Abstract: Fault detection has become an indispensable strategy in the current development of technical features as newer machinery grows increasingly complex, aiming to prevent costly downtime and safety risks. Traditional periodic inspection-based maintenance methods often lead to unexpected breakdowns and inefficient resource utilization. Former studies on predictive maintenance, which were based on sensors such as vibration or current monitors, have been adequate in this regard but lack early anomaly detection and are prone to latency issues in fast-paced industrial settings. The limitations mentioned above are addressed by proposing a real-time fault detection system that combines thermal imaging and machine learning techniques. Thermal cameras stream video from the real-time mapping of temperature distributions, and this process is further enhanced using WebRTC. Temperature variations are analyzed using a Convolutional Neural Network (CNN) and machine detection is performed using the You Only Look Once (YOLOv5) algorithm. Anomaly detection can be accomplished with the Isolation Forest algorithm. The outcome demonstrated an accuracy of 97.5% with a latency of 30 milliseconds. This scalable solution is likely to have industrial applications in the near future.

Key words: thermal imaging, machine learning, IoT integration, real-time monitoring, anomaly detection