AI-enabled thermal monitoring of commercial (PHEV) Li-ion pouch cells with Feature-Adapted Unsupervised Anomaly Detection

Real thermal images had been captured using FLIR T640 infrared camera (FLIR Systems). This system is rated at ± 2% of reading accuracy, capturing 640 x 480 pixel images at 24bit depth using sRGB colour representation. Focal length of 13 mm and exposure time of 1/46s are pre-set. Thermal Images were obtained in a time-lapse mode with 15s intervals, resulting in thousands of images per test. The camera was suspended above the cell being evaluated. To avoid light reflections overlapping the thermal image, the surface of the cell was sprayed with a matte black paint (Ambersil) as advised by the thermal camera manufacturer, and the test chamber was blacked out for the duration of the test. Cell cycling performance data was collected alongside. The combined data obtained was used to prepare this dataset, center cropped a small subset of the images, to serve as training material for AI models. To replicate non-homogeneity, while coating some of the cell with matte black paint, an area was masked, leaving a reflective patch. That patch resulted in spatial anomaly, leaving areas of sharp temperature gradient, which would indicate abnormal cell operation and areas of potential cell failure. The images captured before applying the black paint were also exploited to represent potential anomalies. The captured thermal images have been structured to extend the MVTec [38]. The dataset has two main directories: "clean_pouch" and "noisy_pouch". Each one contains “train”, “test” and "groundtruth" directories. The train directory contains normal data only in a directory called “good”. The “test” directory contains the “good” directory for normal data and remaining directories (“overheat”, “reflection” (reflections before painting) and "spatial_tape" (patch reflections after painting)) to represent anomalies. The "groundtruth" directory contains groundtruth binary masks for the anomalous directories.