SS-DDPM: A novel denoising diffusion probabilistic model for industrial sensor signal data augmentation

Recently, industrial artificial intelligence faces the twin challenges of data scarcity and distribution imbalance, which severely hinder model performance. In comparison with model structure optimization, data augmentation based on generative models has been shown to be an effective solution to the problems of sample scarcity and imbalance. The Denoising Diffusion Probabilistic Model (DDPM) represents the latest generation of generative models. In comparison to alternative methods, the DDPM enhances realism and diversity through a multi-stage generation process. Nonetheless, the conventional DDPM is lacks the capacity to adapt to industrial sensor data. Consequently, this paper proposes an improved DDPM, named SS-DDPM, for industrial sensor signal data augmentation. SS-DDPM introduces a linear modulation strategy to dynamically adjust the feature distribution of the signal samples at each time step. This strategy helps to maintain the semantic balance between the noise intensity and local features during the denoising process, thereby augmenting the capability of the model to extract multilevel features. Furthermore,SS-DDPM incorporates a hierarchical convolution strategy to further enhance the capability of the model to generate complex signals. When generating sensor signals, the hierarchical convolution strategy ensures that the generated samples have highly accurate local details and maintain the integrity and consistency of the overall signal. The effectiveness of the method is evaluated using two public datasets. The results demonstrate that SS-DDPM exhibits greater adaptability to industrial sensor signals compared to state-of-the-art generative models. This study provides a new technological pathway for the acquisition of data and the intelligent decision-making in industrial production.