A multi-layer denoising method for product quality prediction based on key information retention and residual noise elimination

In industrial big data, noise significantly impacts product quality prediction and analysis. Traditional denoising methods often suffer from key information loss and residual complex noise. To address these challenges, this paper proposes a multi-layer denoising method that retains key information and eliminates residual noise, improving data quality and prediction accuracy. First, time-frequency analysis identifies noise types in the raw data. Next, multiple denoising methods are applied to target specific information (trends and key characteristics), and the results are weighted based on multi-dimensional performance evaluation (MDPE), with T-Pearson used to assess trend extraction. The weighted results are fused to obtain preliminary denoised data. This data is then processed by a GWO-DAE model, enhanced by a loss function constrained by key information retention and complex noise elimination. The final denoised data, obtained through reconstruction, offers higher quality and reliability. Experimental results from a tobacco moisture regain machine show a reconstruction error of 0.21, a processing time of 22.83 s, and improved prediction accuracy, with an average absolute error of 0.29 in product quality predictions.