Intelligent distribution board protection system based on ai pattern classification using high frequency earth voltage

Abstract

The present invention relates to an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced-ground voltage signal, which detects high-frequency induced-ground voltage on the surface of a switchboard enclosure, determines an arc using a reference value for a characteristic value of the detected high-frequency induced-ground voltage, and classifies the pattern of partial discharge using a short-time Fourier transform and a two-dimensional convolutional artificial neural network and determines the type of partial discharge when it is not an arc. The system includes a high-frequency induced-ground voltage sensor attached to the enclosure of a switchboard and an analysis device that determines and analyzes arcs and partial discharges based on a signal detected from the high-frequency induced-ground voltage sensor. The analysis device includes: a preprocessing unit that preprocesses a signal output from the high-frequency induced-ground voltage sensor; a signal processing unit that extracts a characteristic value of an arc in the time domain using the signal preprocessed by the preprocessing unit and determines whether an arc has occurred by comparing the characteristic value with a preset reference value; and an intelligent analysis unit that receives data from the signal processing unit and determines the occurrence and type of partial discharge using a wavelet transform and a neural network algorithm.

Description

Technology Field

[0001] The present invention relates to an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced-ground voltage signal, and more specifically, to an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced-ground voltage signal that detects a high-frequency induced-ground voltage on the surface of a switchboard enclosure, determines an arc using a reference value for a characteristic value of the detected high-frequency induced-ground voltage, classifies the pattern of partial discharge using a short-time Fourier transform and a two-dimensional convolutional artificial neural network, and determines the type of partial discharge. Background Technology

[0003] Partial Discharge (PD) is a microscopic discharge phenomenon that occurs only in localized areas without the entire insulator being destroyed. Examples include void discharge occurring in gaps within the insulator, corona discharge occurring at sharp points of terminals, surface discharge occurring on contaminated insulator surfaces, and floating discharge occurring in unfixed metal bodies.

[0004] An arc is a phenomenon in which insulation is completely destroyed and a continuous current path is formed between two electrodes. It occurs when insulation is completely broken due to a short circuit, overheating caused by poor contact, or intensified partial discharge.

[0005] Distinguishing between partial discharge (PD) and arc in switchboards is key to accident prevention; partial discharge is a precursor to an accident, while arc can be viewed as the accident itself.

[0006] In other words, even if partial discharge cannot be detected, arcs must be detected to prevent the accident from propagating not only to the switchboard but also to the loads connected to it.

[0007] In addition, when partial discharge occurs, electromagnetic waves propagate far from the discharge site as transient voltages. Due to the skin effect, transient voltages inside a metal workpiece cannot be directly detected on the external surface. However, electromagnetic waves can propagate into free space where there are electrical discontinuities in the metal workpiece, which generates a series of transient voltage spikes on the surrounding metal surface, known as high-frequency induced ground voltage.

[0009] Meanwhile, as a technology for detecting and diagnosing partial discharge, Registered Patent Publication No. 0-2853979 discloses a switchboard having a deep AI-based partial discharge real-time monitoring and diagnosis function using UHF and HFCT signals.

[0010] The above technology relates to a switchboard having a deep AI-based partial discharge real-time monitoring and diagnosis function using UHF and HFCT signals, which analyzes each of the UHF signal detected through an Ultra High Frequency (UHF) sensor module and the signal detected through an HFCT (High Frequency Current Transformer) sensor module into a Phase Resolved Pulse Sequence (PRPS) and a Phase Resolved Partial Discharge (PRPD), and monitors and diagnoses whether partial discharge occurs through a deep learning Convolutional Neural Network (CNN) based on the analyzed results.

[0011] However, the pulse duration of PD can vary from a few nanoseconds to hundreds of nanoseconds. The method of detecting PD using UHF sensors is widely used due to its major advantages, such as low electromagnetic interference and a high signal-to-noise ratio. However, since this method requires a high sampling rate, hardware costs for processing and storing data increase when applied to power distribution panels.

[0012] As another technology for detecting partial discharge, Registered Patent Publication No. 10-2879452 discloses a switchboard equipped with a partial discharge detection function.

