Imaging Based Fault Detection and Diagnosis of Potassium Dihydrogen Phosphate Batch Crystallization Process

Crystallization is a purification process that is relevant to some industrial fields, such as the pharmaceutical and food sectors. The development of a crystallization process can present some issues related to the occurrence of crystals agglomeration and presence of impurities in the system. As a consequence of these problems, the product of the crystallization process may fail to meet quality requirements. In this sense, the aim of this work is to develop a fault detection and diagnosis (FDD) system based on computer vision. Initially, a potassium dihydrogen phosphate (KDP) batch crystallization experiment in water was performed, presenting nucleation, crystal growth, dissolution, agglomeration and a clog during the batch. After this experiment, images of the KDP crystallization were obtained to train two CNNs. A convolutional neural network (CNN) was developed to identify images of crystals, impurities and agglomerates. The dataset was divided in 64% for training, 20% for test and 16% for validation. Initially, an optimization problem was defined to select the hyperparameters of the CNN. Also, a range of epochs between 10 and 60 was tested to find the best solution. The best CNN was able to identify the images of the test dataset with an accuracy of 93% for the test dataset. The proposed image-based methodology was efficiently applied for fault detection and diagnosis in crystallization, being able to prevent loss of experiments and can also be extended to study the evolution of crystallization batches.