Open Switch Fault Diagnosis in Three-Phase Voltage Source Inverters Using Single Neuron Implementation

Abstract

Fault diagnosis in power converters is essential for keeping electrical systems stable, efficient and long-lasting. Park’s Vector Transform, discrete wavelet transform, Artificial Neural Network, Fuzzy Logic and other methods are used to diagnose faults in the power converter in both single and multiple open switch situations. These methods are implemented on the digital signal processor or controller, which needs additional hardware and consumes more processing time. This paper presents a hardware-based open switch fault diagnostic method in a 3ϕ voltage source inverter to minimize fault diagnosis time and cost. An innovative hardware-based approach that utilizes a single neuron for open switch fault diagnosis in 3ϕ voltage source inverters was successfully implemented without using a digital signal processor or controller. A gradient descent algorithm calculates the weight and bias values of a single processing neuron. Furthermore, a high-speed multiplier and adder circuit seamlessly integrate with the single processing neuron, enabling rapid real-time fault diagnosis. This method is capable of diagnosing single and multiple switch open circuit faults in switching devices under variable load conditions at different frequencies. The proposed system ensures good effectiveness and resistivity, detecting faults in less than one cycle with low implementation effort and no tuning or threshold dependence. It achieves 98% accuracy, 96% precision and 95% recall, with a 2% false positive rate. Unlike traditional methods, it eliminates DSP/controller dependency by using a single neuron-based processing circuit, reducing cost and improving real-time fault diagnosis in three-phase voltage source inverters.

1. Introduction

Nowdays, DC-AC converters are effectively used in industrial equipment like adjustable speed drives, uninterruptible power supplies, flexible AC transmission systems, voltage compensators, photovoltaic inverters and others [1]. DC-AC converters have advantages like high effectiveness, high accuracy, compactness and high consistency. The consistency of DC-AC converters is the most important aspect. Fault-tolerant control to guarantee uninterrupted and correct converter operation is compulsory [2]. This involves a quick and consistent fault diagnosis system. It is observed that the switching devices in power converters are one of the weakest components, causing about 30% of the breakdowns and interruptions [3].

Induction Motors (IMs) are driven by the space vector control approach, and the process to be followed in controls is utilized by Voltage Source Inverters (VSIs) [4]. However, as a result of the complications of the drive systems and the diversity of operation situations, several faults take place in the VSI. When there are faults in the VSIs, the whole IM drive system will not function under normal or usual conditions [5]. Moreover, the faults occurring in control techniques can cause financial losses or terrible accidents [6]. As a result, an investigation of the fault diagnosis in the VSI is most significant.

The basic components associated with power conversion are power switches like Insulated Gate Bipolar Transistors (IGBTs) or Metal Oxide Semiconductor Field Effect Transistors (MOSFETs). It is well-known that there are six power switches in the 3ϕ VSI. The failure of these power switches can be generally classified as short circuit faults (SCFs), Misfiring Faults (MFs) or Open Circuit Faults (OCFs) [7].

SCFs occur quickly and IGBT can tolerate these types of fault for 10 µsec [8]. Therefore, it is difficult to diagnose SCFs using algorithmic methods, as the protective window is very small. MFs result from missing transistor gate pulses, which lead to the reduced efficiency of devices. Several methods have been utilized to diagnose OCFs [9]. As a result, much of that current research is going on in the area of OCF diagnosis in the VSI, which is also the research topic in this paper. Out of several Open Circuit Fault diagnostic Methods (OCFDMs) for switching devices in power converters, Park’s Vector Transform (PVT)-based diagnosis techniques have received a lot of interest in the past few decades. M. Trabelsi et al. proposed OCF diagnosis in the two power converters of a Permanent Magnet Synchronous Generator (PMSG) drive for the online diagnosis of multiple OCFs. This has been implemented signal processing algorithms on the dSPACE DS1103 digital controller that needs only 3ϕ current and speed, which are already sensed for internal control of the drive. The FDM shows high resistance to false alarms and is sustained under variable load or high-speed conditions [10].

The current trajectory’s midpoint [11], Normalized DC Current Method [12], Modified Normalized DC Current Method [13] and Slope Method [14] are based on PVT and have been implemented on MATLAB-R2022b and the Xmath package for the diagnosis of OCFs.