[0013] The above technology relates to a switchgear equipped with a partial discharge detection function that converts the frequency of an oscillation signal to a wide frequency band, determines whether a partial discharge RF signal is received for each frequency of the oscillation signal, calculates the frequency band determined to be received within a reference bandwidth from an initial frequency to a final frequency to generate a measurement bandwidth value, calculates the ratio between the measurement bandwidth and the reference bandwidth to generate a ratio value, and finally determines that a partial discharge has occurred if the ratio value exceeds a set ratio threshold.

[0014] However, since the above technique requires measuring by sequentially sweeping the frequency from the initial frequency to the final frequency, scanning the entire band takes time, making it difficult to detect instantaneously occurring intermittent arcs or initial partial discharges. Prior art literature

[0016] Registered Patent Publication No. 10-2853979 (Aug. 28, 2025) Registered Patent Publication No. 10-2879452 (October 28, 2025) The problem to be solved

[0017] The present invention is devised to solve the above-mentioned problems and provides an AI pattern classification-based switchboard protection system using a coupling capacitor sensor capable of detecting a series of transient voltage spikes on the metal surface of a switchboard using a high-frequency inductive ground voltage sensor and determining the occurrence of an arc from the detected signal.

[0018] Furthermore, the invention provides an intelligent switchboard protection system based on AI pattern classification using high-frequency induced-ground voltage signals, capable of accurately detecting types of partial discharge by generating and providing data optimized for AI learning. means of solving the problem

[0020] An intelligent switchboard protection system based on AI pattern classification using a high-frequency induced-ground voltage signal according to the present invention for solving the above problems includes a high-frequency induced-ground voltage sensor installed attached to the enclosure of a switchboard and an analysis device that determines and analyzes arcs and partial discharges based on a signal detected from the high-frequency induced-ground voltage sensor, wherein the analysis device includes: a preprocessing unit that preprocesses a signal output from the high-frequency induced-ground voltage sensor; a signal processing unit that determines whether an arc has occurred by comparing a preset reference value according to the number of abnormal signal occurrences, the duration of abnormal signal continuity, and sharpness using the signal preprocessed by the preprocessing unit; and an intelligent analysis unit that receives data from the signal processing unit, classifies patterns of partial discharge using a Short-Time Fourier Transform (STFT) and a neural network algorithm, and determines the type of partial discharge based on the classified pattern.

[0021] Here, the high-frequency inductive ground voltage sensor comprises: a detection electrode installed in contact with a detection surface; a ground electrode; a dielectric inserted between the detection electrode and the ground electrode; and a BNC terminal to which the detection electrode is connected to a core wire and which outputs a signal.

[0022] In addition, the preprocessing unit includes a high-pass filter that removes frequency bands below a set lower limit frequency; a low-pass filter that removes frequency bands exceeding a set upper limit frequency; an RF detector that converts the signal passing through the high-pass filter and the low-pass filter into a DC voltage proportional to the power (dBm); an amplifier that scales and outputs a low-frequency voltage signal output from the RF detector to fit the input range of an ADC (Analog-to-digital converter); and a multiplexer that selectively outputs a multi-channel signal output from the amplifier.

[0023] In addition, the signal processing unit may be configured to determine an arc when at least two of the following conditions are satisfied: the number of occurrences of the abnormal signal is greater than or equal to a reference value for the number of occurrences of the abnormal signal, the period time of occurrence of the abnormal signal is less than or equal to a reference value for the period time of occurrence of the abnormal signal, and the sharpness is greater than or equal to a reference value for the sharpness.

[0024] Alternatively, the signal processing unit may be configured to determine an arc when the number of occurrences of the abnormal signal is greater than or equal to a reference value for the number of occurrences of the abnormal signal, the period time of occurrence of the abnormal signal is less than or equal to a reference value for the period time of occurrence of the abnormal signal, and the sharpness is greater than or equal to a reference value for the sharpness.

[0025] In addition, the signal processing unit may be configured to output a trip signal and an alarm signal when it is determined to be an arc.

[0026] In addition, the intelligent analysis unit is characterized by generating a spectrogram through the short-time Fourier transform, wherein the horizontal axis represents time, the vertical axis represents frequency, and the color / intensity represents the energy intensity (magnitude / power) at a specific time and frequency.

