[b]1 Introduction[/b] In a broad sense, electromagnetic interference includes not only interference that enters the monitoring system through the current sensor along with partial discharge signals, but also interference that affects the monitoring system itself, such as interference caused by improper grounding, shielding, and circuit processing. The latter can be solved by improving system design, rationally selecting circuits and components, and improving system manufacturing level. Field electromagnetic interference specifically refers to the former and is the focus of research. It can be divided into continuous periodic interference, pulse interference, and white noise [1]. Periodic interference includes system high-order harmonics, carrier communication, and radio communication. Pulse interference is divided into periodic pulse interference and random pulse interference. Periodic pulse interference is mainly caused by high-frequency inrush current generated by the operation of power electronic devices. Random pulse interference includes corona discharge on high-voltage lines, partial discharge generated by other electrical equipment, discharge generated by tap changer operation, arc discharge generated by motor operation, and floating potential discharge caused by poor contact. White noise includes coil thermal noise, ground grid noise, and various noises coupled into the power supply line and transformer relay protection signal line. Electromagnetic interference generally enters the measurement point through two methods: direct spatial coupling and line conduction. Different measurement points result in different interference coupling paths and have different impacts on the measurement; different measurement points also result in different types and intensities of interference. The principle for selecting transformer partial discharge monitoring points is to ensure high partial discharge signal strength, high signal-to-noise ratio, and ease of measurement. These mainly include the outer casing grounding wire and the bushing end screen grounding wire; some also select the neutral point grounding wire, core grounding wire, and high-voltage output terminal. Sometimes, to suppress interference, a reference interference signal is measured from the transformer's power supply line. Because installing sensors at the neutral point and high-voltage output terminal is inconvenient, and some transformer cores are internally grounded, monitoring systems often choose the outer casing and bushing end screen grounding wires as measurement points. [b]2 Commonly Used Suppression Methods[/b] Interference suppression always considers three aspects: interference source, interference path, and signal post-processing. Identifying the interference source and directly eliminating or cutting off the corresponding interference path is the most effective and fundamental method for solving interference, but it requires detailed analysis of the interference source and path, and generally does not allow changes to the original transformer operating mode; therefore, the measures that can be taken in these two aspects are always very limited. Various signal processing techniques are used to suppress various interferences that enter the monitoring system through the current sensor. Partial discharge signals and interference signals are generally distinguished from the following aspects: power frequency phase, spectrum, pulse amplitude and amplitude distribution, signal polarity, repetition rate and physical location [2], and a large number of anti-interference techniques have been proposed accordingly. There are two different approaches to anti-interference techniques: one is based on narrowband (bandwidth is generally 10kHz to several tens of kHz) signals. It picks up the signal through a narrowband current sensor with a suitable frequency band and a bandpass filter circuit, avoids various continuous periodic interferences, and improves the signal-to-noise ratio of the measured signal. This method is only suitable for a specific substation and is inconvenient to use. In addition, since the partial discharge signal is a wideband pulse, narrowband measurement will cause distortion of the signal waveform, which is not conducive to subsequent digital processing. The other approach is based on wideband (bandwidth is generally 10 to 1000kHz) signal processing. The detected signal contains most of the energy of the partial discharge and a large amount of interference, but the signal-to-noise ratio is low. The general steps for handling these interferences are: a. suppressing continuous periodic interference; b. suppressing periodic pulse interference; c. suppressing random pulse interference. With the development of digital technology and the application of pattern recognition methods in partial discharge, this processing method often achieves good results. Based on the above two approaches, detection signals with different signal-to-noise ratios can be obtained. In the subsequent processing, many processing methods are consistent. They can be summarized as frequency domain processing and time domain processing methods. The frequency domain method utilizes the discrete characteristics of periodic interference in the frequency domain; while the time domain processing method is based on the discrete characteristics of pulse interference in the time domain. There are two implementation methods: hardware and software. They are introduced below. [b]3 Suppression of Periodic Interference[/b] Periodic interference is also known as narrowband interference. It accounts for a large proportion of various interferences, and the suppression and elimination of interference should start from this. Because it has high intensity and fixed phase distribution, it is mostly processed by frequency domain methods. The main methods include FFT threshold filters [3-4], adaptive filters [5], fixed-coefficient filters [6], and ideal multi-band digital filters (IMDF) [7-8]. There are many algorithms for narrowband interference suppression, and they are relatively mature. From the perspective of application effect, fixed-coefficient filters and