Research On Health Monitoring Method Of Gas Turbine Blades Based On Time Series

Driven by strategies such as Industry 4.0 and intelligent manufacturing,industrial digitization is thriving,and the acquisition of operational data for industrial equipment is becoming increasingly easy.The operational data of industrial equipment contains the status information of equipment.Analyzing and mining it can obtain the health status of industrial equipment during operation,monitor the health status of equipment,improve the efficiency of industrial production,and ensure the reliability and safety of industrial equipment operation.Early health monitoring methods were often based on traditional data feature extraction,with poor universality and unsatisfactory monitoring results.At present,algorithms based on time series and deep learning have made significant progress in various health monitoring scenarios,with stronger robustness and universality.However,in practical industrial environments,the problem of data samples generated by industrial equipment not having labels is very common.Without sample labels,general deep learning models are difficult to directly apply.This article takes the health monitoring of gas turbine blades as the application background,conducts research on temperature data of gas turbine blades,extracts and analyzes their features from the perspective of time series,and combines machine learning and deep learning theories to propose an unsupervised turbine blade health monitoring method.The main research content of the paper is as follows:(1)In response to the difficulty in characterizing the temperature signal of turbine blades in the background of this article,three feature extraction methods for turbine blade temperature data were studied,namely time-domain based feature extraction method,VMD based feature extraction method,and GWO-VMD based feature extraction method.These three methods were used for feature extraction of turbine blade temperature data,and the feature extraction effects of each method were analyzed and compared,the superiority of the GWO-VMD feature extraction method has been verified.(2)In response to the problem of difficulty in training general monitoring models due to the absence of labels in temperature data samples of gas turbine blades,this paper improves the Transformer model and designs an unsupervised turbine blade health monitoring model based on association differences and reconstruction errors,named Association Transformer.Convert the attention matrix of the Transformer encoder part into global correlations,convert the Gaussian kernel function that can learn scale parameters into local correlations,and define the symmetric KL distance between global correlations and local correlations as correlation differences.A health deviation degree is defined by combining association differences and reconstruction errors to measure the health status of turbine blades,and a training strategy based on maximum and minimum association differences is proposed.(3)In response to the problem of difficulty in ensuring high accuracy of the evaluation criteria for turbine blade health monitoring models,this paper constructs the evaluation criteria for turbine blade health monitoring models based on fuzzy C-means clustering algorithm and expert scoring method.The Association Transformer model is combined with the constructed evaluation criteria to establish a complete turbine blade health monitoring model.Firstly,PCA was used to perform feature dimensionality reduction on the entropy features of each IMF component extracted by the GWO-VMD method.FCM clustering algorithm was used to cluster the reduced features,and expert scoring method was combined with clustering evaluation indicators.Experimental comparisons were conducted between FCM clustering algorithm and Kmeans,K-medoid,and Kmeans++clustering algorithms to demonstrate the applicability of FCM clustering algorithm to turbine blade temperature characteristics,and it ensures that the evaluation criteria for turbine blade health monitoring models have high accuracy.Subsequently,the optimal number of clusters for FCM was determined to be 5 through the internal evaluation index of clustering.Therefore,the health status level of turbine blades was divided into five categories.Based on expert experience,the samples were divided into five health status levels: excellent,good,average,poor,and extremely poor.Finally,the correspondence between the health status level and the health deviation range output by the Association Transformer was completed through experimental observation,and evaluation criteria for the health monitoring model were established to form a complete health monitoring model.Compared with other existing models,the experimental results showed that the turbine blade health monitoring model based on the Association Transformer has better monitoring effectiveness.