1. Introduction

Arc faults are among the major causes of electrical fires in power systems and can arise from aged wiring, poor contacts, insulation degradation, and other defects. When conductor separation is caused by a broken wire, corrosion, loose crimping, or an imperfect connection, the local air insulation may break down and form a high-temperature arc plasma. Because such faults can damage power equipment and electronic devices and may ultimately lead to fire, early detection is essential. Rapid arc-flash detection and mitigation have likewise been recognized as important issues in power-equipment protection. Arc and insulation safety are also major design considerations for low-voltage electrical apparatus, while online diagnostic methods have been investigated for predicting arc faults in medium-voltage switchgear.

A variety of arc-fault detection methods have been reported. Discrete wavelet transforms have been used to extract time-frequency features from current signals ^[1], and the characteristics and inception conditions of AC series arcing have been investigated in more-electric-aircraft systems ^[2]. Time- and frequency-domain approaches have also been applied to the detection of series and parallel AC arcs ^[3]. Feature-engineering- based approaches can depend heavily on manual feature design. Data-driven arc detection has also been explored by converting current waveforms into time-domain images and applying convolutional neural networks (CNNs) ^[4]. Other methods have combined current-amplitude spectra with sparse representation ^[5], or integrated time- and frequency-domain features through fully connected neural networks ^[6]. High-frequency components, machine-learning methods, and CNN-based approaches have also been investigated for arc-fault and fault diagnosis under changing operating conditions. Related studies have extended arc-fault diagnosis to DC systems. An FFT-based frequency-band analysis combined with RMS current variations has been proposed for DC series-arc detection ^[7].

Output-side AC series arcs in a three-phase PWM inverter are simultaneously affected by the load and control conditions. A change in the fundamental output frequency alters both the duration of one electrical cycle and the locations of the dominant spectral components. As the load inductance increases, current continuity can suppress the interruption signature near the zero crossing. Changes in the switching frequency also redistribute the harmonic components. Consequently, a detector that monitors only a fixed frequency component or applies a single threshold is unlikely to provide stable performance over all operating conditions.

This paper therefore proposes a ResNet-based algorithm that diagnoses output-side AC series arcs using only the output AC current. The primary contribution is the achievement of high diagnostic performance even for weak arcs, whose waveform changes are considerably smaller than those of strong arcs. The proposed architecture also outperforms the comparison models on the same dataset. In addition, its robustness is evaluated under variations in the output fundamental frequency, switching frequency, load inductance, and arc intensity.

2. Experimental Setup and Arc-Current Data Acquisition

To reproduce an output-side AC series arc, an arc generator was inserted between one output phase of a three-phase PWM inverter and a resistor-inductor load. The generator consisted of two electrodes housed in a transparent chamber, with a servo motor used to open the electrical contact in fine increments. Output current was measured using current probes and an oscilloscope. Only the output AC-current signal was used as the input to the artificial-intelligence model. The arc generator was connected in series between one inverter output phase and the corresponding load branch. Each phase was connected to a series resistor-inductor load. Figure 1 presents the physical test setup, comprising the three-phase inverter, RL load, arc generator, and measurement equipment.

The electrical contact inside the transparent arc chamber was gradually separated by the servo motor to generate a controlled series arc. During each test, the currents of the arcing phase and an adjacent healthy phase were measured simultaneously. The diagnostic model developed in this study uses only a single output AC current signal. Figure 2 shows photographs of the discharge between the electrodes. A bright white emission was observed for the strong arc, whereas the weak arc exhibited a dim red emission. Because arc-light intensity has been reported to be related to arc temperature ^[8], these images provide qualitative evidence of differences in the discharge conditions. Strong and weak arcs were distinguished using the arc power and electrode-gap opening condition. For the strong-arc condition, the electrode gap was opened by 0.02 mm at a speed of 0.02 mm/s after the arc-inception gap had been reached. For the weak-arc condition, the gap was opened by 0.002 mm at a speed of 0.002 mm/s. The calculated arc power was greater than 20 W for strong arcs and ranged from 10 to 18 W for weak arcs.

Fig. 1 Experimental setup

Fig. 2 Arc discharge between the electrodes: (a) strong arc and (b) weak arc

The experimental dataset consisted of one normal operating condition and 15 output-side series-arc conditions. Normal data were acquired without actuating the arc generator and were segmented using the same windowing procedure as the arc data, yielding 26,344 normal samples. For the arc data, the series-arc generator was placed between one inverter output phase and the RL load, and the current of the arcing phase and that of an adjacent healthy phase were recorded simultaneously. Strong and weak arcs were labeled according to the electrode-gap and arc-power criteria.

