Flow anomaly detection in harsh industrial environments: A data analytics & machine learning approach

Abstract

Monitoring flow conditions has garnered increased attention with the corresponding increase in pipeline usage and deployment, particularly in industrial settings. More specifically, detecting liquid flow anomalies and disturbances has become a pressing challenge to both governmental and industrial stakeholders due to the associated financial losses and safety concerns it causes. This issue is further highlighted in industrial and manufacturing environments. For example, a fluid flow disturbance in an industrial facility can cause a significant explosion that would threaten both the facility and its operators. One promising approach to adopt is the use of data analysis (DA) and machine learning (ML) algorithms due to their effectiveness in illustrating the behavior of flow anomalies and detecting them in environments such as water transportation systems and underground pipelines. This work proposes a DA and ML-based approached to better understand the characteristics and detect flow anomalies in industrial-installed pipelines. As such, this work first proposes the use of Short-Time Fourier Transform to visualize the variation of the transmitted signal power measured for different frequency components over time. Then, representative and relevant features are extracted, normalized, and fed as input to four different supervised ML classification algorithms for flow anomaly detection. To evaluate the performance of the proposed framework, a real-word dataset collected through an industrial partner representing different working conditions is used. Experimental results show that the proposed ML-based flow anomaly detection framework achieves high accuracy, precision, and recall values. This illustrates the framework’s effectiveness in detecting anomalies in such harsh industrial environments.