A CNN-based intelligent optimization framework for bleaching powder production

Bleaching powder (calcium hypochlorite) production relies heavily on manual experience, which results in low process control precision and poor product quality stability between batches. To address this, an orthogonal experimental design was used to examine the effects of key process parameters, including the Ca(OH) 2 concentration and reaction temperature, on the available chlorine content of the product. A multivariate polynomial regression was employed to establish a predictive model relating product quality to process parameters. After optimization through a grid search and cross-validation, the model achieved an R 2 of 0.9067 and a RMSE of 0.2545. A ResNet-18 convolutional neural network (CNN) was trained to analyze reaction-process images and identify six reaction stages. The model achieved a validation accuracy of approximately 90 %, with precision and recall values in the range of 85–90 %. In particular, the model demonstrated high recall for the key reaction endpoint (Stage 4), enabling reliable identification of the critical process stage. A closed-loop intelligent control system was developed by integrating the predictive model with a visual recognition module. Validation showed that compared with traditional manual control, this system improved the uniformity of product quality, reduced the batch standard deviation by 75 %, and significantly enhanced production stability and efficiency. This study provides an efficient end-to-end solution for intelligent upgrading of complex multiphase reaction processes.