FaultDiffusion: Few-Shot Fault Time Series Generation with Diffusion Model

TL;DR

This work tackles the challenge of scarce fault time-series data by proposing FaultDiffusion, a diffusion-based framework that leverages abundant normal data to synthesize diverse and authentic fault sequences in few-shot settings. The method pretrains a transformer-backed diffusion model on normal data to learn $p_n(x)$, then adapts to fault data through a positive-negative difference adapter that models $p_f(x) = p_n(x) + \\Delta_\\theta(x)$, with a sliding-window attention mechanism to capture local anomalies. A diversity loss complements the standard denoising objective to mitigate mode collapse and promote broad fault-pattern coverage, yielding robust generation quality across multiple industrial datasets. Empirical results show state-of-the-art performance on generation metrics and significant improvements in downstream fault diagnosis accuracy and F1-scores, demonstrating practical impact for predictive maintenance and root-cause analysis in industrial settings.

Abstract

In industrial equipment monitoring, fault diagnosis is critical for ensuring system reliability and enabling predictive maintenance. However, the scarcity of fault data, due to the rarity of fault events and the high cost of data annotation, significantly hinders data-driven approaches. Existing time-series generation models, optimized for abundant normal data, struggle to capture fault distributions in few-shot scenarios, producing samples that lack authenticity and diversity due to the large domain gap and high intra-class variability of faults. To address this, we propose a novel few-shot fault time-series generation framework based on diffusion models. Our approach employs a positive-negative difference adapter, leveraging pre-trained normal data distributions to model the discrepancies between normal and fault domains for accurate fault synthesis. Additionally, a diversity loss is introduced to prevent mode collapse, encouraging the generation of diverse fault samples through inter-sample difference regularization. Experimental results demonstrate that our model significantly outperforms traditional methods in authenticity and diversity, achieving state-of-the-art performance on key benchmarks.

Figures (4)

• Figure 1: Top: Image anomalies are typically static and localized, whereas time-series faults are dynamic and global. Middle: The distribution of fault time series data is significantly different from that of normal time series data, and the distribution within fault classes is diverse, while the distribution of fault images is very similar to that of normal images.Bottom: In experimental comparisons, our model outperformed existing anomaly generation methods in fault generation results. Compared to all other methods, our model produces the most realistic anomalies.
• Figure 2: This is the architecture diagram of the model. The bottom part represents the pre-training process, which uses a large amount of normal data to pre-train the model to achieve normal data distribution parameter learning. The upper left corner represents the fine-tuning architecture, which achieves rapid domain adaptation through local adapters. The adapter architecture uses a local attention mechanism, as shown in the upper right figure.
• Figure 3: This is the tsne visualization of the fault samples generated by timegan. The samples are clustered and the feature space is degraded.
• Figure 4: Visualizations of the time series synthesized by FaultDiffusion and TimeGAN.