A diffusion-based generative model for system failure prediction in industrial predictive analytics

Predicting the time-to-failure (TTF), or remaining useful life (RUL), of engineering assets from degradation signals is central to prognostics and health management (PHM) in aerospace, manufacturing, and energy. Existing methods fall into two families: model-based approaches that impose a parametric degradation trajectory (e.g., Bayesian updating) and data-driven approaches that regress failure time on reduced-dimensional features (e.g., principal component analysis (PCA)-based lognormal regression). Both perform well under smooth, monotonic degradation but degrade under irregular, nonstationary, or multimodal dynamics, and neither provides calibrated uncertainty beyond a single point estimate. We propose a conditional diffusion-based generative model that learns the distribution of future degradation trajectories from a partially observed signal, so that thresholding each generated trajectory yields a full predictive TTF distribution. Two components adapt diffusion to degradation data: a Gaussian-process-correlated forward process that preserves temporal continuity, and a time-aware denoiser that couples a multiscale convolutional encoder–decoder network with a self-attention-based sequence model via cross-attention and dual embeddings of the diffusion step and physical time. On synthetic datasets of controlled difficulty and a public real-world PHM benchmark, the model consistently outperforms both baselines across point, distributional, and trajectory-level metrics, with the largest gains near the failure threshold and under heteroscedastic, multimodal degradation; a controlled comparison further isolates the Gaussian-process-correlated noise as the primary source of this improvement. These results support conditional diffusion as a robust, scalable approach for industrial predictive analytics.