Development of Soft Sensors based on Deep Learning Techniques
Soft sensor, deep learning, semi-supervised learning, system identification
Industrial processes are characterized by strongly correlated variables and high nonlinearity. There are countless processes where it is not possible to accurately measure process variables or process quality variables. When technical limitations, high cost and unavailability of specialist equipment make it difficult to measure the most important variables in real time, soft sensors are an attractive way to deal with the problem. Implementing a system that can infer primary variables of interest by using secondary variables precisely is not simple. In addition to the complexity associated with system deployment alone, the labeled data used in this type of deployment is often unavailable, with the large mass of available unlabeled data. This work proposes the study and development of soft sensors based on deep learning techniques. The goal is to propose a semi-supervised learning technique for building virtual sensors in industrial applications using labeled and unlabeled data. At first, the unsupervised step uses the unlabelled data to extract the main characteristics and initial adjustment of the model weights. The techniques under study are autoencoders and Boltzmann machines. In the supervised stage, model training begins with weights that were generated in the unsupervised phase. The supervised training techniques chosen for the construction of this work are recurrent neural networks and deep networks. It is expected to obtain models capable of representing the entire operating range of the identified system and with great generalization. Soft sensors that well represent the dynamics of complex industrial processes can improve process quality, product quality and hence profitability. To develop this work, we use a benchmark of chemical process controls, the Taneessee Eastman.