Transformer-Based Fault Detection Using Pressure Signals for Hydraulic Pumps

Abstract

In this paper, modified transformer-based fault detection for a hydraulic pump is performed using the pressure signals of the hydraulic pump. The pump is considered a swash plate axial piston pump used in the excavator. Additionally, the outlet pressure data of the pump are extracted based on Amesim. The proposed transformer is a modified transformer, which allows fast fault detection by modifying the transformer and reducing the size of this model. The classes are normal and 6 fault types, and comparison models are long short-term memory (LSTM) and its family models, which are representative time series models. Unlike comparison models, the modified transformer has an average accuracy of 100% and a detection time of 0.00271 s, which is a slight difference of 0.00036 s from the single LSTM that showed the shortest operation time among the models. We also perform fault detection by changing data points and show a stable high accuracy of 99.93% for all data points of 500, 1,000, and 1,500 without any optimization. Various external noises are added because excavators are construction equipment used in rough terrain. Therefore, we conduct detection performance analysis at different additional noise levels with Gaussian noise with zero mean. As a result, we confirm that the modified transformer showed a high detection accuracy of over 98.08% up to standard deviation 4, where data characteristics were well maintained, unlike other time series models. Through the various analyses above, we confirm that fast and accurate fault detection is possible based on the modified transformer.

Keywords

• Hydraulic pump
• pressure sensor
• fault detection
• transformer
• long short-term memory
• attention mechanism