1. Introduction
Wind turbines are one of the most remarkable renewable energy generation systems and many studies have been being conducted to reduce their operating cost. A survey on the total cost of wind power generation cost concluded that the operating cost, which includes maintenance (O&M), training operators and engineers, repair, system upgrades, inventory etc., is larger than capital costs, such as facility design, development planning, and construction. Concretely, the maintenance cost accounts for more than 25% of the total cost [
1,
2].
In practice it is not easy to perform regular inspections for the maintenance of wind turbines that are located at inaccessible places, such as mountaintops, shorelines, or oceans, and deserts. In addition, unexpected faults may happen due to an abrupt change in the environmental conditions including extreme weather events as well as performance degradation. These issues have been addressed by using condition monitoring systems (CMSs). This are a part of predictive maintenance (PdM) [
3,
4] and are installed to detect abnormal conditions of the mechanical parts of wind turbines in advance. PdM, which is also known as condition-based maintenance (CBM), attempts to evaluate the condition of an asset by performing periodic or continuous monitoring [
1]. According to a report by the Electrical Power Research Institute (EPRI) [
5], when a power plant is managed through PdM, its maintenance costs have a minimum five-fold benefit in comparison with the maintenance costs through visual inspection and ten-fold in comparison with the total cost after problems occur. For this reason, it is expected that the economics of wind power generation can be improved if a CMS allows early fault detection of mechanical systems.
A CMS consists of sensors, data acquisition equipment, and signal processing algorithms. Sensors are typically used for vibration analysis, temperature analysis using thermocouples and/or thermography, ultrasonic testing, oil analysis, strain measurement, acoustic emission test, electrical effect measurements, and so on [
6,
7,
8,
9,
10]. Among them, vibration is the most effective quantity measured in rotary wind turbines [
10].
Signal processing algorithms are utilized to examine the health conditions of the mechanical components of wind turbines by extracting important information from the various types of sensors which collect information in two representative dimensions: time and frequency. Time-domain signal processing such as statistical methods, trend analysis, and filtering makes use of overall and/or average quantities and is applied to monitor the health condition of wind turbines. Plots of data, such as rotational speed with time, allow for comparison with pre-defined thresholds: warning and alarm levels. The time-domain signal processing method is very useful for determining whether a wind turbine has crossed a threshold that indicates a dangerous condition.
Frequency-domain signal processing based on the fast Fourier transform (FFT) is employed to diagnose where and how a wind turbine behaves abnormally. Considering frequency-based methods such as spectrum, envelope, and Cepstrum analyses using vibration signals, requires previous understanding of how the frequency is affected when a component behaves abnormally. In the case of variable rotational speed, waterfall analysis using wavelet, short time Fourier transform (STFT) and order analysis [
11] can be applied for detecting abnormal symptoms of wind turbines. These kinds of conventional methods have two main problems. One deals with each sensor’s physical data individually, but does not consider interrelationships with other types of sensors data. The other is a difficulty in analyzing non-linear and non-stationary data, which reflect changes of wind speed and rotation speed. For these reasons, more advanced signal processing techniques such as angular resampling [
12] have been being investigated for monitoring the health condition of wind turbines with greater reliability.
There are numerous expert approaches for the machine diagnosis and prognosis when implementing CBM and they can also apply to fault monitoring of the wind turbines. The ways can be divided into three categories: physical models, artificial intelligence (AI) models, and hybrid models. The physical model is based on mathematical or logical representations of the mechanical system and therefore it is also named a white-box method. On the contrary, AI models are considered black-box methods because they can be applied to diagnose or predict without knowing the internal relationship of the mechanical system. Among the AI models, the artificial neural network (ANN) has been mainly applied to classification and recognition problems because it has a good theoretical background and can perform non-linear feature mapping with high accuracy [
13,
14]. The Hidden Markov model (HMM) is widely used for pattern and/or speech recognition because it has an excellent mathematical basis and is an efficient modeling tool for sequences with temporal constraints [
15]. The hybrid model integrates two or more individual white-box and/or black-box methods for eliminating the limitations inherent in each method. However, it is still difficult to indicate a specific method which can generally apply to any problem because its performance depends on database representing the problem [
13].
The purpose of this study is to propose a fault detection algorithm for the mechanical system of wind turbines in order to upgrade the function of CMS. To this end, wind turbines with a capacity of 3 MW are employed and they consist of a rotor blade, main shaft and bearing, gearbox, and permanent magnetic synchronous generator (PMSG). Long-term data over two years including vibration, wind speed, rotating speed, etc. were collected by their CMS. First, the structure of a wind turbine, its operating phases, and installed CMS are briefly explained in
Section 2. Subsequently relationships among the operating conditions of wind turbine, wind speed, shaft rotating speed, and power are investigated. Trend analysis on vibration signals is conducted as a function of the number of the revolutions of the main shaft in order to find out its distribution and alarm thresholds are determined which indicate abnormalities of the mechanical system in
Section 3. Third, correlation analysis of the vibration signals is used to find coupling of physical components. Then, a HMM is employed to propose the statistical fault detection algorithm in the time domain and take its input sequence considering the correlation between vibration signals in
Section 4.1. Lastly, the performance of the proposed HMM is investigated by introducing in
Section 4.2 some metrics that can account for the imbalance of datasets.
5. Conclusions
This study proposed a fault detection algorithm using HMM to recognize whether mechanical parts of a wind turbine are behaving abnormally or not. A vibration signal was selected to determine the status of the wind turbine and acceleration signals were measured at the bearing of the main shaft, gearbox, and generator for more than 2 years. It was found that the distribution of the long-term vibrations could be approximated with the Weibull probability density function when the vibrations were classified by the rotational speed of main shaft. And then, the probability function was used to determine the threshold values indicating alarm levels.
The input sequence for HMM was obtained by applying the threshold levels and the correlation between the vibration signals. Because HMM took into account the variation of the status during the a given time interval, it could overcome the disadvantage that the conventional methods exhibited, alarm errors due to one-off external impact signals due to either system starts or stops or unexpected electrical noise at a specific channel. As a result, it was found that the proposed HMM algorithm for fault detection achieved 96% accuracy, 0% FPrate, and 100% precision by analyzing the long-term vibration signals.
In fact, it is not easy to obtain the vibration data that is directly related to the fault since there are very few actual fault events in the mechanical parts of the wind turbine. To overcome this, it will be necessary to improve the statistical reliability of the proposed HMM by adjusting the threshold level during continuous data acquisitions and re-training the transient probability distribution and the observation symbol probability distribution. In addition, as further works, the number of detectable states can be subdivided as using the advantage of HMM that can freely expand the number of classes and the performance needs to be compared with the results obtained from other classification methods.