This research presents results and discussion on a novel system for real-time monitoring of the work of industrial belt conveyor, signal classification and load identification. The measurement system is based on the strain gauges measuring the pressing force of the belt on the roller during its work. Automatized operation of the measurement system was designed to minimize operator’s impact on the measurement results. The aims of the research were to create machine learning models for classification of various conditions of conveyor belt and to identify optimal signal length of tensile pressure which enables achieving the best classification accuracy. For this reason, long short-term memory (LSTM) and Transformer neural network models were developed and tested. Both models achieved >90% accuracy in identification of loaded and unloaded dynamic states using pressure raw signal from strain gauges. However, LSTM model using short signals exhibited better classification recall for unloaded conveyor belt condition, and reached the highest classification level. Thus, it was proved that the novel measurement system was able to provide reliable signals, and processing of these signals with machine learning tools gave promising results proving its feasibility for Predictive Maintenance.