Aiming at the problems of high dimension of heavy gas turbine operation data, rapid change of working conditions, unbalanced fault samples and high training cost of traditional deep models, a large fault diagnosis model construction method based on LoRA low-rank modulation was proposed. The method takes multi-source operation data as input, establishes a fault semantic representation module, and maps time series signals such as temperature, pressure, flow, speed, vibration and control feedback into high-dimensional representations with context correlation. On this basis, the LoRA low-rank modulation adaptation mechanism is introduced to efficiently fine-tune the parameters of the pre-trained time series large model, and the feature fusion of time domain, frequency domain and working condition is combined to improve the complex fault recognition ability. Experimental results show that the accuracy, recall rate and macro-average F1 value of the proposed model on the test set reach 96.8%, 95.1%and 95.7%respectively, which are better than those of SVM, CNN, LSTM and standard Transformer models. At the same time, the number of trainable parameters is reduced from 88.4M to 9.6M, which reduces the computational overhead while ensuring the diagnosis performance. The research shows that the proposed method has good accuracy, robustness and deployment feasibility in the intelligent diagnosis scenario of heavy gas turbine.
Povzetek: Raziskava predlaga metodo velikega modela z LoRA nizko-rang modulacijo za diagnostiko okvar težkih plinskih turbin. Model učinkovito združuje večvirovne časovne podatke in dosega visoko natančnost, priklic in F1, ob zmanjšanem številu parametrov.
