 0.9557, respectively. To qualify these results, these accuracies occurred at a recall value of 0. This indicates the model is saturated with false negatives (a wire break actually occurred, but the model predicted no wire break). The level of recall determines how cautious the model should be in predicting no wire breaks. If you predict a break and it actually is fine you've wasted thousands of dollars rehabilitating a good pipe section. If you predict a clean wire but it breaks, the pipe bursts costing millions of dollars in damage As one can see, a high recall is desirable in situations where the cost of a false negative is tremendous. For small pipes the cost of a false negative may be acceptable, but the price of a false negative in large pressurized pipes is unacceptable. Therefore, this model will be reworked to achieve a recall of close 1 and bring false negative predictions close to zero. This project is on an "as-is" basis and makes no warranties regarding any information provided on or through it. It disclaims any liability for damages resulting from the model's use. This model still is not ready for practical application.