summary
Researchers Huang Limei and Zhang Qi from the School of Electrical Engineering and Automation at Fuzhou University pointed out in an article published in the second issue of the journal "Electrical Technology" in 2019 that traditional converter fault diagnosis methods require the establishment of accurate mathematical models in order to identify and locate faults, but the modeling process is complex and cannot establish nonlinear system models.
To address the above issues, this paper employs a data-driven approach to research and apply converter fault diagnosis. The paper primarily categorizes data-driven fault diagnosis into three types: statistical analysis-based converter fault diagnosis methods, signal processing-based converter fault diagnosis methods, and artificial intelligence-based converter fault diagnosis methods. Furthermore, the basic research principles, applications, and limitations of each method are elaborated upon.
Finally, based on the research, we propose to look ahead to converter fault diagnosis from aspects such as the integration of fault diagnosis methods, detection of new fault types, online learning of fault modes, and the setting of data monitoring systems.
With rapid economic development and massive energy consumption, environmental and climate change, as well as the scarcity of fossil fuels, have become increasingly prominent issues. Therefore, microgrids, as one of the main forms of distributed energy, have gained widespread attention from all sectors of society. Converters enable the switching of different voltage levels within a microgrid, but a fault in the converter will affect the stable operation of the microgrid, and if the fault is not cleared in time, it will have serious consequences. Therefore, fault diagnosis of converters has become a research hotspot for ensuring the stable operation of microgrids.
The fault diagnosis process for a converter involves extracting and detecting characteristic variables such as voltage and current under normal operating and fault conditions, and determining whether unacceptable deviations have occurred in these variables. Typically, the fault diagnosis process consists of three main parts: fault detection, fault location, and fault removal. ① Fault detection primarily determines whether a fault has occurred in the converter; ② Fault location involves analyzing the cause of the fault and locating the faulty component; ③ Fault removal involves intervening in the fault and restoring the system to normal operation.
Currently, common classifications of fault diagnosis methods include fault diagnosis techniques based on analytical models, fault diagnosis techniques based on knowledge, and fault diagnosis techniques based on data-driven approaches.
Fault diagnosis technology based on analytical models starts from the essential characteristics of the system and achieves real-time fault detection, making it suitable for systems with simple and easily modelable circuit topologies. This method primarily uses sampling information provided by sensors to establish an accurate mathematical model. The process of establishing this mathematical model requires a deep understanding of the basic structural mechanisms of the circuit and the circuit modes present during operation. However, for real-world complex circuit systems, fault operation involves modes, analytical difficulties, unavoidable errors, and unknown interferences; therefore, it is difficult to ensure the accuracy of the established mathematical model.
Knowledge-based fault diagnosis techniques are suitable for systems with few inputs, few output variables, lack of sensor information, and difficulty in establishing mechanistic models. These techniques primarily include expert systems. Expert system-based fault diagnosis methods rely on the experience and knowledge of experts in the relevant field, and their accuracy is affected by the level of expert knowledge in the knowledge base. Therefore, given the limitations of expert experience and knowledge, and the difficulty in formalizing knowledge rules, knowledge-based fault diagnosis techniques have certain limitations in data processing.
Model- and knowledge-based fault diagnosis techniques are only suitable for systems with relatively few inputs, outputs, and state variables. Faced with today's complex circuit systems, they cannot provide every detail of the complex circuit mechanism model and much of the advanced professional knowledge. Therefore, given the massive, diverse, and rapid nature of the data generated by the long-term operation of complex circuit systems, data-driven fault diagnosis techniques have been widely applied.
Data-driven approaches do not rely on mathematical models or expert knowledge. Instead, they primarily employ various data mining techniques to extract useful information hidden within online and offline data. This information characterizes the normal and fault states of the current system, ultimately enabling fault detection, diagnosis, and isolation. In recent years, with the rapid development of computing power, data-driven fault diagnosis techniques can efficiently extract feature vectors from large amounts of offline and online data and accurately identify faults.
This paper primarily employs a data-driven approach for converter fault diagnosis, encompassing three main aspects: ① statistical analysis-based converter fault diagnosis techniques, ② signal processing-based converter fault diagnosis techniques, and ③ artificial intelligence-based converter fault diagnosis techniques. The paper explores the content, principles, and application background of these three methods, and analyzes their applicability and limitations. Finally, based on converter development trends, the paper provides an outlook on converter fault diagnosis from the perspectives of method fusion, detection of novel fault types, online learning of fault modes, and the establishment of data monitoring systems.
