University of Twente Student Theses

Login

Improving the HUMS capabilities of an EC225 Helicopter by training a model for automatic failure detection of an oil cooler fan shaft

Schonenberg, W. (2015) Improving the HUMS capabilities of an EC225 Helicopter by training a model for automatic failure detection of an oil cooler fan shaft.

[img] PDF
2MB
Abstract:This report is aimed at improving the HUMS capabilities of an EC225 helicopter for an unknown failure related to the oil cooler fan shaft of the helicopter’s main gearbox. HUMS is the acronym for Health and Usage Monitoring System, which is a system that monitors the critical components of a helicopter during the flight using different types of sensors, such as accelerometers. The main goal of HUMS is to detect failures during the flight and warn the helicopter crew as quickly as possible to prevent a crash, whereas HUMS is also useful in optimizing a helicopter’s maintenance program. In this report HUMS acceleration data is used that describes an unknown failure related to the oil cooler fan shaft, located in the main gearbox of the helicopter. The goal of this report is to find out how to extract the vibrational information of the oil cooler fan shaft from the available accelerometer signals and how to determine the type of failure. Subsequently a suitable model has to be found that can be effectively trained to detect this failure in the future. The work starts with selecting the right accelerometer for detecting the vibrations of the oil cooler fan shaft. In order to extract the vibrational information of the oil cooler fan shaft from the accelerometer signal the principle of Time Synchronous Averaging is used to remove vibrations from other components. Condition indicators are identified, whose values are expected to be sensitive for a change in the health state of the oil cooler fan shaft. The first type of condition indicator used is OMx, showing the energy at a frequency of x times the shaft rotational frequency. The second type of condition indicator used is MODx, showing the energy of the first sidebands at x times the shaft rotational frequency. Furthermore, the RMS and Kurtosis are used, showing respectively the average vibration energy of the signal and how peaked the data is. By looking at the behavior of the different indicators, the type of failure is estimated to be an unbalanced fan. The available vibration data of the oil cooler fan shaft is divided into a part for training and a part for testing the model. Different types of clustering algorithms are used for generating healthy and unhealthy clusters from the training data and classifying the health state of the testing points. Only the most relevant condition indicators are used for this modeling procedure: OM1, Kurtosis and RMS. The performances of K-means, Hierarchical Clustering, Support Vector Machine, Multivariate Gaussian and Gaussian Mixture Model are compared using confusion matrices and calculating time. The first three methods use both healthy and unhealthy training data, making them less suitable for detecting unknown failures; failures that were not present in the training data but can occur in real life. The Multivariate Gaussian is evaluated for different types of training data: only healthy, only unhealthy and a combination of healthy and unhealthy training data. This method determines the health state of the testing data by looking at the p-value of the model at the location of the data point. The Gaussian Mixture Model uses only healthy training data and also determines the health state of the testing data by comparing the p-value with a threshold value. The p-value of a single n-dimensional Gaussian can be calculated using the n-dimensional chi-squared distribution. For a Gaussian Mixture Model the p-value cannot be calculated with a chi-squared distribution. Instead, the Gaussian Mixture Model probability density function has to be integrated by a Monte-Carlo or Riemann integration procedure. After creating the p-value boundary, the behavior of the Mahalanobis distance over the p-value boundary is examined. The Mahalanobis distance appears to be constant for the dominant mixture component at that point of the boundary. The Mahalanobis distances of the constant parts belonging to different mixture components are not equal, which is in this case caused by covariance differences, but can also be caused by different weight factors: components with smaller covariance and higher weight factor have a higher Mahalanobis distance. The Mahalanobis distance values of the Gaussian Mixture Model appear to be lower than those of a single multivariate Gaussian with the same p-value, which is caused by overlapping mixture components. Based on the classification performance and the computational time required, Hierarchical Clustering appears to be superior for classifying the investigated failure of the oil cooler fan shaft. In order to improve the model’s capability to detect failure modes that were not present in the training data, it is important to use a model that uses only healthy training data and high dimensional data points, based on many condition indicators. It is then smarter to use an eight-dimensional Gaussian that is capable of detecting a wide variety of failures by looking at disturbances of eight condition indicator values. The Gaussian Mixture Model appears to be too computationally intensive for this relatively easy clustering task, but probably has superior performance in classification problems with a less clear separation between healthy and unhealthy data.
Item Type:Internship Report (Master)
Clients:
ITA, Brazil
Faculty:ET: Engineering Technology
Subject:52 mechanical engineering
Programme:Mechanical Engineering MSc (60439)
Keywords:Clustering, HUMS, k-means, hierarchical, SVM, Gaussian, GMM, p-value, Mahalanobis
Link to this item:http://purl.utwente.nl/essays/69253
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page