Fault detection and classification of industrial processes using lSTM neural networks with data compression techniques

• Ano de defesa: 2021
• Autor(a) principal: Correia, Paulo Victor Queiroz
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Rio Grande do Norte; Brasil; UFRN; PROGRAMA DE PÓS-GRADUAÇÃO EM ENGENHARIA ELÉTRICA E DE COMPUTAÇÃO
• Palavras-chave em Português: Indústria 4.0; Detecção e identificação de falhas; Symbolic Aggregate Approximation; Long Short-Term Memory; Computação de ponta; Compressão por podagem; Tennessee Eastman Process
• Link de acesso: https://repositorio.ufrn.br/handle/123456789/43116

Industry 4.0 set a paradigm shift in industrial process monitoring and control. It installed several sensors in different parts of the plant, connecting these industrial processes with the Internet of Things and Cloud Computing. Although, the data generation growth demanded engineers build proper environments to process and store the growing amount of data. This growth caused an increasing energy consumption, computational complexity and environmental degradation. Therefore to address these demands, this dissertation proposes efficient approaches to perform Fault Detection and Identification in industrial processes. The first approach consists of using Symbolic Aggregate Approximation (SAX) to compress process variables to reduce the load on data warehouses. Then, we train a Long Short-Term Memory (LSTM) neural network with those compressed inputs to perform fault detection. Finally, the second approach addresses efficient edge computing systems, performing LSTM neural network compression with pruning technique. The compression reduces the memory usage and number of operations of these networks, saving energy and accelerating inference speed in edge computation. To assess the performance of both approaches, we use the Tennessee Eastman Process (TEP) as the benchmark with classification metrics of accuracy, precision, recall and F1-Score. We are also going to analyse the compression efficiency of both approaches, studying their viability and parameter reduction in LSTM networks.