1. Current Status of Research on Data-Driven Converter Fault Diagnosis Technology
Data-driven converter fault diagnosis not only acquires data in a timely and accurate manner, but also learns and mines potential connections in historical data. Furthermore, it achieves converter fault detection and diagnosis through the implicit mapping mechanism of data. Currently, the classification of data-driven converter fault diagnosis methods is shown in Figure 1.
Figure 1. Classification of data-driven converter fault diagnosis methods
1.1 A Statistical Analysis-Based Method for Converter Fault Diagnosis
Statistical analysis-based methods were used to extract commonly observed characteristics from historical data, and confidence intervals for normal conditions were established to determine the current normal or fault state of the converter. Statistical analysis-based methods are mainly divided into univariate statistical methods and multivariate statistical methods.
① Univariate statistical methods mainly define a threshold value for a process variable to achieve fault detection and diagnosis. They are simple to implement and suitable for transformers with small data dimensions, but they ignore the correlation between variables.
② Multivariate statistical methods can fully characterize the correlation between multiple variables and are suitable for fault detection and diagnosis of high-dimensional converter systems. Multivariate statistical methods mainly include principal component analysis (PCA), kernel principal component analysis (KPCA), and hidden Markov models (HMM).
1) PCA-based converter fault diagnosis
PCA technology uses a multivariate projection method to map high-dimensional historical data to a low-dimensional space that can fully reveal the characteristics of the original data. It mainly uses the most obvious variables in the low-dimensional space to represent the original historical data, ultimately achieving the purpose of dimensionality reduction and greatly simplifying the data.
PCA technology has been widely used in converter fault diagnosis. Reference [14] mainly extracts the two output line voltages of the three-phase inverter in the permanent magnet synchronous motor as feature vectors, and uses PCA technology to reduce the dimension of the fault feature vector. Reference [15] uses a fault monitoring system to detect and isolate faults in the three-phase inverter, and uses discrete wavelet and PCA technology to detect the current discontinuity caused by the fault.
Reference [16] studies the faults of single and combined switching transistors in the midpoint clamped inverter, using the upper, middle and lower bridge arm voltages as test signals and applying PCA technology to reduce the input of the neural network. Reference [17] uses a PCA-based method to reduce the dimension of the initial feature vector and eliminate redundant data information in order to achieve high-precision fault diagnosis in order to improve the accuracy of nonlinear subway auxiliary signal diagnosis.
In general, the essence of PCA technology is to perform a linear transformation on the input space composed of historical datasets. However, it only has a good extraction capability for data that follows a Gaussian distribution, and its diagnostic effect on nonlinear and non-Gaussian distribution fault data is not ideal.
2) KPCA-based converter fault diagnosis
KPCA employs a kernelization approach to map samples to a higher-dimensional space, then applies basic PCA techniques to achieve the mapping and projection of higher-dimensional data. The mapped and projected feature vectors can fully characterize the converter's operating characteristics, achieving dimensionality reduction of fault feature vectors. Compared to the PCA method, KPCA uses a nonlinear mapping function to map input variables to a high-dimensional linear space, enhancing its ability to process nonlinear data.
Reference [18] uses the KPCA method to reduce the dimension of the feature vector of open-circuit faults in insulated gate bipolar transistors, thus achieving a reduction in dimension. Reference [19] transforms the three-phase current characteristics of the asynchronous motor drive circuit using Concordia technology and then uses KPCA for signal processing. This method is suitable for fault diagnosis under different load conditions. Reference [20] proposes a fault diagnosis method based on heterogeneous information feature fusion to process vibration and current signals, addressing the limitations of single fault signals and the strong nonlinearity of fault characteristics in asynchronous motors. KPCA fully utilizes the redundant and complementary information between different information sources and the nonlinear relationship between feature data to comprehensively describe the operating status of the equipment.
While KPCA-based methods can process raw nonlinear data, they primarily project the data into high- and low-dimensional spaces to achieve dimensionality reduction. This process mainly relies on thresholds to preserve information, resulting in the loss of some original data features. This leads to low accuracy and errors in converter fault diagnosis.
3) HMM-based converter fault diagnosis
Hidden Markov Models (HMMs) utilize historical data that changes over time to build dynamic models. These models are then used to uncover potential information within the data, enabling fault diagnosis of converters. HMMs primarily perform dynamic analysis in the time and frequency domains, making them an important method for dynamic data parsing. Furthermore, the process is simple, easy to implement, and suitable for time-varying systems.
Reference [21] uses the HMM method to diagnose faults in traditional photovoltaic inverters, which reduces the diagnosis time and improves the accuracy. Reference [22] proposes a fault identification and classification method for DC converters based on HMM, which promotes the gradual development of high voltage DC transmission and ensures the stable operation of equipment. Reference [23] establishes an HMM-based model based on the parameter space of the linear time-invariant model, based on the actual operation mode of traditional power grid circuit breakers, and realizes the autonomous learning of data.
