Source: Deep Learning on Medium
Data-set Creation: To get started, one can leverage several open source machining features data-sets. A synthetic data-set of machining features cal also generated using CAD modelling softwares in an automated fashion. The features are generated by randomly selecting feature specific parameter values in a predefined range. For instance, to create models with blind hole features, four parameters can be randomly sampled such as R, Cx , Cy , and D, where R is the radius of the blind hole, (Cx , Cy) are the centre coordinates, and D is the depth. Random vectors of these four parameter values were sampled from the ranges used generally in production. A random valued vector for the parameters of a feature creates one instance of the feature. In order to generate 1000 models of the same feature (blind hole, through hole etc.), the parameter ranges should be uniformly sampled for obtaining 1000
random vectors for each such feature.optimizers
In a similar way, the same approach can be used for all different kind of features required for the recognition. We can also add the different orientations which a feature can take for all the features. The thumb rule for number of total samples required per feature could be to keep it in range 6000–8000 samples per feature. RanFeatureNet: Machining feature recognition based on 3D Convolution Neural NetwoFeatureNet: Machining feature recognition based on 3D Convolution Neural Networdomly sampled sizes and incorporation of different orientations helps making the trained model more robust and accurate.
But, What about the suitable 3D representation for neural networks? The CAD modelling applications generally use boundary representation (B-rep) which are not convenient for use with neural networks. There are various options to represent the 3D CAD models in a way apt for neural network training. It can be broadly classified into euclidean representations (Volumetric voxel, Volumetric octree, Multi-view), and non euclidean representations (Point clouds, Graphs, Meshes). 3D Euclidean data representation has an underlying grid structure that offers global parametrization and a common system of coordinates. Such properties helps extending the already-existing 2-D DL techniques to 3D data a straightforward task. 3D Euclidean representation is more suitable for analyzing rigid objects where the deformations are minimal. On the other hand, 3D non-Euclidean representations do not have the grid array structure where there is no global parametrization. Consequently, it makes extending classical DL techniques to such representations a very challenging task. One of the most convenient representation to use with neural networks in 3D recognition setting is Voxel Volumetric Representation. A 3D shape is represented as a binary valued 3D voxel grid, where 1 indicates that the voxel is inside the shape, and 0 indicates that it is outside. This discretization helps in attenuation of noise.