Publication Date: 2024/01/03
Abstract: The cement manufacturing comprises a series of steps aimed at producing top-tier cement that adheres to industry benchmarks while minimising residual content. Traditional practices involve periodic quality assessments, often hourly, facilitated by sensor-derived data. Cement quality assessment hinges on two critical parameters: residue and reject, which gauge cement fineness. Residue reflects the non-uniformity in the final cement output, influenced by various input factors. Thus, meticulous tracking of pertinent inputs is essential. Yet, this method's drawback is that substandard cement mandates the rejection of entire batches. Even with sensor automation, this approach remains time- intensive, less accurate and detrimental to productivity, culminating in substantial losses encompassing raw materials, time, labour, revenue, and in-demand market fulfilment. To surmount these challenges, the integration of automation in quality assessment with machine learning processes, buoyed by adept algorithms, has emerged as an efficient solution. The pivotal target lies in condensing the quality check time frame from an hour to a mere minute, thereby necessitating computational intelligence. Machine learning models offer a path to automate quality checks, dramatically curtailing the time investment compared to conventional methods. Leveraging historical data from companies, these models are trained to streamline the process. In this context, the deployment of a regression model proves invaluable for predicting and anticipating cement residue, a dependable gauge of its quality. Training the regression model with extensive datasets confers it with the power to discern residue and reject levels accurately, classifying cement quality through the analysis of diverse factors including raw material composition, production parameters, and environmental conditions. The model can unveil hidden patterns and correlations that influence residue levels. This empowers manufacturers to rapidly evaluate the quality of cement batches and expedite corrective actions when necessary. The project employs diverse datasets to train various regression models including multi-linear regression, K-nearest neighbour regression, decision trees, random forest, adaboost, xgboost, and neural network models like multi-layer perceptron. The next step involves evaluating the efficiency and accuracy of these trained models, with a focus on selecting relevant metrics. Given the objective of forecasting residue and reject levels, the mean absolute percentage error (MAPE) is adopted. A lower MAPE value indicates more precise predictions, with a targeted MAPE value set below 10%. To address high MAPE values for the "reject" variable, ensemble stacking of models is employed, involving meticulous hyperparameter tuning for each algorithm. This stacking amalgamates predictions from multiple models, yielding enhanced accuracy. Integrating K-nearest neighbour regression, random forest, and xgboost algorithms in a stacked ensemble capitalises on individual model strengths while offsetting their limitations, ultimately refining the residue and reject level predictions. By implementing the aforementioned strategies, the production process is optimised, quality control practices are heightened, and waste is minimised. The result is the delivery of superior- quality cement that seamlessly aligns with market demand.
Keywords: Cement Manufacturing, Cement Quality Management, Regression Model, Machine Learning, Time Optimization, Quality-Checking.
DOI: https://doi.org/10.5281/zenodo.10453966
PDF: https://ijirst.demo4.arinfotech.co/assets/upload/files/IJISRT23DEC1632.pdf
REFERENCES