Parameter variation caused by manufacturing tolerances and thermal drift makes PMSM fault-severity estimation difficult, because motor-level offsets and fault effects are coupled in the d–q model. This paper proposes a channel-selective BO-Fusion-PINN for parameter-generalized fault diagnosis. A healthy reference window is first used to estimate motor-parameter deviations through an integral least-squares observer, avoiding neural extrapolation of these offsets. A diagnostic window is then processed by a physics-informed LSTM branch and a data-driven LSTM branch, and Bayesian optimization assigns separate fusion weights to stator-resistance degradation and permanent-magnet flux weakening. Experiments over parameter out-of-distribution buckets and non-ideal simulation settings show that the fused estimator consistently improves on either branch alone. The method is especially effective in the flux channel and remains competitive with high-capacity data baselines while preserving physical interpretability. The primary scientific contribution is an identifiability-guided fusion rule that assigns physics and data trust to each fault channel according to its statistical observability rather than through a single global weight; in practical terms, this yields a compact and interpretable estimator that transfers across the parameter-tolerance band of a machine class and can, in principle, support controlled end-of-line screening and scheduled diagnostic assessment under a matched-window acquisition protocol.

Identifying the impact location (“sweet spot”) on a tennis racket is crucial for performance evaluation in tennis training. However, existing approaches typically rely on expensive vision-based systems or specialized sensors, limiting their applicability in real-world scenarios. We propose a sound-sensor-based multi-task framework for racket impact localization using acoustic signals, combining radial region classification with continuous position regression. To effectively model complex acoustic patterns, we design a multi-expert convolutional neural network (CNN) architecture with multi-scale feature extraction and task-specific optimization. Each expert branch operates at a different temporal receptive field and is trained with tailored loss functions, enabling complementary learning of global patterns, class imbalance characteristics, and hard samples. The shared backbone jointly supports both classification and regression tasks, allowing the model to learn more informative and structured representations. Experimental results demonstrate that the proposed framework consistently outperforms conventional methods in radial region classification while achieving accurate impact position estimation. Furthermore, additive noise augmentation significantly improves robustness, enabling stable performance under noisy and practical sensing conditions.