Augmentation of field fluorescence measures for improved in situ contaminant detection.

This research proposes a new method that fuses data from the field and lab-based optical measures coupled with machine learning algorithms to quantify the concentrations of toxic contaminants found in fuels and oil sands process-affected water. Selected pairs of excitation/emission intensities at key wavelengths are inputs to an augmentation neural network (NN), trained using lab-based measurements, that generates synthetic high-resolution spectra. Then, an image processing NN is used to estimate the contaminant concentrations from the spectra generated from a few key wavelengths. The presented approach is tested using naphthenic acids, phenol, fluoranthene and pyrene spiked into natural waters. The spills or loss of containment of these contaminants represent a significant risk to the environment and public health, requiring accurate and rapid detection methods to protect the surrounding aquatic environment. Results were compared with models based on only the corresponding peak intensities of each contaminant and with an image processing NN using the original spectra. Naphthenic acids, fluoranthene and pyrene were easy to detect by all methods; however, performance for more challenging signals to identify, such as phenol, was optimized by the proposed method (peak picking with mean absolute error (MAE) of 30.48 µg/L, generated excitation-emission matrix with MAE of 8.30 µg/L). Results suggested that data fusion and machine learning techniques can improve the detection of contaminants in the aquatic environment at environmentally relevant concentrations.

Application of machine learning methods for rapid fluorescence-based detection of naphthenic acids and phenol in natural surface waters.

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, such as naphthenic acids (NAs) and phenol, accurately and rapidly in the produced oil sands process-affected water (OSPW) is required to ensure the protection of the surrounding aquatic environment. In this paper, fluorescence techniques are investigated to rapidly quantify NAs and phenol concentrations in natural surface waters. Machine learning approaches were applied to identify relevant spectral features to improve detection accuracy in the presence of background interference from organic matter in natural waters. NAs were relatively easy to detect by all methods, however deep convolutional neural networks (CNN) resulted in optimized performance for phenol with mean absolute errors of 1.78 - 1.81 mg/L and 4.68-5.41 µg/L, respectively. Visualization of spectral areas of importance revealed that deep CNNs utilized logical areas of the fluorescence spectra associated with NAs and phenol signals. Results suggest machine learning approaches to interpreting fluorescence data can accurately predict individual toxic components of OSPW in natural waters at environmentally relevant concentrations.

Application of convolutional neural networks for prediction of disinfection by-products.

Fluorescence spectroscopy can provide high-level chemical characterization and quantification that is suitable for use in online process monitoring and control. However, the high-dimensionality of excitation-emission matrices and superposition of underlying signals is a major challenge to implementation. Herein the use of Convolutional Neural Networks (CNNs) is investigated to interpret fluorescence spectra and predict the formation of disinfection by-products during drinking water treatment. Using deep CNNs, mean absolute prediction error on a test set of data for total trihalomethanes, total haloacetic acids, and the major individual species were all < 6 µg/L and represent a significant difference improved by 39-62% compared to multi-layer perceptron type networks. Heat maps that identify spectral areas of importance for prediction showed unique humic-like and protein-like regions for individual disinfection by-product species that can be used to validate models and provide insight into precursor characteristics. The use of fluorescence spectroscopy coupled with deep CNNs shows promise to be used for rapid estimation of DBP formation potentials without the need for extensive data pre-processing or dimensionality reduction. Knowledge of DBP formation potentials in near real-time can enable tighter treatment controls and management efforts to minimize the exposure of the public to DBPs.

Comparison of dimensionality reduction techniques for cross-source transfer of fluorescence contaminant detection models.

