MULTI-LEVEL inverters are widely used in grid-tied energy conversion systems (such as photovoltaic, electric vehicle charging piles, and wind turbines), because of their higher efficiency, less current harmonics, lower overall costs, and inherent fault-tolerance, compared to two-level converters [1],[2]. Among the many types of multi-level inverters (such as T-type, NPC, NPP, and so on), the T-type inverter performs well in the application of medium switching frequency and tens kilowatt so that it is a kind of crucial grid-tied inverter [3],[4].

Transistor fault seriously harms the reliability of the system, causing energy loss and even system shutdown [5]. Unfortunately, transistors are none other than fragile components in converters. Around 38% of the faults in power systems are due to semiconductor devices [6]. From an industrial perspective [7], 40% of the attention should be on semiconductor devices to improve the reliability of the power system. The T-type inverter has twice as many transistors as the two-level converter, so it faces a greater risk of transistor fault.

The transistor fault types include short-circuit faults and open-circuit faults. Hardware protection circuits often deal with instantaneous overvoltage and overcurrent caused by the short-circuit fault. Open-circuit fault results in output distortion and harmonics, which becomes the major concern of transistor faults [8]. The open-circuit fault diagnosis method can quickly locate the fault transistor, provide the basis for fault tolerant control or maintenance decisions, and greatly improve the system’s reliability.

Recently, much research on open-circuit fault diagnosis has been carried out, which can be simply classified as model-based methods, data-based methods, and signal-based methods. Model-based methods collect system parameters and build accurate circuit models to estimate the residuals, to realize fault diagnosis. Data-based methods mine fault information from a large number of system data by techniques of machine learning or deep learning and obtain fault results by classifiers. Signal-based methods use the voltage or current signals of the system to obtain fault information through signal analysis and processing.

Compared with signal-based methods and data-based methods, model-based methods typically require more types of system parameters. In the case of two-level power converters, models such as the pole-to-pole model [9] and phase voltage model [10] are used to realize fault diagnosis. When applying these models, it is necessary to obtain system parameters such as voltages, currents, inductances and capacitances, or currents and switching states. In the case of multi-level power converters, the models built are more complex and there are more fault states. The average voltage model is built to design fault diagnosis methods for T-type inverters [11], which solve the disturbance caused by working conditions in multi-level converters, such as modulation index regulation. A kind of node-path model and the voltage model are used to diagnose the single open-circuit fault for the T-type inverter [12]. In NPC rectifier fault diagnosis, the line voltage model is built to obtain the error value, and then the diagnosis method is designed based on the error [13]. The two diagnosis methods realize the diagnostic speed of switching period levels because of the combination of switching states. Model-based methods are quick to diagnose and adapt to complex working conditions. However, many system parameters are usually required, and these parameters may not be available or accurate, which may limit the application of model-based methods.

Data-based methods are more used in fault diagnosis of two-level converters [14],[15]. There is also precedent for the data-based method in NPC converter [16]. However, it is rarely reported in the diagnosis of T-type converters. Although many of the data-based methods only need to collect a single type of system signal as the input of the learning models, the training of the models needs a lot of training samples. This requires a lot of data acquisition work, and the high computational complexity also increases the difficulty of application.

Signal-based methods have also been widely studied in recent years. In the case of two-level converters, there are many diagnostic methods based only on the three-phase currents [17],[18]. In [19], line voltages are analyzed and used to design the diagnosis method. These methods require only one kind of system signal and are more applicable. Unfortunately, existing signal-based methods in multilevel converter diagnosis still use more system signals or control operations. In the fault diagnosis for grid-tied NPC inverters, the three-phase currents are used as the diagnostic signals, and additional operations such as the underexcited power injection and the switching scheme are performed separately to locate the fault [20],[21]. [22] designs the diagnosis method based on the three-phase currents and grid voltage angle for the T-type rectifier. However, the method is only effective for some transistors. In T-type inverters,[23] uses single-phase line voltage for fault diagnosis, but the performance under modulation index regulation and load regulation is not reported.[24] proposes a fast diagnosis method, but it is based on an additional neutral-point current sensor, three-phase currents, and switching states. In [25], the three-phase currents and two DC-bus capacitor voltages are used to diagnose the open-circuit fault.

