Automatically predictive maintenance (PdM) is critical for minimizing unplanned downtime and reducing operational costs in modern computer numerical control machines. However, traditional cloud-based PdM suffers from high latency, privacy concerns, and heavy infrastructure demands; meanwhile, traditional edge intelligence-based approaches are restricted by the power of edge devices. In order to tackle these issues, this paper proposes a transferable TinyML-based Artificial Intelligence of Things (AIoT) for PdM. First, self-powered piezoelectric sensors in the AIoT are installed for monitoring device vibration. Second, FFT-based feature extraction and quantized TinyML models are deployed on the edge device for real-time, low-power inference on microcontrollers. Third, few-shot transfer learning is incorporated. Experiments on four fault classes—Normal, Misalignment, Bearing Fault, and Idle—demonstrate that our method achieves 94.8% accuracy, 95.1% precision, 94.6% recall, and 94.7% F1-score, outperforming six baselines (LSTM, RF, SVM, KNN, LR, and DT). Ablation studies confirm the critical roles of transfer learning, quantization, self-powered sensing, and FFT features. The proposed framework delivers sub-200 ms inference latency at < 1 mW, making it ideal for always-on AIoT PdM in CNC production.