According to the latest reports, selective soldering plays a critical role in the manufacturing of printed circuit boards (PCBs), especially for through-hole technology (THT) components, where molten solder is dispensed from a nozzle to create a continuous 360^° wave around the part. The consistency of this wave is crucial for reliable electrical and mechanical connections. However, repeated exposure to high temperatures (up to 300^°C) and molten solder (e.g., Sn-Pb or SAC alloys) causes nozzle oxidation, leading to impurity buildup and irregular waves. This degradation results in defective solder joints, compromising product quality, increasing rework costs, and potentially causing production downtime. Current monitoring relies on manual visual inspections, which are subjective and timeconsuming, or rule-based image algorithms that are prone to noise and require frequent recalibration. To overcome these limitations, this paper investigates autoencoders-a type of neural network for unsupervised anomaly detection-that reconstruct input images and flag deviations via reconstruction errors. Data was collected from camera-monitored nozzles under controlled conditions. Preprocessing included grayscale conversion, Gaussian blurring, normalization, and resizing. A Flat (fully connected baseline) was evaluated. Models trained on operable images using mean squared error (MSE) loss; performance assessed with MSE distributions and receiver operating characteristic (ROC) curves, using area under the curve (AUC) as the primary metric. Results show the flat autoencoder achieving  ∼ 96 % AUC with a MSE separation (oxidized > 0.02). This is the first application of autoencoders to real-time nozzle oxidation detection in selective soldering, filling a gap in AI for tool-specific degradation. It enables predictive maintenance, reducing defects.