Application of Machine Learning Algorithms For Crack Detection in PVC Pipes

Most of the underground sewer infrastructure in United States uses Polyvinyl Chloride (PVC) pipes to transport toxic fluids. Cracks in underground PVC pipes are a major cause of effluent discharge in underground sewer systems. Released effluents not only pose risk to the environment, but are also a threat to public health. As current industry standard, utility operators use a closed circuit Television (CCTV) camera mounted crawler to pass through the pipes, and record video to classify condition of the piping network. CCTV based systems are expensive and crew-hour intensive. Recently developed acoustic based pipeline inspection systems are being adopted by the utility operators. These systems, however, do not detect presence of cracks in pipes. This paper reports results of a study to monitor presence of cracks in PVC pipes using acoustic signals. The collected data from extensive laboratory trials is processed using machine learning algorithms to classify the difference between a clean and cracked pipe samples. We use Decision Tree, K-nearest neighbors (KNN), and Naive Bayes (NB) algorithms. The DT and KNN algorithm scores show the highest convergence between acoustic samples from a cracked pipe at frequencies greater than 3.0 kHz. The paper also lays out precision scores obtained from using machine learning algorithms on acoustic data from clean and cracked pipe samples.