In [15], the author proposed the FDM for a 3ϕ Permanent Magnet Synchronous Motor (PMSM) drive based on the average current calculated using PVT. The Fault Diagnosis Variables (FDVs) are computed using the average current. The Fuzzy-based classifier is applied to process the FDVs to diagnose single or multiple OCFs with 96% accuracy. The algorithm is implemented on a Digital Signal Processor (DSP) (TMS320F2812) control board. A combination of PVT and Fuzzy Logic System (FLS) was proposed in [16] for the diagnosis of single and multiple OCFs under variable load conditions. The FLS is utilized for a threshold value adjustment to diagnose the OCF under variable load conditions.

R. B. Dhumale, et al. proposed an FDM based on the combination of signal preprocessing, feature extraction, a feature selection technique and a fault classifier. The PVT is used to normalize current to diagnose the fault under variable load conditions [17]. The features are extracted using discrete wavelet transform (DWT) and the max features are selected to train the Artificial Neural Network (ANN)-based fault classifier. The results are shown for different fault patterns and single and multiple OCFs under variable load conditions at different frequencies. Also, the results are shown for the online fault diagnosis, with the incidence of fault occurrence and the incidence of diagnosis. This method requires a Digital Signal Controller for the implementation of the proposed FDM; as a result, this algorithmic solution requires a long processing time. An easy and reliable method for the diagnosis of OCFs based on a DWT and the FLS has been proposed in [18]. The proposed FDM is checked for a variable load condition and the experimental results are validated on dsPIC30F4011. In contrast, this technique is slower than other FDMs proposed earlier but gives satisfactory results for practical applications. The same combination of DWT-FLS has been implemented on an Intel 80C196KC 16-bit microcontroller for the diagnosis of single switch faults in Pulse Width Modulation (PWM)-VSI. This FDM shows high potential for OCF diagnosis [19].

In [20], the authors proposed an Artificial Neural Network (ANN)-based FDM; four features, the two-phase currents (I_α and I_β), angle to the current pattern (I_θ) and surface difference of the current patterns between healthy and faulty (E_s), are extracted from the current signal using PVT. An ANN of structure 4-15-13 is implemented on the computer for further processing. This FDM is quick, capable and 100% perfect for single or multiple OCF diagnoses. The single and multiple OCF diagnoses of a phase-controlled 3ϕ full-bridge rectifier using the Deep Belief Network have been proposed in [21]. The Wavelet-Neural Network method, Wavelet-ANFI System, Clustering-ANFI method and Model-Based ANN Method were proposed in [22,23,24,25], respectively. Simulation results show a great accuracy for fault classification using ANNs. Some manual methods are proposed in [26,27] that are based on the measured voltage and Bond Graph Model, respectively. In [28], the current spectrum has been examined for the diagnosis of OCF. A Fast Fourier Transform (FFT) has been analyzed for the spectrum analysis. This needs comparatively high computing power.

1.1. Problem Statements and Research Gap

Surviving OCF diagnosis techniques depend on signal processing methods and intelligent algorithms and are generally implemented using DSPs or microcontrollers. Various methods have been evaluated:

• PVT has been commonly adopted for OCF diagnosis because of its capability to normalize current measurements. However, implementations using DSPs need extra processing power and time [10].
• Artificial Intelligence-based Methods like ANNs, FL Systems and Deep Neural Networks have been adopted for fault classification. However, although these techniques provide high accuracy, they need considerable computational resources, increasing complexity in real-time execution [20,21,22,23,24,25].
• DWT-based methods have been analysed for feature extraction in fault detection. While effective, these approaches need long processing times and DSP-driven implementation, rendering them less appropriate for real-time applications [18,19].
• Manual methods and model-based techniques such as Bond Graph Models and FFT-based current spectrum analysis offer knowledge about the fault conditions but demand significant computational power and are not well-suited for real-time fault diagnosis [28,29].

Despite advancements, most surviving approaches need DSPs or microcontrollers, increasing hardware complexity, cost and processing time. Additionally, their real-time applicability is limited, particularly in dynamic operating conditions.

1.2. Proposed Approach

To overcome these difficulties, this research shows an innovative online OCFDM using a Hardware-Based Intelligent Single Neuron Approach (HISNA). In contrast to traditional DSP-based solutions, the suggested method eliminates the requirement for extra digital controllers, considerably minimizing processing overhead. A gradient descent algorithm is employed for dynamic computation of the weight and bias values of the single processing neuron, confirming fast and effective fault diagnosis. A high-speed multiplier and adder circuit seamlessly integrate with the neuron, facilitating real-time fault diagnosis in VSIs under variable load conditions.

