Aura: Universal Multi-dimensional Exogenous Integration for Aviation Time Series

Jiafeng Lin; Mengren Zheng; Simeng Ye; Yuxuan Wang; Huan Zhang; Yuhui Liu; Zhongyi Pei; Jianmin Wang

cs.LGMar 5, 2026

Aura: Universal Multi-dimensional Exogenous Integration for Aviation Time Series

Authors: Jiafeng Lin, Mengren Zheng, Simeng Ye, Yuxuan Wang, Huan Zhang, Yuhui Liu, Zhongyi Pei, Jianmin Wang

Organizations: School of Software, BNRist, Tsinghua University, Beijing, China · Hongshen College, Chongqing University, Chongqing, China · School of Software, Tsinghua University, Beijing, China · China Southern Airlines, Guangzhou, China

Abstract

Time series forecasting has witnessed an increasing demand across diverse industrial applications, where accurate predictions are pivotal for informed decision-making. Beyond numerical time series data, reliable forecasting in practical scenarios requires integrating diverse exogenous factors. Such exogenous information is often multi-dimensional or even multimodal, introducing heterogeneous interactions that unimodal time series models struggle to capture. In this paper, we delve into an aviation maintenance scenario and identify three distinct types of exogenous factors that influence temporal dynamics through distinct interaction modes. Based on this empirical insight, we propose Aura, a universal framework that explicitly organizes and encodes heterogeneous external information according to its interaction mode with the target time series. Specifically, Aura utilizes a tailored tripartite encoding mechanism to embed heterogeneous features into well-established time series models, ensuring seamless integration of non-sequential context. Extensive experiments on a large-scale, three-year industrial dataset from China Southern Airlines, covering the Boeing 777 and Airbus A320 fleets, demonstrate that Aura consistently achieves state-of-the-art performance across all baselines and exhibits superior adaptability. Our findings highlight Aura's potential as a general-purpose enhancement for aviation safety and reliability.

Figures & tables

Figure 1. Illustration of diverse exogenous information influencing aircraft component anomaly detection. Beyond the time series directly associated with the equipment, this detection process incorporates multiple factors, including correlated sensor signals, aircraft register, airport location, and dynamic environmental takeoff conditions, which collectively influence the temporal dynamics, necessitating a unified integration for accurate health monitoring.

Figure 3. Overall architecture of Aura, a multi-dimensional exogenous integration framework. Aura integrates static attributes encoded by LLMs and geospatial embeddings, and prepends them to endogenous tokens. Two cross-attention layers fuse exogenous series via residual connections to regulate the strength of integration. A Mixture of Experts (MoE) module leverages LLM-generated future insights from dynamic events to guide time series forecasting.

Model		Aura	DUET	CrossLinear	Timer-XL	TimeXer	TiDE	TimeLLM	GPT4TS	iTransformer	DLinear	PatchTST	TimesNet	Autoformer	LightGBM
Model		(Ours)	( ( 2025a ) )	( ( 2025 ) )	( ( 2024c ) )	( ( 2024c ) )	( ( 2023 ) )	( ( 2023 ) )	( ( 2023 ) )	( ( 2023 ) )	( ( 2023 ) )	( ( 2022 ) )	( ( 2022 ) )	( ( 2021 ) )	( ( 2017 ) )
Boeing	MSE $\downarrow$	0.075	0.157	0.109	0.140	0.093	0.318	0.358	0.158	0.086	0.203	0.122	0.113	0.823	0.145
777	MAE $\downarrow$	0.180	0.250	0.225	0.232	0.197	0.401	0.427	0.240	0.184	0.286	0.224	0.208	0.711	0.220
Left	TAR $\uparrow$	0.625	0.500	0.500	0.500	0.500	0.375	0.375	0.500	0.500	0.500	0.500	0.500	0.375	0.500
Boeing	MSE $\downarrow$	0.086	0.103	0.117	0.118	0.101	0.308	0.413	0.100	0.090	0.170	0.107	0.106	0.668	0.096
777	MAE $\downarrow$	0.182	0.202	0.230	0.213	0.201	0.396	0.466	0.189	0.189	0.262	0.207	0.199	0.636	0.182

Table 1. Results of time series forecasting and anomaly detection on China Southern Airlines datasets. Bold indicates the best performance. True Alert Rates (TAR) are calculated under a constrained false alarm rate of 5% to ensure practical reliability.

