Studies On Intelligent Diagnosis For Natural Gas Pipeline Under Class Imbalance And Small Sample

With the progressive implementation of the “carbon peaking and carbon neutrality goals”,the natural gas gathering and transmission system is rapidly evolving towards a direction of large-scale,continuous,integrated,and intelligent development.Under this trend,the safety and reliability of pipeline transportation are encountering more significant challenges.As a necessary means to ensure the stable operation of pipelines,industrial monitoring technology that encompasses typical applications like fault diagnosis and early failure warning is able to promptly detect anomalies and precisely identify fault types,thereby mitigating adverse effects.As a result,these techniques have received extensive attention in both academia and industry.In recent years,with the widespread deployment of various advanced sensors and the rapid development of artificial intelligence technology,data-driven intelligent monitoring and diagnosis technology has attracted considerable research attention.However,the class imbalance and small sample problem of big data in the oilfield make research on intelligent fault diagnosis in natural gas pipelines highly challenging.This paper takes natural gas pipelines as the research subject,using the imbalance issue and small sample issue as the breakthrough point,exploring five typical problems in intelligent monitoring and fault diagnosis of pipeline operation states from both “offline” and“online” perspectives,aiming to address the bottlenecks of in the field of natural gas gathering and transportation system maintenance technology.The specific research of this article is presented as follows:Considering the class imbalance issue in the single working condition,an entropyoptimized contrastive adversarial network is proposed for offline small-class data augmentation.First,a dynamic fault semantics encoding mechanism based on priori permutation entropy is devised to characterize and model real-time series.After that,a class-aware mean discrepancy loss is designed to enhance the significance of synthetic data fault features.The proposed algorithm effectively improves the authenticity of the temporal structure and feature discriminability of the synthesized data,thus achieving the natural gas pipeline fault diagnosis in a single operating condition scenario.Considering the class imbalance issue in multiple working conditions,a subdomainalignment adversarial self-attention network is proposed for offline data augmentation.First,a knowledge-sharing structure based on a multi-head self-attention mechanism is designed to make the generative model effectively capture various failure modes.After that,a prototypeassisted subdomain alignment strategy is employed to enhance the clarity of decision boundaries for categories within the augmented dataset.The proposed algorithm effectively improves the diversity of synthetic failure modes and ensures the semantic consistency between synthetic and real data,thus achieving the natural gas pipeline fault diagnosis under multiple working condition scenarios.Considering the small sample issue in the weakly supervised scenario,a weak-shot learning-based offline cross-domain fault diagnosis method is constructed.First,the contrastive adversarial discrepancy criterion is proposed to align the marginal and conditional distributions between the two domains from the global domain level and the fine-grained health state class level.After that,a prototype pseudo-label learning mechanism is introduced to address the issue of label space heterogeneity in knowledge transfer.The proposed algorithm effectively improves the adaptability and discriminability of domain-general features,thus achieving cross-domain natural gas pipeline fault diagnosis in the weakly supervised small sample scenario.Considering the small sample issue in the unsupervised scenario,an evolutionary unsupervised domain adaptation is constructed for offline cross-domain fault diagnosis.First,a local manifold embedding mechanism is constructed to maintain the domain-specific features in shared feature space.After that,a gradient adversarial adaptation method is used to improve the transferability of domain-general features.Finally,a particle swarm optimizationbased control strategy is proposed to dynamically balance the contributions of domain-general and domain-specific features during the training process of the cross-domain diagnosis model.The proposed algorithm synergistically optimizes cross-domain transferability and domainspecific distinguishability to achieve cross-domain natural gas pipeline fault diagnosis in the unsupervised small sample scenario.Considering the issue that offline diagnosis model cannot meet the real-time demand,a transfer learning-based online early warning method for pipeline failure is proposed.First,a cross-domain predictive model based on regular maximum mean discrepancy loss and neighboring point deviation correction loss is designed to predict the future operating status of the pipeline.After that,online early warning of natural gas pipeline failure is effectively achieved by utilizing a dynamic threshold detection method based on predicted operating conditions.