Application of Artificial Intelligence for Data-Driven Prognostic of Finding Remaining Useful Life in Filters

Industry 4.0 has become an established reality thanks to recent technological developments. In this context, Reply Concept, a company where I was employed for a short time to write this thesis, is carrying out a project to build a digital twin for a chemical plant in Sicily, Italy. The main purpose is to establish a link between the production environment and digital information by creating technological models that analyze the system and connect the real and virtual worlds. The aim of the thesis work is to develop an Artificial Intelligence pipeline for the management and predictive maintenance of the chemical filters, analyzing the data collected by sensors (e.g. mixing pumps, flow meters and pressure sensors) are analyzed. The importance of determining the Remaining Useful Life (RUL) of filters lies primarily in the capability to avert the failure of equipment that the filters protect. If a filter becomes clogged and is not replaced in a timely manner, it can lead to a decline in equipment performance and even complete failure. By predicting the RUL of filters, maintenance teams can plan ahead for filter replacement, avoiding unexpected downtime and reducing maintenance costs. The research will analyze the results of various machine learning techniques for predicting the RUL of filters using data-driven prediction. Deep neural networks with specific feature extraction layers, namely Long Short-Term Memory (LSTM) and Implicit Neural Networks (SIREN), are used with fully connected layers for prediction. Moreover, discrete features are represented by continuous values to increase robustness. During training, the data is injected using the sliding window method. As a case study, I used a publicly available dataset created by Ömer Faruk Eker [1] in a realistic experimental rig. The test rig consisted of a pump, a dampener, particles, flow rate and pressure sensors, and a filter. Since its creation, thorough research and public challenges have been conducted to investigate accelerated clogging phenomena and the prediction of RUL. The dataset can be divided into 3 different operational profiles (i.e. the different sizes of the particles) that make it effective for building applicable predictive models. State-of-the-art performance metrics have been used by comparing the RUL diverse models to evaluate the methodology. Results show that using the mentioned neural network architectures achieves high-performance results without having much domain expertise. Quantitatively, test results showed less than 20% mean absolute percentage error as well as more than 90% coefficient of determination which is considered good forecasting [2]. Comparing different setups also yielded similar results, which shows the model is robust.