In general, fewer diagnostic signals are used in diagnosis methods in favor of their application. There have been many researches and applications on the design of two-level converters using only three-phase currents. However, diagnosis methods for multilevel inverters such as T-type still usually use relatively more diagnostic signals or control operations. This paper proposes an online open-circuit fault diagnosis for the grid-tied T-type inverter based on only three-phase currents. Firstly, the fault mechanism of the inverter is analyzed in detail. Then, the design of the proposed method is given, which consists of three steps. The first step is to obtain the Hausdorff distances among the three-phase normalized current accumulations to perform the fault detection. The second step process is to calculate the fundamental frequency values of the three-phase currents to locate the faulty phase. The third step is to locate the faulty transistor by some intermediate conditions. Finally, simulations and experiments verify the robustness and effectiveness of the proposed method. The contributions are:

1) A novel online diagnosis method for the grid-tied T-type inverter only based on the three-phase currents is proposed. The Hausdorff distances and the fundamental frequency values are used as diagnosis features.

2) The proposed method is suitable for all single transistor open-circuit faults, considering modulation index regulation. The fault diagnosis is tolerant of inverter transient conditions.

Section II introduces the system of the grid-tied T-type inverter with the diagnosis unit. Section III is the analysis of the open-circuit fault mechanism in detail. Section IV introduces the design of the proposed method. Section V shows the simulations and experimental verifications. Section VI is the conclusion of this paper.

As a multi-level converter, the grid-tied T-type inverter has a more complex topology than the two-level converter, as shown in Fig. 1. There are twelve transistors $\left({{\mathrm{S}}_{x1},{\mathrm{\;S}}_{x2},{\mathrm{\;S}}_{x3},{\mathrm{\;S}}_{x4}}\right)$ and their respective fly-wheel diodes $\left({{\mathrm{D}}_{x1},{\mathrm{D}}_{x2},{\mathrm{D}}_{x3},{\mathrm{D}}_{x4}}\right).x$ represents the three phases a, b, c. X represents the three phases A, B, C. N is the neutral point. ${u}_{x\mathrm{\;s}}$ are the three-phase grid voltages. ${L}_{x\mathrm{\;s}}$ represent filter inductances. ${i}_{x}$ are the three-phase currents. The positive directions of ${i}_{x}$ and ${u}_{x\mathrm{\;s}}$ are defined in Fig. 1. ${C}_{1}$ and ${C}_{2}$ are DC bus capacitances. E is the DC power supply. $r$ and ${i}_{0}$ are neutral resistance and current. The PWM inverter control system collects ${i}_{x}$ and ${u}_{xs}$, and then outputs the three-phase modulated voltage signals ${u}_{x}$ according to the given power ${p}^{* }$. Drive signals can be calculated by ${u}_{x}$. When ${u}_{x}> 0$, there are switching states P and O in the X phase. When ${u}_{x}< 0$, there are switching states N and O in the X phase. The relationship between the switching state and the switch combination is shown in Table I. The proposed fault diagnosis unit only collects ${i}_{x}$ to perform the fault diagnosis, without any other system signals (system parameters, control parameters, or other electrical signals).

