<p>Conventional predictive maintenance (PdM) methods rely on static datasets, making them less effective in dynamic industrial settings where machine conditions constantly change. This research introduces an adaptive PdM framework for motorized systems that processes continuous data streams while optimizing resource use. The framework uses IoT-connected sensors on a conveyor motor to predict and manage faults by adjusting motor speed in real time. An online Long Short-Term Memory (LSTM) model forecasts sensor values and updates its weights with each new data batch read. To detect faults in both forecasted and real-time data, multiple machine learning models were evaluated on a custom dataset. While Random Forest achieved the highest accuracy (99.45%), k-Nearest Neighbors (k-NN) was chosen for deployment due to its superior real-time performance (97.99% accuracy). Over a 30-min forecast period, the Mean Absolute Percentage Error (MAPE) remained low under normal conditions: temperature (0.03%–0.4%), vibration (0.39%–3.54%), noise (0.5%–1.5%), and power (2.8%–25%). However, during faults, MAPE spiked for vibration and noise, reflecting sudden changes in their corresponding readings. For stepper motors, temperature and torque power increases with decreasing speed and at the time of faults, temperature and power readings also abruptly increased which caused their MAPE values to spike as well. This showcased the robustness of the k-NN model to accurately identify faults since the system was instructed to reduce the speed at the time of fault which contributed to this sudden rise. The system also demonstrated low latency, validating its ability to operate in real time while maintaining efficiency in changing conditions.</p>