1. Introduction
DC power networks are presently finding widespread use as distribution systems in a variety of applications, including data centers, electric vehicles, aircraft, spacecraft, and more. The adoption of DC networks is driven by their potential advantages, including a reduced need for conversion stages and fewer transmission lines. These attributes collectively contribute to an enhanced overall efficiency and reduced space requirements [
1,
2]. Nevertheless, the occurrence of series high-impedance faults remains a prominent challenge that jeopardizes the safe operation of these networks. Within a DC network, there exists a vast network of cables and junction connections. The potential for DC arc faults arises from factors such as ionization transpiring within the air gaps positioned between conductors, constituting a prevalent occurrence within both alternating current (AC) and DC systems. Series arc faults can arise due to various factors, including the vibration of loosely connected or deteriorated terminal junctions, ruptures within electrical circuits, the aging of cables, the wear and tear of conductive components, inadequate maintenance practices, and contamination by substances such as fluids and chemicals [
3,
4,
5].
DC arc faults can manifest in both series and parallel fault scenarios. Detecting a parallel arc fault is relatively straightforward due to a substantial current shift. In contrast, identifying a series arc fault poses a more intricate challenge due to its minimal current fluctuation, heightening the complexity of fault detection in comparison to parallel arc faults [
6]. In contrast to an AC arc fault, a DC arc fault lacks a current zero-crossing point, which means it can persist for a more extended duration, potentially resulting in severe system failures. Moreover, a series arc fault introduces additional arc resistance into the system, causing a reduction in the loop current. Consequently, the fault current fails to activate the conventional overcurrent protection devices [
7]. Arc faults result in energy losses, and this lost energy is transformed into heat energy, which can, in turn, contribute to fire accidents within power systems. DC arc fault detection methods primarily rely on analyzing the physical and electrical characteristics of the arc. Arc faults can be identified through the examination of physical attributes, including sound, light, and electromagnetic radiation signals [
8,
9,
10,
11,
12,
13]. Implementing these detection techniques necessitates dedicated sensors and data acquisition systems. For example, acoustic sensors [
9] and infrared meters [
10] can be strategically positioned near the switchgear to identify arcs resulting from switching operations. Antennas are employed to capture the electromagnetic radiation signals generated by arcs, enabling arc fault detection in photovoltaics (PV) systems [
11,
12,
13].
Several methods for detecting DC series arcs have been investigated, encompassing current sensing, voltage sensing, and physical change detection. Current sensing techniques involve identifying series arc anomalies by monitoring the PV string current using noise sensors and scrutinizing the current’s noise profile during an arc event. Research has explored methods for arc noise analysis and detection in the frequency domain, leveraging the fast Fourier transform (FFT) technique. In [
14], a comparison of FFT waveforms in terms of arc presence is presented, while ref. [
15] demonstrates the application of feature extraction to arc detection. Additionally, ref. [
16] introduces an arc detection algorithm that employs FFT analysis and is implemented in a microcontroller for real-time applications. However, it is worth noting that FFT analysis has limitations, including the requirement for more time to transform the time–domain signals into frequency–domain signals and an inability to analyze the time–domain characteristics effectively. Modern research endeavors have increasingly turned to artificial-intelligence-driven methods for fault detection and diagnosis, a trend that has proven highly beneficial in the context of arc fault diagnosis [
17,
18].
This study leverages these advanced techniques within the domain of DC arc fault diagnosis, yielding noteworthy and affirmative outcomes. Specifically, the research employs a fusion of an empirical-rule-based filtering technique and artificial intelligence learning to diagnose DC arc faults. The process unfolds as follows: during each sampling interval, the arc current signal is extracted and subsequently subjected to rigorous analysis. Next, the empirical-rule-based filtering process is applied to remove undesirable signal components. Subsequently, the filtered signal serves as input for artificial intelligence learning (AIL). The diagnostic outcomes confirm the effectiveness of this proposed detection scheme, particularly in scenarios characterized by low switching frequencies, where it significantly enhances detection accuracy.
This paper is structured as follows:
Section 2 provides an overview of the experimental setup, detailing variations in current characteristics in the time domain during both normal and arcing phases. In
Section 3, we delve into the AIL algorithms used for arc fault detection and explore the empirical filtering techniques incorporated into this study.
