Production line fault prediction and root cause identification using neural networks and graph neural networks

• Autor(a) principal: Bastos, Fábio
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/38818
• topic: Production line failures; Root cause analysis; Neural networks; Shapley values; Feature importance; Graph neural networks; Heterogeneous graphs; GNNExplainer; Real-time prediction; Machine learning

We consider the problem of prediction of production line failures, and identifying their root causes. Modern production lines contain many steps, and integrate many parts and materials. At the same time, sensors monitor every stage and aspect of the process. Limiting product faults, and identifying their causes is an important consideration for improving productivity and reducing costs, but manual approaches are not practical. Automated solutions would therefore be of great benefit. We applied a neural network model to this problem, which achieved results comparable to those achievable using more limited methods. This approach has the advantage of allowing further analysis beyond simple classification. We show that Shapley values can be used to ascribe feature importance, a first step to identifying the root cause of failures. We then developed a Graph Neural Network model, with the aim of producing a more directly interpretable model, as product graphs much more closely represent the interactions between real features. This approach proved to be much more computationally demanding, so that full results could only be obtained for a limited dataset. However, the advantages of this approach can already be seen. By deconstructing product graphs by timestamp, we show that partial graph predictions (representing real time situations where production is still incomplete) can give meaningful results. We also adapted the GNNExplainer model to heterogeneous graphs, allowing a direct calculation of node or feature importance in determining the outcome. These methods, if developed fully, have strong potential for automated, real time root cause analysis solutions.