State Transition Classification for Predictive Maintenance Using Thermal Images and Deep Learning

Predictive maintenance is crucial for reducing industrial downtime, improving efficiency, and preventing unexpected failures through early fault detection.This study introduces a hybrid deep learning framework integrating spatial and temporal analysis to classify induction motor states accurately.The system identifies six operational conditions, including Healthy Full Load, Healthy Medium Load, Healthy No Load, Corroded Full Load, Corroded Medium Load, and Corroded No Load, by utilizing high-resolution thermal imaging and sequential temperature-time data to capture both static and evolving fault patterns.Unlike conventional methods focusing only on image classification or time-series modeling, the proposed approach combines VGG16 (Visual Geometry Group) -based thermal image classification, Gradient Boosting Machine-based state transition modeling, and YOLO (You Only Look Once) -based hotspot detection for improved fault diagnosis.Initially comprising 488 thermal images, the dataset was expanded to 1,952 images using augmentation techniques to enhance model generalization.Sequential temperaturetime data sampled every ten seconds improves fault detection by capturing state transitions over time.The system achieves 98.21 percent accuracy on the primary dataset and 98.92 percent on an external benchmark dataset, surpassing existing models such as ResNet50 (Residual Network ), ANFIS (Adaptive Neuro-Fuzzy Inference System), and LSTM (Long Short-Term Memory) -based methods.The Gradient Boosting Machine achieves R scores above 0.99, with a mean squared error as low as 0.0013, effectively modeling motor state transitions.A dynamic weighted averaging mechanism optimally balances VGG16 and Gradient Boosting Machine predictions, further enhancing accuracy.This framework is scalable for real-time applications, with future work focusing on real-time monitoring and adaptive fusion mechanisms to advance predictive maintenance strategies.