While RAG has greatly enhanced LLMs, extending this paradigm to Time-Series Foundation Models (TSFMs) remains a challenge. This is exemplified in the Predictive Maintenance of the Pressure Regulating and Shut-Off Valve (PRSOV), a high-stakes industrial scenario characterized by (1) data scarcity, (2) short transient sequences, and (3) covariate coupled dynamics. Unfortunately, existing time-series RAG approaches predominantly rely on generated static vector embeddings and learnable context augmenters, which may fail to distinguish similar regimes in such scarce, transient, and covariate coupled scenarios. To address these limitations, we propose RAG4CTS, a regime-aware, training-free RAG framework for Covariate Time-Series. Specifically, we construct a hierarchal time-series native knowledge base to enable lossless storage and physics-informed retrieval of raw historical regimes. We design a two-stage bi-weighted retrieval mechanism that aligns historical trends through point-wise and multivariate similarities. For context augmentation, we introduce an agent-driven strategy to dynamically optimize context in a self-supervised manner. Extensive experiments on PRSOV demonstrate that our framework significantly outperforms state-of-the-art baselines in prediction accuracy. The proposed system is deployed in Apache IoTDB within China Southern Airlines. Since deployment, our method has successfully identified one PRSOV fault in two months with zero false alarm.
Figures & tables
Figure 1. Illustration of the PRSOV Scenario. (a) The PRSOV operates under strict pneumatic control logic where the target Manifold Pressure (MP) is primarily influenced by the Engine Speed (N2) and Intermediate Pressure (IP). (b) Example of real world PRSOV: data scarcity (one sample per flight), short transient sequences (18 points in 10 seconds), and complex covariate coupling.
Figure 2. Overall RAG Pipeline. Tail numbers are masked for privacy.
Figure 3. Tree-Structured Knowledge Base. Unlike vector stores, it preserves raw sequences following their physical hierarchy. Tail numbers are masked for privacy.
Figure 4. The Time-Series Native Retrieval Mechanism. It employs a bi-weighted coarse-to-fine strategy. Tail numbers are masked for privacy.
Figure 5. The agentic context augment process. The top-1 sample is used as an agent to self-calibrate the optimal number of context fragments ( k∗ ). Tail numbers are masked for privacy.
Dataset
Planes
Dim
Len
KB Samples
Query
Total
Min
Max
Total
B777L
30
3
18
17,573
91
1,050
7,380
B777R
30
3
18
17,573
91
1,050
7,380
A320L
41
3
18
112,932
1,707
3,244
30,013
A320R
41
3
18
112,932
1,707
3,244
30,013
Table 1. Statistics of the CSA-PRSOV Datasets. KB denotes Knowledge Base (2023-2024). Min/Max denotes the minimum and maximum sample counts per plane in KB.
Methods
B777L
B777R
A320L
A320R
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
Learning-based
DLinear
0.254
0.336
0.430
0.467
0.562
0.560
0.600
0.578
TimeMixer
0.467
0.452
0.580
0.538
0.509
0.512
0.570
0.538
Pyraformer
0.085
0.176
0.108
0.168
0.376
0.427
0.429
0.457
iTransformer
0.171
0.252
0.181
0.270
0.471
0.491
0.554
0.530
PatchTST
0.486
0.461
0.580
0.533
0.564
0.518
0.594
0.538
Table 2. Main performance comparison. The best results are highlighted in bold and the second best are underlined .
Configuration
RAG4CTS (ours)
Chronos-2
MSE
MAE
MSE
MAE
0 covariate
0.254
0.315
2.289
1.189
Only N2
0.249
0.319
1.958
0.972
Only IP
0.214
0.296
1.729
1.020
Full covariates
0.187
0.268
1.781
1.020
Table 3. Impact of covariate availability on forecasting performance. Full physical covariates yield the highest accuracy.
KB Strategy
B777L
B777R
MSE
MAE
MSE
MAE
Same Plane
0.103
0.205
0.216
0.314
B777
0.058
0.153
0.095
0.203
Full KB
0.057
0.154
0.096
0.217
Table 4. Impact of Knowledge Base scope on B777. “B777” indicates the same aircraft type as the query.
Retrieval Metric
MSE
MAE
Cosine + DTW
0.103
0.203
Cosine + Euclidean
0.085
0.186
Cosine + Matrix Profile
0.077
0.178
Euclidean + DTW
0.111
0.208
Euclidean + Matrix Profile
0.080
0.181
Matrix Profile + DTW
0.099
0.197
Table 5. Sensitivity of retrieval metrics. The hybrid Cosine + Matrix Profile strategy is the most effective.
Method Setup
Uniform
Wpoint
Wpoint⊙Wcov
MSE
MAE
MSE
MAE
MSE
MAE
k=0 (Chronos-2)
1.564
0.910
1.564
0.910
1.564
0.910
k=1
1.992
1.145
2.052
1.107
2.132
1.158
k=2
0.459
0.501
0.431
0.458
0.433
0.455
k=3
0.151
0.306
0.144
0.294
0.141
0.289
k=6
0.095
0.220
0.086
0.213
0.084
0.185
Table 6. Ablation on weighting schemes and context length ( k ) on B777. “ Wpoint ” for Point Weighting and “ Wcov ” for Covariate Weighting.
