Combining classifiers to detect faults in wastewater networks

Abstract

This work presents a methodology for automatic detection of structural faults in sewers from CCTV footage, which has been improved by combining the outputs of different machine learning techniques. The predictions of support vector machine and random forest classifiers are combined using three distinct techniques: 'both', 'most likely' and 'stacking'. Each technique is tested on CCTV data taken from real surveys covering a range of pipes at locations in the south-west of the UK. The best tested technique, stacking, offers a 5% increase in accuracy for minimal impact in efficiency, proving useful for future development and implementation of the fault detection methodology.

Autoencoders for Semi-Supervised Water Level Modeling in Sewer Pipes with Sparse Labeled Data

“…Myrans et al [14] use GIST descriptors and Random Forest to classify scenes with root intrusion. An interesting approach is presented by Myrans et al [15], which is the first attempt to use different classifiers trained using GIST descriptors to detect fault samples, and then each classifier is combined to a single one using Hidden Markov Models.…”

“…An interesting topic for further research would be the combination of different latent spaces extracted from different methods, for example, combining the latent space from the classic AE, VAE and VQ-VAE as performed by Myrans et al [15]-in this case using different classifiers combined with Hidden Markov Models.…”

A defect classification methodology for sewer image sets with convolutional neural networks

“…They use 25-fold cross validation and provide the ROC curve along with the misclassification rates for various defect types, but work with a dataset that consists of approximately 37% images with defects, which is not representative of a realistic scenario. In [17] they combine both the SVM and the random forest on a dataset in which "approximately half " the images contained defects, and obtain results superior to either individual classifier. Again, the validation results are not representative of a real-world scenario because of the high prevalence of defects.…”