Key Technologies of a Domain-Specific Large Model for Telescope Array Fault Diagnosis Based on the Fusion of Low-Rank Adaptation and Retrieval-Augmented Generation

doi:10.11871/jfdc.issn.2096-742X.2026.04.003

Abstract

Abstract:

[Objective] To address the bottleneck issues of obscure specialized terminology, semantic heterogeneity of multi-source fault information, and scarcity of domain knowledge in the fault diagnosis of telescope arrays, this study aims to construct a high-precision and high-reliability intelligent diagnosis model to enhance the operation and maintenance efficiency and system stability of large-scale scientific facilities. [Methods] This paper proposes a technical framework for a domain-specific large model for fault diagnosis based on the fusion of Low-Rank Adaptation (LoRA) and Retrieval-Augmented Generation (RAG). Taking the DeepSeek-R1-Qwen3-8B large model as the backbone, the framework utilizes a constructed high-quality instruction fine-tuning dataset for telescope arrays and employed LoRA technology to efficiently adapt and enhance domain diagnostic reasoning capabilities. Simultaneously, a RAG mechanism is introduced to establish an external knowledge base based on a vector database, enabling the real-time retrieval and context injection of fault-related knowledge fragments to strengthen the model's coverage of long-tail knowledge and implicit causal logic. [Results] Experimental verification using the DAocheng Radio Telescope (DART) as a case study demonstrates that the fused model significantly outperforms the original base model in BERTScore evaluation metrics, with an F1 score improvement of 22.6%. The model exhibits superior diagnostic accuracy, reasoning stability, and output structuredness when handling complex faults. [Conclusion] The synergy of LoRA and RAG effectively addresses the “hallucination” phenomenon and logical gaps faced by general large models in domain-specific applications. This study validates the effectiveness of the proposed technical framework, providing a feasible technical solution and reference for constructing intelligent fault diagnosis systems for radio telescope arrays and other complex equipment operation and maintenance scenarios.

Key words: telescope arrays, fault diagnosis, low-rank adaptation, retrieval-augmented generation

LI Jiashan, DENG Li. Key Technologies of a Domain-Specific Large Model for Telescope Array Fault Diagnosis Based on the Fusion of Low-Rank Adaptation and Retrieval-Augmented Generation[J]. Frontiers of Data and Computing, 2026, 8(4): 30-41, https://cstr.cn/32002.14.jfdc.CN10-1649/TP.2026.04.003.

Figures/Tables 12

Fig.1

Table 1

Fig.2

Table 2

Fig.3

Table 3

Table 4

Fig.4

Table 5

Fig.5

Table 6

Comparison of model inference results for polarization response anomalies"

维度	问题（用户）	原始模型	LoRA+RAG 模型
现象归纳	在分析数字终端观测数据时,发现32号天线的极化响应异常。涉及状态数据: 前端电压=7.5 V（在正常范围[7 V，8 V]内）；前端电流=220 mA（在正常范围[150 mA,300 mA]内）；变频模块电压=5.01 V（正常）；但终端数据分析显示左右旋极化分量相位差错位180°（极化反）。可能的原因是什么?应该如何解决?	32号天线在观测过程中出现极化响应异常，具体表现为左右旋极化分量的相位差偏离正常值180°，导致极化反转。同时,前端电压、电流及变频模块电压均在正常范围内，未见供电异常	数字终端观测数据中32号天线极化响应异常，表现为左右旋极化分量相位差错位180°（极化反）。涉及状态数据: 前端电压=7.5 V（正常）；前端电流=220 mA（正常）；变频模块电压=5.01 V（正常）
原因推断		由于供电参数正常，问题可能出在信号处理环节。变频模块电压虽然正常，但变频模块通常涉及射频信号的处理，如果变频模块的输出相位调整出现错误,可能导致极化反转。另外，极化反转也可能与信号的调制或解调过程有关，比如在信号处理芯片中相位校准错误，或者天线馈电网络的相位匹配出现问题	前端供电与变频模块参数均在正常范围，但极化响应异常,指向前端输入极化接反或变频端极化映射配置错误导致相位反转
排查建议		需要具体检查变频模块的输出相位设置，确认是否有错误配置。同时，检查信号处理链路中的相位校准参数，确保没有漂移。如果涉及天线馈电网络,需要量其相位匹配情况,排除馈电网络的故障	1）检查前端与变频模块的极化接线,确认连接无误；2）核对变频模块的极化映射配置，确保与前端极化方向一致；3）更换线缆后重新执行极化标定，更新相关配置参数并复核相位一致性

Table 6

Table 7

Comparison of model inference results for pointing offset errors"

