With the widespread application of space traveling wave tube amplifiers (TWTAs) as core final-stage amplification devices in satellite communication systems, their operational reliability directly affects satellite signal transmission capability, and any failure may even lead to service interruption. Currently, fault statistics for space TWTAs are recorded in text form, encompassing a large number of fault phenomena, causes, and corresponding solutions. However, as the volume of fault events continues to grow, traditional analysis methods relying on manual statistics and empirical judgment suffer from low efficiency and lack of knowledge reasoning capabilities, making it difficult to comprehensively analyze multi-dimensional fault information. Although artificial intelligence technologies, particularly knowledge graphs (KGs) and large language models (LLMs), have demonstrated substantial potential in fault diagnosis, research on effectively integrating these two approaches and applying them to the specific fault diagnosis of space TWTAs remains unexplored. Therefore, how to construct an intelligent diagnosis framework capable of automatically understanding natural language fault descriptions and performing knowledge reasoning has become a key challenge in improving the troubleshooting efficiency and decision support capabilities for space TWTAs.
In a recent study published in Space: Science & Technology , the research team from the School of Electronic Science and Engineering, University of Electronic Science and Technology of China, proposed a fault diagnosis framework for space TWTAs based on the integration of knowledge graphs and large language models. The study constructs an intelligent diagnosis framework comprising three modules: a named entity recognition (NER) module, a knowledge graph module, and a graph retrieval module. The Qwen2-7B-Instruct large language model is fine-tuned using QLoRA technology to automatically extract three categories of entities—stage, component, and fault—from user natural language inputs. Five entity types and five relationship types are defined to construct the space TWTA fault knowledge graph. A two-step self-instruction method is adopted to align the LLM with the knowledge graph, enabling automatic translation from natural language to graph query language, and fault causes and treatment solutions are obtained through a retrieval-augmented generation (RAG) mechanism. In real-scenario testing, the framework provided responses to 92 out of 100 fault descriptions, of which 87 were adopted by domain experts, achieving an adoption rate of 94.56%, whereas the non-fine-tuned baseline model attained only a 23% adoption rate. The framework maintained semantic consistency in responses to identical questions, while the baseline model exhibited instability. In interference experiments introducing three types of noise, the framework maintained an accuracy above 95% and effectively rejected irrelevant data, demonstrating robust stability; on a larger test set, it still achieved an accuracy of 88.13%, with an average response time of 19 seconds per query, validating its effectiveness as a decision-support tool. This research provides an efficient technical solution for intelligent fault diagnosis of space TWTAs, offering significant engineering application value for enhancing the on-orbit reliability and operational maintenance efficiency of spacecraft.
First, this paper focuses on the intelligent demand for fault diagnosis of space traveling wave tube amplifiers (TWTAs) and proposes a fault diagnosis framework that integrates knowledge graphs with large language models. Space TWTAs, by virtue of their high power, high efficiency, and ultra-wideband advantages, have long served as the core final-stage amplification devices in satellite communication systems, and their operational status directly affects satellite signal transmission capability. With the continuous increase in the number of on-orbit applications, text data on fault events are accumulating steadily. Traditional analysis methods relying on manual statistics and empirical judgment suffer from low efficiency and are increasingly unable to cope with the growing complexity of fault knowledge bases. As illustrated in Fig. 1, the proposed framework consists of three modules—the named entity recognition (NER) module, the knowledge graph module, and the graph retrieval module—which collectively implement retrieval-augmented generation (RAG) functionality. The user inputs a fault description in natural language; after the NER module extracts relevant entities, the graph retrieval module automatically generates graph query language to perform retrieval in the knowledge graph; the retrieved contextual information, together with the original user query, constitutes the prompt, and a general-purpose large language model subsequently generates the diagnostic response. The requirements analysis, graph retrieval, information judgment, and answer generation of this framework are all accomplished by a single large language model, enhancing the coupling among modules and the overall stability of the framework.
Second, this paper elaborates on the specific construction and implementation methods of the three core modules in the framework. In the named entity recognition (NER) module, the research team constructs three types of corpora—component corpus, stage corpus, and fault description corpus—for data alignment. As shown in Fig. 2, the implementation process of NER using the large language model comprises three steps: data augmentation, data pair preparation, and model training. Data augmentation enriches the diversity of training data by replacing component name aliases, rephrasing fault descriptions, and adjusting sentence structures. As illustrated in Fig. 3, taking one raw data entry as an example, multiple augmented data entries are obtained through the above augmentation methods; the component name, stage, and fault description information are extracted to form the ASSISTANT, which is combined with the PROMPT to constitute training data pairs. Model training employs QLoRA technology to fine-tune Qwen2-7B-Instruct, freezing the 4-bit quantized pre-trained model and updating only the parameters of the low-rank adaptation modules, thereby achieving efficient fine-tuning under limited computational resources. In the knowledge graph module, as shown in Fig. 4, five entity types—stage, component, fault, analysis, and correction—and five relationship types—causal relationship, corrective relationship, occurrence relationship, involvement relationship, and composition relationship—are defined, comprehensively restoring details of fault events including occurrence time, involved components, phenomenon descriptions, cause analysis, and corrective solutions. In the graph retrieval module, as illustrated in Fig. 5, a two-step self-instruction method is adopted to generate training data pairs from natural language to graph query language, enabling automatic query translation through alignment between the large language model and the knowledge graph.
Finally, this paper validates the performance of each module and the overall framework through a series of experiments. The test results of the named entity recognition (NER) module, as shown in Table 3, demonstrate that the Qwen2-7B-Instruct model fine-tuned with QLoRA achieves F1-scores of 97.56%, 79.99%, and 90.00% for the three entity types—stage, component, and fault, respectively, with a macro-average F1-score of 89.18%, significantly outperforming the non-fine-tuned version and GPT-4o. The graph retrieval module attains a hit rate of 89% on 100 natural language queries. In real-scenario testing, as presented in Table 4, the framework provides responses to 92 out of 100 fault descriptions, of which 87 are adopted by domain experts, corresponding to an adoption rate of 94.56%; in contrast, the non-fine-tuned baseline model, although generating 100 responses, achieves only 23 adoptions. The framework maintains semantic consistency in responses to identical questions, whereas the baseline model exhibits persistent inconsistency. In interference experiments, the framework achieves accuracies of 97.22%, 95.00%, and 100% under three types of noise—spelling errors, distracting sentences, and irrelevant data, respectively—and effectively rejects irrelevant queries, demonstrating excellent robustness. In larger-scale experiments involving 1,618 test data entries, the framework still maintains an accuracy of 88.13%. The average response time per query is 19 seconds, validating its effectiveness as a decision-support tool. This framework provides an efficient technical solution for intelligent fault diagnosis of space traveling wave tube amplifiers.
Space Science & Technology
Fault Diagnosis Framework for Space TWTAs by Combining the Knowledge Graph and Large Language Mode
30-Jun-2026