Wind Turbines Fault Detection System Using Convolutional Neural Networks
As a renewable source of energy and an alternative to burning fossil fuels, wind power industry is rapidly growing with more than 539 GW cumulative installed capacity at the end of 2017. However, because of the harsh climate conditions, wind turbines (WTs) still faces many failures which increase the cost of energy and reduce wind power reliability. Hence, using reliable monitoring and diagnosis systems of wind turbines are of a great importance.
Operation and maintenance costs can account for 30% of the total cost of large wind farms. Installing offshore and remote wind farms have increased the necessity of an efficient monitoring and fault detection system (CMFD). These systems acquire data from various components of WT including bearings, generator, gearbox, etc. and detect the possible faults using an appropriate data analysis method. Current methods for interpretation of WT signals require expertise and high engineering knowledge. In recent years, the attempts has been made to use computers ability to correctly identify the defects without human intervention. This has led to new maintenance challenges to develop more intelligent CMFD algorithms which will be able to perform classification and prediction operations.
Furthermore, to achieve a reliable automatic diagnosis method and increase operator confidence in alarm signals, consideration of multiple signals is essential. Data processing requires robust algorithms that can extract maximum useful information from available data. Machine learning methods have a great ability to deal with large amount of data. Artificial neural networks (ANN) are the most frequent machine learning algorithm used in WT CMFD systems. An ANN learns the patterns of different WT faults from the input data which might be raw signals, pre-processed signals or the extracted features. There are some models that evaluate whole WT subsystems and detect the defects, but most of the algorithms are developed for fault detection of specific WT components mainly include bearings and gearbox, rotor and blades, and generator. Multilayer perceptron, radial basis function NN, backpropagation NN, self-organized-map NN, recurrent NN, convolutional NN, and neuro-fuzzy networks are some of the popular types of ANN can be distinguished
CNNs are a combination of biology, mathematics and computer science that are particularly powerful for analysis of images. In the last few years, there has been a growing interest in using CNNs to detect bearing and gear faults, but it is not specifically focused on WTs and their subsystems.
One of the first examples of CNN-based feature learning model for CMFD used a shallow structure with one convolutional layer and one fully connected layer to diagnose bearing health conditions. The Discrete Fourier Transform (DFT) of normalized vibration signal collected by two accelerators placed perpendicular to one another is the network input, and the outputs are four classification categories.
Sun et al. utilized a CNN approach to automatically recognize a gear fault feature from the multi-scale signal features which are acquired by a dual-tree complex wavelet transform (DTCWT) of vibration signals. Jing et al. developed a 1 dimensional (1-D) CNN to learn features from frequency data of vibration signals directly and detect faults of gearboxes. They also compared the performance of feature learning from raw data, frequency spectrum and combined time-frequency data. In another research a method named CNN with training interference (TICNN) is proposed for bearing fault diagnosis under noisy environment and different working load. The article used a 6-layer 1-D CNN with dropout in the first-layer kernel and very small batch training whose input is raw expriment data obtained from the accelerometers of the motor driving mechanical system from the Case Western Reserve University (CWRU) Bearing Data center. Furthermore, there are four fault types of the bearing: normal, ball fault, inner race fault and out race fault.
Recent engineering applications have begun to use digital images for fault detection. For instance, Shahriar et al. employed digital grayscale images for fault diagnosis of induction motors using local binary pattern-based texture analysis. Also, Hoang and Kang and Wei et al. eveloped a bearing fault detection method based on 2-D CNN with vibration signals texture images as input. In the study a methodology for WT actuators and sensors fault detection and classification, based on texture image processing, is proposed. Four types of texture characteristics including statistical, wavelet, granulometric, and Gabor features are extracted.
Finally, faults are detected and distinguished between them using five classification algorithms except CNN approach and the results are compared. To the authors´ best knowledge, no publication can be found in the literature that utilize a multichannel CNN (MC-CNN) for the issue of CMFD and so far, only recorded vibration data has been used as CNN input to detect the defects. However, in this paper, multiple signals such as electrical generated power, pitch angles, and tower acceleration are considered to enhance the reliability of automatic diagnosis method. The aim of this article is to introduce an autonomous feature extractor and classifier algorithm for WT fault detection that does not require human expertise or prior knowledge of the problem. A sophisticated benchmark model for a modern 5MW WT presented by Odgaard and Johnson, which is implemented in the FAST (Fatigue, Aerodnamics, Structures, and Turbulence) software, is used to simulate different fault scenarios. Then, the time domain signals are stored and converted to different digital image formats. Finally, an MC-CNN with independent 2-D local convolutions and concatenated fully connected layers is applied on image format of the signals and a process of training and validation of the classification technique is developed. Thus, the methodology for fault diagnosis is completed and calibrated.
