Research on the generation method of missing data for soft measurement based on GAN

To solve the problem of low precision in soft sensor models caused by sensor data loss in industrial processes, a new method of sensor data generation based on generative adversarial nets (GAN) is proposed. Firstly, the missing area of sensor data is detected by the isolated forest algorithm. Secondly, conditional generative adversarial nets (CGAN) are training using the attributes of missing data. By adding random sequences to the input conditions of CGAN as additional information, the data is generated iteratively in CGAN. The wasserstein generative adversarial nets gradient penalty (WGAN-GP) cost function is used to improve the stability of network training. Finally, a sampler is introduced to fill the sampled data into the missing region and form a complete data set to improve the accuracy of the soft sensing model. In this paper, the temperature sensor data of a nickel flash furnace is used as the target variable for soft-sensing modelling, and the feasibility and effectiveness of the proposed method to improve the accuracy of the soft-sensing model are verified.