Fluorescence spectroscopy shows promise as a tool for monitoring water quality due to its real-time capabilities and sensitive detection of several compounds of interest. Previous work has shown the possible use of fluorescence to detect and quantify low levels of polycyclic aromatic hydrocarbons and fluorescing pesticides. However, the fluorescence-based contaminant detection models are highly source-specific and require significant effort and resources to build and calibrate them for each source water of interest. In this study, the novel application of data processing techniques was investigated to enable the transfer of fluorescence detection models from one water source to another. A contaminant detection model from a relatively consistent and low organic background source (Lake Ontario, TOC: 2.07-2.26 mg L-1) was transferred to the Otonabee River, which has higher organic concentrations and distinct characteristics (TOC: 5.20-5.66 mg L-1). Only a few additional fluorescence spectra of the background water quality and contaminants of interest were required to successfully transfer the model, without the need for labelled samples in the new source. Notable differences in peak location and spectral shape of identical compounds were found in source-specific models between the two water sources, implying variability in fluorescence signals resulting from environmental conditions. Despite the impact of environmental conditions, features identified by principal component analysis (PCA) and an autoencoder produced sensitive transferred models capable of addressing the spatial and temporal source diversity with mean absolute error (MAE) < 0.5 µg L-1 for quantification of PAHs and pesticides at concentrations between 0.1 and 7 µg L-1. The results of this study show the potential of the cross-source transferred model to be implemented in a wide range of environmental conditions.

Integration of weather conditions for predicting microbial water quality using Bayesian Belief Networks.

Levels of fecal indicator bacteria (FIB) provide a surrogate measure of the microbial quality of water used for a wide range of applications. Despite the common use of these measures, a significant limitation is a delay in results due to the time required for cultivation and enumeration of FIB. Testing requires at least 18-24 h, and therefore, FIB cannot be used to identify current or real-time microbial water quality. An approach of nowcasting or empirical modelling approaches that incorporate water quality, environmental, and weather variables to predict FIB levels in real-time has been developed with some success. However, FIB levels are dependent on a complex interaction of numerous variables, which can be challenging to model with ordinary linear regression or classification methods most commonly applied. In this study, novel use of Bayesian Belief Networks (BBNs) that allow for a probabilistic representation of complex variable interactions is investigated for real-time modelling of FIB levels surface waters. In particular, the integration of both water quality measures and current/historical weather for prediction of fecal coliforms and Escherichia coli levels is achieved using BBNs. For 4-bin classification of fecal coliform levels, BBNs increased prediction accuracy by 25%-54% compared to other previously used techniques including logistic regression, Naïve Bayes, and random forests. Binary prediction of E. coli levels exceeding a threshold of 20 CFU/100 mL was also significantly improved using BBNs with prediction accuracies >90% for all monitoring sites. Advantages of the BBN approach are also demonstrated identifying the ability to make predictions from incomplete monitoring data as well as probabilistic inference of variable importance in FIB levels. In particular, the results indicate that water quality surrogates such as conductivity are essential to real-time prediction of FIB. The results and models described in this work can be readily utilized to provide accurate and real-time assessments of FIB levels in surface waters utilizing commonly monitored parameters.

Towards real-time detection of wastewater in surface waters using fluorescence spectroscopy.

The presence of municipal wastewater at the intake of a major drinking water treatment facility located on Lake Ontario was examined using fluorescence data collected during a period of continuous monitoring. In addition, controlled mixing of lake water and wastewater sampled from a local treatment facility were conducted using a bench-scale fluorescence system to quantify observed changes in natural organic matter. Multivariate linear regression was applied to components derived from parallel factors analysis. The resulting mean absolute error for predicted wastewater level was 0.22% (V/V, wastewater/lake water), indicating that wastewater detection at below 1.0% (V/V) was possible. Analyses of sucralose, a wastewater indicator, were conducted for treated wastewater as well as surface water collected at two intake locations on Lake Ontario. Results suggested minimal wastewater contribution at the drinking water intake. A wastewater detection model using a moving baseline was developed and applied to continuous fluorescence data collected at one of the drinking water intakes, which agreed well with sucralose results. Furthermore, the simulated addition of 1.0% (V/V) of wastewater/wastewater was detectable in 89% of samples analyzed, demonstrating the utility of fluorescence-based wastewater monitoring.

Neural networks for dimensionality reduction of fluorescence spectra and prediction of drinking water disinfection by-products.