1.3. Contributions of the Work

The following are the paper’s contributions:

• Removal of DSP/Controller Dependence: In contrast to traditional methods requiring DSPs or microcontrollers, this technique depends on hardware-based computation, minimizing cost and operational complexity.
• Real-Time Fault Diagnosis: The addition of a single processing neuron with a high-speed multiplier and adder enables fast diagnosis of OCFs, making it appropriate for real-time applications.
• Robustness Under Variable Load Conditions: The suggested method maintains high accuracy and stability in OCF diagnosis despite fluctuating load conditions, confirming consistency in practical implementations.

By reducing the difference between hardware simplicity and real-time fault diagnosis, this research offers a groundbreaking alternate to OCF diagnosis techniques, improving efficiency and applicability in industrial environments.

2. Fault Diagnostic Method

The structure of a VSI consists of six transistors, i.e., two transistors for each phase. The supply from the AC mains is converted into DC using a bridge rectifier and filter. This DC input is provided to each of these transistors in a typical switching manner to generate the 3ϕ voltage used to drive the load. The circuit diagram of the 3ϕ inverter is shown in Figure 1. The current signals at the output of each phase of VSI are sensed and converted into voltage using a V-I converter. In the proposed HISNA system, the extracted signal is given by Equation (1).

$${\mathrm{V}}_{\mathrm{p}}={\mathrm{V}}_{\mathrm{m}}\times \mathrm{s}\mathrm{i}\mathrm{n}\left(\mathsf{\omega }\mathrm{t}\right)$$

(1)

where V_p is the output of VSI and p is phase R, Y or B. V_m is the peak value of the signal, t is intntaneous time and ω is the angular frequency. The sensed signals are normalized so that the system can diagnose faults under variable load conditions. In the normalization process, the sensed signals are divided by the peak value of that signal, which is given by Equation (2).

$${\mathrm{V}}_{\mathrm{p}\left(\mathrm{n}\right)}={\displaystyle \frac{{\mathrm{V}}_{\mathrm{m}}\times \left(\mathrm{s}\mathrm{i}\mathrm{n}\left(\mathsf{\omega }\mathrm{t}\right)\right)}{{\mathrm{V}}_{\mathrm{m}}}}$$

(2)

where V_p(n) is the output of the normalizer. The normalized signal is further applied to the processing unit; the processing unit is called a neuron cell. It receives input through the synaptic terminals, with each having a weight associated with it. A single artificial neuron cell has been used for the classification of the faulty switch. In the HISNA system, six neuron cells are implemented for six switches, as shown in Figure 1. A neuron cell classifies faulty switches from the same phase. I_p(n) and bias ($\mathrm{b}$) are two inputs to the neuron cell, with their corresponding weights as W_p and W_0p, respectively.

The neuron cell or processing unit uses the equation to calculate output, as given in Equation (3).

$${\mathrm{Y}}_{\mathrm{p}\left(\mathrm{s}\right)}={\mathrm{W}}_{\mathrm{R}}\times {\mathrm{V}}_{\mathrm{R}\left(\mathrm{n}\right)}+{{\mathrm{W}}_{\mathrm{Y}}\times {\mathrm{V}}_{\mathrm{Y}\left(\mathrm{n}\right)}+{\mathrm{W}}_{\mathrm{B}}\times {\mathrm{V}}_{\mathrm{B}\left(\mathrm{n}\right)}+\mathrm{b}\times \mathrm{W}}_{0\mathrm{p}}$$

(3)

Y_p(s) is the output of the ANN for the individual switch. S is a parameter showing the upper or lower switch. U stands for the upper switch and L for the lower switch in phase. ${\mathrm{W}}_{\mathrm{R}},{\mathrm{W}}_{\mathrm{Y}}\mathrm{a}\mathrm{n}\mathrm{d}{\mathrm{W}}_{\mathrm{B}}$ are the calculated weights of neuron cells of the R, Y and B phases, respectively. ${\mathrm{W}}_{0p}$ is the pre-determined synaptic weights for bias input (b).

