Machine learning models for measurement of pH using a low-cost image analysis strategy

Soft sensor use has been growing in many industries as an alternative to conventional physical ones due to their disadvantages, such as prices and maintenance. One difficult measurement is the pH values in a pressurized system, which requires specialized equipment. This work aims to develop machine learning models to determine the pH in pressurized systems (up to 6 MPa) directly from the reaction medium using image analysis as an initial step to create an applicable soft sensor. Classification and prediction models were developed using different techniques: Convolution Neural Networks (CNN), support vector machines (SVM), and decision trees (DT). All of them were explored in the classification models, but CNN was the only used for the regression ones. The models were trained with buffer solutions images and then tested using two scenario cases (acid-base titration curve and pressurized CO2-H2O systems). The best classification models were the CNN ones, once the SVM and DT had good results for the test group (accuracy (ACC) above 90 %), but ACC mean values below 75 % for the scenario cases. Model CNN_class_10 presented the best results (accuracy above 70 % in all the cases). The CNN prediction models showed the best results, where the models CNN_pred_7 and CNN_pred_9 had present R2 values higher than 80 % for all experiments in both scenarios cases. In conclusion, improving the CNN regression models appears to be the most promising approach to creating the soft sensor in future works.