In the existing Manufacturing Execution System (MES), process design relies on manual experience, equipment selection is not accurate enough, and the error in working time prediction is significant, which seriously restricts production efficiency and intelligence level. This paper constructs a process design method based on machine learning (ML). First, equipment usage records, process flow data, and production efficiency data are extracted from MES, and the mean interpolation method is used to repair missing values. Outliers are detected and handled using box plots, and standardization is applied to ensure feature consistency. In the feature engineering stage, the expressiveness of categorical variables is improved through unique-hot encoding, and time-series trends are extracted via time-window aggregation. The comprehensive performance score for equipment is constructed in combination with key features to enhance the expressiveness of the feature space. At the same time, a generative adversarial network (GAN) is introduced to generate synthetic data for equipment failure and maintenance records, thereby optimizing the model's adaptability to sample scarcity and category imbalance. In the equipment selection stage, based on the constructed process tree, the random forest (RF) model is used to perform multidimensional mapping of complex feature interactions and to predict equipment selection. In the working time prediction stage, a multilayer perceptron (MLP) model is used to learn a nonlinear mapping from input features to working time, and the weights are continuously adjusted via the backpropagation algorithm to improve prediction accuracy. The experimental results show that in the equipment selection task, the RF model has a maximum accuracy of 95.1% and a maximum F1 value of 94.3%; in the working time prediction task, the mean absolute error (MAE) of the MLP regression model is 1.45 h, the root mean square error (RMSE) is 1.87 h, the mean absolute percentage error is 9.8%, and the determination coefficient is 90.2%. The research results verify the effectiveness and practical value of the proposed method in improving the accuracy of equipment selection and working time prediction.
Yin et al. (Thu,) studied this question.