Research On Fault Recognition Method Of Wind Turbines In Semi-supervised Environment

With the decrease of global fossil energy and the increasingly serious environmental problems,more and more countries and regions are using wind power technology to gradually reduce the use of fossil fuel.Due to the continuous increase of the capacity and cost of the wind turbine,once the wind turbine has a serious failure,it will increase the huge cost of operation and maintenance,and even more serious may make the wind turbine scrapped.There is SCADA system in the unit which can alarm certain faults of the unit,but it can only be used as auxiliary diagnosis for faults of mechanical parts.In order to reduce the cost of operation and maintenance and the downtime of the unit,it is necessary to adopt appropriate methods for fault diagnosis of the unit.Wind turbines of similar failure completely repetition probability is low,even if the same site failure,the fault category and degree is different,therefore,its monitoring and diagnosis are half supervision environment,namely cannot rely on sample all the existing fault diagnosis model is established and a diagnostic test units,aiming at this problem,in this paper,the main work is as follows(1)Preprocess the SCADA data:Preprocess the unit data through the operating mechanism of wind speed,active power,pitch angle and generator speed.At this stage,the data of the unit’s wind abandonment and power rationing caused by human intervention will be screened out;After that,denoising processing is performed based on statistical law data.In view of the irregularity of SCADA in diagnosis,the SCADA data with a sampling frequency of 10 minutes will be flattened for each day,and the flattened data of each day will be used as a data point,so that it is hoped that it contains enough information,and then Classify it;Since the preprocessing will delete some data,the dimensions of the flattened data will be different every day,so this paper uses CWGAN-GP to generate the missing data.(2)Use Simaese Networks to discover new classes:build a Simaese network based on two and three tuples according to the characteristics of the two-tuple loss function and triple loss function in the Simaese Network,so as to build multiple feature extractors.Save the model by combining the output results of the two-tuple network and the three-tuple network.The two-tuple network machine is used to select the cosine similarity,and the cosine similarity is used to determine the threshold of each category,so as to realize the diagnosis of whether a new category of data appears.(3)Use Prototype Learning to discover new classes:Research the characteristics of the Convolution Prototype Learning framework,use the structure of the combination of convolution and fully connected layers to operate on the data,and add self-encoding and hierarchical data extractors on the basis of the framework to extract the features of the data.The dimension of the data output by this party is 2 dimensions,so the data can be directly visualized,and the output data can be transformed into the t-space using the TSNE algorithm for re-visualization operation,so as to realize the distance between the classes and the aggregation of the data.