As an important part of wind turbine, gearbox is mainly composed of gear, rolling bearing and other parts. The common faults include wear, pitting, fracture and so on. Its health has a great impact on the normal operation of the unit. At present, the fan fault alarm mostly depends on the upper and lower thresholds set by SCADA system, but the threshold setting is usually broad, which leads to the failure of timely alarm. The main component analysis method is used to select the parameters which have a great influence on the bearing temperature of the gearbox, and the BP neural network prediction model is established. The improved particle swarm optimization algorithm is used to optimize the model, so as to realize the effective early warning of bearing failure. Firstly, the operating conditions of the unit are identified, and the corresponding two-way cycle neural network prediction model is established in each working condition. The random forest is used to analyze the prediction residual of the model. The simulation results show that the method has high accuracy for bearing early warning of the unit.
A layer by layer weighted average error model for bearing fault prediction is proposed. The sliding window is used to count the model residuals, and the corresponding alarm threshold is set. Combined with information entropy and kernel extreme learning machine, the abnormal condition of the unit is early-warning. The self coding network is established, and a method to determine the adaptive alarm threshold by using the extreme value theory is proposed, which can realize the early warning of fault. Compared with the traditional model algorithms (naive Bayes, random forest, decision tree), the results show that the method has high robustness in unit fault diagnosis.
An algorithm combining recurrent neural network and self encoder is proposed to predict the remaining service life of bearings. The vibration signal of the unit bearing is extracted, and the bearing fault is diagnosed by using the logic regression algorithm. The method is simple and time-saving. In reference [10], genetic algorithm is used to optimize the relevance vector machine. The fault diagnosis results show that this method is superior to BP neural network and support vector machine. A W-type structural element feature extraction method is proposed, which can obtain more feature information from the original signal and improve the accuracy of model diagnosis results. Convolution neural network is used for feature extraction of unit operation data. Soft Max classifier is used to diagnose different types of gearbox faults. Compared with traditional algorithms, this method is more efficient. The stator and rotor current signals of the unit are collected, and the support vector machine model is used to effectively detect the gearbox fault.
The above research methods promote the research on bearing fault early warning of fan gearbox, but there are still some research blind spots, such as failure to consider the relationship between the time of operation data; unreasonable alarm threshold setting, poor generalization ability, complex model, large amount of parameters, difficulty in parameter optimization, etc. In this paper, the bearing fault of fan gearbox is taken as the research object. Firstly, the mutual information method is used to extract useful feature information from high-dimensional original data, and then the long-term and short-term memory neural network is used to establish the bearing temperature prediction model. The model can extract the depth feature of the relationship between the running data before and after time. In order to prevent the false alarm of the model, the sliding window is used to analyze the trend of the prediction error, and the corresponding alarm threshold and alarm rules are set to ensure the accuracy and effectiveness of the bearing fault early warning of the fan gearbox.
Based on the operation data of wind farm, a bearing fault early warning method of wind turbine gearbox is proposed
1) The mutual information method can effectively select the features of the preprocessed data, and screen out the feature information with high correlation with the target object (i.e. gearbox bearing temperature).
2) The long-term and short-term memory neural network model can extract the depth features of the relationship between the running data before and after the time, and accurately and effectively predict the gearbox bearing temperature based on the associated variables.
3) Through the analysis and processing of the prediction residual by sliding window, the appropriate alarm threshold and alarm rules can be set to give early warning of the bearing fault of the fan gearbox, so as to avoid further aggravation of the fault.
