Application of machine learning techniques on outlier detection of piezometer data

Hugo A. Brandão*, Tatiana B. Santos2, Christiane S. de Oliveira 3 and Ana P. F. Costa

1. Geotechnical Department, Pimenta de Ávila, Brazil, Innovation Coordinator and Geotechnical Engineer, hugo.assis@pimentadeavila.com.br and (+55 31) 99912-1993
2. Mining Engineering Department, Federal University of Ouro Preto, Brazil, Professor, tatiana.santos@ufop.edu.br, (+55 31) 97566-7590
3. Geotechnical Engineering Department, Gorceix Foundation, Brazil, Graduate Student, christiane.oliveira@samarco.com and (+55 31) 98735-2113
4. Geological Department, Pimenta de Ávila, Brazil, Geological Engineer, ana.felix@pimentadeavila.com.br and (+55 31) 99148-6084

ABSTRACT

The number of reported incidents associated with tailings storage facilities (TSFs) has increased in the past few years. It is estimated that the reported failures between 1915 and 2022 resulted in the loss of life of more than 2.650 people and more than 250 million of m3 of residue released to the environment. Dam monitoring provides engineers important information that can be used to better understand the TSF behaviour and the current practice relies primarily on visual and manual analysis. Due to instruments automation, large sets of data are being generated and visual/manual analysis leaves much to desire when applied to anomaly detection on large data sets. Machine learning and statistical models can be used to pre-screen anomalies and provide information to engineers, helping the TSF operational team with real-time monitoring. This paper presents a methodology to identify anomalies based on predictive models and the obtained residues. First, the k-nearest neighbours (KNN), a machine learning model, was used to predict future readings of a piezometer based on the last reading. The calibrated model presented R2 values over than 0.91, RMSE near to 0.16 m and MAE values equal to 0.10 m. After the prediction of pore pressure values, the residues based on the difference between the predicted and real values were calculated. They were used to identify the anomalies in the data. A statistical distribution associated with a 95% interval of confidence and local outlier factor (LOF) method were applied. The LOF method presented the best results and not only it was able to identify the same anomalous readings of the statistical distribution, but it also identified other anomalous data points.

INTRODUCTION

Mining operations have a large impact in the environment and society and generate large amounts of tailings during the ore extraction process that are commonly stored in Tailings Storage Facilities (TSF). Between 1915 and 2022, 257 incidents related to TSF were reported that resulted in the release of approximately 250 Mm3 and loss of approximately 2.650 lives (Piciullo et al., 2022).

The monitoring of a TSF is crucial to promote the safe operation of the structure as it provides critical information about the TSF behaviour. This allows a better risk management and helps the operational team to reduce the probability of failure. The implementation of a monitoring plan is a legal requirement in many countries and recommended by the Global Industry Standard on Tailings Management (GISTM).

The pore water pressure (PWP) is one of the many variables that can be monitored in a TSF and it is related to several failures modes. By monitoring the PWP, the dam safety team can identify anomalous behaviour in advance that could lead to a TSF failure and providing the owner with time to act and reduce the probability of occurrence of the failure mode or reduce its severity (ICOLD, 2018).

The monitoring data can be used to better understand dam ́s behaviour based on deterministic and statistical models (ICOLD, 2018). Li et al. (2019) also presented the applicability of predictive models, such as time series and machine learning, in the analysis of monitoring data.

The automation of monitoring instruments has become a common practice as it provides more reliable and near real time information. It resulted in the generation of large sets of data that must be analysed in order to support the TSF operation. This task is often performed visually through chart analysis with low efficiency and high subjectivity.

This paper presents a general methodology applicable to anomaly identification based on two machine learning models, the k-nearest neighbours (KNN), a predictive model, and local outlier factor (LOF), an anomaly detection model.

METHODOLOGY

Figure 1 presents the applied methodology in this study. First, an exploratory data analysis was performed on the piezometer dataset. The dataset was then split into 80% training data and 20% testing data. The KNN model was calibrated and evaluated based on the evaluation metrics (R2, MAE and RMSE). The best fitted model was than selected and used to obtain the predicted values (yp) based on past readings. The absolute residue values ware calculated and the LOF model was applied in order identify the anomalies in the data.