Multiple-conditioned welding monitoring is a challenging issue in complex robotic welding tasks. In practice, the monitoring system has to be sensitive to different welding conditions (WCs) and weld quality changes. In this paper, a swing high temperature sensor system is used in order to obtain the temperature distribution curve under different WCs. A sigmoid feature extraction (SFE) method is proposed to obtain the geometric features of the temperature distribution curve, and a weld monitoring algorithm is proposed for multiple-conditioned welding tasks using multi-layer perceptron (MLP) classifier and Kalman filter-based Gaussian probability density function (PDF) prediction for the probabilistic weld quality estimation. When there are unknown WCs, the proposed method uses an efficient incremental learning for the MLP and an online maximum likelihood estimation for the Gaussian models of the unknown WCs. The experimental results show that the proposed framework can accurately reflect the weld quality changing in both single-WC and multiple-WC tasks. In addition, the proposed adaptive updating methodology can achieve comparable performance with unknown WCs, as compared to the results with knowing all the WCs.