Safety Risk Assessment Of Steam Turbine Based On Semi-supervised Learning

In the era of big data,safety risk assessment and early warning of steam turbines under the condition of massive operation data are the premise to ensure the normal operation of the electric generation system,which is of great significance to ensure the stable development of electric enterprises.However,the traditional risk assessment models based on expert knowledge have been far from meeting the needs of accurate assessment,while the data-driven safety assessment methods have attracted more and more attention from scholars and management departments in enterprise.With the rapid development of data mining,machine learning models have been applied to and have certain practical value for fault diagnosis of steam turbine.However,there still exist shortcomings in identification results and interpretability of unsupervised learning,and the availability of samples required for supervised learning is too low.Therefore,this paper proposes a hierarchical early warning mechanism based on semi-supervised learning and applies a variety of machine learning methods to detect steam turbine anomalies.The specific research contents of this paper are as follows:Firstly,the Principal Component Analysis(PCA)is used to identify risk factors in view of the redundancy and high dimension of turbine units’ features,and thus the original features are transformed into the four principal components that can explain the security events,namely unit power,unit load,actual flow and valve threshold.Then the importance of the given risk features is ranked by Fault Tree Analysis(FTA).Finally,an example is given to verify the effectiveness and rationality of the method.Secondly,for the lack of anomalies and labeled samples of steam turbine,the clustering methods are used to self-learn the features and judge the normal and abnormal samples based on the given safety risk factors.Since there is no scientific evaluation standard for the quality of results of clustering methods,this paper will integrate the results of clustering methods of different principles such as K-means,Isolation Forest and Local Outlier Factor.Then the samples will be labeled with real anomalies,suspected anomalies and normal data,and therefore form a hierarchical warning mechanism.In the end,the effectiveness of the safety risk assessment model will be proved by comparing the experimental results with the real anomaly nodes.Finally,in order to effectively predict the security risk events during the operation of the steam turbine unit,classification models such as K-Nearest Neighbor,Support Vector Machine,Decision Tree,Random Forest and Gradient Boosting Decision Tree are applied to train and learn the real anomaly samples,suspected anomaly samples and normal samples.Then the prediction results show that the Gradient Boosting Decision Tree and Random Forest can more accurately detect the real anomalies of the steam turbine.In the end,the results of a new sample set which is applied to validate the trained model demonstrate the superiority of the above two methods again.It can be seen that the semi-supervised learning method designed in this paper can identify anomalies through hierarchical early-warning.The F1-score can reach 98% and obtain the early-warning models with high accuracy.It not only increases the interpretability of the results of anomaly detection,but also provides a new idea for turbine safety risk assessment in electric power enterprises.