Evaluating the Use of Artificial Neural Networks and Graph Complexity to Predict Automotive Assembly Quality Defects

Abstract

This paper presents the use of subassembly models instead of the entire assembly model to predict assembly quality defects at an automotive original equipment manufacturer (OEM). Specifically, artificial neural networks (ANNs) were used to predict assembly time and market value from assembly models. These models were converted into bipartite graphs from which 29 graph complexity metrics were extracted to train 18,900 ANN prediction models. The size of the training set, order of the bipartite graph, selection of training set, and defect type were experimentally studied. With a training size of 28 parts, an interpolation focused training set selection with a second-order graph seeding ensured that 70% of all predictions were within 100% of the target value. The study shows that with an increase in training size and careful selection of training sets, assembly defects can be predicted reliably from subassemblies' complexity data.

Comparing function structures and pruned function structures for market price prediction: An approach to benchmarking representation inferencing value

“…In this manner, the prediction error can be greater than 100% for the percent error and the normalized error. It is fully recognized that improvements to the predictive power of the approach can be made based on larger training sets, more similar training samples, and including information beyond the graph connectivity of the models (see Patel et al, 2016, for a comparison on factors influencing the inferencing power). However, improving the predictive capabilities of the approach is not the focus of this paper; rather, comparing the representations using the same prediction framework is.…”

“…To understand the accuracy from each of the four prediction models, it is important to understand the amount of error in each of these models. To calculate the amount of error, three different types of error analysis formulae, listed in this section, were used based on those found in Patel et al (2016). To handle the wide range of values calculated by the ANNs, it is important to have a specialized error calculation formula.…”

“…From the previous discussions on pruning and comparison of error values obtained from the graph connectivity complexity method, which will be discussed further in Section 7, it is clear that there is a marked difference in how well each of the various pruned function structure sets behave as prediction models. This method has been used for market price prediction (Mathieson et al, 2011; Mohinder et al, 2014; Sridhar et al, 2016 b ), for assembly time estimation (Owensby et al, 2012; Namouz & Summers, 2013, 2014), for assembly quality defect prediction (Patel et al, 2016), and for using requirements to predict life cycle costs (Visotsky et al, 2017). As per the classification framework developed in Summers and Shah (2004), a design representation can be classified as being distinct on the basis of these five characteristics: vocabulary, structure (grammar), expression, purpose, and abstraction of the representation.…”

Towards the Rapid Design of Engineered Systems Through Deep Neural Networks

“…The utilization of deep learning, and specifically autoencoders, also led to the creation of a computational framework that models the curiosity of a given user in order to provide surprising examples [21]. Neural networks have also been utilized to automatically predict quality defects in automotive parts [22] and to support design for additive manufacturing [23][24][25]. These examples, while not exhaustive, serve to highlight potential utility of neural networks for design and the need for a standardized approach to implementing them.…”

Function in engineering: Benchmarking representations and models

“…Several studies have been undertaken to study the authorship and consistency and interpretability of models (Kurfman et al, 2003; Caldwell, Ramachandran, et al, 2012; Caldwell, Thomas, et al, 2012), as well as elucidating the correctness of model construction (Nagel et al, 2015). Alternatively, some approaches have been proposed that automatically reason on function models from database collections (Lucero et al, 2014; Patel, Andrews, et al, 2016; Sridhar et al, 2016). Finally, some approaches entail the support of first principle based physics reasoning (Goel et al, 2009; Sen et al, 2011 b , 2013 a ).…”