计算机科学
判别式
稳健性(进化)
图形
桥接(联网)
语言模型
数据挖掘
代表(政治)
可识别性
人工智能
故障检测与隔离
水力发电
实时计算
模式识别(心理学)
断层(地质)
遮罩(插图)
分布式计算
一套
软件部署
标识
DOI:10.1088/2631-8695/aea5e5
摘要
Abstract Hydropower unit rotating components operate under variable load regulation, frequent condition switching, and coupled hydraulic, mechanical, and electromagnetic excitation. Multi-source vibration measurements acquired at distributed structural locations are non-stationary, and the same health state may exhibit different signal patterns across operating conditions. These characteristics reduce the transferability of purely discriminative diagnostic models and make it difficult to associate a predicted label with physically traceable sensing evidence. This study develops a multi-source sensing sparse-masked heterogeneous graph-constrained large language model framework for cross-condition fault diagnosis. Local condition-adaptive time–frequency encoding extracts node-level signal representations, while a heterogeneous sensing graph organizes spatial adjacency, structural transmission, and signal-correlation relations. Type-aware message passing and condition-gated aggregation produce topology representations, which are aligned with local time–frequency representations through optimal transport and maximum mean discrepancy constraints. Physics-guided sparse masking then retains a compact evidence subgraph. Signal quality, degradation trend, and cross-condition transfer evidence are organized into structured prompt blocks before entering a frozen large language model that serves as an evidence-constrained diagnostic reasoning layer rather than a raw-signal encoder. Experiments on five public rotating-machinery datasets and a multi-unit hydropower dataset evaluate cross-condition diagnosis, multi-condition health-state discrimination, module contribution, representation alignment, and controlled Gaussian-noise robustness. Across the reported comparisons, the proposed configuration maintains favorable cross-condition performance relative to the evaluated baselines, while the sparse evidence path links the diagnostic case to selected sensor nodes and retained graph relations. The evaluation protocol further separates predictive robustness from evidence traceability and defines reproducible assessments for simulation augmentation, diagnostic-head controls, and deployment cost.
科研通智能强力驱动
Strongly Powered by AbleSci AI