Implementation of an Intelligent Weighing Error Compensator for a Copper Concentrate Belt Scale using Neural Networks

Abstract

Inaccuracies inherent to the dynamic weighing of mineral concentrates during container filling operations can result in substantial economic losses. This study addresses this problem by developing and implementing an intelligent compensator based on a Long Short-Term Memory (LSTM) recurrent neural network. The proposed system processes real-time sensor data—namely, load cell voltage, conveyor belt speed, and inclination angle—as a multivariate time series to predict and correct weighing errors on-the-fly. Following a systematic hyperparameter optimization, the optimal architecture (20 LSTM units, learning rate of 0.001) reduced the Mean Absolute Percentage Error (MAPE) from 8.5% to 3.01% on the validation dataset, representing a 64% improvement in accuracy. Subsequent validation on an independent test set confirmed the model's robustness and generalizability, yielding a coefficient of determination (R2) of 0.97. A feature importance analysis revealed that the load cell signal is the primary contributor, accounting for 55% of the predictive power, thereby aligning the model's behavior with the underlying physical principles of the process. This research offers two primary contributions: (1) a methodological framework for integrating LSTM networks into industrial weighing systems, and (2) a prototype validated under real-world operating conditions, thereby providing a scalable solution for process optimization in the mining sector.