Improved Virtual Sample Generation Method Based on Conditional Regions Denoising Diffusion Probabilistic Model for Soft Sensors with Small Data

Soft sensors have excellent adaptability to complex environments and are widely used in complex industrial processes. However, its construction relies on high-quality samples, which poses a challenge to construct reliable soft sensors. To solve this challenge, this paper proposes a conditional regions denoising diffusion probabilistic model-based virtual sample generation (CRDDPM-VSG) method. It aims to make the generated samples better to fill the low-density regions. In CRDDPM-VSG, density-based spatial clustering with noise application (DBSCAN) is used to find low-density samples. Then, the improved CRDDPM is used to generate samples in the specified low-density sample regions. Finally, the generated samples are added to the original samples to enrich the sample information, thus improving the performance of the soft sensor. We used two sets of data to validate the performance of CRDDPM-VSG. Results show that CRDDPM-VSG not only generates samples that match the original distribution, but also better boosts the prediction ability of the soft sensor.