Abstract: To detect hidden welding defects efficaciously with image information detection methods, a new detecting and identifying method was proposed based on multi-source data fusion to improve the detection types and accuracy of defect detection. Firstly, to form a original feature space of welding quality, several processes was arranged, including the eigenvalue calculation of multi-dimensional information such as sound, voltage, spectrum and temperature collected in the welding process, and the combination of the calculated eigenvalues with the image eigenvalues of weld pool. Then, to reduce the dimension of feature space, a linear discriminant analysis (LDA) method was used to form a low-dimensional feature space of welding information. And then, screening the neighborhood search space with the isolated forest method, dividing the welding data points of the neighborhood search space into multiple overlapping subsets, and searching new data points in the multiple overlapping subsets with the local outlier factors (LOF) method to detect the anomalies easily in the welding process, the method was designed to fully consider the global characteristics of welding quality data and great reduction of the computational complexity. Finally, a probabilistic neural network (PNN) optimization was carried out based on the artificial bee colony (ABC) algorithm for the accuracy of subdivided welding quality data and identified anomalies, enhancing the global search capability and avoiding falling into local optimality. The experimental results show that the proposed method can achieve 97.44% welding anomaly detection accuracy and 96.03% comprehensive welding anomaly recognition accuracy, proving the effectiveness of the proposed method.