Optimizing Industrial Systems with Machine Learning Anomaly Detection

Document Type



The United Arab Emirates’ (UAE) goal is to be among the top globally recognized countries in machine learning, technological innovation, advanced technology, and artificial intelligence (AI). As H.H Sheikh Mohammed Bin Rashid, the vice president of the UAE said, ”We want the UAE to become the world’s most prepared country for Artificial Intelligence,” thus, the country’s AI national strategy is to establish a world leader in AI by 2031 [40]. Globally, AI and machine learning are already having a transformative impact on many industries, and their potential benefits will likely grow as technology continues to evolve. The government has made significant investments in AI and machine learning technology. It actively promotes their use in various sectors, such as healthcare, transportation, and finance. These efforts are expected to drive economic growth and innovation while improving UAE residents’ lives.

This research contributes to finding the most effective machine-learning techniques that can be utilized for predictive maintenance in different industrial fields, including Abu Dhabi National Oil Company (ADNOC). Furthermore, this research would benefit from achieving the UAE’s Vision for AI and machine-learning technology by offering comprehensive research outcomes with the slightest errors in various sectors, i.e., education, transportation, and healthcare. This research presents recommendations that could be considered as a reference for future researchers to find diverse outcomes with lower errors that could be used to conduct profound research that was not considered in the thesis’ scope.

The goal of this research is tackled by utilizing machine learning techniques for predicting outliers, such as Local Outlier Factors (LOF), Isolation Forests (IF), Exploratory Data Analysis (EDA), Gaussian Mixture Model (GMM), K-Means and One-Class Support Vector Machines (OCSVM) are used for global and contextual outliers, and Long Short Term Memory (LSTM), Total Convolutional Networks (TCN), and ARIMA for collective outliers. The objective was to find the best F1 Score. The second LSTM model (which predicts instead of reconstructs) and the filtered Isolation Forest model for global outliers achieved the highest scores. Classical machine learning techniques such as GMM also predicted the anomalies effectively. However, OCSVM didn’t get the best results because it is assumed that a single cluster. LSTM was also applied to calculate the mean squared error (MSE) for univariate and multivariate cases. Finally, explainable AI (XAI) was employed to identify features relevant to the targeted feature, resulting in improved performance of the LSTM model through the reduction of MSE.

This research offers industry sectors already using AI in predictive maintenance a way to increase the accuracy of their results and reduce errors by introducing Explainable AI to be used in predicting the anomalies correctly by only using features highly correlated to the targeted feature. On the other hand, it encourages industry sectors not utilizing AI to use it by showing the benefits of such technology and taking a real-life national example (ADNOC) that is concluded in the experiment and results.

First Page


Last Page


Publication Date



Thesis submitted to the Deanship of Graduate and Postdoctoral Studies

In partial fulfillment of the requirements for the M.Sc degree in Machine Learning

Advisors: Dr. Martin Takac, Dr. Le Song

Online access available for MBZUAI patrons