0 Introduction Online monitoring of partial discharge (PD) in high-voltage electrical equipment operating on the power grid is of great significance for timely detection of potential hazards in insulation. Many methods are used for online monitoring, among which electrical measurement is widely adopted due to its speed, accuracy, and high precision. Electrical measurement utilizes sensors to introduce the electrical signal generated by partial discharge (PD) into the data acquisition and analysis system through a detection circuit. Because the spatial electromagnetic field is coupled with the detection circuit, the detection circuit introduces external electromagnetic interference signals along with the PD signal. Among these, carrier communication interference, also known as discrete spectral interference (DSI), is the most significant, often accounting for more than 80% of the total signal energy. Therefore, the effectiveness of DSI suppression directly affects the detection accuracy and reliability of online monitoring equipment. Currently, many digital filtering methods for suppressing DSI have emerged [1-3], but practical experience has shown that most methods lack strong adaptability and robustness. When used in complex PD online monitoring environments, they result in low signal-to-noise ratios (SNR) and often cause significant PD signal distortion. To achieve satisfactory DSI suppression and lay a solid foundation for subsequent PD identification, we propose a novel digital filtering method based on the neighbor inhibition effect of competitive subnetworks, which has been practically applied in the PD online monitoring system. 1 Competitive Subnetworks Self-organizing networks are an important type of artificial neural network. These networks often employ competitive learning methods during the self-organization process. The competition process is organized in a "winner takes all" manner [4], meaning that after the competition process ends, the winning network node outputs 1, and all non-winning network nodes output 0. In neural network systems, this process can be implemented using various competitive subnetworks. MAXNET is a typical competitive subnetwork, and its structure is shown in Figure 1. The connections between nodes are of the self-excitation type (forming a positive self-loop from the node to itself) and the neighbor inhibition type (a negative inhibition connection from one node to another) to complete the competition between nodes. [align=center] Fig.1 Structure of MAXNET[/align] MAXNET consists of m nodes. Let uik be the connection weight from the i-th node to the k-th node. We can take: (1) MAXNET iterates through the interaction of the connection weights based on the output values ​​of each node. The output yi(t+1) of node i at time t+1 depends on the output yi(t) of each node at time t, that is: (2) i,k∈{1,2,…,m} where (3) The node function f(x) described by equation (3) is nonlinear. It has the function form shown in Fig.2 and is usually called the threshold logic activation function. [align=center] Fig.2 Threshold logic activation function[/align] Equation (2) shows that each node tries to maintain its own value and suppress other nodes through the connection weights. This mode of action is called neighbor suppression. If the independent variable x of the function f(x) is positive, then the output yi(t+1) is also positive; while the outputs of all nodes whose independent variables are non-positive will be 0. Through the neighbor suppression effect, the competitive subnetwork will go through several iterations to make all inputs except the maximum input negative, and make their corresponding outputs 0. Finally, the remaining positive output will be located exactly at the node with the maximum initial input. Transforming the network output with a step function, we can obtain the "winner takes all" result: yj(t) = 1, while yk(t) = 0, k ≠ j. 2 Application of Neighbor Suppression in Digital Filtering Through the above analysis of MAXNET, we have gained some understanding of the structure and function of the competitive subnetwork. It should be noted that the digital filtering method introduced in this paper does not aim to find the node corresponding to the maximum value, but only utilizes the neighbor suppression effect of the competitive subnetwork. The characteristic of this effect is that nodes with larger initial input values ​​are less inhibited by other nodes and have a stronger ability to maintain their own values; nodes with smaller initial input values ​​are more inhibited by other nodes and have a weaker ability to maintain their own values, and their output may become 0 after only one or a few iterations. How can we apply the neighbor suppression effect to the digital filtering stage? We know that the raw data signals acquired in the field contain frequency components of discharge signals such as DSI and PD. In the frequency domain, the amplitude spectrum of DSI is a pulse waveform centered on the main frequency; while the discharge signals such as PD form a flat waveform that covers almost the entire frequency domain. The energy of the DSI signal is concentrated, while the energy of the discharge signal is dispersed; therefore, from the amplitude spectrum perspective, the amplitudes of the two differ significantly. If we perform a Fourier transform on each segment of the raw signal and use the resulting amplitude spectrum value as the input to the competing sub-network, then each node in the network corresponds to a different frequency component in the input signal. By appropriately selecting the number of iterations and the neighbor suppression coefficient ε, we can make the output of each node corresponding to the frequency component of the discharge signal such as PD 0, while the output of each node corresponding to the frequency component of DSI is not 0. In this way, the DSI frequency components in the original signal can be determined. By setting these frequency components to 0 and then returning from the frequency domain to the time domain, filtering can be achieved. 