A Lightweight and Explainable Deep Learning Framework for Cross-Load Bearing Fault Diagnosis with Zero-Shot Cross-Machine Evaluation

Rolling-element bearing failures cause a large share of unplanned downtime in rotating machinery. Convolutional neural networks (CNNs) reach near-perfect accuracy on single-machine benchmarks such as the Case Western Reserve University (CWRU) bearing dataset, but their behavior under load and machine shift is rarely tested, many are too large for edge hardware, and black-box predictions give engineers little diagnostic context. This paper evaluates a lightweight one-dimensional CNN built from depthwise-separable convolutions, with a Domain-Adversarial Neural Network (DANN) branch for load-invariant feature learning, Gradient-weighted Class Activation Mapping (Grad-CAM) for post-hoc explanation, and knowledge distillation with dynamic INT8 quantization for compression. A file-grouped protocol separates two settings: cross-load generalization within CWRU, training on three loads and holding out the fourth, and strict zero-shot cross-machine evaluation from CWRU to the Society for Machinery Failure Prevention Technology (MFPT) bearing dataset. On the unseen CWRU load, the DANN reaches 98.0% accuracy (macro-F1 0.978), against 100.0% for a plain CNN and 76.9% for a hand-crafted-feature SVM, so the adversarial branch gave no accuracy gain in this saturated setting. Zero-shot CWRU-to-MFPT transfer reaches only 4.5% accuracy (macro-F1 0.036), below the 48.2% majority-class baseline, and t-SNE shows a persistent source-target feature gap. A distilled 0.25-width student reduces the parameter count from 29,942 to 3,551 and model size from 0.128 MB to 0.025 MB; dynamic INT8 quantization reduces size to 0.019 MB without changing the parameter count. Within-machine accuracy on the single evaluated load holdout is high, but load-adversarial training on raw signals does not close the larger machine-domain gap.