Anomaly Detection Method for Harmonic Reducers with Only Healthy Data.

Area of Science:

• Robotics
• Mechanical Engineering
• Machine Learning

Background:

• Harmonic reducers are critical in industrial robots, but acquiring sufficient anomaly data for supervised training is challenging.
• The sensitivity of regular features to anomaly detection in harmonic reducers remains unclear.

Purpose of the Study:

• To propose an effective anomaly detection framework for harmonic reducers utilizing exclusively healthy operational data.
• To investigate a method for enhancing anomaly detection sensitivity without relying on anomalous samples.

Main Methods:

• An auto-encoder (AE) was trained on healthy harmonic reducer features to create a new high-dimensional feature space sensitive to anomalies.
• Feature mapping was employed, where the difference between the AE's output and input highlighted anomalous information.
• The mapped features were then processed using One-Class Support Vector Machine (OCSVM) to preserve anomaly details.

Main Results:

• The proposed method demonstrated superior performance in detecting anomalies compared to using AE or OCSVM alone.
• Validation using multiple datasets from harmonic reducers confirmed the framework's effectiveness.
• The enhanced feature space proved more sensitive to abnormal data than the original feature space.

Conclusions:

• The developed anomaly detection framework effectively identifies anomalies in harmonic reducers using only healthy data.
• The combination of AE-based feature mapping and OCSVM offers a robust solution for anomaly detection in industrial components.
• This approach addresses the challenge of limited anomaly data in practical applications.