edfas.org 19 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 28 NO. 3 under various magnifications (20X, 50X, and solid immersion lens (SIL)) across different bias conditions and device configurations. Images captured at 1X magnification were intentionally excluded due to excessive noise levels that would adversely impact training quality and model performance. A key component is the stream-aware training architecture that handles two distinct image types: raw (grayscale) images capturing pure photon emission data, and overlay (RGB) images superimposing emissions onto circuit layouts. These streams possess fundamentally different statistical properties and noise characteristics. Raw images exhibit high dynamic range with hotspots as bright regions against dark backgrounds, while overlay images present complex visual landscapes with potential interference patterns. Strict stream separation is maintained throughout training, computing separate normalization parameters to ensure appropriate feature learning without cross-contamination. Given the scale of the dataset, manual annotation would be prohibitively time-consuming. A hybrid labeling strategy was developed combining automated proposal generation with expert curation. The auto-seeding component utilizes classical computer vision techniques including multiscale blob detection using difference of Gaussians (DoG) filters, intensity-based thresholding, and morphological operations. Connected component analysis identifies candidate regions filtered by size, aspect ratio, and intensity characteristics. For overlay images, association rules and clustering algorithms were applied to identify pixels with similar spectral characteristics, achieving 70-80% recall. The human-in-the-loop curation process involves expert FA engineers reviewing and refining the auto-generated annotations. This curated subset of 100-200 high-value images per stream corrects false positives, captures subtle emission patterns, and establishes ground truth for edge cases. The curated dataset becomes the foundation for fine-tuning operations that enhance model sensitivity and precision. Data augmentation is implemented through streamaware pipelines designed to preserve emission signal integrity. Conservative geometric transformations are applied, including horizontal and vertical flips, controlled cropping, limited-angle rotations, and carefully constrained photometric adjustments. Raw images undergo contrast and gamma modifications to simulate acquisition variability, while overlay images further incorporate saturation adjustments to account for differences in colormapping schemes. The training pipeline utilizes TensorFlow’s TFRecord format for efficient data handling and stratified sampling for balanced representation (80:10:10 train/validation/ test split). The multi-stage approach begins with autoseeded dataset training using standard hyperparameters (batch size 4, learning rate 0.001, 25,000 steps, 4 hidden layers with 256 filters). Finetuning employs the curated subset with reduced learning rate (0.0001) and regularization techniques. Hyperparameter optimization includes weighted sigmoid focal loss tuning (alpha=0.25, gamma=2.0) to address class imbalance. Model selection combines precision, recall, and inference speed metrics for optimal production deployment. Figure 2 summarizes the end-to-end training data pipeline and model training workflow. AI MODEL SELECTION AND PRINCIPLES For emission hotspot detection, an SSD/VGG16 detector based on the single shot multibox detector (SSD) architecture[7] was selected, consistent with prior computer-vision work applied to PEM analysis.[8] This selection was driven by a careful evaluation of the specific requirements of PEM analysis. PEM hotspots are typically small, circular or near-circular blobs with low contrast against the background. The SSD architecture is particularly well-suited for this morphology due to its multiscale feature maps, which allow it to detect objects of various sizes effectively in a single pass. The VGG16 backbone provides a robust feature extractor that balances depth and computational efficiency. Unlike two-stage detectors like Faster R-CNN, which can be computationally expensive, the SSD/VGG16 combination offers fast inference times, making it ideal for a bench-side application where engineers need near real-time results. The principle behind Fig. 2 Machine learning process flow chart illustrating data extraction from FA database, auto-seeded proposal generation, manual curation, augmentation, TFRecord generation, and model training using SSD/VGG16 architecture.
RkJQdWJsaXNoZXIy MTYyMzk3NQ==