Research On Feature Extraction And Fault Detection Technology Of Rotating Machinery

Rotating machinery in power plants includes turbines,generators,all kinds of pumps,all kinds of fans,all kinds of motors and so on.The structure of rotating machinery is becoming more and more complex,and a series of complicated problems appear in operation.It is necessary to strengthen the research of rotating machinery feature extraction and fault detection,to find out the fault accurately in the early stage,avoid the fault to the later development,and eliminate the occurrence of malignant accidents such as shaft system breakage.Firstly,the vibration feature extraction model of power plant rotating machinery is established,and the vibration features are divided into two categories: time domain and frequency domain.Considering the influence of the vibration of the rotating speed change unit,the time domain features are subdivided into two categories: fixed speed features and start-stop features.Brownian exponential method is proposed to extract the vibration creeping features and compare with the traditional least squares method.Compared with the least squares method,the Brownian index method extracts more distinct features.Three turbine-generator units are used as the application objects,and the corresponding features are extracted from the SIS system by obtaining the fixed-speed and variable-speed process data.The features extracted by the feature extraction algorithm are in good agreement with the actual situation.A vibration and process parameter correlation feature extraction model was developed,and the differences between Pearson correlation and partial correlation features were analyzed.The process parameter correlation leads to the difference of the two correlation features and has a great influence on the correlation features.A correlation feature extraction algorithm is proposed to group process parameters by physical significance and eliminate redundant process parameters within the group by using determination coefficients.A turbine generator set is used as the application object and the correlation features are extracted.Compared with Pearson correlation features and partial correlation features,the correlation features extracted after eliminating parameters can better explain the vibration phenomenon of this unit.Secondly,a long short-term memory(LSTM)vibration deviation prediction model is established to realize the vibration parameter early detection by the deviation of the predicted value from the measured value.The influence of the number of input measurement points on the output deviation of the LSTM model was compared for a 135 MW turbine generator unit.A vibration deviation prediction model based on similarity is established to realize vibration parameter detection by state matching.The deviation of the state from the normal state is calculated for a 135 MW turbine generator set.The calculation results of LSTM and similarity model are compared.Under the condition that the number of input measurement points is selected appropriately,the detection trigger time of LSTM model is earlier.Finally,the mechanisms of three types of typical faults of turbine-generator sets,namely,turn-to-turn short circuit,dynamic and static touch-grinding,and vapor flow excitation,are analyzed,and the fault characteristics are summarized.The feature extraction and deviation prediction models are integrated to establish the early detection model.A fault cause inference model is established,and the cause of the fault is obtained with the help of rule base and inference machine after triggering the early detection.The detection trigger time is calculated for three faulty units,and the support degree of each fault is obtained by the inference machine with inter-turn short circuit as an example,and the results are basically consistent with the actual faults.