Target-specific Adaptation and Consistent Degradation Alignment for Cross-Domain Remaining Useful Life Prediction

AI summary

Overview

Research area: Unsupervised domain adaptation (UDA) for Remaining Useful Life (RUL) prediction in Prognostics and Health Management (PHM), applied to multi-sensor time series from turbofan engines.

Technical level: Advanced. The paper assumes familiarity with adversarial domain adaptation, auto-encoders, dynamic time warping, and clustering, though its core ideas are intuitive.

Scope: The paper proposes TACDA, a deep network that transfers RUL prediction knowledge from a labeled source operating condition to an unlabeled target condition while preserving target-specific information and aligning data within matching degradation stages.

What This Paper Is About

Data-driven RUL prediction normally assumes plenty of labeled failure data and that training and test data come from the same operating condition. In real industrial settings neither holds: failures are expensive to observe, and a model trained on one condition degrades badly on another due to domain shift. Existing UDA methods try to fix this by forcing target features to look globally like source features, but this can erase information unique to the target domain and can wrongly match, for example, early-stage target degradation with late-stage source degradation. TACDA addresses both problems by preserving target-specific information with a soft-DTW auto-encoder and by aligning source and target data stage by stage.

Key Contributions

1. TACDA framework. A new approach for cross-domain RUL prediction that jointly considers target-specific adaptation and consistent degradation alignment, rather than pursuing domain-invariant features alone.

2. Target-specific decoder with soft dynamic time warping (soft-DTW). An auto-encoder is used to preserve target-specific information while invariant representations are learned. Instead of mean squared error, soft-DTW is used as the reconstruction loss because it is robust to shifts and dilatation along the time dimension.

3. Clustering and pairing strategy. Source and target data are each clustered into degradation stages, then paired by similar status and aligned within the same stage, giving a second round of domain adaptation that avoids cross-stage misalignment.

4. Unsupervised target stage identification. Because target labels are absent, target degradation stages are recovered using k-means with soft-DTW as the distance metric, and clusters are ordered by variance (highest variance = terminal stage, lowest = sluggish stage).

Main Findings

• Best average performance across 12 cross-domain scenarios. TACDA achieved the best averaged RMSE and Score on the C-MAPSS benchmark, improving over the second-best method by more than 10% for RMSE and more than 32% for Score. Its average RMSE was 20.66 and average Score 3,077.

• Wins in most individual scenarios. TACDA outperformed all compared state-of-the-art methods in 7 scenarios for RMSE and 9 scenarios for Score, and was second-best in 4 scenarios for RMSE and 1 scenario for Score. It beat the second-best method by 55% in FD003→FD001, by 39% in FD002→FD004, and by 32% in FD003→FD002.

• Both components matter. In the ablation study, TACDA improved Score by around 35% over the variant without the clustering/pairing second round ("w/o C"), by around 39% over the variant without the target decoder ("w/o D_T"), and by around 72% over the variant with neither ("w/o C-D_T", equivalent to ADARUL). The unadapted "Source" variant performed worst, averaging RMSE 33.74 and Score 29,986.

• Soft-DTW beats MSE as a reconstruction loss. In the version without degradation alignment, soft-DTW outperformed MSE in 9 scenarios for RMSE and 10 scenarios for Score.

• Loss behaviour supports the design. Soft-DTW loss dropped sharply in early epochs of the first adaptation round and then stabilized. In the second round, initial values were higher for the sluggish and accelerated stages than the moderate stage, indicating that information from those stages was not well preserved in the first round, and the loss decreased across all stages during the second round.

• Sensitivity to the soft-DTW weight. Across λ values from 0.001 to 1 (tested in 10× intervals), performance was stable in RMSE but degraded in Score when λ was too small or too large. The recommended value is λ = 0.1. The initial soft-DTW magnitude exceeds the adversarial loss by a factor of 1,000 during training, motivating a weight no larger than 1.

• Not reported in the available text. The section on insufficient degradation data analysis ("IV-F") is truncated in the provided content, so its results cannot be summarized.

Methodology in Plain English

TACDA uses three building blocks: a source encoder, a target encoder, and a shared predictor.

1. Pretraining. The source encoder and the RUL predictor are trained on labeled source data using a mean squared error loss, giving the model basic knowledge of the RUL task. The target encoder is then initialized with the pretrained source encoder's weights.

2. First adaptation round (global). A target-specific adaptation module runs adversarial training: a domain discriminator tries to tell source features from target features while the target encoder tries to make them indistinguishable. At the same time, a target decoder reconstructs the target input from the target features, and the reconstruction error is measured with soft-DTW. This keeps target-only information from being wiped out by the alignment pressure.

