<p>Auto test equipment is widely used in semiconductor manufacturing for high-throughput electrical testing and device classification. During tray handling, vacuum suction abnormalities can occur because of empty trays, residual devices, improper chip placement, or package warpage. These events usually require operator intervention and may reduce equipment utilization. This study proposes a computer vision-based diagnostic framework for identifying tray conditions related to suction abnormalities in IC tray handling. A simulated handler inspection platform was constructed using top-view image acquisition, controlled illumination, representative tray geometries, and chip-like samples. A convolutional neural network model was developed using transfer learning to classify tray images into with-chips and without-chips conditions. The effects of illumination uniformity, tray format diversity, class imbalance, and error-driven retraining were systematically investigated. Compared with a single-light configuration, the three-light configuration increased the illumination level and improved spatial uniformity, with the relative standard deviation of illuminance decreasing from 22.63% to 5.17%. Expansion of the dataset from a single tray format to four tray formats improved the ability of the model to learn representative tray features. Furthermore, incorporating representative misclassified samples into the training set substantially improved the recognition of empty trays. Under the optimized internal evaluation condition, the without-chips class achieved 99.86% class-wise accuracy, 99.44% precision, and 98.89% sensitivity. The results indicate that a vision-based CNN framework can support automated diagnosis of suction abnormalities and reduce manual inspection effort in semiconductor ATE tray-handling applications.</p>