Ice Detection For Wind Turbine Blades Via Multivariate Statistical Analysis Method

With the continuous expansion of industrial production scale,the industrial process becomes more and more complex.It is very difficult to monitor the safety of production processes and establish an effective fault alarm model based on traditional rules or mechanism-dependent modeling.On the other hand,mass industrial production data are preserved with the decrease of sensor and storage costs.At the same time,the improvement of artificial intelligence technology has been exposed in many general fields.The industry has received increasing attention as a depression of degree of intelligence and its potential huge commercial value.The fault diagnosis of wind turbine blade icing is taken as an application of industrial intelligence in this paper.Detection of ice formation on wind turbine blades is still in the research and development stage.Generally,wind turbines have to be shut down to deice the blades when the icing is serious.It is of great practical significance to extend the wind turbine life,improve the efficiency of power generation and ensure the safety of field workers if icing process can be accurately predicted so as to be able to open the deicing system as soon as possible.In this article,a large amount of data generated by SCADA system of wind power plant combined with machine learning algorithm to predict the icing condition of wind turbine blades.Firstly,the characteristics of industrial big data are expounded and they indicate that industrial intelligence is not a simple reuse of artificial intelligence technology in industrial scenarios.In the data preprocessing stage,the culling operation is performed for invalid and outlier data.To deal with the severe imbalance of positive and negative samples,this paper first removes some large sample through scene analysis and then under-sampling and oversampling,SMOTE and other methods are used to balance the data set.In the feature construction stage,in addition to the use of dominant features that have a significant correlation with the icing phenomenon,a series of invisible features are constructed to improve the accuracy of prediction combined with the mechanism analysis.In terms of algorithm selection,this paper first applies traditional machine learning methods,such as support vector machine,K-proximity algorithm,BP fully connected neural network.Secondly,the XGBoost algorithm is adopted which is based on the boost theory,but many improvements are made contrast with the traditional gradient boost algorithm,including loss function,regularization,etc.It can solve the problem of industrial scale very well.Finally,the paper uses deep learning theory to build a CNN-LSTM deep learning network to train the big data and also obtain good results.By using different data-driven methods to analyze wind turbine blade data and build a detection model,the ideas and methods how to deal with industrial big data are grasped preliminarily which verify that artificial intelligence technology has broad application prospects in the industrial field.