3 Application of Neighbor Suppression Digital Filtering Method in PD Online Monitoring System The preamplifier frequency band of the PD online monitoring system is 5 kHz to 500 kHz, and the sampling frequency of the data acquisition card is set to 2 MHz. For information on the system hardware, please refer to reference [5]. The software part of the monitoring system is based on a digital filtering program for various interferences, among which DSI suppression is the first step in the filtering process. Since we started developing the partial discharge online monitoring system, we have accumulated a large amount of field data. Through amplitude spectrum analysis of these digital signals after segmentation and windowing, we have drawn the following conclusions: a. In the 0 to 500 kHz frequency band, the spectral lines of almost all DSI signals are more than 15 dB higher than the baseline of the amplitude spectrum formed by the discharge pulse signal; b. In the 0 to 500 kHz frequency band, the average amplitude of the amplitude spectrum is comparable to the amplitude corresponding to the smallest DSI and is higher than the baseline formed by PD and other pulse signals. The above conclusions make it possible to determine and separate DSI using the neighbor suppression effect of competing subnetworks in the amplitude spectrum. The following example uses a segment (1024 sampling points) of field data from a transformer in Dalian. First, the data segment is windowed (hamming window), and then an FFT is performed to obtain the amplitude spectrum corresponding to the 0–500 kHz frequency range, as shown in Figure 3(a). The amplitude spectrum values ​​of each point are input into MAXNET. The structure of MAXNET is: number of nodes m = 512; transverse suppression coefficient ε = 1/800. After one iteration (approximately 1.5 s), the output waveform is shown in Figure 3(b). Figure 3(b) clearly shows that the outputs of nodes corresponding to the narrow pulse of DSI in the amplitude spectrum are not 0, while the outputs of the remaining nodes are all 0. This indicates that although only one iteration was performed, the method has a significant suppression effect on small signals, but not much suppression effect on large signals. After this processing, only the frequency components with non-zero amplitude in Figure 3(b) need to be filtered out, while the frequency components with zero amplitude are retained. Then, an IFFT is performed to complete the digital filtering of the DSI signal in this data segment. [align=center] Figure 3 Comparison of the magnitude spectrum of field signal before and after the lateral inhibition procedure[/align] To verify the filtering effect, the time-averaged power spectral density analysis was performed on 128 kB (256 segments) of data obtained from a single sampling by the acquisition system. Figure 4 shows the time-averaged power spectral density diagram of the field data before and after DSI filtering. By comparison, it can be seen that the filtering method has a satisfactory suppression effect on DSI, and there is almost no residual spectrum of DSI in the filtered signal. This makes the DSI filtering process not cause too much PD signal distortion. After the digital filtering process for all interferences is completed, the PD signal distortion is indeed small, and its waveform is shown in Figure 5. From the waveform details, it still has a pulse shape with significant oscillation attenuation. [align=center] Fig. 4 Time-average power spectrum density comparison of field signal before and after the filtering procedure[/align] [align=center] Fig. 5 Waveform of PD[/align] Additionally, Fig. 6 shows the time-domain waveforms of the ground signal before and after filtering within half a power frequency cycle. By comparing the correction pulse and baseline height, it can be seen that the SNR improved by nearly 10 dB after filtering, significantly improving the detection accuracy of the monitoring system. [align=center] Fig. 6 Waveforms of field signal before and after the lateral inhibition filtering procedure[/align] Through observation of online monitoring, it was noted that the amplitude of the DSI signal varied greatly, the generation and disappearance of DSI were irregular, and the external discharge conditions also differed significantly depending on the time and weather conditions. However, based on long-term operation results, the filtering method based on the neighbor suppression effect of the competing subnetwork exhibits strong adaptability and robustness, which can be verified by time-averaged power spectrum analysis of data from randomly sampled data at different times. [b]4 Conclusion[/b] The digital filtering method based on the neighbor suppression effect of the competing subnetwork has many unique features: a. The MAXNET neighbor suppression effect is nonlinear, therefore this method has strong adaptability and can automatically determine the frequency components to be filtered based on changes in the DSI frequency and modulation frequency. b. It thoroughly filters out DSI with large amplitudes and also has a strong suppression effect on smaller DSIs with amplitudes higher than the baseline and at the average amplitude level, making up for the main defects of other similar methods. This allows the online monitoring system to identify PDs as small as 3750 pC with minimal PD waveform distortion. c. The neighbor suppression effect is significant at points in the amplitude spectrum not higher than the baseline level. Even if the discharge intensifies and the baseline increases to a certain extent, it can still effectively suppress the discharge without parameter adjustment. This method has achieved satisfactory results in a large number of field operations, well meeting the requirements of online monitoring of partial discharge under strong interference conditions. References [1] Feser K, Konig G, Ott J, et al. An Adaptive Filter Algorithm for On-Site Partial Discharge Measurements. In: Conference Record of the 1988 International Symposium on Electrical Insulation. Boston: 1988 [2] Borsi H, Hartje M. New Methods to Record the Disturbance Influences on the In-Situ Partial Discharge Measurement and Monitoring. In: 6th International Symposium on High Voltage Eng. New Orleans: 1989 [3] Nagesh V, Gururaj B I. Evaluation of Digital Filters for Rejecting Discrete Spectral Interference in On-Site PD Measurements. IEEE Trans, 1993, EI-28(1): 73～85 [4] Zhou Jicheng, Zhou Qingshan, Han Piaoyang. Artificial Neural Networks - Implementation of the Sixth Generation Computer. Beijing: Science Popularization Press, 1993 [5] Wang Zhe, Cai Weizheng, Chen Xueyun. Online monitoring of partial discharge in high-voltage transformers based on wavelet analysis. Automation of Electric Power Systems, 1998, 22(4): 19-23