Multi-condition Quality Prediction of Production Process Based on Hybrid Transfer Learning

The modern production process is gradually developing towards multi-variety and small-batch direction, and the production process presents the characteristic of multi-conditions. Under this characteristic, the data distribution differences of different working conditions are significant and the sample size is limited, resulting in low accuracy of the constructed quality prediction model. To address this problem, this paper proposes a method for multi-condition quality prediction of the production process based on hybrid transfer learning. Firstly, during the feature extraction stage, transfer component analysis (TCA) is used to extract the features which have little difference in distribution between the historical working conditions and the working conditions to be tested, so as to reduce the difference in quality data distribution between different working conditions. Secondly, during the instance transfer stage, the TrAdaBoost.R2 algorithm is used to transfer knowledge of historical working condition quality data to help build an effective quality prediction model for the working conditions to be tested. Finally, the effectiveness of the proposed method was validated through a case study. The results demonstrate that compared to traditional machine learning algorithms and instance transfer algorithm algorithms that do not reduce the difference in working condition data distributions, the prediction accuracy of the method presented in this article has been improved by an average of 49.2 % and 46.4 % respectively, indicating superior predictive performance.