ideal multi-band digital filters are more ideal. Since IMDF needs to perform multiple FFT and IFFT when processing data, it will consume a lot of computation time, which is not conducive to real-time processing. However, based on the optimal monitoring frequency band found by IMDF, a fixed-coefficient finite impulse response (FIR) digital filter can be formed to process directly in the time domain, which simplifies the operation and speeds up the processing speed. Specific applications are discussed in detail in references [7-8]. All of the above methods can be implemented by software or hardware circuits. Although hardware filtering is not flexible in adjustment, it can effectively suppress narrowband interference after selecting the optimal frequency band through field experiments. Although software methods are more flexible in adjustment, they have the disadvantage of slow real-time calculation speed. [b]4 Suppression of periodic pulse interference[/b] After the periodic interference is removed from the signal, other interferences become the main problem. There are two main types of methods for suppressing periodic pulse interference: analog methods and digital methods. Analog methods include differential balancing, directional coupling, and reference signal methods; the first two methods are also applicable to the suppression of random pulse interference, which will be introduced later. Reference [9] selected a distribution line containing only pulse interference and not discharge pulses to measure the pulse interference signal. The measured interference pulse was used as a control signal. When the signal level exceeded the set threshold and was determined to be interference, the analog-to-digital converter (ADC) was stopped to eliminate the interference pulse from the distribution line. The principle of digital methods is to process the interference and partial discharge signals by utilizing the different phase distribution characteristics. For example, KONIG, G. and KOPF, U. proposed a method [10], which first records the signal for multiple cycles, then averages the data in the same phase of each cycle, and uses this to form a template to subtract from the original signal, thereby eliminating the periodic interference signal. This method is more effective in removing interference when there are fewer partial discharge signals and the distribution characteristics are relatively clear, but less effective when there are more and stronger partial discharge signals. V. Nagesh and BIGururaj from India proposed a method [11] that draws on some achievements in biosignal processing. Its basic principle is based on the fact that partial discharge signals and periodic interference signals have different shapes. First, data is segmented to separate pulses from the waveform signal, forming individual pulse sequences. The FFT algorithm is used to perform cross-correlation calculations on each pulse in the frequency domain, determining their similarity and grouping them according to certain criteria. Based on these groups of pulses, class signal templates are obtained, and then each class of signals is synthesized in the time domain. Analysis revealed that the phase of the partial discharge signal is relatively dispersed, while the interference is very concentrated. Using this characteristic, the periodic pulse interference signal class is eliminated, and the remaining signal is reconstructed to obtain the signal after removing the periodic pulse interference. Therefore, it is feasible to use the differences in waveform and phase between partial discharge and periodic pulse interference for interference suppression. This method can also be used for localization, identifying the pulse waveforms caused by different discharge points by analyzing their characteristics. The disadvantages of this method are: when the partial discharge repetition rate is high, it is possible to treat two adjacent pulses as one, which affects the recognition effect; in addition, when there are many pulse waveforms, the calculation speed is affected, but with the significant improvement of the computing power of microcomputers, this effect will be increasingly ignored. [b]5 Suppression of Random Pulse Interference[/b] This type of interference is the most difficult to eliminate. Since the characteristics of interference and partial discharge signals in the frequency domain are similar, most existing methods are based on the time domain. Commonly used methods include hardware circuit method, software waveform recognition method and artificial intelligence method. 5.1 Hardware Circuit Method Its basic idea is to use the characteristic that the external pulse interference in the output signals of the two measurement points is in the same direction, while the internal discharge pulse is in the opposite direction to remove the pulse interference. Specifically, it is implemented as a hardware circuit. Commonly used circuits include differential balance method [12], pulse polarity identification method [13] and directional coupling method [14]. In practical applications, the first two methods are not ideal. This is because for the differential balance method, due to the different propagation paths, the two signals that make up the differential often cannot correspond well, so the differential effect is not good. Reference [15] proposed the concept of differential "balanced pair" to improve this, which can eliminate interference and simultaneously obtain the amplitude and number of partial discharge pulses. The limitation of pulse polarity identification is that due to the analog delay and the influence of external factors on the polarity discriminator, it can cause electronic gate malfunctions and reduce the accuracy of polarity identification. The directional coupling method was proposed by Borsi H et al. in Germany in 1987. The schematic diagram is shown in Figure 1. It uses a specially wound Rogowski coil to couple the partial discharge signal at the bottom of the high-voltage bushing near the flange, and judges whether it is a partial discharge signal or external electromagnetic interference based on the magnitude of the