Application of optimized CNN to fixture layout in automotive parts. Training Dataset.

Fixture layout has aroused substantial interest in the research community during the last decades. It affects the design and production of fixtures themselves, and therefore manufacturing costs. While fixturing may seem simple in conception, it requires expertise and well-trained engineers. 

A general principle, usually called the 3-2-1 locating principle, ensures fixing an object, restringing its degrees of freedom. Fixtures must always comply with the principle. While most research approaches automation of the fixture layout using optimization or rule-based frameworks, this paper proposes supervised learning. The presented framework solves the 3-2-1 locating principle for sheet metal designs based on the experience of previous designs, using automotive b-pillars as a test study.

There are three main contributions. 1. A novel idea to introduce sheet metal design data in a Convolutional Neural Network (CNN), projecting the geometry over a plane. The Z coordinate transforms into gray-scale pixel values, generating a topographic map. 2. The framework reuses knowledge about fixturing to layout new workpieces. The framework is an add-in integrated with the CAD environment. 3. A hyperparameter-tuned CNN for regression generates the final output.

The results show high accuracy (≈ 100%) in classifying b-pillars and fast convergence in regression, proving model usability for industrial cases.

Source Code: https://colab.research.google.com/drive/18lS83mZEdwY5S-41pvcGXgUP6GbKpun_?usp=sharing