Dynamic Vision-Enabled Machine Condition Monitoring: A Point Cloud-Based Diagnostic Methodology

Tongmiao Xu, Dehao Cai, Shaojie Yang, Shuai Gu, Xiang Li

Xi'an Jiaotong University · 2024 IEEE Global Reliability & PHM Conference (PHM-Beijing)

Paper Talk Code

Test rig: motor, coupler and bearing, with an event-based camera aimed at the bearing; below, photos of inner race, outer race and ball faults — An event camera watches a rotating bearing on the test rig. We classify four states: healthy, inner race fault, outer race fault and ball fault.

Abstract

This work is one of the first to use a 3D point cloud algorithm on event data for machine fault diagnosis. Event cameras (Dynamic Vision Sensors) report per-pixel brightness changes with microsecond timing, instead of full frames. We represent each slice of the event stream as a geometric point cloud over (x, y, t), then classify it with a deep point cloud network (PointNet++). We test the method on diagnosing faults in a rotating machine with an event camera.

SensorEvent camera (DVS)

RepresentationEvents as (x, y, t) point clouds

ModelPointNet++ classifier

ClassesHealthy · inner race · outer race · ball

Events as point clouds

Randomly sampled regions of interest from the event stream, shown as 3D point clouds — Randomly sampled regions of interest from the event stream. Each one shows how events from the sensor spread across space and time.

Network

Talk

Citation

@inproceedings{xu2024dynamic,
  title     = {Dynamic Vision-Enabled Machine Condition Monitoring:
               A Point Cloud-Based Diagnostic Methodology},
  author    = {Xu, Tongmiao and Cai, Dehao and Yang, Shaojie and Gu, Shuai and Li, Xiang},
  booktitle = {2024 Global Reliability and Prognostics and Health Management Conference (PHM-Beijing)},
  year      = {2024}
}