The normalized signal at the output of the normalizer is multiplied with a pre-determined synaptic weight. The product obtained has a pre-determined bias added to implement Equation (3). The values of the synaptic weights are calculated by training the ANN using the Delta Learning Rule. The gradient descent algorithm is used to calculate optimum values of parameters by minimizing the cost function (C_F) given by Equation (4).

$${\mathrm{C}}_{\mathrm{F}}={\displaystyle \frac{1}{\mathrm{n}}}\sum _{\mathrm{d}=0}^{\mathrm{n}}{({\mathrm{Y}}_{\mathrm{p}\left(\mathrm{d}\right)}-{\mathrm{Y}}_{\mathrm{w}})}^2$$

(4)

Y_W is the actual output for ${\mathrm{d}}^{\mathrm{t}\mathrm{h}}$ training sample. The term $\left(Y_{p\left(n\right)}-Y_w\right)$ is for calculating the difference between the actual value and fitted value of the ${\mathrm{d}}^{\mathrm{t}\mathrm{h}}$ training example. The reason for squaring is to eliminate the sign earlier than this error, as it can be both positive and negative. The new weights are calculated by using the weight updating equations given in Equations (5) and (6).

$${\mathrm{W}}_{\mathrm{p}}={\mathrm{W}}_{\mathrm{p}}+\mathrm{ɳ}\times {\displaystyle \frac{1}{\mathrm{n}}}\sum _{\mathrm{d}=0}^{\mathrm{n}}({\mathrm{Y}}_{\mathrm{d}(\mathrm{p}\left(\mathrm{s})\right)}-{\mathrm{Y}}_{\mathrm{w}})\times {\mathrm{V}}_{\mathrm{p}\left(\mathrm{n}\right)}$$

(5)

$${\mathrm{W}}_{0\mathrm{p}}={\mathrm{W}}_{0\mathrm{p}}+\mathrm{ɳ}\times {\displaystyle \frac{1}{\mathrm{n}}}\sum _{\mathrm{d}=0}^{\mathrm{n}}\left({\mathrm{Y}}_{\mathrm{d}(\mathrm{p}\left(\mathrm{s})\right)}-{\mathrm{Y}}_{\mathrm{w}}\right)$$

(6)

where ɳ is the learning rate, which is a parameter that decides to what extent the new weights change concerning the old values. The activation function is a function that is used to map the output of ANN Y_P(S) into a response variable based on a decision. The threshold activation function is used to convert the process value ${\mathrm{Y}}_{\mathrm{p}\left(\mathrm{S}\right)}$ into a decision, which is given by Equation (7).

$${\mathrm{T}}_{\mathrm{j}}={\mathrm{f}(\mathrm{Y}}_{\mathrm{P}\left(\mathrm{S}\right)})=\left\{\begin{array}{c}0{\mathrm{Y}}_{\mathrm{p}\left(\mathrm{S}\right)}<\theta \\ 1{\mathrm{Y}}_{\mathrm{p}\left(\mathrm{S}\right)}>\theta \end{array}\right.$$

(7)

J is 1, 2, 3, ..., 6 and θ is the threshold value. The value of $\mathsf{\theta }$ is determined by observing the output ${\mathrm{Y}}_{\mathrm{p}\left(\mathrm{S}\right)}$, which is given by Equation (3). The output of the processing unit is provided as an input to the activation function, which compares it with a threshold value. The output of the activation function is ‘1’ if the fault is present in the switch and ‘0’ for the healthy condition.

3. Implementation

An OCF in the VSI is introduced by opening the collector terminal in the experimental setup for the proposed HISNA system. Such a facility is generated in the test box. The Data Acquisition System (DAS) consists of a VSI, IM and real-time interface. A protection circuit is provided to avoid damage to the IGBTs caused by various faulty conditions. The specifications of VSI used in the experiments are given in

Table 1. Specifications of VSI.

Parameters	Values
DC Link Electrolytic Capacitor	5000 µF
Load Inductance	10 mH
Load Resistor	0.20 Ω
Output AC Voltage	230 Vp
Output Current	3.0630 Amp
Output Frequency	40–70 Hz
Load Power (Variable)	500 W–1.5 kW

. The variable load conditions are implemented in the test bench for its application during the proposed methodology assessment using a pulley wheel mechanism, which helps to increase or decrease the load on the IM. An IM with 0.75 kW, 1415 rpm and 1.8 Amp current rating is used for experimentation. Features like the analog-to-digital converter and Pulse Width Modulation of the DSP processor TMS320C2407 are used. The experimental setup is shown in Figure 2. The 3ϕ AC signal is normalized within the range of ‘±1’ to avoid the effect of load variation.

The normalized signal is used to train the ANN that is implemented by Equation (3). The values of weights W_p and W_0p are calculated using gradient descent algorithm implemented using MATLAB-R2022b, as given below.

Load training data as input vector ${\mathrm{V}}_{\mathrm{p}\left(\mathrm{n}\right)}$ and their expected output vector $Y_w$. Initialize weights (${\mathrm{W}}_{\mathrm{p}}$ and ${\mathrm{W}}_{0\mathrm{p}}$), number of iterations, learning rate (η) and cost function (C_f).

