Noise Adaptive Filtering Neural Network Under Multiscale Features

Various uncertain disturbances in industrial processes bring noise to industrial process data, which brings great challenges to industrial soft sensor modeling. Traditional soft sensor models focused on removing noise in the process data, but it is almost impossible to remove all noise in actual engineering. Therefore, a novel noise adaptive filtering method integrating the multiscale neural network (NAF-MSNN) is proposed for the soft sensor, which can incorporate a noise processing mechanism that adaptively removes noise at different scales during feature extraction. The MSNN extracts overall trend and local trend features through the multiscale convolution. Then, the NAF converts multiscale features into frequency domain features, and constrains the noise filter matrix through proposed piecewise regularization to select important frequency domain components at different scales. Moreover, the multiscale fusion module controls denoised multiscale features exchange fusion between different scale based on the important measurement of each corresponding scale. Finally, the gated recurrent unit (GRU) establishes the dynamic relationships between the fused multiscale features and the key indicator. The proposed NAF-MSNN is compared with state-of-the-art soft sensor models in three datasets. In terms of R2 metrics, the accuracy improvement of NAF-MSNN reaches 4%, 8% and 3% in the public dataset and two industrial datasets. Note to Practitioners—Due to uncertain production environments, there may be a large amount of noise in industrial process data. The focus of this paper is to build a noise adaptive filtering method integrating the multiscale neural network (NAF-MSNN) to improve the robustness to noise. The proposed model can adaptively eliminate noise that has a large impact on the prediction results according to different scale features during the training process, thereby achieving high-precision soft sensing in an uncertain noise environment. In addition, we conducted comparative experiments and noise injection experiments on two real industrial data sets. The proposed method has higher soft sensing accuracy and stronger noise robustness than baseline methods.