Assembly Faults Diagnosis Using Neural Networks and Process Knowledge

Traditional manufacturing process fault monitoring and detection methodologies have been based on Statistical Process Control (SPC) charts and rules. In most cases, SPC is used to detect a fault or an out-of-control condition while fault diagnosis relies on operator expertise to identify the potential root causes. Current sensor developments allow for the acquisition of large amounts of data from parts and processes in a manufacturing environment. In addition, new modeling tools have increased the efficiency and accuracy of process modeling, providing useful knowledge about product-processes interaction. This paper presents a new methodology for fault diagnosis using a Feed Forward Back-Propagation Neural Network. The proposed neural network is trained using process knowledge and then applied to the detection of manufacturing process faults. The methodology results in a modified control chart that uses measurement data from the assembly components and plots an indicator representing the presence or absence of a predefined fault. Two case studies are presented: a diagnosis system for fixture faults in a generic assembly process, and a diagnosis tool for fault detection and identification in an automobile door assembly.