voltage across the coil. This method connects the middle tap of the Rogowski coil to the measuring terminal of the transformer bushing end screen. At this time, the measuring terminal of the end screen is grounded in series with a small resistor, which can be regarded as the low-voltage arm of the capacitive voltage divider formed by the end screen and the end screen to ground capacitor. After being grounded by the small resistor, it forms a high-pass filter, and only high-frequency signals can pass through. The Rogowski coil and the measuring terminal of the high-voltage bushing end screen are connected to form a directional coupling circuit. [img=273,179]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/dgdnxjs/dgdn99/dgdn9904/image4/25.gif[/img][align=left] Figure 1. When the current I in the directional coupling circuit is in the direction shown in the figure, U(1) = Uc + U1, U(2) = Uc - U2 = Uc - U1. At this time, U(1) > U(2); if the current I is in the reverse direction, then U(1) 5.2 Software Waveform Recognition Method With the development of computer technology and digital signal processing technology, logical judgment using pulse signal characteristics can also suppress interference. Its premise is pulse recognition, that is, determining the existence of the pulse, its duration, and its corresponding start and end points, in order to accurately determine the discharge phase and acoustic delay. Currently, pulse recognition mostly uses the threshold recognition method. However, the pulses measured on-site are mostly attenuated oscillating waves, and this method is prone to misjudgment and cannot determine the pulse duration. Reference [9] proposes a method combining pulse amplitude threshold and waveform characteristics to identify oscillating pulses, and it has achieved good results in practical applications. 5.3 Application of Pattern Recognition The essence of this method is still to distinguish signals by their phase characteristics. Although the amplitude of partial discharge signals varies greatly, their phases are concentrated around 45° and 225°, respectively. For example, because the phase of arc discharge differs from that of partial discharge, the amplitude variation is smaller, and the pulse shape is slightly different, based on these characteristics, an experienced expert can easily distinguish this interference as an arc discharge signal. Pattern recognition methods are software implementations of expert experience, as confirmed in CIGER's report, and some corresponding software has emerged [17-20]. Common methods include fuzzy logic, Kohonen network classification, KLT transform, and artificial neural network based on minimum distance. Generally speaking, the difficulty of pattern recognition methods lies in the need to accumulate a large amount of prior knowledge and identify specific differences between interference and partial discharge. In online measurements, it is difficult to find these differences in strong interference signals. Several methods are introduced below. 5.3.1 Karhunen-Loeve-Transform Method Research has found that when the dimension of the input vector used for pattern recognition is high, classification is difficult and the effect is poor; reducing the dimension improves the classification effect. In other words, to improve the recognition rate and highlight the characteristics of the signal, interference or noise information in the signal must first be removed. The principle of KLT transform is shown in Figure 2. As can be seen from the figure, if the x1-x2 coordinate system is used, x1 and x2 coordinates must be used simultaneously for classification; if an orthogonal transformation is performed, it is transferred to the w1-w2 coordinate system. Therefore, only the w2 coordinates are needed for classification. It can be seen that the KLT transform can remove interference. [/align][img=244,186]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/dgdnxjs/dgdn99/dgdn9904/image4/26-1.gif[/img] Figure 2 KLT Transform Principle 5.3.2 Kohonen Network This algorithm is an unsupervised algorithm (as shown in Figure 3). Its principle is to find the node with the shortest Euclidean distance from the input vector to the output layer, use this as the output, and modify the weights of this node and its neighbors through a self-organizing algorithm, so that these nodes have a greater response to the current input. The Kohonen algorithm can perform adaptive classification, distinguishing between partial discharge signals and interference signals, thereby achieving the purpose of interference cancellation and suppression. [img=219,178]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/dgdnxjs/dgdn99/dgdn9904/image4/26-2.gif[/img] Figure 3 Self-organizing feature mapping network 5.3.3 Pulse sequence analysis method It is said that this method is simple, effective and has a high recognition rate [21]: It uses the discharge voltage difference or phase difference between partial discharges to form an analysis sequence, and uses these features to distinguish different discharge modes and interferences to achieve the purpose of interference suppression; in addition, it can also be used for fault location. [b]6 Summary[/b] A large number of research results show that with the improvement of A/D conversion rate and the development of computer technology, the transformer partial discharge online monitoring system using wideband (10k-1000kHz) sensors combined with high-speed sampling has become the mainstream of development. Signal processing has developed from traditional spectral analysis to time-domain analysis of partial discharge waveforms. Several advancements in digital processing technology and artificial intelligence have been widely applied to interference suppression in online monitoring, with the potential for groundbreaking results. To further improve the effectiveness of anti-interference measures, research should be strengthened on the propagation patterns of interference and pulses, including their propagation in substations and within transformers. This research may reveal differences in their waveform, phase, and direction characteristics.