Investigation of machine learning approaches in process industry

Abstract

Today, industry is in continuous development, with digitalisation occurring at an increasing rate. Industry 4.0 and Internet of Things are two common expression which both concern the development of industry with focus on artificial intelligence (AI). Studies have shown the benefits of applying various AI techniques in industry, e.g. more accurate fault diagnostics and better estimation of remaining useful life of process equipment. The main aim of this master’s thesis project was to investigate and apply machine learning methods as solutions to challenges identified in the process industry, and evaluate the outcome. A literature review was first conducted to gain insight in the process industry and its current issues for which machine learning techniques could be applied. Thereafter, a study visit to Södra Cell’s market pulp mill in Mönsterås was carried out, during which a case study suitable for data-driven modeling was identified. The case study involved one of the cooling towers at the mill, as well as process units in close proximity to the cooling tower, for which input and output variables were defined. The plant operators suspected fouling to negatively affect the performance of the cooling tower, and asked if the system could be modeled to identify the fouling effect. This hypothesis was investigated by modeling the system with two neural network architectures; Multilayer Perceptrons (MLP) and Long Short-Term Memory (LSTM), the latter being a type of recurrent neural network (RNN). The neural networks were modeled based on data extracted from four years with a time resolution of one hour. Comparing the results of the MLP networks with the LSTM networks led to the identification of recurrency, which in this case referred to the fouling effect. Out of all output variables, conductivity was set as the target variable, or the variable of extra interest, as it was assumed to directly correlate with the amount of fouling occurring in the cooling tower system. As fouling occurs, there is an increase of metal ions in the recirculating cooling stream, leading to an increase in the measured conductivity. The results showed that all LSTM networks, except for one, obtained better model accuracy than the MLP networks. The best MLP network yielded a value of Mean Squared Error (MSE) MSEMLP2 = 0.003067 while the best LSTM network, LSTM 7, yielded MSELSTM7 = 0.001335. Furthermore, all LSTM networks, regardless of overall performance, modeled the conductivity output better than the MLP networks. The results give a clear indication that there is some recurrency present in the modeled system, which confirms the plant operators’ hypothesis of fouling ocv curing in the system. The statistical model yielded in this thesis work could then be used as groundwork for future projects. By accurately predicting the performance of the cooling tower system, investment decisions and optimisation of operational conditions could be carried out.