This paper proposes a cloud-edge collaborative framework supported by a lightweight CNN for aero-engine health monitoring and real-time diagnosis. In this framework, multi-sensor streams are organized into synchronous time Windows, and compressed convolutional models are deployed at edge nodes to achieve low-latency state inference, while parameter aggregation, threshold calibration and model update are performed in the cloud. In order to enhance the stability of diagnosis, the inference backbone jointly introduces depth-wise separable convolution, channel recalibation and residual feature reuse mechanisms. The experiment is carried out on 16800 labeled samples, covering four types of states: normal, degradation, surge precursor and fault. The results show that the proposed method achieves 94.7% accuracy and 0.931 F1-score, the inference delay of the single window on the edge side is 21 ms, and the model size is 3.6 MB. Under the condition of cloud-edge collaborative operation, the alarm consistency reaches 92.4%, while maintaining good online diagnosis response ability. Experimental results verify the computational efficiency, deployment suitability, and diagnostic reliability of the proposed framework in aero-engine monitoring tasks.