Calculate the values of new weights using Equations (5) and (6).

Calculate the values of ${\mathrm{Y}}_{\mathrm{p}\left(\mathrm{s}\right)}$ using Equation (3).

Calculate the values of the cost function using Equation (4).

Repeat steps 3, 4 and 5 for the given number of iterations or until obtaining the required value of C_f.

A separate processing unit for each switch is implemented to diagnose an OCF in the switch. To implement the HISNA system, healthy condition samples and samples with faulty conditions have been combined. For the information that is collected to diagnose the fault in T1, combinations like T1, T1-T2, T1-T3, T1-T4, T1-T5 and T1-T6 have been used. Such information is collected for other switches. For example, the combinations of T2, T2-T1, T2-T3, T2-T4, T2-T5 and T2-T6. In this way, six sets for the six switches have been created. Different values for ${\mathrm{W}}_{\mathrm{p}}$ and ${\mathrm{W}}_{0\mathrm{p}}$ are calculated for each set. The algorithm above is used to find it. The calculated weights are substituted into Equation (3) to obtain the output.

The training data for the neural network is generated using MATLAB simulation. The ANN model is trained for 1250 iterations with a 0.007 learning rate, which is obtained by trial and error and a 98.00% cost function is achieved. The parameters ${\mathrm{W}}_{\mathrm{p}}$ and ${\mathrm{W}}_{0\mathrm{p}}$ obtained during the implementation of the HISNA system are given in

Table 2. Calculated values of synaptic weights and C_f using gradient descent algorithm.

Switch	WR	WY	WB	W0	Cf (%)
T1	0.36	0.10	0.09	0.58	0.89
T4	0.34	0.06	0.08	0.65	0.90

. By putting these values of weights in Equation (3) and looking at the healthy condition of the switch, the maximum value of ${\mathrm{Y}}_{\mathrm{p}\left(\mathrm{S}\right)}$ for the good condition is 0.86. Therefore, the value of θ for OCF diagnosis is greater than 0.86. The output of the processing unit is compared with the threshold value to decide whether the switch is healthy or faulty using Equation (7). The threshold activation function is implemented for each switch by observing the values of output. The Operational Amplifier (Opamp)-based comparator circuit is used to implement the activation function. The sensed signal was the current parameter (I_P) that has changed to voltage (V_P) using the V-I converter. V_P is divided by the voltages (V_PC) stored in the capacitors. In this case, R_P1, R_P2 and R_P3 were used to fine-tune the sensed signals from VSI, as shown in Figure 3. The values of R_P1, R_P2 and R_P3 are kept, such that the sensed signals given as input to the multiplier will always be between −1 and 1. To obtain normalized values within the range of ±1, V_P is stored in the capacitor. The detail steps of HISNA implementation for IGBT T1 are given below.

Step 1.

Current normalization: For the diagnosis of faults under variable load conditions, 3ϕ currents are normalized within the range of ±1. The peak voltage Vm is stored as V_RC, V_YC and V_BC in the capacitor, as shown in Figure 2. The input voltages V_R, V_Y and V_B are divided by V_RC, V_YC and V_BC, respectively, and normalized like V_RN, V_YN and V_BN within the range of ±1.

Step 2.

Data Collection: The 3ϕ current waveforms are collected under healthy and different faulty conditions. An open circuit fault in the VSI is introduced by the opening collector terminal. Such a facility is generated in the test box. A protection circuit is provided to avoid the damage to the IGBTs caused by the various faulty conditions. A data packet consists of an accumulation of samples for the angular frequency of 360° or one fundamental period of the current cycle. The training data set consists of 5000 samples for healthy and faulty conditions. Such samples are separately collected for six processing units, which are shown in Figure 1. The testing data set consists of 1250 samples.

Step 3.

Weight Calculations: The collected data samples for every processing unit are used to calculate the weights of that processing unit using MATLAB. For every processing unit weight (${\mathrm{W}}_{\mathrm{p}\mathrm{s}}$) and bias (${\mathrm{W}}_{0\mathrm{p}}$) are calculated using Equations (5) and (6).

Step 4.

Multiplier: The normalized voltages, like V_RN, V_YN and V_BN, are multiplied by their respective weights that are calculated in Step 4.

Step 5.

Adder: The result of the multiplication in step 4 is added using adder circuitry, as shown in Figure 2.

Step 6.

