In-Line Measurement Technology and Quality Control

Developing a supervised machine‐learning model capable of distinguishing fiber orientation of polymer composite samples nondestructively tested using active ultrasonics

“…The images of repeated trials for the different orientations were split into training and testing sets on a 75:25 ratio respectively. Optimization of hyperparameters, being the model optimizer (RMSprop, Adam, SGD), the STFT image size (150 Â 150 pixels to 500 Â 500 pixels in +50 Â 50 pixel steps), convolutional layer filter size (32 and 64 filters), convolutional layer kernel size (3 Â 3 and 5 Â 5 pixels), dense layer neuron size, [32,64] dropout size (0.1, 0.25, 0.5, and 0.75) and number of epochs (50 to 500) were done to improve model accuracies. A net optimum was found, with the following CNN specifications; RMSprop optimizer used, image size of 300 Â 300 pixels, convolutional layer filter size of 32 for the first four convolutional layers and 64 for the final two convolutional layers, convolutional kernel size of 3 Â 3 for all layers, dense layer neuron size of 32, a dropout value of 0.25 and the epoch number to be 300.…”

“…[12][13][14][15][16][17][18][19][20][21][22][23][24][25][26][27][28][29][30][31] They can be embedded in the process to act as in-line, online or offline testing apparatuses, but need to be robust and inexpensive (relative to the costs of production). [32] The use of artificial intelligence (AI) paired with these technologies offers a new approach to creating more discriminating quality assurance monitoring tools due to its ability to continually learn from historical data.…”