An Intelligent Micromachine Perception System for Elevator Fault Diagnosis.

Area of Science:

• Engineering
• Artificial Intelligence
• Sensors

Background:

• Elevator fault diagnosis requires precise sensing of physical states.
• Micro-Electro-Mechanical System (MEMS) sensors offer high precision but face challenges with data streams, noise, and interpretation.
• A gap exists between raw sensor data and actionable maintenance insights.

Purpose of the Study:

• To propose a collaborative edge-cloud intelligent diagnosis framework for elevator systems.
• To address the limitations of MEMS sensor data in fault diagnosis.
• To create a closed loop from micro-signal acquisition to decision support.

Main Methods:

• Developed a lightweight temporal Transformer model (ELiTe-Transformer) for edge deployment on Jetson platforms.
• Integrated industrial positional encoding, linear attention, and INT8 quantization for efficient edge processing.
• Utilized retrieval-augmented generation (RAG) on the cloud to combine edge-extracted features with domain knowledge.

Main Results:

• Achieved an overall system accuracy of 96.0%.
• The edge-cloud framework improved complex fault diagnosis accuracy to 92.5%.
• Retrieval-augmented generation (RAG) reduced report hallucination rates by 71.4% and achieved a real-time inference latency of 21.4 ms.

Conclusions:

• The proposed framework effectively overcomes MEMS sensor perception bottlenecks in elevator fault diagnosis.
• The edge-cloud collaboration enhances diagnostic accuracy and interpretability.
• This approach provides robust decision support for elevator maintenance.