Research On General Method Of Rotating Machinery Operation Trend Prediction And Fault Diagnosis

Large rotating machines such as gas turbines,steam turbines and compressors have been widely used in many modern industrial fields such as energy,oil,gas,and petrochemistry.The failure of a large rotating machine occurs may not only result in the huge economic loss,but also pose a threat to the life safety.In the past decade damage diagnosis and prognosis has become an increasingly important tool in improving the safety and reliability of rotating machines while reducing the maintenance cost.In particular,the operation trend prediction of the machine based on historical time series data plays a key role on the early detection of equipment anomaly.As the advance of engineering and manufacturing technologies the rotating machines become increasingly larger and more complicated,which makes the operation and maintenance of rotatory machines more challenging.In the meantime,this brings new opportunity to integrate the advanced digital technologies such as artificial intelligence and big data analysis with the damage diagnosis of rotatory machines.This study attempts to develop advanced deep learning methods to predict the operation trend of the rotating machine,and diagnose its possible fault.This study mainly includes three parts of fault diagnosis for rotating machines:1.In order to handle the uncertainties in the sensor data and prediction models,an improved Bayes-LSTM prediction model is proposed by seamlessly integrating Bayesian variational inference with long short-term memory model.This method models each of the weight and bias parameters as a probabilistic distribution,thus considering the uncertainties of the data and model.The multivariate sensor data collected from a real-world centrifugal compressor and a comparison study with traditional machine learning models are employed to illustrate the feasibility and accuracy of the proposed model.2.The traditional model for fault diagnosis of rotating machinery is established by using the data collected from one unit,which results in the model lack of robustness and generalization.This study proposes a cross-unit fault diagnosis model based on quadratic feature extraction of vibration signals by combining statistical analysis,principal component analysis and convolutional neural network.Multidimensional features are extracted from vibration signals of rotating machines in terms of both frequency and time domains.The proposed method and procedure are established from the features collected from one unit with the accuracy of more than 90%.The validated model is then used to conduct cross-unit fault diagnosis.A comparison study with various conventional machine learning models like support vector classification and multivariate vibration data from two different centrifugal compressors with unbalance issues are employed to illustrate the effectiveness and accuracy of the proposed model.3.The multivariate operation data of rotating machinery contain spatial and temporal patterns as they are collected from different locations in the form of time series.In order to capture the spatio-temporal characteristics of the data,a new convolutional LSTM prediction model is developed in this study.This method converts the raw data into time series image data to represent the spatio-temporal feature and then predict the image series.A novel Conv-LSTM model is constructed to predict the trend by using the images.The data collected from two different centrifugal compressors with unbalance issue are used to demonstrate the effectiveness of the proposed model.The data from one unit is used to establish the model during the healthy period.The performance of the model is evaluated by calculating the structural similarity between the predicted and actual images.The validated model is then employed to predict the status of another unit.