维度	问题（用户）	原始模型	LoRA+RAG 模型
现象归纳	在 DART 望远镜调试期间，发现部分天线指向误差较大，扫描方向图显示波束中心偏离太阳理论坐标约2°-3°。这种情况严重影响了高精度成像。根据 DART 三轴座架系统的特点,这种明显的指向偏移最可能的根本原因是什么?应该采取什么标准化的标校流程来将误差控制在指标要求（0.5°）以内?	用户描述的现象很明确: 天线指向存在明显偏差，波束中心偏离理论位置2-3度。这个误差在调试阶段出现，且影响成像质量，说明问题比较严重。需要强调的是,这种偏差是系统性的，因为扫描方向图显示整体偏离，而不是随机误差	天线扫描方向图显示波束中心偏离太阳理论坐标约2°-3°,指向误差明显，严重影响高精度成像
原因推断		根据 DART 三轴座架系统的特点，可能的原因包括: 机械结构误差（如轴间间隙、装配误差）、控制算法问题（如伺服环增益设置不当）、环境因素（如温度变化导致的结构变形）等。考虑到误差范围（2-3度）远超指标要求（0.5度），最可能的是机械系统存在累积误差，或者控制参数未正确配置	根据背景信息，三轴编码器零点未准确标定是导致指向偏移的主要原因。三轴编码器（赤经、赤纬、方位）的零点误差会导致天线整体指向偏移，且误差范围与用户描述的情况相符
排查建议		标准化的标校流程应该包括:首先进行静态标校，检查各轴的零位和重复性；然后进行动态标校，测试伺服系统的响应特性；接着进行全天空扫描,对比实际指向与理论模型的差异；最后通过参数优化和补偿算法来修正误差	1. 漂移扫描: 使用漂移扫描数据获取天线在不同方向上的指向偏差，分析误差来源。2. 重新标定三轴编码器零点: 根据漂移扫描结果,调整三轴编码器的零点位置,确保其准确。3. 更新控制软件: 将校正后的数据输入控制软件，使系统能够自动补偿指向误差。4. 验证校正效果: 通过多次扫描和对比，确保指向误差降至0.5°以内。5.定期维护: 建议定期进行漂移扫描和零点校准，以防止误差累积