The Hidden Markov Model (HMM) method can establish simple fault diagnosis models that are easy to learn, but its disadvantages are that the accuracy of the established dynamic models is low, and the classic algorithms used in the learning process do not take into account the complexity of the model and cannot solve the problems of overfitting and underfitting.
1.2 Signal Processing-Based Converter Fault Diagnosis Method
When a converter malfunctions, the amplitude, phase, and frequency of characteristic quantities such as voltage and current at its measurable points will change significantly. Signal processing-based fault diagnosis methods primarily process and analyze feature vectors to obtain a comprehensive evaluation of the converter under both normal and fault conditions. Currently, the main signal processing-based converter fault diagnosis methods are wavelet transform (WT) and Hilbert-Huang transform (HHT).
WT is a new transform analysis method that not only inherits the idea of localization of short-time Fourier transform, but also overcomes the disadvantage that the window size does not change with frequency. WT refines the time and frequency domain signals step by step through scaling and translation operations, and finally achieves the effect of time refinement at high frequency and frequency refinement at low frequency, which can meet the requirements of arbitrary detail time and frequency signal analysis. Reference [24] uses WT to extract the output voltage fault features of a three-level inverter and uses this as the input signal of support vector machine (SVM).
Reference [25] proposes a method combining WT and extreme learning machine for fault diagnosis of grid-connected photovoltaic inverters, in which the WT method is mainly used to analyze the output current signal of the inverter. Reference [26] performs Fourier transform and WT analysis on the output current signal of a three-phase grid-connected inverter. The results show that WT has the advantages of flexibility and dual domain, and can accurately provide fault characteristics, thereby realizing fault diagnosis and location.
HHT primarily analyzes non-stationary signals and is divided into two aspects: Empirical Mode Decomposition (EMD) and Hilbert Spectral Analysis (HHT). Unlike Fourier and wavelet transforms, it does not select fixed basis functions to expand the signal. Instead, it adaptively decomposes the signal into several eigenmode functions derived from the signal itself and obtains the corresponding Hilbert spectrum.
Reference [27] uses the HHT fault diagnosis method to identify short-circuit faults by decomposing the three-phase current characteristic signals. This method is simple and easy to implement and can accurately detect fault information. Reference [28] makes full use of the HHT fault diagnosis method to extract faults from the characteristic quantities of the voltage source-based high-voltage DC converter system. This method is simple to implement and has a high diagnostic rate.
1.3 Artificial Intelligence-Based Converter Fault Diagnosis Method
Artificial intelligence-based fault diagnosis methods do not require the establishment of quantitative mathematical models. They only need to train AI algorithms on data from both normal and fault states of the converter to diagnose and locate faulty components within the converter. This method applies fault features to establish a relationship between features and a classifier, enabling fault diagnosis and identification of complex mode converters. Currently, AI-based algorithms mainly include SVM-based methods and neural network (NN)-based methods.
1) SVM-based method
SVM is a machine learning algorithm based on statistical learning theory and structural risk minimization. It can perform data analysis and pattern recognition, and is suitable for small sample, nonlinear and high-dimensional pattern recognition, classification and regression analysis.
Figure 2. SVM classification principle diagram
Based on the above principles and characteristics, SVM has been widely used in converter fault diagnosis. Reference [30] proposed a method combining Fourier transform, relative principal component analysis and SVM to diagnose the output voltage of the inverter, targeting the characteristics of cascaded H-bridge multilevel inverters. Reference [31] used discrete orthogonal wavelet transform to decompose the output voltage of a three-phase voltage source inverter and obtained the corresponding wavelet coefficient matrix. Hybrid SVM was used to train, diagnose and classify the wavelet coefficient matrix.
Reference [32] uses genetic algorithm and SVM to diagnose faults in four parameters of power electronic inverters. Reference [33] detects the voltage of the upper, middle and lower bridge arms of the neutral clamp inverter, and trains the feature vectors through Fourier transform using multi-level SVM, which has high fault diagnosis accuracy. Reference [34] applies the relative principal component analysis and SVM method to diagnose faults in cascaded H-bridge multilevel inverters. Compared with the traditional backpropagation neural network and SVM, the method proposed in the reference reduces the computation time and improves the diagnostic accuracy.
