Machine learning approach to inline monitoring of apple puree consistency through process data and fruit characteristics

Consistency is a key attribute for apple purée producers, since it affects the product quality and acceptability, and can be used to regulate the process settings. This study aims to build a statistical model to predict puree consistency using data collected from an industrial production line. The dataset includes 14 variables across 524 samples, comprising process parameters (e.g., pump performance, temperature) and characteristics of the apples. Three predictive models were developed incrementally. Model 1 relied on the mechanical energy conservation law to establish a relationship between a measured pressure drop and the consistency. In this study, we quantify consistency using the Bostwick flow distance, a widely adopted practical proxy for apparent viscosity in the food industry. Model 2 incorporated the machinery-related parameters, and Model 3 further integrated the characteristics of the apples to account for raw material variability. For models training and validation, generalized linear models (GLM), gradient boosting machines (GBM), and deep learning (DL) architectures were compared. The effect of the different input variables on the predictive performance was also assessed. In all models, pressure difference emerged as the most influential variable. Model 3, particularly with GBM, resulted in the highest predictive accuracy (R 2 = 0.78; MAPE = 9.2 %), demonstrating the importance of incorporating ripening information of the apples in the model. The model outperforms traditional laboratory-based viscosity measurements, which typically have greater variability. The model considers the interactions between processing conditions and raw material properties that influence puree rheology, enabling a real-time inline consistency monitoring of apple puree in industrial settings, offering manufacturers operational benefits. These include the potential for reducing manual testing time and decreasing production costs through minimized batch rejections, and the ability to implement timely process adjustments. These advantages collectively contribute to reduced product waste, enhanced quality control, and improved economic outcomes in commercial apple puree production. • Inline apple purée consistency prediction models developed. • Comparison of GLM, GBM, Deep Learning. • Pressure drop is the most influential variable. • Apple data improves prediction accuracy significantly. • Basis for real-time consistency monitoring established.