Research And Application Of Fault Diagnosis Method Of Neutral Point Clamped Three-level Inverter Based On LSTM

With the continuous application of high-voltage and high-power equipment,and the rapid development of related technology of frequency conversion speed regulation,how to improve the stability of the inverter has become a widespread concern in the society.As an important part of the inverter,it is prone to failure due to the complex structure of the equipment and the harsh working environment.The failure of the inverter will not only affect normal production operations,but also threaten the personal safety of workers.Therefore,the research on fault diagnosis of commonly used Neutral Point Clamped(NPC)three-level inverters has important engineering value and practical significance.The three-phase current signal of the inverter can effectively reflect the characteristic information of the equipment state,so the three-phase current signal is usually used for fault diagnosis.Three-phase current vector analysis can extract current signal fault feature information.However,there will be various noise interferences in the current data in actual engineering,and it is necessary to study the time-frequency analysis method for nonlinear non-stationary signals.The original feature set obtained by time-frequency analysis will have a lot of redundant and conflicting information,which needs to be eliminated by using sensitive feature screening methods.For high-dimensional feature sets,the dimension reduction method needs to be used to reduce the dimension of the feature sets to obtain low-dimensional feature sets with better discrimination performance.For fault pattern recognition,a stable and effective classifier model is needed to distinguish the fault state.In order to solve the above problems,the specific research work is as follows:(1)By analyzing the working principle of Neutral Point Clamped three-level inverter,we understand its loop structure and working mode to determine that the main research object is the IGBT power switch tubes.This paper analyzes the open circuit fault of the power tube and the three-phase current data waveform.This article reveals the fault mechanism,analyzes the corresponding characteristics,and obtains the preliminary characteristics of the current data.(2)In this paper,the current signal feature extraction method based on Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(CEEMDAN)is used to extract the first four order IMF components.It forms signal sample sequence with time domain signal and Hilbert envelope spectrum.Six statistical feature values of the sample sequence are calculated to construct a high-dimensional original feature set.This paper proposes a fault-sensitive feature extraction method that combines adjusted rand index(ARI)and random forest(RF)feature importance.The ARI and difference of error rate outside the bag describe the distinguishing degree and feature importance between feature classes.This method solves the problem of redundant and insensitive information in the fault feature set.(3)In order to further extract the low-dimensional feature expression of the high-dimensional feature space,a nonlinear dimensionality reduction method based on Kernel Principal Component Analysis(KPCA)is studied to extract the mapping relationship between principal features of nonlinear features and realize the reduce of high-dimensional feature sets.This paper constructs a three-level inverter fault diagnosis model based on Long Short Term Memory Network(LSTM),and uses Particle Swarm Optimization(PSO)algorithm to adaptively optimize the model hyperparameters and improve the stability of model.A comparison experiment using simulation data and experimental bench data verifies the effectiveness and applicability of the proposed three-level inverter fault diagnosis model.(4)In order to better realize inverter fault monitoring and diagnosis,the fault diagnosis model has more practical engineering significance.Based on the analysis of the multi-dimensional state association relationship and knowledge structure of each loop and component of the inverter,this paper builds the ontology structure of the health elements of the inverter and the semantic rules of the state relationship.This article combines the Neo4 j graph database and the Protégé ontology description language to design the knowledge base of the inverter health status.This paper develops a three-level inverter state monitoring and fault diagnosis system based on.net platform.The system can realize efficient data storage,intelligent reasoning of accident warning information and real-time monitoring of open circuit faults of IGBT power switch tubes.The system has strong data analysis,processing capabilities and a good human-computer interaction interface.It provides user management,real-time data query,rule reasoning,fault diagnosis and other functions.The experimental results show that the feature selection method proposed in this paper can effectively filter out the fault-sensitive features,and the constructed LSTM-based inverter fault diagnosis model has good adaptability and significantly improves the inverter under the same and variable conditions.The diagnosis and recognition accuracy rate,the three-level inverter state monitoring and fault diagnosis system designed in conjunction with the relevant theories such as ontologydescription and graph database can effectively describe the fault type and have strong operability.The paper has 44 pictures,21 tables,and 91 references.