edfas.org 33 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 exciter narrows the bandwidth of the exciting light and the emitter blocks reflected and stray light. To measure the fluorescence lifetime, the magnification of the microscope was set to 20X. Furthermore, the camera had to be referenced, which was done using a Starna Scientific microscope slide, which has a fluorescence lifetime of 3.75 ns. After referencing, 45 FD-FLIM measurements of the 14 samples were done. The measurements were saved as ND2 files using different class names, e.g., ABS_M1, for further evaluation and the training of the NNs. DATASET PREPARATION From each ND2 file, the five layers, I, P, M, Pl, and Ml are extracted to prepare the datasets for the training of the NNs. To extract the layers, a python function is programmed. The function imports the ND2 files using a for-loop and the ND2 reader function from the “bio- formats” package. The function then saves the five extracted matrices as single tagged image file (TIF) files. As the MLP classification uses tabular data, the mean and standard deviation assuming a gaussian normal distribution is calculated for all I, P, M, Pl, and Ml channels of each measurement and the calculated values are saved into a comma-separated value (CSV) file. Therefore, a function was created using the “pandas” library in python. As four measurements of 630 measurements were erroneous, the data set consists of 626 FD-FLIM measurements. The mean and standard deviation were calculated for each of the five layers. The data was split in a ratio of 80% (500 measurements) to train and 20% (126 measurements) to test the NN. An example of the extracted data of the mean and standard deviation derived from the phase and modulation dependent fluorescence lifetime image of ABS and SAN is given in Table 1. One point to mention is that the modulation lifetime of SAN has a really high standard deviation, which comes from gaussian noise. The obtained noise is in the region of 0 to 330 ns. However, this has not a huge impact on the classification. DATAFLOW FOR MLP CLASSIFICATION A data flowchart for the MLP classification is shown in Fig. 3. In data-based classification, two approaches are investigated: MLP Classifier using Scikit-Learn and TensorFlow based MLP training. The MLP is trained with different combinations of these five layers using mean and Table 1 Mean and standard deviation of the phase and modulation dependent fluorescence lifetime of ABS and SAN Plastic type Phase dependent fluorescence lifetime Modulation dependent fluorescence lifetime ABS 2.47 ± 0.07 3.84 ± 0.20 ABS 2.47 ± 0.07 3.84 ± 0.19 SAN 1.54 ± 0.40 4.00 ± 0.93 SAN 1.53 ± 0.28 4.00 ± 0.95 Fig. 2 Schematic representation of the measurement setup containing a microscope, a laser diode, a FD-FLIM camera and two optical filters. Fig. 3 Data flowchart of the MLP Scikit-Learn and TensorFlow.
RkJQdWJsaXNoZXIy MTMyMzg5NA==