Parametric Generalized Adaptive Moment Features (PG-AMF) for Bearing Fault Diagnosis and Machine Health Monitoring

Traditional vibration-based fault diagnosis relies on fixed statistical features like RMS, skewness, and kurtosis. These descriptors are limited by predefined exponents that may not be optimal for all fault types. The PG-AMF framework addresses this by parameterizing these exponents as learnable scalars, allowing the model to optimize feature sensitivity directly from vibration data. The system extracts three families of moments—absolute, signed, and AC-coupled—to capture energy distribution, waveform asymmetry, and dynamic fluctuations, respectively. These features are then processed through a cross-channel bilinear fusion module that models interactions between horizontal and vertical sensor inputs.

The proposed PG-AMF model achieved 100% classification accuracy on the XJTU gearbox bearing dataset, outperforming the statistical baseline, which achieved 96%. The baseline struggled specifically with distinguishing between Inner Race and Compound faults, where fixed features showed significant distribution overlap. By contrast, the learned exponents in PG-AMF effectively spanned a range of sensitivities, allowing the model to capture subtle spectral modulations that fixed-order statistics missed. The inclusion of diversity regularization ensured that the learned features remained distinct, preventing redundancy and improving class separability in low-dimensional projections.

This research provides a bridge between interpretable classical signal processing and the high performance of deep learning. By using a compact MLP architecture with learnable parameters, the PG-AMF framework avoids the high data requirements and "black-box" nature of large-scale convolutional neural networks. This makes it a highly efficient and interpretable solution for industrial predictive maintenance, where computational resources may be limited and diagnostic transparency is critical for safety.