Section 4 systematically presents the results of the arc fault detection process, employing four AIL algorithms in conjunction with the filtering technique. This evaluation encompasses scenarios with varying current amplitudes and operational frequencies. Finally,
Section 5 summarizes the key findings and insights derived from AIL-driven arc fault detection, bringing this study to a close.
4. Intelligence Diagnosis of Series DC Arc Fault with Empirical Filtering
Figure 10 presents a schematic block diagram outlining the conceptual framework for the diagnosis of arc faults. The continuous current data undergo a precise sampling procedure and are subsequently partitioned into subdatasets, each containing 200 data points. These individual subsets of data then undergo a specialized filtering process, meticulously designed to remove data points that fall outside predefined ranges. To elucidate further, the data are initially sampled at a rapid rate of 250 kHz and segmented into smaller datasets, each with a duration of 0.8 ms, amounting to 200 data values per subset. Following this initial segmentation, the empirical filtering technique is meticulously applied, resulting in smoother and clearer signals. These refined signals subsequently serve as inputs for the AIL models within the context of DC arc fault diagnosis. It is important to note that the dataset encompasses two distinct current amplitudes, namely 5 A and 8 A, and four different switching frequencies, which include 5, 10, 15, and 20 kHz. Consequently, there are a total of eight distinct cases, with each case utilizing 3000 datasets for training and 1000 datasets for testing, resulting in a cumulative dataset of 24,000 for training and 8000 for testing. It is noteworthy that the distribution ratio between the normal and arc instances is meticulously maintained at a 1:1 proportion for both the training and testing phases. On the other hand, the ripple components increase with the increase in the switching frequency regardless of whether empirical filtering is employed or not, as shown in
Figure 5,
Figure 6,
Figure 7 and
Figure 8. It is noteworthy that the presence and magnitude of ripple components in the current signals could have a notable impact on the performance of the AIL models, especially at higher switching frequencies. Ripples introduce noise into the current signal. This noise may not be very significant at lower switching frequencies, but as the switching frequency increases, the amplitude of these ripples becomes more prominent. This can make it harder for AIL models to discern the underlying patterns in the data, as the ripples can mask the relevant information. The influence of ripple components can result in the AIL models making incorrect predictions or diagnoses, especially when the ripples coincide with the frequency ranges associated with fault signatures.
In the evaluation of AML algorithms, various metrics are employed as the primary criteria. The
Accuracy detection rate, in particular, quantifies the proportion of correctly predicted datasets relative to the total number of test datasets. It is expressed as
The
False Detection rate is calculated as the ratio of normal state datasets that are incorrectly predicted as the arcing state to the total number of normal state datasets. It is obtained as follows
Conversely, the
Missing Detection rate is determined by the ratio of arcing state datasets that are inaccurately predicted as the normal state to the total number of arcing state datasets. It is determined as follows
Figure 11 and
Figure 12 depict the arc failure diagnosis rates for a three-phase inverter operating at 5 A and 8 A, respectively, across various switching frequencies. These rates are determined using empirical filtering and SVM. When the empirical filtering signals within the three σ ranges are utilized as inputs, there is a noticeable improvement in the accuracy of the fault detection compared to using the raw signals alone. Moreover, as the empirical filtering ranges are further refined to two and one σ ranges, the diagnosis rates for arc failure show additional enhancements. As elucidated in the filtering section, narrowing down the filtering ranges to two and one σ ranges renders the distinctions between normal and arcing states more discernible than the broader three
σ range, thereby resulting in improved diagnosis accuracies. Furthermore, the adoption of the empirical filtering process concurrently leads to a reduction in false detection rates. This, in turn, mitigates false alarms, thereby bolstering the overall stability of the system. On the other hand, the diagnosis accuracies are highest at a 10 kHz switching frequency and gradually decrease with an increase in the switching frequency. The impact of ripple components is tied to the switching frequency and remains consistent, whether empirical filtering is utilized or not. While these ripples might be relatively inconspicuous at lower switching frequencies, their significance amplifies with increasing switching frequencies. As the switching frequency rises, so does the amplitude of these ripples. This heightened presence of ripples poses a notable challenge for AIL models. It complicates the task of uncovering the core data patterns since the ripples have the potential to overshadow the relevant information. This influence of ripple components can lead to inaccurate predictions or diagnoses being made by the AIL models, undermining their effectiveness in scenarios with higher switching frequencies.
