Fault Classification through Machine Learning in Fluid Technology
Machine learning techniques are continuously gaining attention and importance in several technical domains. Especially in the area of condition monitoring they can potentially provide manifold advantages compared to established approaches. Nevertheless, the practical implementation of data-based condition monitoring systems can be time-consuming and costly. The main challenges are the availability of an extensive database on regular and faulty machine operation and the selection of suited data features and classification algorithms.This research project aims at the development of a method for an efficient implementation of data-based condition monitoring systems. Hence, an automated generation of training data through parameter variation in a 1‑D simulation tool is examined. Furthermore, different data features and classification algorithms are assed regarding their suitability for hydraulic applications. Finally, the performance and robustness, as well as the transferability of the obtained solution are investigated on a test bench.
|Fault classification by considering effects at system level||Construction of a hydraulic reference system in test bench and simulation|
|Reduced effort for the generation of training data||Execution of stack simulation for fault-free and fault-prone system states|
|Definition of data characteristics for fluid technology applications||Investigation of knowledge-based features and different machine learning approaches|
|Extension of the possibilities of condition monitoring||Validation and evaluation on the test bench|
Challenges for Condition Monitoring in Fluid Technology
Despite the meanwhile good availability of measurement data and computing power, the development of an effective and comprehensive Condition Monitoring (CM) is rarely economically feasible with currently common approaches. Established approaches usually consist of the manual definition of limit values by experts or of a comparison of the machine behaviour with realistic simulation models, which is ultimately associated with high development and implementation costs. In addition, such solutions are individually adapted to the respective systems and have a low tolerance to system changes or inaccurate measurement data.
An alternative to the preceding approaches is the use of methods from the field of machine learning (ML). The algorithms summarized under this term enable systems to independently recognize patterns and relationships in large amounts of data. Instead of manually programming rules and parameters, as is necessary with conventional rule-based and model-based error detection methods, ML allows fault patterns to be extracted from data sets through algorithmic training.
The challenge in introducing ML-based approaches for automated fault detection is to generate a sufficiently large amount of training data. A purely experimental collection of training data on machines and test benches is in most cases connected with a disproportionate economic and technical effort, especially if a comprehensive mapping of faulty system states is required.
Method for Efficient Fault Classification through Machine Learning
The challenges for the introduction of data-supported CM systems are addressed in the research project "Fehlerklassifizierung durch Machine Learning in der Fluidtechnik". The aim is to develop a methodology for the efficient implementation of data-based condition monitoring by generating training data using 1-D simulation tools and investigating suitable data characteristics and classification algorithms in the context of fluid technology applications.
Reproduction of faults on the demonstrator and in the simulation model
The present research questions are investigated on a representative servo-hydraulic reference system, which is set up as a physical demonstrator and as a simulation model. For the representation of both faultless and faulty operating states of the reference system, different component faults are integrated into the system simulation. For this purpose, the possibilities of currently available simulation tools for component parameterization are used. Likewise, a selection of errors is implemented on the demonstrator and compared with the simulated results.
Automated Generation of Training Data
Using methods of statistical experimental design, large parts of the parameter space of the simulation model are run in a stack simulation. Data for a large number of failure scenarios and failure levels of the reference system are automatically generated and simultaneously labelled.
Fault Classification with Approaches of Machine Learning
Based on the generated database, different approaches of monitored machine learning are examined for their suitability for the detection of fault patterns for the classification of machine condition. In combination with the classification algorithms, the selection of suitable data characteristics, also called features, is a central issue. Not only directly measurable state variables of the system are considered, but also those that are obtained from preprocessing and linking of information.
Validation and Robustness Analysis
Finally, the methodology obtained is examined on the physical demonstrator with regard to its performance and the limits of its applicability. For this purpose, sensor data on fault-free and fault-afflicted operating states are processed by the test bench through the trained model and the state and fault detection is tested. In addition, the model is checked for its robustness with regard to parameter variations in the simulation model. Of interest is the effect of different component scaling in simulation and demonstrator. Also the classification quality under the influence of external disturbance variables or the extension of the demonstrator by further subsystems is evaluated.
The project is funded by the Forschungskuratorium Maschinenbau e.V. - Fluid Technology Research Fund of the VDMA. The ifas would like to thank all project participants.Copyright: VDMA