DEVELOPMENT OF A PREDICTIVE MAINTENANCE MODEL FOR A CASSAVA PULVERISER

Abstract

Cassava is an important staple crop that serves as a primary food source in Nigeria. Nigeria produces about 20% (60 million tonnes) of the total world cassava production (302.6 million tonnes) of cassava. To process cassava tubers into edible products, machines and mechanical tools are frequently used. Cassava pulverisers are essential machines for shearing cassava cakes into fine particles. Determining a method to predict the time to next failure occurrence of a cassava pulveriser is desirable for reducing maintenance costs, ensuring equipment availability, and reducing the Mean Time Between Failure (MTBF). This study presents a model that uses a pulveriser’s vibrational, sound, temperature, and feed capacity data to determine its time to next failure occurrence and maintenance actions.  Primary data of the pulveriser’s conditions was obtained over eight months, characterised, and assessed using Pearson’s R-Correlation Coefficient and Principal Component Analysis (PCA). Two multivariate prediction models were developed. While the Multiple Linear Regression (MLR) was the base predictive maintenance model, the Support Vector Regressor (SVR) was the improved model that considered the non-linearity of the independent variables with the dependent variable. The performance of the MLR and SVR models were evaluated using the Root Mean Square Error (RMSE), Multiple R-squared score, Mean Absolute Deviation (MAD), and Adjusted R-squared score metrics. Subsequently, the SVR model was deployed to a web-based application. The results of this study revealed MTBF of approximately 12 days and a correlation of all independent variables with the cassava pulveriser’s time to next failure occurrence. For the MLR model, a Multiple R-squared score of 0.2823, RMSE of 4.023, MAD of 3.0462, and Adjusted R-squared value of 0.14939 were obtained. The MLR model underfit the data set, and predictions were approximately three days away from the actual failure days on average. In comparison, the SVR model produced a Multiple R-squared score of 0.7524, Adjusted R-squared value of 0.7065, MAD of 1.1538, and RMSE of 2.3632 when implemented with its parameters, C value of 24, Gamma value of 0.042, Epsilon of 0.03, and a Radial Basis Function (RBF) Kernel. These results validated the SVR model as a reliable regression model for the pulveriser data set. The Multiple R-squared score indicates that 75% of the independent variables selected sufficiently explained the values of the dependent variable and the MAD outlines that the predictions are one day away from the actual failure time on average. In conclusion, the developed SVR model successfully modelled the deterioration trends of the cassava pulveriser and provided reliable predictions of its time to next failure occurrence to aid maintenance planning. This study recommends continuous data collection and retraining of the SVR model to gain ongoing insights into the dynamics of the cassava pulveriser.