When analyzing the open-circuit fault mechanism, the working mode of the T-type inverter should be analyzed first. The inverter has three switching states(P, O, N). Considering the current direction, there are six current flow paths, as shown in Fig. 2. They are ${\mathrm{p}}_{+ }\left({{C}_{1}\rightarrow {\mathrm{S}}_{x1}\rightarrow {L}_{x\mathrm{\;s}}\rightarrow {u}_{x\mathrm{\;s}}}\right),{\mathrm{p}}_{- }\left({{u}_{x\mathrm{\;s}}\rightarrow {L}_{x\mathrm{\;s}}\rightarrow {\mathrm{D}}_{1}\rightarrow {C}_{1}}\right)$, ${\mathrm{O}}_{+ }\left({{\mathrm{S}}_{x3}\rightarrow {\mathrm{D}}_{x2}\rightarrow {L}_{x\mathrm{\;s}}\rightarrow {u}_{x\mathrm{\;s}}}\right)$, ${\mathrm{O}}_{- }\left({{u}_{x\mathrm{\;s}}\rightarrow {L}_{x\mathrm{\;s}}\rightarrow {\mathrm{S}}_{x2}\rightarrow {\mathrm{D}}_{x3}}\right)$, ${\mathrm{n}}_{+ }\left({{C}_{2}\rightarrow {\mathrm{D}}_{x4}\rightarrow }\right.{L}_{x\mathrm{\;s}}\rightarrow {u}_{xs})$, and ${\mathrm{n}}_{- }\left({{u}_{x\mathrm{\;s}}\rightarrow {L}_{x\mathrm{\;s}}\rightarrow {\mathrm{S}}_{x4}\rightarrow {C}_{2}}\right)$. In path ${\mathrm{p}}_{- },{\mathrm{o}}_{- }$, and ${\mathrm{n}}_{- }$, the current direction is negative. In paths ${\mathrm{p}}_{+ },{\mathrm{o}}_{+ }$, and ${\mathrm{n}}_{+ }$, the current direction is positive. In paths ${\mathrm{p}}_{- }$ and ${\mathrm{p}}_{+ }$, the switching state is $\mathrm{P}$. In paths ${\mathrm{o}}_{- }$ and ${\mathrm{o}}_{+ }$, the switching state is O. In paths ${\mathrm{n}}_{- }$, and ${\mathrm{n}}_{+ }$, the switching state is N.

During the healthy operation of the inverter, the energy is transferred from the DC side to the grid side, which means that both ${i}_{x}$, and ${u}_{x\mathrm{\;s}}$ are in the positive direction. The pole-to-pole voltage ${U}_{\mathrm{{XN}}}$ is determined by the switching state and the capacitor voltages. The relationship between ${i}_{x}$ and ${U}_{\mathrm{{XN}}}$ is shown in Fig. 3. The fault mechanisms of the current path under the health condition, the ${\mathrm{S}}_{x4}$ fault condition, and the ${\mathrm{S}}_{x2}$ fault condition are analyzed below. The ${S}_{x1}$ fault condition and the ${\mathrm{S}}_{x3}$ fault condition are similar to the previous two fault conditions, except that the direction of each parameter is opposite.

Only the three-phase currents are used to design the opencircuit diagnosis method in this paper. The method is a step-by-step process. The first step is to obtain the Hausdorff distances among the three-phase normalized current accumulations to perform the fault detection. The second step process is to calculate the fundamental frequency values of the three-phase currents to locate the faulty phase. The third step is to locate the faulty transistor. The ${\mathrm{S}}_{x1}/{\mathrm{S}}_{x3}$ or ${\mathrm{S}}_{x2}/{\mathrm{S}}_{x4}$ fault is located by the three-phase normalized current accumulations and the ${\mathrm{S}}_{x2}/{\mathrm{S}}_{x3}$ or ${\mathrm{S}}_{x1}/{\mathrm{S}}_{x4}$ fault is located by the fundamental frequency values between the phases.

The experimental platform used for verification is shown in Fig. 5. The main circuit uses 10-FY12NMA160SH01 IGBT modules and a TMS320F28335 controller. DC source is FTG100-800. The current sampling frequency f is 5kHz. The grid source frequency f_s is 50Hz and RMS is 50V. The switching frequency f_k is 10kHz, and the dead time t_d is 2 $\mu \mathrm{s}$. The capacitances ${C}_{1}$ and ${C}_{2}$ are both ${550\mu }\mathrm{F}$, and the filter inductances L_xs are all 2.632mH. The open-circuit fault is realized by the host shielding the driver signal. The thresholds k and ${k}_{1}$ are set to 0.3 and 1.08, respectively. The experiments verify the robustness and effectiveness of fault diagnosis under the modulation index regulation.

A novel online open-circuit fault diagnosis method for grid-tied T-type inverters based on only three-phase currents is proposed in this paper, which is suitable for all single transistor fault diagnosis. The distortion mechanism of the currents is analyzed in detail. The Hausdorff distances among the three-phase normalized current accumulations and the fundamental frequency values of the three-phase currents are used as diagnosis features. Compared with the existing methods, fewer system signals are used, which reduces the cost of diagnosis and the difficulty of implementation. Experimental results show the robustness under inverter transient conditions and effectiveness of fault diagnosis even with modulation index regulation.