The strong-arc dataset was designed to reflect the effects of the output fundamental frequency, switching frequency, and load inductance on the observed arc waveform. The baseline operating point was defined by a 60-Hz output fundamental frequency, a 10-kHz switching frequency, and a 10-mH load inductance. Around this baseline, the output fundamental frequency varied among 30, 60, 90, and 120 Hz while the switching frequency and load inductance were held at 10 kHz and 10 mH, respectively. In the switching-frequency tests, the switching frequency was changed to 5, 15, and 20 kHz while the output fundamental frequency and load inductance were maintained at 60 Hz and 10 mH. Load-inductance values of 2, 10, 15, and 18 mH were included to examine how current continuity affects the arc signature near the zero crossing.

The weak-arc dataset was acquired to assess whether the model could distinguish the arc state when waveform distortion and spectral variation were substantially smaller than in the strong-arc case. During the weak-arc tests, the output fundamental frequency, switching frequency, and load inductance were fixed at 60 Hz, 10 kHz, and 10 mH, respectively. The complete dataset contained 29,904 arc samples in addition to the 26,344 normal samples. Most individual arc conditions contributed 2,136 windows; for several strong-arc conditions, 1,424 windows were retained based on the available stable acquisition interval. For model evaluation, the test set was fixed at 142 samples per experimental condition. Because the dataset comprised one normal operating condition and 15 AC-arc conditions, the test set contained 142 normal samples and 2,130 AC-arc samples, for a total of 2,272 samples. All models were evaluated using this same fixed test set for both the time-domain and frequency-domain inputs.

3. Analysis of Output-Current Characteristics under Arc-Fault Conditions

An AC series arc is extinguished as the alternating voltage approaches zero and reignites after the voltage rises again. This repeated extinction and reignition can produce a brief interruption, a localized slope change, or shoulder-like distortion in the current waveform near the zero crossing ^[2]. In a strong arc, these zero-crossing signatures are relatively pronounced and are accompanied by a reduction in current magnitude and changes in the harmonic content. Figure 3(a) shows the phase-A and phase-B currents during normal operation, whereas Figure 3(b) shows the strong arcing phase A and the adjacent healthy phase B during the same test interval.

Fig. 3 Output AC-current waveforms in the time domain for the strong-arc test: (a) normal operation and (b) strong arc

Following arc inception, the phase-A current exhibits a lower overall amplitude and a localized shoulder near the zero-crossing relative to its normal waveform and the healthy-phase current. This behavior is consistent with the characteristic extinction and reignition process of an AC series arc. Accordingly, the normal and arcing states can be differentiated relatively clearly in the time domain under the strong-arc condition. Figure 4 compares the output-current spectra of the strong-arcing phase A and the adjacent healthy phase B acquired during the same test interval.

Fig. 4 Frequency-domain comparison of the output-current spectra of the strong-arcing phase A and the adjacent healthy phase B

The spectrum is presented from 0 to 1 kHz, and the region around the 60-Hz fundamental component is enlarged in the inset. At the fundamental frequency, the magnitude of the strong-arcing phase A is approximately 1.8 A lower than that of the healthy phase B. This pronounced interphase difference indicates that the strong arc produces a readily distinguishable spectral signature.

By contrast, the waveform change caused by a weak arc is smaller, and the shoulder near the zero crossing is far less distinct. Figure 5(a) shows the phase currents during normal operation, whereas Figure 5(b) shows the weakly arcing phase A and the adjacent healthy phase B. Although the phase-A current decreases after arc inception, the abrupt interruption and shoulder-like distortion observed in the strong-arc case are only weakly expressed. Thus, classification based solely on a local time-domain signature may be unreliable for weak arcs. Figure 6 compares the output-current spectra of the weak-arcing phase A and the adjacent healthy phase B acquired during the same test interval. As in Figure 4, the 0-1-kHz spectrum is shown with an enlarged view of the 60-Hz fundamental component. The magnitude difference between the weak-arcing phase A and the healthy phase B is less than 0.1 A, which is substantially smaller than the approximately 1.8-A interphase difference observed in the strong-arc test. Although the spectral difference is also subtle, the frequency domain enables small changes in the fundamental and neighboring components to be quantified more directly than the raw waveform. The reduced zero-crossing shouldering under the weak-arc condition makes the arc and normal currents less distinguishable in the time domain. Therefore, a frequency-domain representation that preserves changes distributed across the spectrum is expected to provide more reliable diagnostic features than the raw time-domain waveform.

