Fault Diagnosis Of Heavy-duty Gas Turbine Based On Deep Learning

The heavy-duty gas turbines have a significant proportion in current power generation industry.Due to the characteristics of fast startup,stable operation,and high thermal efficiency,the heavy-duty gas turbines have great advantages in suppressing grid fluctuations caused by large-scale grid connection of new energy.However,due to its complex structure and harsh working environment,gas path failures usually lead to huge economic losses.Therefore,it is necessary to pay attention to gas path fault diagnosis for heavy-duty gas turbines.In addition,the continuous emergence and development of artificial intelligence and big data have been expanding the improvement space for deeplearning-based fault diagnosis methods.Therefore,this paper conducts the research on deep-learning-based fault diagnosis methods for heavy-duty gas turbine.Among many fault diagnosis methods based on deep learning,pattern recognition methods have greatly improved the degree of automation and intelligence by simplifying fault detection and isolation.Therefore,this paper put the research on the fault diagnosis methods based on deep learning and pattern recognition.However,there existing the first two challenges for gas path fault diagnosis scheme based on deep learning and pattern recognition:the improvement of feature extraction ability and the extraction of fault invariant features across operating conditions.In addition,it also needs to consider the existing situation of insufficient labeled samples in reality.The main research content is stated as follows:Firstly,in order to improve the abstract feature extraction ability and ensure the faultinvariant feature extraction when labeled samples are sufficient,the fault diagnosis method across operating condition for heavy-duty gas turbines is proposed,which is based on selfcalibrated convolution and adversarial domain adaptation.Specifically,in order to further improve the abstract feature extraction ability,the multi-scale convolution kernel’s output feature fusion and local feature adjustment are performed based on self-calibrated convolution and dual attention mechanism respectively.In the meanwhile,the improved self-calibrated convolution is used for encoder designment.Based on the designed feature encoder,considering the alignment of conditional distribution,the conditional adversarial domain adaptation is used to achieve fault-invariant feature extraction for heavy-duty gas turbines.Finally,the designed encoder and domain adaptive scheme are validated through single mode and cross mode fault diagnosis comparative experiments,and the experimental results show that the designed scheme is effective.Finally,the designed encoder and domain adaption scheme are validated by contrastive experiments for singleoperation and cross-operation gas path fault diagnosis respectively,and the results show that the designed scheme is effective.Secondly,in order to extract fault-invariant features across operating conditions when labeled samples are insufficient tag samples,the few shot-fault diagnosis method across operating conditions is proposed,which is based on prototypical self-supervised learning.This method consists of three parts.Aiming at the shortcomings of traditional prototypelearning methods that prototype computation and prototype similarity computation are limited to label samples and Euclidean distances,the few-shot in-domain prototypical selfsupervised learning is performed as the first module,which is based on fuzzy C-means clustering and contrastive learning.Secondly,the traditional domain adaption methods’rigid requirements on labeled sample size make it difficult to combine with the few-shot learning.To overcome this trouble and ensure the conditional distribution alignment,the few-shot cross-domain prototypical self-supervised learning is performed as the second module,which is based on prototype transfer and contrastive learning.Then,in order to maximize the accuracy of calculated prototype,the adaptive cosine classification is performed as the third module,which is based on limited labeled samples.In addition,the multi objective task balancing strategy based on uncertainty is adopted to coordinated above modules.Finally,the model comparison experiment and module ablation experiment are conducted to verify the scheme,and the results show that the designed scheme is effective.