Auxiliary Feature Based Domain Adaptation for Reciprocating Compressor Diagnosis

Abstract

At present, machine learning is widely used for classification, such as automatic speech recognition, image identification, text classification and numbers of researches for fault diagnosis besides. Generally, most of the models used for fault diagnosis are based on the same data distribution, while the applications of the equipment in actual production and operation are mostly under unstable conditions, which may make data distribution different and the model unavailable. For example, various operating conditions (e.g. variable speed) in reciprocating compressor may cause difference of data distribution, so the present model established under a stable condition is no longer applicable to fault diagnosis of the compressor under other conditions. Therefore, a model should be established to reduce the differences caused by different operating conditions as much as possible. And at the same time, the model is supposed to synthesis representative features under different conditions containing defects. Domain adaptation is widely used for cross-domain data mining and setting up a learning model applied to source domain and target domain sampled from discrepant distributions. So it can be used to reduce the cross-domain discrepancy by learning joint feature representation. Nonetheless, when it comes to the random assimilation, the feature representation assimilated for each category makes it impossible to distinguish. Hence we propose a strategy that auxiliary feature, as another type of abstract feature which is adept in representation of respective domain for category, is embedded to enhance the representative features for classification. And we also establish an actual complex model. This model can take advantage of rational weakening and strengthening from domain adaptation and assisted training features to ensure a high classification accuracy in the target domain. Experimental results of a reciprocating compressor under different operating conditions demonstrate the effectiveness of the proposed method.