Auxiliary-model-based domain adaptation for reciprocating compressor diagnosis under variable conditions

Machine learning is widely used for fault diagnosis research. In general, most models used for fault diagnosis are based on the same data distribution, whereas applying equipment to practical productions and operations are mostly done under variable conditions. This often produces changes in data distribution and makes the model unavailable. As one of the most commonly used pieces of equipment in industry, a reciprocating compressor operates under various operating conditions (e.g., variable speed), which may produce changes in data distribution. Thus, the current model established under stable conditions is no longer applicable for fault diagnosis under variable conditions. To solve this problem of variable conditions, a model should be established that 1) reduces the differences caused by different operating conditions as much as possible, and 2) learns representative fault features under different working conditions. Thus, a new strategy that employs an auxiliary model is proposed that combines a convolutional neural network (CNN) and a marginalized stacked denoising autoencoder (mSDA). In our method, 1) the pre-training model CNN is used for feature learning, and 2) the learned features are transformed by mSDA to eliminate data distribution differences between different conditions. A statistical measure based on kernel maximum mean discrepancy is used to evaluate the differences across different domains. Experimental results of a reciprocating compressor under different operating conditions demonstrate that the proposed method can learn class sensitive features and eliminate differences with changing working conditions. It also obtains higher classification accuracy for reciprocating compressor diagnosis under different working conditions.