Comparator: The processed value in step 5 is converted into the decision of whether the IGBT is healthy or faulty. Such a decision is taken by comparing the processed value with the threshold value as given in Equation (7). By observing the waveforms in Figure 4, Figure 5, Figure 6, Figure 7 and Figure 8, the threshold value (θ) is decided as 0.86. The threshold values, i.e., in this application of fault diagnosis the normalized signals, are used; hence, it is not required to change the threshold value (θ) as per load variation.

4. Results and Discussion

In this section, the results for some faulty conditions are presented to elaborate on the performance of the HISNA system. Analysis of the system is performed with the help of waveforms under different conditions in the time domain.

A.

Single switch open circuit fault (T1):

In this case, an open circuit fault is generated in T1 at time instant T_Fault. It is required to avoid false interpretation of the waveforms of 3ϕ signals. Hence, evaluation of the output is conducted with the help of a processed value. All the waveforms are shown in Figure 5. In this case, a sample value is taken from the normalized 3ϕ current waveform, R = 0.9162, Y = 0. 315, B = 0.8746, giving the processed value ${\mathrm{Y}}_{\mathrm{R}\left(\mathrm{U}\right)}$= 1.0727. As the processed value is greater than the threshold value, it indicates that switch T1 is faulty.

B.

Double switch OCF in the upper part of different legs (T1-T3):

Switches T1 and T3 are present in the upper part of the different legs in an inverter. In an inverter, the fault is introduced in T1 and T3 switches. The time domain waveforms of 3ϕ current in healthy as well as in faulty conditions at different phases are shown in Figure 5. For analysis purposes, sample values from the waveforms of normalized 3ϕ currents are taken at time instant T_Fault, as shown in Figure 5b, R = 0.721, Y = 0.996, B = 0.980. At this point, the processed value for T1 is ${\mathrm{Y}}_{\mathrm{R}\left(\mathrm{U}\right)}$ = 1.0436, as shown in Figure 5c. As this output value is greater than the threshold value Y_out = 1, it indicates transistor T1 is faulty; this is shown in Figure 5d. The same analysis is performed at time instant T_Fault-2, R = 0.887, Y = 0.7631, B = −0.5125. For this processed value, we get ${\mathrm{Y}}_{\mathrm{R}\left(\mathrm{U}\right)}$ = 0.9465. Hence, a fault is indicated with Y_out = 1.

C.

Double switch OCF in the same phase (T1 and T4):

The faults are generated in switches T1 and T4 lay in the same phase. This is a single-phase open fault. The 3ϕ signal, normalized output and processed signal waveforms are shown in Figure 6. For the analysis, a sample value is taken from the normalized 3ϕ current waveform at time instant T_Fault, as shown in Figure 6b, R = 1.0014, Y = −1.1556, B = 0.9881. For this sample, we obtain ${\mathrm{Y}}_{\mathrm{R}\left(\mathrm{U}\right)}$ = 0.9228, which indicates transistor T1 is faulty, as shown in Figure 6c. The same procedure is repeated for R = 0.8248, Y = 0.3526, B = 0.9648, giving the processed value ${\mathrm{Y}}_{\mathrm{R}\left(\mathrm{U}\right)}$ = 1.0149. This shows that switch T1 is defective.

D.

Double switch OCF in different legs (T1 and T6):

In this case, an open circuit fault is generated in T1 and T6 at time instant T_Fault. All the waveforms are shown in Figure 7. In this case, a sample value is taken from the normalized 3ϕ current waveform, R = 0.9162, Y = 0.8746, B = 0. 315, giving a processed value output = 1.0599. As the processed value is greater than the threshold value, it indicates switch T1 is faulty.

E.

OCF diagnosis under variable load conditions:

The time-domain waveform caused by a transient load variation from a light to a heavy load during a T-1 fault condition is shown in Figure 8a. Due to load variation, the amplitude of 3ϕ voltage is changed and generates variations in sensed 3ϕ voltage. The 3ϕ currents are normalized in the range of ±1, as shown in Figure 8b. An unimportant change in the amplitude is observed in load voltage during load variation, which is no sudden rise or fall in the current signal. Hence, the results of HISNA will not be affected. The values of ${\mathrm{Y}}_{\mathrm{R}\left(\mathrm{U}\right)}$ after an IGBT fault are very high compared to the values of the threshold, as shown in Figure 8c.

