Quality Prediction Of Multi-variety Small-batch Products Based On Two-layer Feature Selection And Stacking Integrated Learning

With the intensification of global market competition,enterprises gradually switch from mass and single production mode to flexible production mode of multi-variety and small batch in order to meet the current market demand for diversified and personalized products.However,the multi-variety and small batch production mode is characterized by many varieties of products,small batches and complex production process,which makes the prior quality control of multi-variety and small batch products have many problems.The implementation of "Made in China 2025" strategy has brought unprecedented opportunities for China’s manufacturing industry,prompting it to shift to digital and information-based intelligent manufacturing mode,of which industrial big data,as one of the core elements,provides data basis for establishing data-driven multi-variety and small-lot ex ante quality control.Based on this,this paper takes the ex ante quality control of multi-variety and small batch products as the starting point,and carries out the research of multi-variety and small batch product quality prediction based on two-layer feature selection and Stacking integrated learning by using industrial big data and Machine Learning(ML)and other related technologies,and the research contents are as follows.(1)Through the analysis of the current research status of product quality control,critical quality characteristics identification and integrated learning quality prediction of multi-variety and small batch products at home and abroad,the general research framework of the data-driven ex-ante quality control method with critical-to-quality characteristics(CTQ)identification and quality prediction of multi-variety and small batch products as the core is established by combining the characteristics of multi-variety and small batch production mode and the difficulties of quality control.(2)To solve the problem that the validity and analysis accuracy of the data-driven model are directly affected by the abnormal data in the collected quality data,an abnormal data repair model combining the box plot method and Multiple Imputation by Chained Equations(MICE)is proposed,which identifies the abnormal data identified by the box plot method The anomalous data are removed as vacant data,and then the vacant data are interpolated by MICE,thus realizing the repair of the anomalous data.It is shown that the proposed method can repair the anomalous data into data closer to the true value without affecting the overall distribution state of the data,which ensures the input data quality of the subsequent data-driven ex ante quality control model.(3)Aiming at the problem of high dimensionality of multi-variety and small batch product quality features and redundancy of quality information among features,the data is processed by feature selection method to reduce dimensionality,so as to achieve the recognition of CTQ.Combined with the repaired product quality dataset,a CTQ recognition model based on a two-layer feature selection method is proposed.Examples show that the model can identify fewer product CTQs while ensuring the accuracy of cross-validation prediction compared with the single feature selection method.(4)In order to obtain better quality prediction results for multi-variety small-lot products,a multivariety small-lot product quality prediction model based on Stacking integrated learning algorithm is constructed based on the study of Stacking integrated learning algorithm and combined with CTQ dataset of multi-variety small-lot products,and optimized in two aspects of base model combination method and metamodel multi-dimensional hyperparameters.The model combines multiple ML prediction algorithms and uses a two-layer structure to generalize and correct the prediction results of multiple ML prediction algorithms to achieve improved prediction performance.Experiments show that the established quality prediction model has superior prediction performance compared with the traditional single ML prediction algorithm.The above-mentioned research content utilizes the relevant technology in ML field to explore the common problems in the historical quality data of multi-variety and small-lot products,giving full play to the value of industrial big data and providing a quantifiable basis for the data-driven ex ante quality control method of multi-variety and small-lot products,which has certain application value for upgrading the quality of multi-variety and small-lot products,as well as achieving the goals of "cost reduction and efficiency increase" and "strong manufacturing country".