Fault Diagnosis For Switch Components In Modular Multilevel Converter Based On Machine-learning

MMC has been widely promoted and applied in the fields of flexible DC transmission,electric drive,new energy,and Static Var Generator(SVG),Active Power Filter(APF)and Unified Power Flow Controller(UPFC)because of its low harmonics,good expandability,low switching frequency and low loss.With the increase of the MMC voltage level and capacity,the number of sub-modules connected in series with the bridge arm has reached hundreds or several hundreds.Long-term operation,environmental impact,and electromagnetic stress will inevitably lead to sub-module fault,which seriously affects system operation.Domestic and foreign scholars have conducted a series of studies on the DC side faults and AC side faults of the MMC system,but there are few studies on the diagnosis and submodule faults location.Most of the existing work is based on the premise that each sub-module is equipped with a sensor and a separate diagnostic unit,which increases the hardware cost.Therefore,how to obtain fault diagnosis and location of MMC sub-module in low sensor mode has important theoretical significance and application value.Based on the analysis of the basic topology,working principle and failure mode of MMC,this paper focuses on the analysis of several fault modes of MMC sub-modules and introduces the impact on the MMC system by its fault.Aiming at the fault of the sub-module switching device,this paper proposes machine learning methods to detect and locate the open-circuit fault of the sub-module switching device.The specific research contents are as follows:(1)Whether the MMC system is in normal operation or there is an open-circuit fault of the sub-module,there will always be a three-phase circulation existing inside the system.But this circulation cannot be completely eliminated by the general suppression strategy.Similar to the current data in the three phases,there is a significant difference between the two operating states.Therefore,using the three-phase circulating current and current signal of MMC system as the fault characteristic signal,a fault diagnosis method based on hybrid core-supported tensor machine is proposed.The MMC simulation model is established in MATLAB/Simulink,and the system of normal system and single-phase sub-module failure is constructed for simulation research.The correctness of the proposed method is verified based on the simulation results.(2)The effect of the MMC sub-module failure on the three-phase AC current is investigated at a higher level state.According to the analysis,the failure of the sub-modules between different bridge arms will bring different changes to the three-phase AC current.Besides,the higher the level,the weaker the effect of a single sub-module fault on the AC current,and the less the fault current output change is less obvious than the normal phase.Therefore,more accurate and sensitive fault detection methods are needed for fault diagnosis.After the collected AC signal is filtered,the envelope mean is obtained,and the obtained signal is used as a training set of the least squares mutual information spectrum clustering to obtain a classification label.The training samples are then input to the overall least squares support vector machine for training to obtain a decision function,and the fault phase can be judged by deciding the output value.In the RT-LAB hardware-in-the-loop simulation platform,the normal and single-phase sub-module faults of the 201-level MMC system are simulated.The simulation results verify the correctness and effectiveness of the classification.(3)Targeting at a multi-phase sub-module failure in the MMC,due to the amount of time-domain data information of the voltage and current is too small,accurate fault detection cannot be achieved.As a result,the time domain signal of each phase voltage is first converted into a frequency domain signal that is easy to analyze by using Fast Fourier Transform(FFT).According to the changing characteristics of the frequency domain signal,the multi-class correlation vector machine is used to classify the obtained data into multi-phase faults.The MMC simulation platform is established to verify the effectiveness and correctness of the proposed method.(4)After a submodule fault is located on a specific bridge arm,it is necessary to locate the fault of the specific submodule.In general,each submodule is equipped with a voltage sensor.In order to avoid the use of a large number of voltage sensors,the machine learning method is used to directly locate the voltage changes of the bridge arms to the specific bridge arms.Based on the analysis of the variation of bridge arm voltage data,this paper proposes a method based on deep learning to train and test the bridge arm voltage data when there is an open-circuit fault in normal and sub-modules,which locates the fault to the specific sub-module.