Sensor fusion with Kalman filter and support vector machine for fault detection in automated guided vehicle
Industries are moving towards automation and the usage of machines such as Automated Guided Vehicle (AGV) is increasing. Thus, demands for the reliability of AGVs are increasing as they have various complex tasks to carry out. Unfortunately, AGVs are still susceptible to faults and breakdown. Theref...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2023
|
| Subjects: | |
| Online Access: | http://eprints.utm.my/102527/1/MarvinDaresMSChE2023.pdf |
| Tags: |
Add Tag
agv
1
|
| Summary: | Industries are moving towards automation and the usage of machines such as Automated Guided Vehicle (AGV) is increasing. Thus, demands for the reliability of AGVs are increasing as they have various complex tasks to carry out. Unfortunately, AGVs are still susceptible to faults and breakdown. Therefore, fault detection is important to provide means of self-diagnosis on AGV. However, fault detections are generally threshold based which are unsatisfying in terms of accuracy and are prone to false triggering. Extended Kalman Filter (EKF) has limitations in handling nonlinear models while Unscented Kalman Filter (UKF) seems promising. Support Vector Machine (SVM) was used as a fault detection method. Thus, this research proposes a sensor fusion enhanced with SVM for fault detection on AGV. The first objective of this research is to develop a test AGV. This AGV is a two-wheeled differential driven mobile robot with multiple sensors and able to make various types of movements to emulate an industrial AGV. Next objective is to develop an enhanced sensor fusion method using EKF and UKF for fault detection with SVM on AGV. The last objective is to evaluate the performance of the developed method. Experiments were carried out where the AGV was used as a test bed for sensor fusion and fault detection. The AGV was tested in different experiment setups such as different track layout, different wheel condition, and different castor conditions. Result shows that UKF handles changes and non-linearity better than EKF. The average residual generated during the test for UKF is 0.0083 meter while for EKF is 0.0129 meter. With sensor fusion, deviations in odometry data can be compensated with the usage of a LiDAR sensor as reference. Using UKF parameters to detect fault, the accuracy achieved with SVM is 64.2% compared to 37.9% without SVM. Fault detection accuracy using EKF parameters with SVM is 82.5% while without SVM is 41.0%. As a conclusion, the results show SVM improves fault detection accuracy regardless of using UKF or EKF. |
|---|