Fig. 5 Output AC-current waveforms in the time domain for the weak-arc test: (a) normal operation and (b) weak arc

Fig. 6 Frequency-domain comparison of the output-current spectra of the weak-arcing phase A and the adjacent healthy phase B

Arc-current characteristics were also examined under changes in operating and hardware conditions. As the output fundamental frequency increases, the duration of one electrical cycle decreases, and the dominant spectral components shift accordingly. A detector that monitors only fixed frequency bands may therefore be vulnerable to operating-point changes. Figure 7 compares strong-arc current waveforms at 60 and 120 Hz. In each case, the arcing-phase current is shown together with the current of an adjacent healthy phase to illustrate the change in the time-domain arc signature caused by the output-frequency variation. An increase in load inductance increases current continuity and attenuates the zero-crossing shouldering associated with arc extinction and reignition.

Fig. 7 Strong-arc current waveforms at different fundamental output frequencies: (a) 60 Hz and (b) 120 Hz

Fig. 8 Strong-arc current waveforms at different load inductances: (a) 2 mH and (b) 18 mH

Fig. 9 Strong-arc current waveforms at different switching frequencies: (a) 5 kHz and (b) 20 kHz

Figure 8 compares the strong-arc currents for load inductances of 2 and 18 mH. At 18 mH, the shouldering near the zero crossing is reduced compared with that at 2 mH, making the arc signature less distinctive in the time domain. Like the weak-arc condition, the reduced time-domain distortion limits waveform-based discrimination and supports the use of frequency-domain features for robust diagnosis under varying load conditions. The switching frequency affects the output-current ripple and harmonic distribution. Figure 9 compares strong-arc current waveforms at switching frequencies of 5 and 20 kHz. Because arc inception alters components around the switching frequency as well as the fundamental and its harmonics, the training data must include switching-frequency variations to support robust diagnosis.

4. Proposed Diagnostic Algorithm

The final input to the proposed algorithm is a frequency-domain representation derived from the output AC current. The measured current waveform is first divided into fixed-length windows, after which an FFT is applied to each window. The resulting magnitude spectrum is normalized and supplied to the proposed Ghost-ResNet, which classifies each sample as either normal or AC arc. The time-domain waveforms in Section 3 were used to interpret the physical changes associated with arc inception, while the time- and frequency-domain inputs are compared in Section 5. The frequency-domain input, which produced the higher overall accuracy, was used for the final performance analysis. As illustrated in Figure 10, the model receives a single output AC-current signal, learns FFT-based spectral features, and produces a final normal-or-arc classification through a softmax layer. A residual-block-based CNN was adopted to capture both local spectral patterns and the global distribution of the input. To improve computational efficiency, the initial Conv2D layer of the baseline ResNet was replaced by a Ghost module. The Ghost module reduces feature redundancy by first generating intrinsic features through a 1 x 1 Conv2D operation and then producing additional feature maps at low computational cost using a 3 x 3 depthwise convolution. The intrinsic and ghost features are concatenated to obtain the required number of output channels ^[9]. Because the initial convolution operates on the largest input tensor, reducing its cost can substantially lower the computational burden of the full network. In the proposed model, the initial 3 x 3 Conv2D layer with 16 output channels is replaced by the Ghost module. Three subsequent residual blocks are retained to learn the subtle spectral variations associated with arc inception in a hierarchical manner. Each residual block consists of Conv2D, batch normalization, and Leaky-ReLU operations. Max pooling is applied between blocks, and the number of channels is progressively increased from 16 to 32 and then to 64. The residual connections help preserve small spectral perturbations during training and allow the network to capture both localized frequency changes and global spectral-pattern variations ^[10]. A softmax output layer performs the final classification. This architecture is particularly suitable for weak arcs because residual feature learning mitigates the loss of small arc-related perturbations during feature extraction, while the Ghost module reduces the computational cost of the initial feature extractor.

Fig. 10 Proposed Ghost-ResNet-based AC series-arc diagnostic framework: (a) overall architecture, (b) Ghost module, and (c) residual block

For performance comparison, the proposed model and three conventional machine-learning or hybrid models were evaluated using the same data. All models used inputs derived from the same output AC-current records, enabling direct comparison between classifiers that operate on frequency-domain features and models that employ deep feature extraction. The support vector machine (SVM) was selected as a representative margin-based classifier, and the random forest (RF) was used as a nonlinear decision-tree ensemble. The CNN+RF model combines CNN-based feature extraction with an RF classifier. The proposed Ghost-ResNet uses the Ghost module and residual blocks to hierarchically learn the subtle spectral changes produced by weak arcs.

