Developing Graph Invertible Neural Network for high-dimensional Inverse Problems (a part of AutoKval: Automatic quality assurance and process feedback in running production)
In recent years, the development of computer vision and deep learning technologies has promised significant advancements in the production processes, such as the sheet metal forming process, where these can provide an opportunity to analyze high dimensional 3D scanned data of complex geometries in order to obtain a more comprehensive understanding of the underlying causes of defects and failures. These technologies offer the potential to transform the industry by reducing waste, improving quality, and ultimately, increasing profitability.
To achieve these benefits, this project aims to develop a deep learning model that can analyze 3D scanned data of sheet metal components to perform root cause failure analysis by determining the potential causes that could lead to inspected defects and unacceptable geometrical deviations. To approach this problem, the root cause failure analysis will be reframed as an inverse/inference problem, which involves working backward from the defects and geometrical deviations observed in the 3D scanned data so as to identify the most likely causes among various factors, including material properties, manufacturing process parameters, or tooling issues. By representing the root cause failure analysis as an inverse/inference problem, it becomes possible to provide the manufacturer with valuable informative feedback, enabling them to optimize the process and take corrective actions in order to prevent similar issues from occurring in the future.
In our project, we aim to develop and use a proper deep generative model to solve the aforementioned inverse/inference problem for the following reasons:
1. The high dimensionality of the observation data which is 3D point clouds or mesh of the scanned parts
2. Our objective for accelerating the required time for solving the inverse/inference problem and make it possible to conduct an almost real-time inference task
3. Reducing the sensitivity of the solution of the inverse/inference problem to the noisy observations
Since RealNVP (Real-Valued Non-Volume preserving method), which is a type of normalizing flow model, has the potential to perform exact likelihood inference, we aim to incorporate graph neural network (GNN) blocks in the architecture of the conditional invertible neural network and develop an invertible graph neural network based on the RealNVP concept. This approach will enable us to effectively process high-dimensional mesh data with graph-like structures, such as the geometrical deviation data in our application, and estimate the density of the parameters of interest like material and process parameters.