Model	Aura		iTransformer		TimeXer		TimerXL		TiDE		CrossLinear		DUET		PatchTST		TimesNet		DLinear		Autoformer
Metric	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
BE	0.338	0.239	0.356	0.250	0.367	0.231	0.608	0.413	0.411	0.288	0.417	0.269	0.463	0.333	0.434	0.336	0.481	0.351	0.392	0.280	0.829	0.549
DE	0.266	0.322	0.268	0.324	0.301	0.328	0.921	0.633	0.392	0.399	0.429	0.416	0.336	0.351	0.302	0.323	0.291	0.333	0.375	0.387	1.080	0.695
FR	0.346	0.187	0.371	0.205	0.361	0.197	0.573	0.350	0.457	0.299	0.409	0.244	0.420	0.238	0.420	0.219	0.408	0.236	0.411	0.273	0.763	0.480
NP	0.183	0.227	0.200	0.253	0.197	0.241	1.194	0.799	0.252	0.287	0.549	0.554	0.313	0.367	0.466	0.478	0.231	0.296	0.237	0.273	0.819	0.657
PJM	0.069	0.165	0.078	0.166	0.082	0.179	0.339	0.323	0.098	0.215	0.098	0.193	0.080	0.176	0.109	0.167	0.103	0.177	0.091	0.206	0.334	0.441

Table 2. Exogenous forecasting results on the EPF benchmark with future exogenous series. We set the input length to 168 and the prediction length to 24, following TimeXer ( Wang et al., 2024c ) experimental setup. AVG means the average results from all five datasets.

Figure 4. Statistical distribution of gated weights for history and future exogenous series.

Design		Boeing 777 Left			Boeing 777 Right			Airbus A320 Left			Airbus A320 Right
Design		MSE	MAE	TAR	MSE	MAE	TAR	MSE	MAE	TAR	MSE	MAE	TAR
Aura		0.075	0.180	0.625	0.086	0.182	0.625	0.048	0.138	0.345	0.063	0.146	0.414
Data	w/o Static Attributes	0.107	0.213	0.500	0.094	0.200	0.625	0.050	0.141	0.345	0.152	0.152	0.310
	w/o Dynamic Events	0.085	0.190	0.500	0.096	0.209	0.625	0.050	0.139	0.345	0.064	0.147	0.310
	w/o Exogenous Series	0.542	0.528	0.375	0.422	0.451	0.500	0.055	0.145	0.276	0.068	0.154	0.241
Arch	Token Concat	0.094	0.196	0.500	0.094	0.194	0.625	0.054	0.142	0.276	0.064	0.149	0.310

Table 3. Ablation Results on the design of Aura. Data refers to the removal of three distinct types of exogenous information. Arch represents architectural modifications, where the first three experiments employ uniform modeling for all exogenous information, and the final experiment evaluates the removal of the gated residual module.

Figure 5. Deployed model inference visualization within the health monitoring system of China Southern Airlines.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Samples	Vars	En. Input	Ex. Input	Output	En. Description
Boeing 777 Left	(28486, 4181, 7827)	3	6	24	18	Bleed Air Manifold Pressure
Boeing 777 Right	(28486, 4181, 7827)	3	6	24	18	Bleed Air Manifold Pressure
Airbus A320 Left	(97011, 16401, 30880)	3	6	24	18	Precooler Pressure
Airbus A320 Right	(97011, 16401, 30880)	3	6	24	18	Precooler Pressure
NP	(36500, 5219, 10460)	3	168	192	24	Nord Pool Electricity Price
PJM	(36500, 5219, 10460)	3	168	192	24	Pennsylvania-New Jersey-Maryland Electricity Price

Appendix

Table 4. Statistics of the datasets used in our experiments.