The use of fluorescence data coupled with neural networks for improved predictability of drinking water disinfection by-products (DBPs) was investigated. Novel application of autoencoders to process high-dimensional fluorescence data was related to common dimensionality reduction techniques of parallel factors analysis (PARAFAC) and principal component analysis (PCA). The proposed method was assessed based on component interpretability as well as for prediction of organic matter reactivity to formation of DBPs. Optimal prediction accuracies on a validation dataset were observed with an autoencoder-neural network approach or by utilizing the full spectrum without pre-processing. Latent representation by an autoencoder appeared to mitigate overfitting when compared to other methods. Although DBP prediction error was minimized by other pre-processing techniques, PARAFAC yielded interpretable components which resemble fluorescence expected from individual organic fluorophores. Through analysis of the network weights, fluorescence regions associated with DBP formation can be identified, representing a potential method to distinguish reactivity between fluorophore groupings. However, distinct results due to the applied dimensionality reduction approaches were observed, dictating a need for considering the role of data pre-processing in the interpretability of the results. In comparison to common organic measures currently used for DBP formation prediction, fluorescence was shown to improve prediction accuracies, with improvements to DBP prediction best realized when appropriate pre-processing and regression techniques were applied. The results of this study show promise for the potential application of neural networks to best utilize fluorescence EEM data for prediction of organic matter reactivity.

Characterization of UF foulants and fouling mechanisms when applying low in-line coagulant pre-treatment.

Fluorescence spectroscopy was used as a characterization method to examine organic fouling of single ultrafiltration (UF) fibres at bench-scale. Low doses of coagulant were applied to modify organic properties, without significant formation of precipitates. This approach compliments previous studies investigating coagulation as a pre-treatment method for UF fouling control, which have principally focused on reduction of foulant concentrations. Using a continuous system, short time-scale fluorescence results demonstrated significant adsorption of humic components to virgin membrane fibres. Following an initial adsorption phase, protein-like material was the only organic component to be significantly removed by UF. Low doses of coagulant (<1 mg/L as alum; < 0.043 mg/L as Al3+) were observed to significantly reduce irreversible fouling rates for two different surface waters. Paralleling reduced irreversible fouling, surface tryptophan fluorescence resulting from material adsorbed to the fouled membrane increased, as measured using a fibre optic probe. Analysis of peak shifts in the protein-like component revealed a red-shift at low coagulant dose, possibly indicative of greater exposure of tryptophan residues resulting from conformational changes in the protein structure. It is hypothesized that low coagulant doses modified membrane-foulant interactions, resulting in increased adsorption of protein-like matter to the surface. Subsequent interactions of bulk foulants with the adsorbed organic monolayer discouraged further adsorption and reduced irreversible fouling potential.

Investigation of ozone and peroxone impacts on natural organic matter character and biofiltration performance using fluorescence spectroscopy.

Impacts of ozonation alone as well as an advanced oxidation process of ozone plus hydrogen peroxide (H2O2 + O3) on organic matter prior to and following biofiltration were studied at pilot-scale. Three biofilters were operated in parallel to assess the effects of varying pre-treatment types and dosages. Conventionally treated water (coagulation/flocculation/sedimentation) was fed to one control biofilter, while the remaining two received water with varying applied doses of O3 or H2O2 + O3. Changes in organic matter were characterized using parallel factors analysis (PARAFAC) and fluorescence peak shifts. Intensities of all PARAFAC components were reduced by pre-oxidation, however, individual humic-like components were observed to be impacted to varying degrees upon exposure to O3 or H2O2 + O3. While the control biofilter uniformly reduced fluorescence of all PARAFAC components, three of the humic-like components were produced by biofiltration only when pre-oxidation was applied. A fluorescence red shift, which occurred with the application of O3 or H2O2 + O3, was attributed to a relative increase in carbonyl-containing components based on previously reported results. A subsequent blue shift in fluorescence caused by biofiltration which received pre-oxidized water indicated that biological treatment readily utilized organics produced by pre-oxidation. The results provide an understanding as to the impacts of organic matter character and pre-oxidation on biofiltration efficiency for organic matter removal.

Investigation of fluorescence methods for rapid detection of municipal wastewater impact on drinking water sources.

