With the development of complex industries, soft sensors have extensive application prospects. For the optimization of intricate industrial processes, precise models are essential. However, due to the insufficient and poor-quality training data in industrial processes, the established models frequently exhibit low accuracy. We propose an effective method for virtual sample generation based on the Score Based Generative Model (SGM) to address this challenge. In this approach, the Local Outlier Factor (LOF) algorithm is initially employed to detect outliers in the data. Subsequently, the Score Based Generative Model generates virtual input samples around the identified outliers. Following this, the Mean Teacher approach for semi-supervised learning is utilized to forecast the outputs of the virtual samples. The student model’s prediction accuracy is improved by incorporating the virtual samples into its updates. Finally, the synthetic dataset is formed by combining the input and output components of the virtual samples, augmenting the original dataset. In order to prove the efficiency and superiority of this approach, three-dimensional numerical simulations and industrial data purified terephthalic acid (PTA) were used for experiments. The results show that SGM-VSG can improve the prediction accuracy of soft sensor better than other methods of generating virtual samples.