3. Clustering into degradation stages. The source domain is split into three stages using its labels and a health index derived as a linear combination of critical sensor readings. The normalized RUL labels are discretized into K = 100 equally spaced bins, averaged per bin, smoothed with a Gaussian filter, and the second derivative of the averaged health index is computed. This yields three stages: sluggish (0%–33%), accelerated (33%–85%), and terminal (85%–100%). For the target domain, where no labels exist, k-means with soft-DTW distance groups the data, and each cluster's total sensor variance is used to assign the stage: highest variance is treated as terminal, lowest as sluggish.

4. Second adaptation round (stage-wise). Paired source and target groups from the same stage are fed back into the target-specific adaptation module in a loop (stage 1 → stage 2 → stage 3) to fine-tune the model. The final target encoder plus the predictor is then used to estimate RUL in the target domain.

Experimental setup. Evaluation uses the C-MAPSS benchmark: FD001 (100 training engines, 100 test engines, 1 operating condition, 1 fault mode), FD002 (260 / 259, 6 conditions, 1 fault mode), FD003 (100 / 100, 1 condition, 2 fault modes), and FD004 (249 / 248, 6 conditions, 2 fault modes). Though 21 sensors are deployed, preprocessing selects 14 sensors. All experiments were run five times and averaged. Batch size was 256, optimizer Adam, learning rate 5e-5 for the target encoder and domain discriminator, 5e-3 for the decoder (trained from scratch), and γ = 0.1 for soft-DTW. Models were built in PyTorch on an NVIDIA GeForce RTX A4000 GPU. Metrics are RMSE (equal weighting of early and late errors) and Score (a larger penalty for late predictions), both lower-is-better.

Why This Matters

Impact on research. The paper challenges the assumption that simply forcing target features to resemble source features is always beneficial. It shows that deliberately retaining target-specific information, and aligning only within matching degradation stages, produces measurable gains, and it offers soft-DTW as a concrete alternative to MSE for reconstruction in time-series domain adaptation.

Real-world applications:

• Aerospace: turbofan engine health monitoring, where engine failures account for a significant share of aircraft breakdowns and where engines run under differing conditions.
• Energy production: avoiding unplanned downtime of equipment, which the paper notes can cause substantial financial losses and disrupt supply.
• Manufacturing: scheduling proactive maintenance for machinery whose degradation spans prolonged periods.
• Fleet-wide deployment: transferring a model trained on one machine or fleet with labeled failures to another fleet with no failure labels.

Industry relevance. Operators rarely have labeled failures for every asset or operating condition, and manual stage labeling does not scale. TACDA's unsupervised target stage identification and label-free target adaptation reduce the annotation burden, and its stage-aware alignment targets a failure mode (mismatched degradation phases) that practitioners recognize in the field.

Future Directions

• Source-free domain adaptation. The authors identify TACDA's requirement for simultaneous access to source and target data during training as its current limitation, and state that future work will explore adaptation using only unlabeled target data plus a model initially trained in the source domain.
• Relaxing the fixed three-stage assumption. The paper defines exactly three degradation stages and empirically fixed boundaries; whether the number and boundaries of stages could be learned or adapted per dataset is left open.
• Reducing clustering cost. K-means with soft-DTW has a stated time complexity of O(I·N·K·M·L²); further efficiency work could make the second adaptation round cheaper on large fleets.
• Robustness under very limited degradation data. The paper includes a section on insufficient degradation data analysis, but its results are not available in the provided content, so how TACDA behaves when degradation records are scarce remains to be examined.

Target Audience

Researchers and graduate students working on domain adaptation, transfer learning, and time-series prognostics; industrial AI and PHM engineers who need RUL models that transfer across operating conditions with few or no target labels; and practitioners in aerospace, energy, and manufacturing evaluating stage-aware adaptation methods against simpler global-alignment baselines.

Authors’ abstract

Accurate prediction of the Remaining Useful Life (RUL) in machinery can significantly diminish maintenance costs, enhance equipment up-time, and mitigate adverse outcomes. Data-driven RUL prediction techniques have demonstrated commendable performance. However, their efficacy often relies on the assumption that training and testing data are drawn from the same distribution or domain, which does not hold in real industrial settings. To mitigate this domain discrepancy issue, prior adversarial domain adaptation methods focused on deriving domain-invariant features. Nevertheless, they overlook target-specific information and inconsistency characteristics pertinent to the degradation stages, resulting in suboptimal performance. To tackle these issues, we propose a novel domain adaptation approach for cross-domain RUL prediction named TACDA. Specifically, we propose a target domain reconstruction strategy within the adversarial adaptation process, thereby retaining target-specific information while learning domain-invariant features. Furthermore, we develop a novel clustering and pairing strategy for consistent alignment between similar degradation stages. Through extensive experiments, our results demonstrate the remarkable performance of our proposed TACDA method, surpassing state-of-the-art approaches with regard to two different evaluation metrics. Our code is available at https://github.com/keyplay/TACDA.