Fluorescence spectroscopy as a means to detect low levels of treated wastewater impact on two source waters was investigated using effluents from five wastewater facilities. To identify how best to interpret the fluorescence excitation-emission matrices (EEMs) for detecting the presence of wastewater, several feature selection and classification methods were compared. An expert supervised regional integration approach was used based on previously identified features which distinguish biologically processed organic matter including protein-like fluorescence and the ratio of protein to humic-like fluorescence. Use of nicotinamide adenine dinucleotide-like (NADH) fluorescence was found to result in higher linear correlations for low levels of wastewater presence. Parallel factors analysis (PARAFAC) was also applied to contrast an unsupervised multiway approach to identify underlying fluorescing components. A humic-like component attributed to reduced semiquinone-like structures was found to best correlate with wastewater presence. These fluorescent features were used to classify, by volume, low (0.1-0.5%), medium (1-2%), and high (5-15%) levels by applying support vector machines (SVMs) and logistic regression. The ability of SVMs to utilize high-dimensional input data without prior feature selection was demonstrated through their performance when considering full unprocessed EEMs (66.7% accuracy). The observed high classification accuracies are encouraging when considering implementation of fluorescence spectroscopy as a water quality monitoring tool. Furthermore, the use of SVMs for classification of fluorescence data presents itself as a promising novel approach by directly utilizing the high-dimensional EEMs.

Fluorescence spectroscopy for monitoring reduction of natural organic matter and halogenated furanone precursors by biofiltration.

The application of fluorescence spectroscopy to monitor natural organic matter (NOM) reduction as a function of biofiltration performance was investigated. This study was conducted at pilot-scale where a conventional media filter was compared to six biofilters employing varying enhancement strategies. Overall reductions of NOM were identified by measuring dissolved organic carbon (DOC), and UV absorbance at 254 nm, as well as characterization of organic sub-fractions by liquid chromatography-organic carbon detection (LC-OCD) and parallel factors analysis (PARAFAC) of fluorescence excitation-emission matrices (FEEM). The biofilter using granular activated carbon media, with exhausted absorptive capacity, was found to provide the highest removal of all identified PARAFAC components. A microbial or processed humic-like component was found to be most amenable to biodegradation by biofilters and removal by conventional treatment. One refractory humic-like component, detectable only by FEEM-PARAFAC, was not well removed by biofiltration or conventional treatment. All biofilters removed protein-like material to a high degree relative to conventional treatment. The formation potential of two halogenated furanones, 3-chloro-4(dichloromethyl)-2(5H)-furanone (MX) and mucochloric acid (MCA), as well as overall treated water genotoxicity are also reported. Using the organic characterization results possible halogenated furanone and genotoxicity precursors are identified. Comparison of FEEM-PARAFAC and LC-OCD results revealed polysaccharides as potential MX/MCA precursors.

Comparison of three-dimensional fluorescence analysis methods for predicting formation of trihalomethanes and haloacetic acids.

This work investigated the application of several fluorescence excitation-emission matrix analysis methods as natural organic matter (NOM) indicators for use in predicting the formation of trihalomethanes (THMs) and haloacetic acids (HAAs). Waters from four different sources (two rivers and two lakes) were subjected to jar testing followed by 24hr disinfection by-product formation tests using chlorine. NOM was quantified using three common measures: dissolved organic carbon, ultraviolet absorbance at 254 nm, and specific ultraviolet absorbance as well as by principal component analysis, peak picking, and parallel factor analysis of fluorescence spectra. Based on multi-linear modeling of THMs and HAAs, principle component (PC) scores resulted in the lowest mean squared prediction error of cross-folded test sets (THMs: 43.7 (µg/L)(2), HAAs: 233.3 (µg/L)(2)). Inclusion of principle components representative of protein-like material significantly decreased prediction error for both THMs and HAAs. Parallel factor analysis did not identify a protein-like component and resulted in prediction errors similar to traditional NOM surrogates as well as fluorescence peak picking. These results support the value of fluorescence excitation-emission matrix-principal component analysis as a suitable NOM indicator in predicting the formation of THMs and HAAs for the water sources studied.