A Hybrid Deep Learning Framework for VibrationBased Machine Health Classification

Predictive maintenance systems increasingly rely on machine learning to monitor equipment condition and anticipate failures. This paper introduces a hybrid framework that combines principal component analysis (PCA), oversampling, and a convolutional long short-term memory (CNN-LSTM) network to classify machine health states using vibration data from real machinery data. Unlike prior studies that depend on simulated datasets or manually labeled fault data, this work derives an unsupervised PCA-based health index from real-world CNC machining data collected over several years in production environments. The approach integrates automatic index generation, class balancing, temporal modeling, and comprehensive performance evaluation. Experimental results demonstrate robust classification accuracy across four health states, with additional latency and throughput analysis confirming the feasibility of real-time deployment.