summary
Researchers Wang Feifei, Ruan Aimin, Wei Gang, and Sun Haibo from Nanjing Institute of Technology and Zhenjiang Power Supply Branch of Jiangsu Electric Power Co., Ltd., published an article in the 4th issue of "Electrical Technology" in 2019 (titled "Partial Discharge Fault Identification in Switchgear Based on Convolutional Neural Networks"). The article points out that most current classification algorithms for partial discharge faults are shallow learning algorithms, and manually extracted features directly affect the classification results. In contrast to shallow learning algorithms, deep learning has a deeper architecture and can automatically learn features from samples; convolutional neural networks are a typical deep learning algorithm.
This paper aims to study the application of convolutional neural networks (CNNs) in partial discharge detection of switchgear, demonstrating that deep learning architectures can effectively improve the recognition rate. Two types of audible sound signals were collected: normal and faulty. Features were extracted from these two types of sound signals, and then fed into an SVM model and a CNN for classification, respectively. Experimental results show that the CNN improves the accuracy of sound recognition compared to the traditional SVM.
In power systems, 10kV metal-enclosed switchgear is the power equipment directly facing users, and the safe operation of the power system is closely related to the condition of the switchgear. According to relevant data, insulation degradation and poor contact are the main faults of switchgear. Partial discharge phenomena often precede these faults. Therefore, by effectively monitoring partial discharge signals, measures can be taken before the insulation layer of the switchgear deteriorates, reducing the accident rate.
Currently, fault detection methods for high-voltage switchgear include electrical and non-electrical methods. Electrical methods mainly include pulse current method, ultra-high frequency detection method, and radio interference voltage method. Non-electrical methods mainly include acoustic emission detection method, optical detection method, and infrared detection method. Compared with electrical methods, non-electrical methods have stronger resistance to electromagnetic interference. Among non-electrical methods, ultrasonic waves have the widest range of applications; however, ultrasonic waves attenuate quickly and are difficult to penetrate the metal casing of equipment. Therefore, we detect the low-frequency band of acoustic signals, that is, fault detection based on audible sound signals. Diagnostic technology based on acoustic characteristics does not require contact with the equipment, does not interfere with the normal operation of the equipment, can collect signals at any time, and is convenient to install.
Compared with the research status abroad, my country's online audible sound monitoring technology started late, but developed rapidly. Reference [3] proposed a partial discharge detection algorithm for switchgear based on spectrogram features, which improved the fault detection rate by 2.5% compared with the fault detection rate based on MFCC features. However, due to the single feature selection, the algorithm has a certain false recognition rate. Reference [4] constructed a complete binary tree by using fuzzy C-means clustering and split hierarchical method, which effectively solved the problem of inseparability between traditional one-to-one support vector machine and one-to-many support vector machine, and improved the efficiency of partial discharge diagnosis.
However, because support vector machines rely on manually constructed features, their error rate remains high when dealing with scenarios with complex environmental features (such as speech). Therefore, this paper proposes a fault identification system for high-voltage switchgear based on multimodal features and convolutional neural networks, and compares it with a detection system based on support vector machines.
1. Acoustic signal manifestations and characteristics of partial discharge
During normal operation, switchgear exhibits mechanical vibration, producing a low-pitched "humming" sound. However, when partial discharge begins to occur, a "sizzling" discharge sound emerges in the discharge area. This sound is slightly louder than the normal "humming" sound, and as the discharge deepens, it eventually leads to insulation breakdown. A brief, sharp "pop" sound can be heard upon insulation breakdown. Therefore, changes in the audio characteristics of the switchgear's sound signal can help determine if a fault has occurred.
Figures 1 and 2 show the time-domain waveforms and spectrograms during normal and fault operation. In the time-domain waveform diagram, the waveform is stable during normal operation. When a partial discharge fault occurs, the amplitude increases compared to before, but the waveform remains stable. However, when the insulation layer breaks down, the signal amplitude shows a trend of instantaneous increase followed by a slow decrease.
The spectrogram shows that the frequency distribution is uniform during normal operation, while during a fault, the signal frequency suddenly increases and then decreases, resulting in a lighter color in the spectrogram, indicating a reduction in signal energy. This comparison demonstrates a significant difference between normal and faulty operation; therefore, signal analysis can be used to determine the operating status of the equipment.
Figure 1. Time-domain waveform and spectrogram of normal operation signal
Figure 2. Time-domain waveform and spectrum of the signal during discharge.
2-classifier (omitted)
Currently, most classification methods in fault detection still utilize well-established shallow learning algorithms, such as SVM. Shallow learning uses neural networks with relatively few hidden layers, resulting in a simple algorithm structure. Since the concept of deep learning was proposed, scholars have begun applying it to various fields. This paper attempts to apply convolutional neural networks (CNNs) from deep learning algorithms to the field of partial discharge detection in switchgear, demonstrating that deep learning achieves a higher recognition rate than shallow learning algorithms. Compared to shallow learning, deep learning has a deeper architecture and more complex computational layers, thus exhibiting significant advantages in feature processing.
Figure 3. Structure diagram of convolutional neural network
3. Experiments and Analysis (omitted)
in conclusion
Experimental results show that the fault identification rate of convolutional neural networks (CNNs) is 1.81% higher than that of SVMs, demonstrating significant superiority. With the same input sample data, the higher recognition rate of CNNs proves that the features extracted by CNNs are more discriminative and more effective in classification than those extracted by SVMs.
This experiment used a relatively small number of samples, yet the convolutional neural network still achieved high resolution, effectively demonstrating that deep learning architectures do not depend on the number of samples for feature extraction. Therefore, it can be concluded that this deep learning architecture of convolutional neural networks is of research significance in switchgear partial discharge fault detection systems.
However, the above fault diagnosis system still has certain shortcomings, such as the lack of detailed classification of partial discharge faults. In the future, an experimental platform can be established to simulate various faults, employing multiple detection methods in conjunction. Data will undergo feature dimensionality reduction and eigenvalue optimization via a host computer, and finally, the results from multiple classifiers will be fused. Based on this method, the classification accuracy can be further improved.
Disclaimer: This article is a reprint. If it involves copyright issues, please contact us promptly for deletion (QQ: 2737591964). We apologize for any inconvenience.