5. Results and Discussion

This section compares the diagnostic accuracy obtained from time- and frequency-domain current inputs and analyzes the precision and recall of each model. All performance metrics reported in Figs. 11-13 were calculated using the fixed test set described in Section 2, comprising 2,272 samples, including 142 normal samples and 2,130 AC-arc samples. Because the test set is class-imbalanced owing to the inclusion of one normal condition and 15 AC-arc conditions with the same number of test samples assigned to each experimental condition, class-wise precision and recall are reported together with the overall accuracy. Accuracy represents the overall classification performance, whereas precision and recall reveal the detailed misclassification tendencies for the normal and AC-arc classes. Figure 11 compares the diagnostic accuracies of the four models for the frequency- and time-domain inputs. On the fixed test set of 2,272 samples described above, the proposed Ghost-ResNet achieved the highest accuracy of 96.89% with the frequency-domain input. The corresponding accuracies of the SVM, RF, and CNN+RF models were 83.05%, 82.72%, and 94.21%, respectively. With the time-domain input, the SVM, RF, CNN+RF, and Ghost-ResNet models achieved accuracies of 80.47%, 87.54%, 87.30%, and 95.88%, respectively. The SVM and RF directly classify FFT features and showed lower frequency-domain accuracies than the deep-feature-based models. CNN+RF benefited from CNN-based feature extraction and achieved 94.21%, but its accuracy remained lower than that of Ghost-ResNet. These results show that the proposed architecture achieved the highest accuracy among the evaluated models for both input domains, with its best performance obtained from the frequency-domain representation. Figure 12 compares the class-wise precision and recall values obtained from the same fixed test set for the frequency-domain input. Ghost-ResNet achieved a normal-class precision of 0.7038, an AC-arc precision of 0.9909, a normal-class recall of 0.8662, and an AC-arc recall of 0.9757. Compared with the other models, Ghost-ResNet provided the most balanced classification performance while maintaining particularly high precision and recall for the AC-arc class. These results indicate that the proposed model can effectively distinguish AC-arc samples, including weak-arc cases, from normal operating samples. Figure 13 presents the corresponding class-wise results obtained from the same fixed test set for the time-domain input. Ghost-ResNet achieved a normal-class precision of 0.6221, an AC-arc precision of 0.9908, a normal-class recall of 0.8662, and an AC-arc recall of 0.9649. Although the model maintained high AC-arc precision and recall, its normal-class precision was lower than that obtained with the frequency-domain input. The frequency-domain representation therefore provided more consistent discrimination between the normal and AC-arc classes. Considering the overall accuracy and class-wise metrics, the proposed model performed more consistently with the frequency-domain input. As shown in Figs. 5 and 8, weak arcs and increased load inductance reduce the zero-crossing shouldering that distinguishes arc and normal currents in the time domain. The FFT input instead represents small changes distributed across the fundamental and neighboring spectral components. Consequently, the frequency-domain input achieved a higher diagnostic accuracy of 96.89%, compared with 95.88% for the time-domain input, and improved the normal-class precision from 62.21% to 70.38%. In addition, replacing the initial convolution of the baseline ResNet with the Ghost module reduced the number of multiply-accumulate operations from 223 million to 78 million and decreased the FLOP count from 446 million to 157 million.

Fig. 11 Diagnostic accuracy of the proposed Ghost-ResNet and comparison models for frequency-domain and time-domain inputs

Fig. 12 Class-wise precision and recall of the proposed Ghost-ResNet and comparison models for the frequency-domain input

Fig. 13 Class-wise precision and recall of the proposed Ghost-ResNet and comparison models for the time-domain input

6. Conclusion

This paper proposed a Ghost-ResNet-based method for detecting output-side AC series arcs in a three-phase PWM inverter using a single output-current signal. Strong arcs produced distinct zero-crossing shouldering, whereas weak arcs exhibited only slight shouldering. Increased load inductance also improved current continuity and reduced the zero-crossing shouldering, making the arc and normal currents less distinguishable in the time domain. These characteristics support the use of frequency-domain inputs that capture small changes distributed across the fundamental and neighboring spectral components. Using the FFT magnitude spectrum, the proposed Ghost-ResNet achieved an accuracy of 96.89%, an AC-arc precision of 99.09%, and an AC-arc recall of 97.57%. The frequency-domain input provided higher diagnostic accuracy and normal-class precision than the time-domain input under varying operating and load conditions. These results demonstrate the effectiveness of the proposed method for detecting weak AC series arcs when time-domain arc signatures are attenuated by weak discharge conditions or increased load inductance.