Data-driven Fault Diagnosis Method For Power Devices In Inverter

As the bridge for energy conversion,inverters are widely used in traditional industry,electrified transportation,new energy and power engineering.Driven by both the energy revolution and the digital revolution,electrification has penetrated into all walks of life,which indirectly leads to the growth of the number of inverters worldwide.The reliability of inverters is an important condition to ensure the stable system operation,which has also led to the development of inverter monitoring and maintenance technology.With the rapid development of data science,smart sensors,Io T,edge computing,cloud computing,digital twins and big data analytics have facilitated the deep integration of industrial technology and information technology,and laid the foundation for the collection and transmission of data at different stages of the power electronic system life cycle.The growing amount of data has changed the traditional actively monitoring paradigm of industrial equipment and brought new opportunities for the widespread application of intelligent diagnosis techniques based on data-driven approaches.However,current inverter fault diagnosis based on data-driven approaches have many limitations in practical industrial applications,such as poor economics and weak generalization.To solve the above problems and realize intelligent fault diagnosis methods for industrial applications,this research explores the potential of intelligent fault diagnosis methods in practical applications from multiple perspectives,and provides solutions for the intelligent diagnosis of power devices in inverters.The main research content of this article includes the following four aspects:1)Aiming at the problem that it is difficult to locate IGBT faults accurately with a single sampled signal in the complex applications,an offline fault diagnosis framework combining signal processing methods and deep learning methods is proposed to achieve fault IGBT localization from a single sampled signal.Specifically,adaptive chirp mode decomposition is used to extract and reconstruct sub-signal components from the original sampled signals for adaptive mining of deep fault information.Then,to avoid artificial selection of sub-signal components,silhouette coefficients are introduced to characterize the importance of each component and automatically filter out the signal components that are more important in fault diagnosis.Finally,the temporal convolutional network is used to extract the features of the filtered signal components and output the results.The experiments prove that the method still has good robustness under certain noise conditions.The working mechanism of the ACMD-SC-TCN framework is explained by feature visualization of different residual blocks and channels.The joint diagnostic framework can mutually complement the online diagnostic method to improve the reliability of the multilevel converter.2)Aiming at the limitation that traditional methods can only diagnose a fixed number of faults categories,an end-to-end open-set diagnosis method is proposed.First,batch normalization and layer normalization are introduced into the backbone model to accelerate convergence and improve the generalization ability of the model for different tasks.Second,to enhance the feature extraction capability of the model,a multi-scale coordinate residual attention mechanism is designed for one-dimensional current signals to improve the performance and stability of the model.Experiments demonstrate that the proposed multi-scale coordinate residual attention shows better performance compared to the latest developed attention mechanism in the end-to-end multilevel converters fault diagnosis task.Finally,the additive angular margin loss and local outlier factor algorithms are introduced into the above framework to identify the density differences between known and unknown fault clusters by controlling the intra-class similarity and inter-class variance of the samples.The experimental results demonstrate the feasibility of the proposed fault diagnosis framework for the diagnosis of known and unknown faults,and provide a reference for the implementation of open-set inverters fault diagnosis.3)Aiming at the data shortage problem in the real applications,a knowledge transfer approach using inexpensive simulation data from simulation domain to build models is proposed.First,by analyzing the reasons for the domain adaptation under large cross-domain gaps,a multipath,weakly shared feature extractor with the domain-specific normalization layer and the domain-specific channel attention layer is proposed to separate domain-specific information.Then,an auxiliary classifier for the target domain is introduced to explicitly optimize the separation paths in weakly shared feature extractor and to reduce the bias of the feature extractor towards the source domain data during parameter optimization.Finally,subdomain adaptation is used to adjust the subdomain distribution of the relevant fault categories.Physical domain data with different levels of noise are used to represent different levels of domain shifts,and the effectiveness of the method is evaluated.The results show that the domain-specific adaptation network can realize the end-to-end knowledge transfer task even under very large domain biases,which provides a successful case for solving the data shortage problem of inverter fault diagnosis based on data-driven methods.4)Aiming at the problem of poor generalization ability in unknown environments or working conditions,a domain generalization network exploiting causal and non-causal representations is proposed.Firstly,a basic inverter fault diagnosis model is established,and the effects of common perturbations in inverters on the model generalizability are analyzed,as well as explained how batch normalization weakens the model generalization ability from the visualization.Then,a structural causal model is introduced to formalize the causal mechanism in current data.Then,a causal decomposition module based on a causal matching strategy is designed to extract causal representations,and a domain classifier is used to capture domain-specific non-causal representations.Finally,both representations are further purified by the decoupling module and the extraction of non-causal representations is optimized by meta-learning to achieve stable generalization to unknown working conditions.Extensive comparative and ablation experiments confirm the effectiveness and superiority of domain generalization networks and verify the potential of data-driven methods in real industrial applications.