A comparison of various OCF diagnosis techniques based on key performance parameters is given in Table 3. The PVT is ambiguous at small currents and shows poor resistivity under such conditions, with a detection time of 20.00 ms and requiring medium implementation effort, high tuning effort and high threshold dependence. The phase current-based diagnostic algorithm also performs poorly with small currents with limited resistivity, having a maximum detection time of 63.5% of the electrical period and low implementation effort but a high tuning effort and threshold dependence. The Wavelet Fuzzy Method offers good effectiveness when fuzzy rules are carefully designed, with good resistivity and a detection time of 75.19 ms, though it demands high implementation effort, medium tuning effort and low threshold dependence. The Wavelet-Neural Network method offers a diagnosis error of less than 5% and good resistivity when the neural network is well-trained. However, it requires high implementation effort due to neural network training, low tuning effort and is not threshold dependent. The Hybrid Approach with combiniation PVT and Wavelet-Neural Network gives good effectiveness and resistivity, with a detection time of one cycle and needing high implementation effort, though tuning effort is unspecified. Lastly, the proposed HISNA system achieves good effectiveness and resistivity, with a detection time of less than one cycle, and requires low implementation effort and does not depend on tuning effort or thresholds.

A comparative analysis of various OCFDMs based on accuracy, precision, recall and false positive rate is given in Table 4. A traditional method like PVT shows 85% accuracy and a 12% false positive rate. Advanced approaches significantly improve fault detection: Wavelet-Fuzzy shows 92% accuracy and a 7% false positive rate and Wavelet-Neural Network shows 95% accuracy and a 5% false positive rate. The Hybrid Approach further enhances reliability by combining PVT and Wavelet-ANN, achieving 96% accuracy and a 4% false positive rate. However, the proposed HISNA system surpasses all methods, with 98% accuracy, 96% precision and the lowest false positive rate of 2%, demonstrating its superior effectiveness in fault diagnosis.

A comparison of various FDMs based on the processor or controller used is given in Table 5. The FPGA implementation of AI-based inverter IGBT OCF diagnosis of IM drives (2022) shown in [30] employs an FPGA (Xilinx Zynq-7000 SoC) for AI-driven fault diagnosis, reducing reliance on DSPs. In contrast, the joint fault diagnosis of IGBT and current sensor in LLC resonant converter module based on reduced order interval sliding mode observer (2024) shown in [31] uses a DSP (Texas Instruments TMS320F28335), which introduces additional processing time and power consumption. Similarly, the two-step process-based OCF diagnosis for three-level NPC converters (2025) shown in [32] also utilizes the TMS320F28335 DSP but suffers from higher processing time, increased power consumption, complex implementation and limited scalability, making it less efficient for real-time, low-cost fault diagnosis. The proposed hardware-based approach eliminates the need for a DSP or controller by utilizing a single neuron-based processing circuit, offering cost and speed benefits.

Table 3. A comparison of various OCFDMs.

Methods	Effectiveness	Resistivity	Detection Time	Implementation Effort	Tuning Effort	Threshold Dependence
Park’s Vector method [10]	Ambiguous at small currents	Poor at small currents	20.00 ms	Medium	High	High
Phase current-based diagnostic algorithm [33]	Poor at small current	Poor	maximum of 63.5%	Low	High	High
Wavelet fuzzy Method [34]	Good if the fuzzy rules are carefully designed	Good	75.19 ms	High	Medium	Low
Wavelet-Neural Network [35]	Diagnosis error < 5%	Good if NN is thoroughly trained	-	High due to NN training	Low	N/A
Hybrid Approach (Park’s Vector method and Wavelet-Neural N/w) [17]	Good	Good	1 Cycle	High	-	N/A
Proposed HISNA system	Good	Good	Less than 1 Cycle	Low	-	N/A

Table 3. A comparison of various OCFDMs.

Methods	Effectiveness	Resistivity	Detection Time	Implementation Effort	Tuning Effort	Threshold Dependence
Park’s Vector method [10]	Ambiguous at small currents	Poor at small currents	20.00 ms	Medium	High	High
Phase current-based diagnostic algorithm [33]	Poor at small current	Poor	maximum of 63.5%	Low	High	High
Wavelet fuzzy Method [34]	Good if the fuzzy rules are carefully designed	Good	75.19 ms	High	Medium	Low
Wavelet-Neural Network [35]	Diagnosis error < 5%	Good if NN is thoroughly trained	-	High due to NN training	Low	N/A
Hybrid Approach (Park’s Vector method and Wavelet-Neural N/w) [17]	Good	Good	1 Cycle	High	-	N/A
Proposed HISNA system	Good	Good	Less than 1 Cycle	Low	-	N/A

Table 4. Statistical performance comparison of OCFDMs.

