A Novel Bearing Fault Diagnosis Method based on Stacked Autoencoder and End-edge Collaboration

The deep learning based fault diagnosis methods show excellent performance. However, cost and delay factors make it difficult for their widespread industrial application. Microcontroller units (MCUs) in industrial equipment have the advantages of real-time response and high reliability and usually have some redundant computational resource. However, even lightweight deep learning models cannot be deployed in MCUs due to severely limited computational resources. This paper proposes an end-edge collaborative fault diagnosis framework, by combining real-time decision-making at the end with dynamic adaptive diagnosis at the edge to improve inference performance. The model's minimum input size is deduced through theoretical analysis of the bearing working mechanism, and to make the model suitable for MCUs, we leverage the differential characteristics of the bearing vibration data and proposed a TinyML model based on stacked autoencoders. The pre-autoencoder extracts differential features, while the post-autoencoder performs fault diagnosis based on pooled differential features. Finally, the stacked-autoencoder model and collaborative framework were evaluated using the CWRU bearing dataset, achieving 384x compression in parameter size and 100% accuracy for binary fault classification, requiring only 6.44kB RAM. With the dynamic adaptive collaboration mechanism, the proposed fault diagnosis framework can reduce the edge load by approximately 94%.