While SVM has demonstrated many advantages in converter fault diagnosis, it struggles to handle large-scale data samples, and the accuracy of fault diagnosis is closely related to the completeness and representativeness of the samples. Furthermore, classic SVM only provides a binary classification algorithm, diagnosing faults solely from a classification perspective without delving into the structural information of the data.
2) NN-based methods
Neural Networks (NNs) are generalized mathematical models that mimic the behavioral characteristics of animal neural networks to establish distributed information data. The transformer uses NN methods to monitor and diagnose normal and faulty data. By adjusting the relationships between neuron nodes, it achieves self-learning and adaptive capabilities. NNs establish a mapping between fault symptoms and fault types through learning between network layers, so that nodes in the input layer correspond to fault symptoms, and nodes in the output layer correspond to fault types, ultimately realizing the reasoning process from fault symptoms to fault types.
Most current neural network models adopt the MP model, jointly proposed by psychologist W. Mc. Cuoooch and mathematical logician W. HPitts. Figure 3 shows a neuron model.
Figure 3. A model of a neuron.
Numerous studies have shown that neural networks (NNs) have powerful pattern classification and recognition capabilities. Reference [37] studied the open-circuit fault of IGBT switching transistors in a three-phase voltage-source static converter and used an algorithm combining discrete wavelet and NN to analyze the feature vectors. Reference [38] compared various NN algorithms and gave the most basic NN model, in which the input layer represents the original feature vector and the output layer represents the corresponding fault type.
Reference [39] combines principal component analysis, genetic algorithm and NN method to diagnose faults in cascaded H-bridge multilevel inverters. This method overcomes the complex and nonlinear system composed of multiple switches. Reference [40] uses genetic algorithm, particle swarm optimization algorithm and NN method to detect drive faults caused by multilevel inverters. This method reduces system harmonics and improves the diagnostic efficiency of the system.
Neural networks (NNs) are large-scale, highly nonlinear systems. Their high-speed complexity makes it impossible to accurately analyze various performance indicators, and the diagnostic process requires a large number of fault samples, limiting their application to small-sample systems. Currently proposed NN types are only applicable to certain converter types, and there is no NN architecture as simple and universal as the von Neumann architecture.
2. Prospects for Data-Driven Converter Fault Diagnosis
The fault diagnosis method for converters has effectively processed fault characteristics and detected fault categories, preventing the serious impact of converter operation with faults on the power grid waveform quality. However, as mentioned above, the fault diagnosis method still has limitations. Therefore, data-driven converter fault diagnosis methods are continuously being improved and developed, with the main development directions likely being:
1) Integration of Data-Driven Converter Fault Diagnosis Methods. Data-driven fault diagnosis methods do not require the establishment of precise mathematical models; they mainly rely on reasoning and analysis of historical data. However, each method has its own advantages and limitations. To address these limitations, a novel converter fault diagnosis method overcomes these shortcomings by combining the strengths of multiple methods, achieving the integration of various diagnostic approaches and ensuring the accuracy, reliability, and effectiveness of fault diagnosis.
2) Currently, most converter fault diagnosis methods focus on learning and training for known faults in the system, neglecting the detection of novel fault types, which can easily lead to misidentification of these fault samples. Therefore, the detection and identification of novel fault categories is of great significance in fault diagnosis applications.
3) Data-driven fault diagnosis methods primarily process and train large amounts of offline historical data to achieve fault diagnosis. However, for complex systems with multiple scales and levels, obtaining such massive amounts of data is difficult, and the training process consumes a significant amount of time. Converter systems, in actual operation, generate a large amount of online data. If data-driven algorithms can utilize online data for learning, it not only reduces reliance on historical data but also shortens training time, ultimately achieving real-time online fault diagnosis.
4) The implementation of data-driven methods is inseparable from the application of data monitoring systems in actual engineering. Fault diagnosis results need to be detected, stored, and fault protection measures implemented through the monitoring system. The improvement and promotion of monitoring systems is also a future trend in converter fault diagnosis.
in conclusion
This paper presents a data-driven approach to converter fault diagnosis, preventing serious consequences from converter failures. It reviews commonly used methods for converter fault diagnosis, categorizing existing methods into statistical analysis-based, signal processing-based, and artificial intelligence-based methods, and focuses on analyzing the applicability and limitations of each type.
Finally, the trends in converter fault diagnosis are discussed from the perspectives of integrating fault diagnosis methods, detecting new fault types, online learning of fault modes, and setting up data monitoring systems. Currently, research on the application of data-driven methods in converter fault diagnosis is still in its early stages, and further research is needed to explore the issues more deeply.
Disclaimer: This article is a reprint. If there are any copyright issues, please contact us promptly for deletion (QQ: 2737591964). We apologize for any inconvenience.
