Inteligência artificial aplicada à detecção e diagnóstico de falhas em processos químicos

Abstract

Fault detection methods are used to detect abnormal operations in processes caused by disturbances, equipment degradation, defects in sensors and equipment. These methods can be based on residue analysis, analytical redundancy, or physical redundancy. The methods with analytical redundancy are among the most industrially used approaches. The analytical redundancy in processes can be obtained through analytical mathematical models of the process through the identification of models or artificial intelligence through machine learning. Failure detection methods through artificial intelligence have been increasingly studied in the literature in the last decade. The methods that use neural networks, fuzzy logic and support vector machines are fault detection techniques with high usability and reliability. The detection procedure is done in three steps: model training for detection, using normal and abnormal process data; validation of the model; and fault pattern detection. Currently, one of the paradigms in the area of fault detection is the extraction of important characteristics to the procedure of detection and patterns classification of process abnormal operations. The present work investigates these methods of fault detection based on machine learning and proposes a new method based on deep learning. Deep learning is an innovative approach to machine learning in order to dispense part of the necessary preprocessing of classic machine learning methods by automatically generating invariant properties in its hierarchical representation layers. These methods have presented promissing results in different applications such as speech recognition and natural language processing. Deep-learning algorithms can solve problems found in fault-detection methodologies, especially in the extraction of properties and preprocessing of data. The results obtained by the proposed technique linguistics neural network show that learning failure detection methods can be easily applied in different processes, only with the use of historical data and/or process models, resulting compatible detections with other studied methodologies.