Robust Mechanical Fault Diagnosis Using Time-Frequency Feature Fusion and Deep Convolutional Neural Networks

Abstract

We study fault feature extraction and diagnosis for rotating machinery using a compact 1D CNN. The dataset comprises 322,800 labeled windows (2,048 points at 12–20 kHz) collected from laboratory benches and an industrial pumping station, covering normal operation and faults in bearings (inner/outer race, rolling-element spalling), gears (pitting, cracks), and motors (imbalance). The model stacks 5–8 convolutional layers with small kernels (3–5), Batch Normalization, ReLU, max-pooling, and two fully connected layers with dropout, followed by a softmax classifier. On a held-out test set, the method attains 96.8% overall accuracy and 95.0% macro-F1, outperforming support vector machines and a deep feed-forward network by +8.3 and +5.6 percentage points in accuracy, respectively. Robustness tests under additive noise maintain 92.8% accuracy at 5 dB. In a three-month on-site validation at a pumping station, the system detected incipient bearing cracks early and issued timely warnings with 96.5% online accuracy, reducing unplanned shutdown risk. These results indicate that the proposed CNN delivers accurate, robust, and field-ready diagnosis for predictive maintenance.