Degradation Modeling and Prognostic Analysis Under Unknown Failure Modes

TL;DR

This work tackles prognostic forecasting when multiple failure modes are present but unlabeled. It introduces an unsupervised FM identification pipeline using UMAP to embed multisensor degradation trajectories and DTW-based clustering to define FM groups, followed by a joint FM and RUL predictor with a monotonicity constraint to ensure RUL declines over time. The approach demonstrates FM discovery and FM-aware RUL prediction on the C-MAPSS turbofan dataset, showing improved accuracy and more interpretable degradation trends, including real-time FM probability estimates. By integrating FM diagnosis with RUL forecasting and enforcing monotonicity, the method offers practical benefits for maintenance planning in complex systems with unknown failure modes.

Abstract

Operating units often experience various failure modes in complex systems, leading to distinct degradation paths. Relying on a prognostic model trained on a single failure mode may lead to poor generalization performance across multiple failure modes. Therefore, accurately identifying the failure mode is of critical importance. Current prognostic approaches either ignore failure modes during degradation or assume known failure mode labels, which can be challenging to acquire in practice. Moreover, the high dimensionality and complex relations of sensor signals make it challenging to identify the failure modes accurately. To address these issues, we propose a novel failure mode diagnosis method that leverages a dimension reduction technique called UMAP (Uniform Manifold Approximation and Projection) to project and visualize each unit's degradation trajectory into a lower dimension. Then, using these degradation trajectories, we develop a time series-based clustering method to identify the training units' failure modes. Finally, we introduce a monotonically constrained prognostic model to predict the failure mode labels and RUL of the test units simultaneously using the obtained failure modes of the training units. The proposed prognostic model provides failure mode-specific RUL predictions while preserving the monotonic property of the RUL predictions across consecutive time steps. We evaluate the proposed model using a case study with the aircraft gas turbine engine dataset.

Figures (9)

• Figure 1: The proposed architecture is composed of three components: the failure mode identifier (red), failure mode predictor (green), and RUL predictor (orange). The high-dimensional sensor signals from $n$ training units are initially processed through UMAP into low-dimensional representations regarding the trajectories of each unit. Then, we employ time series clustering to obtain the failure mode of each training unit. The training unit's sensor signals are first processed via a sliding window method and then simultaneously plugged into both a feed-forward failure mode predictor (green) and an RUL predictor (orange). The joint loss, which combines failure mode classification loss and RUL prediction loss at each iteration, is backpropagated to update the neural network weights.
• Figure 2: UMAP projections of the C-MAPSS dataset are visualized through 2D scatter plots and corresponding 3D line plots (with the RUL as the third dimension). Within the 2D scatter plots, each point represents the low-dimensional representation (2D) of each record (one cycle on a unit). The colors in the 2D scatter plot indicate the RUL, with blue denoting a greater RUL and orange representing a smaller RUL. Each arrow indicates a failure trend. As for the 3D line plots, each distinct-colored line traces the low-dimensional trajectory of a unit over time.
• Figure 3: UMAP projections of C-MAPSS dataset on 3D. Each point represents the three-dimensional representation of each record (one cycle on a unit). Each color corresponds to one working condition.
• Figure 4: FD003: Trajectories of all units after time series clustering. Each line represents a unit. The color reveals the failure modes.
• Figure 5: FD003: Mean trajectories of two failure modes. Each color represents one failure mode. The solid line represents the mean trajectory, while the tube represents one standard deviation of trajectories of all units.