Table 7

References 24

[1]	CAIAZZO M, LABATE M G, TIRONE L, et al. The latest Aperture Array Verification System AAVS 3 for the world’s largest radio telescope[C]// 2024 IEEE International Symposium on Antennas and Propagation and INC/USNC-URSI Radio Science Meeting (AP-S/INC-USNC-URSI). Piscataway: IEEE, 2024: 297-298.
[2]	张荣玉, 阎敬业, 武林, 等. 综合孔径射电望远镜冗余基线可见度噪声相关性对灵敏度的影响[J]. 空间科学学报, 2021, 41(5): 808-818.
[3]	路光, 王冰, 陈耀, 等. 稻城太阳射电望远镜(DSRT)天线遮挡效应仿真[J]. 空间科学学报, 2022, 42(2): 294-305.
[4]	DING N, QIN Y J, YANG G, et al. Parameter-efficient fine-tuning of large-scale pre-trained language models[J]. Nature Machine Intelligence, 2023, 5(3): 220-227. doi: 10.1038/s42256-023-00626-4
[5]	JIAN C X, PENG Y H, MO G P, et al. Open-set domain generalization for fault diagnosis through data augmentation and a dual-level weighted mechanism[J]. Advanced Engineering Informatics, 2024, 62: 102703. doi: 10.1016/j.aei.2024.102703
[6]	秦一凡, 付翔, 张智星, 等. 基于LoRA微调与RAG融合的煤矿专业大模型应用关键技术[J]. 工矿自动化, 2025, 51(8): 34-42.
[7]	唐雷, 陈子逸, 梁锶翰, 等. Llama2-70b模型的微调技术及其在材料领域的应用研究[J]. 数据与计算发展前沿(中英文), 2025, 7(1): 163-174.
[8]	吴建华, 刘振宇, 曾睿, 等. 基于LoRA微调大模型在网络安全等级保护测评报告质量检测中的应用研究[J]. 数据与计算发展前沿(中英文), 2025, 7(6): 111-123.
[9]	MIAO R, WU T, ZHANG Z. Graph RAG-based Fault Diagnosis for Train Bogies Using Knowledge Graphs and Large Language Model[J]. Knowledge-Based Systems, 2025: 114855.
[10]	PENG C, PENG J, WANG Z, et al. Adaptive fault diagnosis of railway vehicle on-board controller with large language models[J]. Applied Soft Computing, 2025: 113919.
[11]	ZHANG H, ZHAO Y, SUN B, et al. Large Language Model Based Intelligent Fault Information Retrieval System for New Energy Vehicles[J]. Applied Sciences, 2025, 15(7): 4034. doi: 10.3390/app15074034
[12]	LIN L, ZHANG S, FU S, et al. FD-LLM: Large language model for fault diagnosis of complex equipment[J]. Advanced Engineering Informatics, 2025, 65: 103208. doi: 10.1016/j.aei.2025.103208
[13]	LI D, PANG Z, CHEN Y, et al. FD-MVLLM: Fault diagnosis based on multimodal vibration data and large language model for bearing system[J]. Mechanical Systems and Signal Processing, 2025, 239: 113226. doi: 10.1016/j.ymssp.2025.113226
[14]	WANG H G, PING B W, WANG S, et al. LoRA-Flow: Dynamic LoRA Fusion for Large Language Models in Generative Tasks[C]// Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, Vol 1:Long Papers. Bangkok, Thailand: Association for Computational Linguistics, 2024: 12871-12882.
[15]	ZHONG H, YU S, SHI M, et al. GAFT: a small-sample bearing fault diagnostic method based on LoRA fine-tuning and GAN-TrAdaBoost[J]. Structural Hea-lth Monitoring, 2025.
[16]	ZHANG Y, PENG Y, TU C, et al. Exploring efficient optimization techniques in online retrieval-augmented generation application[C]// Proceedings of the 24th ACM/IEEE Joint Conference on Digital Libraries. New York: ACM, 2024: 1-11.
[17]	BAI J Z, BAI S, CHU Y F, et al. Qwen technical report[R/OL]. 2023 [2025-12-15]. https://arxiv.org/abs/2309.16609.
[18]	WEI J, WANG X Z, SCHUURMANS D, et al. Chain-of-thought prompting elicits reasoning in large language models[C]///KOYEJO S, MOHAMED S, AGARWAL A, et al. Advances in Neural Information Processing Systems 35. Curran Associates, Inc., 2022: 24824-24837.
[19]	HU E J, SHEN Y, WALLIS P, et al. LoRA: Low-Rank Adaptation of Large Language Models[C]// Proceedings of the 10th International Conference on Learning Representations (ICLR). Virtual Event: OpenReview.net, 2022.
[20]	XIAO S T, LIU Z, ZHANG P T, et al. C-Pack: packed resources for general Chinese embeddings[C]// Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. New York: Association for Computing Machinery, 2024: 641-649.
[21]	JOHNSON J, DOUZE M, JÉGOU H. Billion-scale similarity search with GPUs[J]. IEEE Transactions on Big Data, 2021, 7(3): 535-547. doi: 10.1109/TBDATA.2019.2921572
[22]	ZHANG T, KISHORE V, WU F, et al. BERTScore: Evaluating Text Generation with BERT[C]// Proceedings of the 8th International Conference on Learning Representations (ICLR). Addis Ababa: OpenReview.net, 2020.
[23]	吐力帕尔·吐鲁洪. DSRT三轴低频射电望远镜指向误差研究[D]. 北京: 中国科学院大学, 2023: 1-6.
[24]	吐力帕尔·吐鲁洪, 阎敬业, 武林, 等. DSRT三轴低频天线指向误差批量测量和校正方法[J]. 天文学报, 2024, 65(3): 54-65.

统计维度	类别	样本数（条）	占比（%）
子系统分布	天线分系统	353	29.4
	接收链路分系统	341	28.4
	数字分系统	237	19.8
	辅助设施及其他	269	22.4
故障类型	参数异常	652	54.3
	物理损坏	412	34.3
	逻辑错误	136	11.3
总计	全集	1,200	100.0

参数名称	参数取值
LoRA的秩	16
LoRA缩放系数	32
学习率	2e-4
批处理大小	2
梯度累积步数	4
优化器	AdamW (8-bit)
权重衰减系数	0.01
学习率调度策略	Linear

模型	Precision	Recall	F1 Score
基线模型	0.5780	0.6787	0.6239
RAG-only	0.6071	0.7243	0.6602
LoRA-only	0.7538	0.7162	0.7333
RAG+LoRA	0.7710	0.7606	0.7648

Top-K	Precision	Recall	F1 Score	平均推理耗时
1	0.7643	0.7351	0.7486	3.43 s
3	0.7650	0.7478	0.7552	3.99 s
5	0.7710	0.7606	0.7648	4.09 s
7	0.7641	0.7518	0.7570	4.17 s
9	0.7602	0.7508	0.7546	4.22 s