Method	Accuracy (%)	Precision (%)	Recall (%)	False Positive Rate (%)
Wavelet-Fuzzy Method [36]	92	90	89	7
Wavelet-Neural Network [35]	95	93	91	5
Hybrid Approach (Park’s Vector + Wavelet-NN) [17]	96	94	92	4
Proposed HISNA System	98	96	95	2

Table 4. Statistical performance comparison of OCFDMs.

Method	Accuracy (%)	Precision (%)	Recall (%)	False Positive Rate (%)
Wavelet-Fuzzy Method [36]	92	90	89	7
Wavelet-Neural Network [35]	95	93	91	5
Hybrid Approach (Park’s Vector + Wavelet-NN) [17]	96	94	92	4
Proposed HISNA System	98	96	95	2

Table 5. Comparison of OCFDMs based on processor or controller used.

Paper/Study	Processor or Controller Used	Remarks
FPGA Implementation of AI-Based Inverter IGBT Open Circuit Fault Diagnosis of IM Drives (2022) [30]	FPGA (Xilinx Zynq-7000 SoC)	FPGA-based implementation for AI-driven fault diagnosis, reducing reliance on DSPs.
Joint Fault Diagnosis of IGBT and Current Sensor in LLC Resonant Converter Module Based on Reduced Order Interval Sliding Mode Observer (2024) [31]	DSP (Texas Instruments TMS320F28335)	DSP-based implementation; requires additional processing time and power.
Two-Step Process-Based OCF Diagnosis for Three-Level NPC Converters (2025) [32]	DSP (Texas Instruments TMS320F28335)	Higher processing time, power consumption, complex implementation, limited scalability, making it less efficient for real-time, low-cost fault diagnosis.
Proposed Hardware-Based Approach	No DSP or Controller Used	Uses a single neuron-based processing circuit, eliminating DSP/controller dependency for cost and speed benefits.

Table 5. Comparison of OCFDMs based on processor or controller used.

Paper/Study	Processor or Controller Used	Remarks
FPGA Implementation of AI-Based Inverter IGBT Open Circuit Fault Diagnosis of IM Drives (2022) [30]	FPGA (Xilinx Zynq-7000 SoC)	FPGA-based implementation for AI-driven fault diagnosis, reducing reliance on DSPs.
Joint Fault Diagnosis of IGBT and Current Sensor in LLC Resonant Converter Module Based on Reduced Order Interval Sliding Mode Observer (2024) [31]	DSP (Texas Instruments TMS320F28335)	DSP-based implementation; requires additional processing time and power.
Two-Step Process-Based OCF Diagnosis for Three-Level NPC Converters (2025) [32]	DSP (Texas Instruments TMS320F28335)	Higher processing time, power consumption, complex implementation, limited scalability, making it less efficient for real-time, low-cost fault diagnosis.
Proposed Hardware-Based Approach	No DSP or Controller Used	Uses a single neuron-based processing circuit, eliminating DSP/controller dependency for cost and speed benefits.

5. Conclusions

In this work, a single neuron is constructed using basic electronics components. The delta learning rule was used to find the values of the weights and biases. Using this algorithm, three weights and one bias are calculated for each switching device. The current signal is converted into voltages using a V-I converter. Normalized 3ϕ voltages are given as an input to a single neuron model. This paper explains how all the training samples needed to train a single neuron were collected. The built-in single neuron model was tested with all combinations of switching devices like single switch faults, double switch faults in the same leg, and double switch faults in different legs. The proposed HISNA system outperforms the existing open switch fault diagnosis methods by offering good effectiveness and resistivity with a fast detection time. Unlike the Hybrid Approach, which requires high implementation effort, HISNA achieves similar diagnostic accuracy with lower effort. Traditional methods like Park’s Vector and Slope suffer from poor resistivity and longer detection times, while ANN-based techniques demand intensive training. HISNA ensures efficient multiple switch fault diagnosis without threshold dependence. The proposed system diagnoses faults in less than one cycle with 98% accuracy, 96% precision and 95% recall, maintaining a 2% false positive rate. It eliminates DSP or controller dependency using a single neuron-based circuit, reducing costs and improving real-time fault diagnosis. The proposed system effectively diagnoses OCFs with minimal hardware requirements. However, expanding its adaptability to diverse fault conditions and scaling for high-power applications could enhance its applicability. Future work may explore AI integration and adaptive learning for improved performance.