AI-based Condition Monitoring of Electro-Hydraulic Systems for Non-Road Mobile Machine Applications

Abstract

The electrification of non-road mobile machines has led to implementation of electro-hydraulic systems in numerous applications. These systems combine electrical, mechanical, and hydraulic components to offer efficient and versatile solutions. However, the components are subject to failure over time and may incur significant economic loss due to substantial downtime and can also lead to potential hazard to humans and environment. To overcome this challenge, it is essential to implement intelligent condition monitoring solution that can monitor real-time data. Continuous surveillance can identify potential issues in the components of a system in their incipient stage. Artificial intelligence (AI)-based techniques such as the implementation machine learning or deep learning algorithms are gaining popularity in the domain of condition monitoring and predictive maintenance as they support real-time monitoring of the system. Despite that, these algorithms require large amounts of data which is often difficult to obtain in all the possible fault scenarios of each component within the system. This problem can be solved by utilizing a simulation model as a tool to generate data source for training and testing the algorithms.

In this thesis, an in-depth analysis of the steps involved in the development of AI-based condition monitoring of electro-hydraulic systems is performed by utilizing information from existing sensors in the system. In this scope, the condition monitoring of various components such as a hydraulic check valve, axial piston pump, external gear pump, and hydraulic actuator are studied. The simulation models are validated with experimental data in their healthy state. The data generated is pre-processed by implementing methods such as feature extraction for a sliding window, data normalization, and feature selection. Several algorithms for developing a classifier are trained and tested to identify the best fit for the data at hand. Furthermore, the algorithm is optimized to improve the performance of the classifier and to save on computational resource requirements. in addition, the adaptability of the developed classifier by implementing transfer learning approaches for a broader range of electro-hydraulic systems or those with similar components is investigated.

The overall performance of the classifier provides accuracy scores over 70% in multi-class classification scenarios. In the case of binary classification, the accuracy scores are over 85%. The process independent condition monitoring study also demonstrated balanced classification accuracy scores over 85% up to 98%. Thus, the methodology and framework proposed in this work may be implemented for real­time condition monitoring of electro-hydraulic systems in non-road mobile machine applications.