| CPC G05B 19/4099 (2013.01) [G06N 3/04 (2013.01); G06N 3/08 (2013.01); G06N 5/04 (2013.01); G06N 20/00 (2019.01); G05B 2219/49023 (2013.01)] | 10 Claims |
|
1. A method for automatic prediction of porosity appearance generated during Laser Powder Bed Fusion (L-PBF), performed by an additive manufacturing (AM) system (1) from at least one material (2), the method comprises steps for training a neural network comprising:
reading a kinematic laser file describing laser exposure over a layer of a part slice, the kinematic laser file comprising at least an X coordinate, an Y coordinate of each of a plurality of points in a laser path, the plurality of points being situated on a 2D mesh representing the layer, a point of said plurality of points being associated with an exposure time representing a time which a laser passes through the point and an exposure type indicating whether a point of said plurality of points belongs to a part contour or to an inner laser path, and
converting the kinematic laser file into a dataset comprising an array of pixels discretizing the layer, a pixel of said array being associated with local coordinates and with an exposure time representing a time at which a laser passes through the local coordinates of the pixel;
generating (S1) labels of pores, representing presence or absence of pores, in each pixel of the dataset using a porosity simulator (50) wherein the porosity simulator is image-based and the generated label of pores for a pixel of the dataset is in the form of 1 or 0, representing respectively presence or absence of pores with respect to the pixel;
pre-training (S2) a machine learning (ML) model, wherein the pre-training comprises:
a first pre-training sub-step (S21) comprising creating a first data-set, by for each pixel of the dataset, processing input features and pore label features based on the labels of pores generated by the porosity simulator, said input features and processing pore label features constituting feature vectors associated with said pixel; and
a second pre-training sub-step (S22) comprising using the feature vectors of the first data-set to generate a pre-training ML model (200);
training (S3) a ML model, wherein the training comprises:
a first training sub-step (S31) comprising creating a second data-set by, for each pixel of the dataset, processing input features and pore label features based on tomography of manufactured coupons (20), said input features and pore label features constituting feature vectors associated with said pixel; and
a second training sub-step (S32) comprising using the feature vectors of the second dataset to train the pre-trained ML model (200) to generate a trained ML model (300);
generating a predicted porosity appearance during Laser Powder Bed Fusion (L-PBF), performed by the AM system (1) from at least one material (2), by applying (S5) the trained ML model on new data continuously provided by the AM system (1); and
post-processing the predicted porosity appearance so as to generate a quality report,
wherein the input features being processed in the first pre-training sub-step (S21) and first training sub-step (S31) comprise Correlation Density Spectrum (CDS), and the input features are obtained by preprocessing steps (SA, SB, SC) comprising:
kinematic processing (SA) to generate a distance matrix based on kinematic laser data, by calculating a Euclidean distance of each pixel of the dataset, belonging to an inner laser path, referred to as inner pixel, with all the pixels of the dataset through which the laser passes;
calculating (SB) a distance correlation of pixels comprising covariance matrix based on the distance matrix; and
generating (SC) CDS based on the covariance matrix.
|