State Assessment For Valve Leakage Of Reciprocating Compressor Based On Transfer Learning Under Multiple Working Conditions

Reciprocating compressor plays an important role in the production of petrochemical enterprises.Due to the complex structure,many parts of reciprocating compressor are prone to damage,even with high probability of failure.The faults are not easy to be found by the staff,and hence cause immeasurable economic loss,even personnel casualties.In order to reduce economic loss and casualties,it is imperative to carry out equipment state assessment.The traditional equipment state assessment is mostly based on the data-driven model,assuming that the training data set and the test data set meet the same probability distribution;however,the unstable operating conditions of the reciprocating compressor cause the vibration signal to be volatile,and the data does not meet the same distribution assumption,which reduces the stability and adaptability of the assessment model.In this paper,with the reciprocating compressor valve as the object,the association rules,transfer component analysis,pattern recognition and other technologies are employed to conduct an in-depth study of the multiple working conditions feature change law,multiple working conditions transfer assessment model and model stability,and a method of state assessment for valve leakage of reciprocating compressor under multiple working conditions is proposed,the main works are as follows:(1)A method for mining valve leakage signal feature under multiple working conditions is proposed based on equal probability association rules.Map the feature data to different interval grades and symbolize each interval grade by dividing the interval with equal probability.Then,the Apriori algorithm with additional constraints is used to mine the association rules between features and states.At last,the association rules are used to draw a state-feature association map to intuitively display the distribution law of the features with different working conditions and the features with the fault degree.In the example analysis,the features of the valve leakage under multiple working conditions and the rules of the features changing with the state are excavated.(2)A method of state assessment for valve leakage under multiple working conditions is proposed based on TCA.The data under one working condition is used as training data and the data under another working condition as test data.Then,the feature selection of the two working conditions is carried out by the transfer component analysis method.Finally,the data is projected into the potential space to reduce the data distribution differences,and improves the accuracy of the model.In the example analysis,the valve operating status under different working conditions are recognized.The accuracy of the proposed method reached 84.2%,which is superior to the three comparison methods such as using SVM to recognize directly,and realizes the accurate assessment of the operating status of the valve under different working conditions.(3)A quantitative analysis method for the auxiliary data in the working conditions of the target domain is proposed.The "excellent" state data and a small number of labeled data are added to the training data set under different proportions of target working conditions to observe the state assessment results of the model established based on different training data sets.The evaluation index of transfer contribution rate is established,and the stability and adaptability of the state assessment model are quantitatively analyzed.The research results show that the state assessment model based on trasnsfer component analysis has the best trasnsfer effect,the assessment accuracy rate varies steadily between 85% and 100%,and the trasnsfer contribution rate reaches more than 85%.The state assessment model has excellent stability and adaptability,and can effectively use historical data.