Abstract: Wave soldering online quality inspection serves as a crucial method for identifying defective products. High-precision inspection can reduce production costs and support early warning mechanisms during the wave soldering process.Conventional inspection techniques are effective in detecting common defects such as solder skips,cold solder joints,and solder bridging.However,various other anomalies may also occur on the production line,characterized by their suddenness,rarity,unpredictability,and heterogeneity. These anomalies are difficult to detect using standard inspection approaches. To address this challenge,an anomaly detection method based on reconstruction error metrics was proposed.Latent feature vectors generated by existing quality inspection models are reconstructed,and reconstruction errors were used to determine whether a sample was anomalous. This enabled online detection of production anomalies while maintaining high-accuracy detection of common defects. Furthermore,a feature extraction model was developed by integrating the Convolutional Block Attention Module (CBAM) with the EfficientNet architecture,and a Gated ensemble Generative Adversarial Network (GGAN) was introduced to further enhance anomaly detection performance. Empirical verification using production data from an electronics enterprise in Nanjing demonstrated that an anomaly detection rate of up to 98.5 % was achieved.Compared to existing methods,the proposed method significantly improved the detection of anomalous samples in wave soldering production lines while ensuring high-precision identification of typical defects.

Key words: wave soldering, anomaly detection, latent feature vector, reconstruction error, gated ensemble generative adversa-rial network