Transfer learning-based variational Laplace autoencoder for industrial process monitoring under varying working conditions

Industrial processes usually operate under varying conditions, making it essential to effectively utilize historical data to establish monitoring model that can accurately and robustly adapt to new operating conditions. To address this challenge, a transfer learning-based variational Laplace autoencoder (T-VLAE) is proposed in this study. The Laplace distribution is utilized instead of Gaussian distribution to enhance the robustness of the variational autoencoder. Then, the historical original domain data is categorized into different clusters, whose similarities to the current operating condition is assessed to calculate the weighting factors. After that, the obtained weights are adopted to design the loss function of the proposed model, including both VLAE loss and alignment loss. By assigning varying weights to the original domain data according to their respective degrees of similarity, a more meticulous and effective utilization of historical data can be achieved. Besides, the attention mechanism is also employed in the proposed T-VLAE model to enhance its performance. Based on the trained model, a monitoring framework comprising two statistics is established. It is noted that resampling operation is deactivated after model training, so that the stability of reconstruction results can be ensured. The monitoring performance of the proposed method is validated using a simulation process and a real industrial application. Experimental results verify the cross-domain adaptability and monitoring sensitivity of the proposed method. Specifically, the proposed method can effectively adapt to new operation condition and accurately identify abnormal status.