Figure 13 and
Figure 14 illustrate the arc failure diagnosis rates for a three-phase inverter operating at 5 A and 8 A, respectively, across various switching frequencies. These rates are determined using empirical filtering in conjunction with an RF. The utilization of empirical filtering signals within the three σ ranges as inputs yields a discernible enhancement in the accuracy of arc fault detection when compared to relying solely on raw signals. Furthermore, as the empirical filtering ranges are further refined to two and one σ ranges, there are additional improvements in the diagnosis rates for arc failure.
Figure 15 and
Figure 16 illustrate the arc failure diagnosis rates for a three-phase inverter operating at 5 A and 8 A, respectively, across various switching frequencies. These rates are determined using empirical filtering in conjunction with the KNN algorithm. When the empirical filtering signals within the three σ ranges are utilized as inputs, a notable improvement in the accuracy of arc fault detection is observed compared to using the raw signals alone. Furthermore, as the empirical filtering ranges are further refined to two and one σ ranges, there are additional enhancements in the diagnosis rates for arc failure.
Figure 17 and
Figure 18 provide a visual representation of the arc failure diagnosis rates for a three-phase inverter operating at 5 A and 8 A, respectively, across different switching frequencies. These rates have been computed by employing empirical filtering in tandem with the NB algorithm. Similarly to the mentioned AIL models, it is evident that when the empirical filtering signals within the three σ ranges are employed as inputs, a marked improvement in the accuracy of arc fault detection is discernible when compared to using the raw signals in isolation. Furthermore, as we further fine-tune the empirical filtering ranges to two and one σ ranges, we observe additional enhancements in the diagnosis rates for arc failure.
Figure 19 and
Figure 20 provide a visual representation of the arc failure diagnosis rates for a three-phase inverter operating at 5 A and 8 A, respectively, across different switching frequencies. These rates have been computed using empirical filtering in tandem with the DT algorithm. It is evident that when the empirical filtering signals within the three σ ranges are employed as inputs, a marked improvement in the accuracy of arc fault detection is discernible when compared to using the raw signals in isolation. Furthermore, as we further fine-tune the empirical filtering ranges to two and one σ ranges, we observe additional enhancements in the diagnosis rates for arc failure.
Figure 21 illustrates the comprehensive accuracy of the AIL models in detecting DC arc faults. Remarkably, among the inputs, the employment of empirical filtering with a one σ range yields the highest accuracies across all AIL models and switching frequencies. Notably, all the empirical filtering signals outperform the raw signals in terms of accuracy. Additionally, when considering the AIL models, an RF consistently emerges as the top-performing model, consistently achieving the highest diagnosis rates regardless of the inputs or switching frequencies. These diagnostic results unequivocally validate the efficacy of the proposed DC arc failure diagnosis scheme.
5. Conclusions
This research has successfully introduced an efficient and less complex diagnostic scheme for DC arc failure. The implementation of the empirical filtering procedure notably enhances signal clarity, resulting in more pronounced and distinct manifestations of arc distortions. This improvement in signal quality holds the potential to enhance the effectiveness of AIL models in arc diagnosis. The diagnostic results provide strong evidence for the efficiency of the proposed detection scheme. The analysis of the probability distribution of raw signals reveals an overlapping region between normal and arcing states, which can potentially lead to confusion when directly classifying these states. Using such mixed data for classification can reduce accuracy and result in incorrect predictions being made by machine learning classifiers. To address this issue, an empirical filtering process is applied to enhance the classification accuracy. By reducing this overlap, the empirical filtering technique significantly contributes to improved diagnostic performance. Furthermore, as the empirical filtering ranges are further refined to two and one σ ranges, the diagnosis rates for arc failure show additional improvements. Narrowing down the filtering ranges to two and one σ ranges makes the distinctions between normal and arcing states more discernible than the broader three σ range, leading to enhanced diagnosis accuracies. In addition, when considering AIL models, the RF consistently emerges as the top-performing model, achieving the highest diagnosis rates across various inputs and switching frequencies. The diagnostic evaluations conducted across all switching frequencies corroborate the effectiveness of the proposed arc detection scheme when integrated with the empirical filtering technique, notably enhancing the accuracy rates of all AIL models.