Thesis Defence: Machine Learning Applied for Contaminant Detection in Natural Surface Waters

Maria Claudia Rincón Remolina, supervised by Dr. Nicolas Peleato, will defend their thesis titled “Machine Learning Applied for Contaminant Detection in Natural Surface Waters” in partial fulfillment of the requirements for the degree of Master of Applied Science in Mechanical Engineering.

An abstract for Maria’s thesis is included below.

Defences are open to all members of the campus community as well as the general public.

To access a zoom link to attend this defence please contact the supervisor at nicolas.peleato@ubc.ca.

---

ABSTRACT

Approximately 1.4 billion m3 of fluid tailings produced from oil sands mining operations are currently being held in Alberta, Canada and pose a significant risk to the environment if not properly treated and managed. The ability to quantify levels of toxic compounds present in oil sands tailings, such as naphthenic acids (NAs), phenol, and polycyclic aromatic hydrocarbons (PAHs), is required to ensure the protection of the surrounding aquatic environment. In this thesis, machine learning techniques are applied to rapidly detect toxic contaminants from the production of petroleum products and crude oil extraction in natural surface waters. The first study implemented modern neural network structures for improving supervised regression of NA and phenol concentrations in water. Multilayer perceptron (MLP) and convolutional neural networks (CNN) were both capable of identifying relevant spectral features of the pollutants and were accurate in estimating concentrations in the presence of background interference from organic matter in natural waters. NAs were relatively easy to detect by MLP and CNN, however, deep CNNs resulted in optimized performance for phenol with mean absolute errors (MAE) of 1.78 – 1.81 mg/L and 4.68 – 5.41 μg/L for NAs and phenol, respectively. Visualization of spectral areas of importance revealed that both MLPs and deep CNNs utilized logical regions of the fluorescence spectra associated with NAs and phenol signals. The second study aimed to improve available methods for in situ quantification of NA, phenol, pyrene, and fluoranthene concentrations in natural waters. An augmentation network that fuses data from field and lab-based optical measures was used to improve the quality of field measurements. NAs, fluoranthene and pyrene were easy to quantify by the augmentation network, however, performance for phenol was more challenging due to the low intensity of the signal and overlap with other more dominant signals from NAs. Detection and quantification of phenol was optimized by the augmentation network proposed compared to standard in situ methods such peak-picking (peak-picking MAE of 30.48 μg/L, augmentation network MAE of 8.30 μg/L). Finally, an unsupervised approach to detecting pollutants in water was investigated, where no knowledge of contaminant levels is needed to train the model. A variational autoencoder was used to reduce data dimensionality and, through unsupervised methods, NAs and phenol could be identified in natural surface waters. According to the unsupervised classification methods, water samples contaminated with only NAs had the highest accuracy scores. The model classified water samples contaminated with NAs and phenol as if they were only contaminated with NAs, since NAs signals overshadowed phenol signals.