Figure 6. The distribution of k selected dynamically by our agentic splicing context augment.
Figure 7. The PRSOV predictive maintenance pipeline of China Southern Airlines.
Tail
Fault Type
Identified Date
Actual Fault
Type
B-2**8
PRSOV R
2023-10-15
Yes
Historical
B-2**0
PRSOV R
2023-11-04
Yes
Historical
B-7**5
PRSOV R
2023-12-26
Yes
Historical
B-2**1
PRSOV L
2024-04-01
Yes
Historical
B-2**0
PRSOV R
2024-08-22
Yes
Historical
B-2**9
PRSOV R
2024-09-11
Yes
Historical
Table 7. Predictive maintenance records from the CSA deployment (Starting Nov. 2025). Online visualizations are presented in Appendix A.7 . Tail numbers are masked for privacy.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8. Visual comparison between the baseline and our RAG approach.
Figure 9. Visualization of covariate ablation studies. We compare prediction trajectories under four settings: (1) No Covariates, (2) Only IP, (3) Only N2, and (4) Full Covariates.
Figure 10. Impact of Knowledge Base (KB) Scope on retrieval and prediction. From top to bottom: Scope 1 (Single Plane), Scope 2 (B777, same aircraft type), and Scope 3 (B777 + A320, cross-Type mixed aircraft).
Figure 11. Comparison of weighting strategies: Uniform (Unweighted), Temporal Weight Only, and our proposed Variable + Temporal strategy.
Figure 12. Visualization of prediction performance with varying context lengths (Top- K ). The plots display the concatenated input sequences and resulting predictions for K∈{1,2,3,4,5} .
Figure 13. The MSE of the B-27 PRSOV L in time order. PRSOV L fault notified and detected with our method on plane B-27.
Large-scale multivariate time series from heterogeneous IoT sensors demand accurate long-term forecasting for resource scheduling and predictive maintenance. While recent time series foundation models exhibit strong generalization, they rely on static parametric knowledge and lack dynamic access to external historical patterns during inference. Retrieval-Augmented Generation (RAG) offers a potential remedy, yet its application to time series forecasting is challenged by magnitude variations across heterogeneous sources and the mismatch between historical similarity and future consistency. We propose CrossRAG, a retrieval-augmented forecasting framework that integrates Shape-Aware Memory (SAM) with RevIN normalization for magnitude-robust shape-level retrieval, Future-Consistent Contrastive (FCC) learning to distinguish informative references from hard negatives with similar history but divergent futures, and Cross-Attention Temporal Fusion (CATF) to fuse retrieved historical--future reference pairs into the backbone's representations at the representation level. Experiments on seven public benchmarks show that CrossRAG consistently outperforms both parametric-only baselines and existing retrieval-augmented forecasting methods.
Yu Sun, Yuan Chang, Xiaohou Shi +1
Beijing University of Posts and Telecommunications Beijing, China · China Telecom Research Institute Beijing, China
Remaining Useful Life (RUL) prediction is essential for industrial predictive maintenance, yet many learning-based approaches rely on extensive feature engineering or large labeled datasets to train task-specific sequence models. In this work, we introduce a lightweight learning approach, in which we leverage a frozen pretrained time-series foundation model (TSFM) and combine it with a small regression head for RUL estimation from multivariate sensor streams. More specifically, we use Chronos-2 as a frozen backbone to extract context window features and train a lightweight regression neural network for RUL prediction. Experiments on real-world industrial sensor data from two device types show that Chronos-2 features consistently improve over recurrent, convolutional, Transformer-based, and gradient-boosting baselines under the same preprocessing and evaluation protocol. We further analyze the impact of context length and find that performance improves significantly with longer histories, indicating that TSFM representation offer a practical and data-efficient alternative for RUL estimation in industrial settings.
Amir El-Ghoussani, Michele De Vita, Ronald Naumann +1
Friedrich-Alexander-Universität Erlangen–Nürnberg (FAU), Germany · Nokia Solutions and Networks GmbH & Co. KG, Germany
Retrieval-augmented generation (RAG) enhances zero-shot time series (TS) forecasting by leveraging external knowledge bases, yet existing approaches overlook input-level relevance when fusing retrieved samples with the query. We argue that not all retrievals are equally useful, and irrelevant ones can degrade performance. To this end, we propose Cross-RAG, a zero-shot RAG-based forecasting framework that selectively attends to query-relevant retrieved samples via query--retrieval cross-attention. By modeling input-level relevance between the query and retrieved samples, Cross-RAG jointly incorporates three sources of information: 1) the query itself, 2) the retrieved samples, and 3) their relational interactions. In particular, this input-aware design enables Cross-RAG to remain stable as the number of retrieved samples k grows, whereas prior methods without cross-attention require careful k tuning to avoid degradation from irrelevant retrievals. Extensive experiments demonstrate that Cross-RAG consistently improves zero-shot forecasting performance across multiple TSFM backbones and various RAG methods, with additional analyses confirming its effectiveness across various retrieval scenarios. Code is available at https://github.com/seunghan96/cross-rag/.