Carbon fiber PAEK prepreg micrograph analysis using Weka deep learning methodology.
Topic(s) : Special Sessions
Co-authors :
Adrien LE REUN , Arthur LEVY (FRANCE), Henri-Alexandre CAYZAC , Vincent SOBOTKA (FRANCE), Steven LE CORRE (FRANCE)Abstract :
One of the major challenges in composite manufacturing is the identification, control and
minimization of defects in the final parts. In continuous fiber laminates, defects mainly refer
to fiber misalignment, waviness or wrinkle, fiber content variability, and porosities. Many
destructive or non-destructive experimental techniques can be used to assess these defects
[2]. Porosity or fiber content, for instance, can be quantified by density measurement,
ultrasonic techniques, matrix digestion, or X-Ray tomography. Another classical laboratory
method to investigate microstructure is by micrograph analysis. It consists in cutting crosssections, embedding, polishing and observing by optical microscopy. The obtained
micrographs are analyzed with image processing methods to quantify magnitudes such as void content, fiber volume fraction. The optical microscopy can reach resolutions below 1 micron which appears useful in the field of carbon fiber composites for instance.
Prior to understanding the manufacturing of continuously reinforced PAEK/carbon fiber
laminates, the quality of the prepreg tape is to be assessed. Two types of prepregs were
investigated. In a standard way, samples were embedded in a potting resin, polished and
observed under optical microscope Keyence VHX-7000.
The segmentation step to identify the three phases (fiber, resin or void) is not trivial. Besides
the classical thresholding methods used in the literature, novel deep learning tools have
appeared in the domain of image processing. In this work, the Fiji trainable segmentation tool Weka was used [2]. After the learning sequence provided by the user, Weka applies filtering techniques and applies data mining to suggest a segmentation based not only on thresholding (see Figure 1).
For the two types of prepreg tapes studied here, the variability due to the learning process
was investigated. We found that the number of images and the region of interest used for the learning affects the final suggested segmentation. In addition, different operators were asked to perform the learning sequence to investigate the variability due to the human contribution to the training of the Weka tool.
The segmentation gave insight on the microstructure morphology of the prepregs. Moreover, profilometer measurement was performed to discriminate actual porosities from polishing residues (Figure 2). In addition to the average void and fiber content, their distribution in space was found to differ from a prepreg to another. This is suspected to be related to the preimpregnation process. This variability in void content and fiber content will affect the subsequent manufacturing of the laminate.
minimization of defects in the final parts. In continuous fiber laminates, defects mainly refer
to fiber misalignment, waviness or wrinkle, fiber content variability, and porosities. Many
destructive or non-destructive experimental techniques can be used to assess these defects
[2]. Porosity or fiber content, for instance, can be quantified by density measurement,
ultrasonic techniques, matrix digestion, or X-Ray tomography. Another classical laboratory
method to investigate microstructure is by micrograph analysis. It consists in cutting crosssections, embedding, polishing and observing by optical microscopy. The obtained
micrographs are analyzed with image processing methods to quantify magnitudes such as void content, fiber volume fraction. The optical microscopy can reach resolutions below 1 micron which appears useful in the field of carbon fiber composites for instance.
Prior to understanding the manufacturing of continuously reinforced PAEK/carbon fiber
laminates, the quality of the prepreg tape is to be assessed. Two types of prepregs were
investigated. In a standard way, samples were embedded in a potting resin, polished and
observed under optical microscope Keyence VHX-7000.
The segmentation step to identify the three phases (fiber, resin or void) is not trivial. Besides
the classical thresholding methods used in the literature, novel deep learning tools have
appeared in the domain of image processing. In this work, the Fiji trainable segmentation tool Weka was used [2]. After the learning sequence provided by the user, Weka applies filtering techniques and applies data mining to suggest a segmentation based not only on thresholding (see Figure 1).
For the two types of prepreg tapes studied here, the variability due to the learning process
was investigated. We found that the number of images and the region of interest used for the learning affects the final suggested segmentation. In addition, different operators were asked to perform the learning sequence to investigate the variability due to the human contribution to the training of the Weka tool.
The segmentation gave insight on the microstructure morphology of the prepregs. Moreover, profilometer measurement was performed to discriminate actual porosities from polishing residues (Figure 2). In addition to the average void and fiber content, their distribution in space was found to differ from a prepreg to another. This is suspected to be related to the preimpregnation process. This variability in void content and fiber content will affect the subsequent manufacturing of the laminate.