Figure 6. Visualization results of Aura, TimeLLM and TimeXer on Boeing 777 Datasets.

Data	MSE	MAE	TAR
Boeing 777 Left	$0.075\pm 0.004$	$0.180\pm 0.004$	$0.625\pm 0.000$
Boeing 777 Right	$0.086\pm 0.005$	$0.182\pm 0.010$	$0.625\pm 0.000$
Airbus A320 Left	$0.048\pm 0.001$	$0.138\pm 0.000$	$0.345\pm 0.000$
Airbus A320 Right	$0.063\pm 0.004$	$0.146\pm 0.002$	$0.414\pm 0.000$

Appendix

Table 5. Error Bar Analysis of Aura. The results are obtained from three random runs.

Figure 7. Visualization Prior to PRSOV Anomaly.

Figure 8. The Prompt Design of Aura for dynamic contextual events.

Exogenous Time Series Forecasting

See all 14 papers

TiWeaver: Unified Temporal Dynamics Modeling via Contextual Patching
Jun 2, 2026Zhe Li, Jindong Tian, Hao Miao +3Irregular Time-Series Modeling Multivariate Time Series Forecasting
East China Normal University Shanghai, China · Hong Kong Polytechnic University Hong Kong, China · Aalborg University Aalborg, Denmark
Aurora-X: Built for Extreme Time Series Forecasting
Sep 25, 2026Xingjian Wu, Chenjuan Guo, Xiangfei Qiu +6Time Series Forecasting Time-Series Foundation Models
East China Normal University
SDMixer: Sparse Dual-Mixer for Time Series Forecasting
Feb 27, 2026Xiang AoMultivariate Time Series Forecasting Time Series Forecasting
School of Software Engineering, Beijing Jiaotong University, Beijing, China

Jun 2, 2026J/K move · Enter open · S save

TiWeaver: Unified Temporal Dynamics Modeling via Contextual Patching

Zhe Li, Jindong Tian, Hao Miao, Zhi Lei+2

East China Normal University Shanghai, China · Hong Kong Polytechnic University Hong Kong, China · Aalborg University Aalborg, Denmark

Irregular Time-Series Modeling Multivariate Time Series Forecasting

Open paper AskPDF arXiv

Multivariate time series forecasting plays a critical role in real-world applications, including weather prediction, stock analysis, and health monitoring. Due to the diversity of data sources, time series exhibit diverse temporal dynamics, often accompanied by various irregularities such as missing values and non-uniform sampling frequencies. Such irregularities lead to complex and asynchronous temporal dependencies across channels. Thus, a single model with a fixed patching scheme often fails to adapt well to diverse multivariate time series, hindering accurate forecasting. In this paper, we propose TiWeaver, a unified framework designed to handle temporal dynamics and fine-grained inter-channel dependencies adaptively. Specifically, we introduce a Graph-Guided Adaptive Tokenizer (G

^2

AT) that divides time series into high contextually coherent patches by jointly considering temporal density and representation consistency. In addition, we propose a Fine-grained Asynchronous Dependency Extractor (FADE), which is designed to model fine-grained asynchronous inter-channel dependencies while incorporating long-term historical dependencies. We evaluate TiWeaver on 12 real-world time series datasets, where it achieves state-of-the-art performance, outperforming existing methods up to 25%. These results demonstrate its robustness and effectiveness across diverse domains and data characteristics.

Abstract

Exogenous Time Series Forecasting

Explore similar work

TiWeaver: Unified Temporal Dynamics Modeling via Contextual Patching

Aurora-X: Built for Extreme Time Series Forecasting

SDMixer: Sparse Dual-Mixer for Time Series Forecasting