Reasearch On Intelligent Diagnosis Of Reciprocating Compressor Based On Topic Models

The reciprocating compressor plays an important role in the process of industrial production process. Once faults happen, it will affect the efficiency of production, and sometimes even heavy accidents happen.Therefore, it is significant to realize the abnormal detection and fault diagnosis of reciprocating compressor. The traditional monitoring method is to extract one or a few features, constructing simple relationships between operating conditions and features. So it is difficult to mine deep information from monitoring data, leading to low accuracy of anomaly detection and fault diagnosis. In order to solve the above problems, a variety of discrete and continuous topic models are applied to abnormal detection and fault diagnosis of reciprocating compressor. The content of the research is outlined below:(1) Aiming at the problem that the single feature is difficult to describe the signal characteristics of a unit, high dimension features of the reciprocating compressor is extracted. The structure and working mechanism of reciprocating compressor are introduced, based on which,the characteristic of reciprocating compressor is analyzed. The high-dimensional sensitive feature parameters of the signal are extracted to fully contain the signal feature information of the unit.(2) Aiming at the difficulty in modeling high dimensional features, the research studies the anomaly detection method of reciprocating compressor based on discrete topic model. The feature is discrete and coded and the discrete topic model is constructed. The LDA model is used to detect abnormal data of the unit. On the basis of the model, HMM-LDA discrete topic model is proposed to improve the quality of the model and increase the accuracy of abnormal detection. Engineering data is usedto verify and compare the above two methods in engineering application and the results indicate that the two methods can achieve anomaly detection, but the HMM-LDA method works better.(3) Aiming at the disadvantage of discretization and setting the parameters of discrete topic model depending on the manual experience,the abnormal detection method based on HMM-LDA model is proposed .Constructing phase space for the high dimensional featuresand establishing the continuous topic model for the phase space. A series of continuous topic models are considered, and the validity of the model is verified by the engineering case data. The most suitable continuous topic model for reciprocating compressor anomaly detection is chosen by comparing anomaly detection effect, data analysis result shows that the standard Dirichlet-Gaussian mixture model, stick-breaking Dirichlet-Gaussian mixture model and variational inference Dirichlet-Gaussian mixture model can all realize reciprocating compressor anomaly detection, but the effect of variational inference Dirichlet-Gaussian mixture model is the best one among all the continuous topic models.(4) Compare the applying effect of anomaly detection methods based on single feature, discrete topic model and continuous topic model for the reciprocating compressor. The data analysis demonstrates that the continuous topic model has the best quality on abnormal detection of reciprocating compressor.(5) Fault diagnosis method based on continuous topic model is proposed. Variational inference Dirichlet-Gaussian mixture model is applied to construct models for real-time data and fault data.According to the topic model, Bayesian Inference Contributions(BIC) is calculated for real-time data and all kinds of faults as the determination models ,then the distances between real date BIC and fault data BIC model are calculated to determine fault type. The results show that the method can effectively realize the fault diagnosis of reciprocating compressor.