[0027] In addition, the neural network algorithm of the intelligent analysis unit is characterized by applying a 2D Convolutional Neural Network (2D-CNN) to infer partial discharges through probabilistic calculations.

[0028] In addition, the above probability calculation is characterized by inferring partial discharge through the following formula.

[0029] formula)

[0030]

[0031] Here, is the softmax function of the k-th node, n is the number of nodes in the output layer, z kis the input of the k-th node of the output layer, z i is the input of the i-th node of the output layer.

[0032] In addition, the prediction based on the inference of the above partial discharge is characterized by being noise, defective insulator discharge, floating electrode discharge, and corona discharge.

[0033] In another embodiment, an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced-ground voltage signal according to the present invention may comprise a high-frequency induced-ground voltage sensor installed attached to the enclosure of a switchboard and an analysis device that determines and analyzes arcs and partial discharges based on a signal detected from the high-frequency induced-ground voltage sensor, wherein the analysis device may comprise a preprocessing unit that preprocesses a signal output from the high-frequency induced-ground voltage sensor; and an intelligent analysis unit that receives data from the preprocessing unit, classifies patterns of partial discharge using a Short-Time Fourier Transform (STFT) and a neural network algorithm, and determines the type of partial discharge based on the classified patterns. Effects of the invention

[0035] According to the present invention, an arc is quickly determined from the detected high-frequency induced ground voltage, and if it is determined to be an arc, a trip signal is output to prevent secondary accidents caused by the arc. If it is not an arc, the occurrence of partial discharge can be diagnosed, thus allowing for selective response to arcs and partial discharges and improving reliability.

[0036] In addition, since the type and risk level of partial discharge can be accurately determined from the initial stages, there is an advantage in being able to take appropriate measures before it progresses into a major accident such as an arc. Brief explanation of the drawing

[0038] FIG. 1 is a configuration diagram of an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention applied to a switchboard. FIG. 2 is a schematic diagram of an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention. FIG. 3 is a schematic structural diagram and equivalent circuit of a high-frequency induced-ground voltage sensor applied to an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced-ground voltage signal according to the present invention, FIG. 4 is a circuit diagram of a preprocessing unit applied to an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention. FIG. 5 is a configuration diagram of a signal processing unit and an intelligent analysis unit applied to an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention. FIG. 6 is a spectrogram of one embodiment generated through a short-time Fourier transform in an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention, FIG. 7 is a structural diagram of a two-dimensional convolutional artificial neural network in an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention. FIG. 7a is a structural diagram of a one-dimensional convolutional neural network explaining a change in data form in an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention. FIG. 8 is a diagram showing the process of extracting a feature map from a convolutional layer according to an embodiment of a one-dimensional convolutional neural network in an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention. FIG. 9 is a diagram illustrating the process of extracting a feature vector using maximum pooling in a pooling layer according to an embodiment of a one-dimensional convolutional neural network in an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention. FIG. 10 is a schematic diagram showing the configuration of a signal processing unit and an intelligent analysis unit of another embodiment in an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention. FIG. 11 is a configuration diagram of a simulation for an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention. FIG. 12 is a verification graph of the accuracy in the process of learning partial discharge applied to an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention. FIG. 13 is a verification graph of the standard deviation in the process of learning partial discharge applied to an intelligent switchboard protection system based on AI pattern classification using high-frequency induced ground voltage signals according to the present invention. Figure 14 is a table showing the learning results of a simulation for an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention. Specific details for implementing the invention

[0039] Next, a preferred embodiment of an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention will be described in detail with reference to the drawings.

[0040] In the following, the same reference numerals are used for technical elements that perform the same function, and repeated detailed descriptions are omitted to avoid redundant explanations.

[0041] Furthermore, the embodiments described below are provided as examples to effectively demonstrate preferred embodiments of the present invention and should not be interpreted to limit the scope of the present invention.

[0043] The present invention relates to an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced-ground voltage signal, and more specifically, to an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced-ground voltage signal that detects a high-frequency induced-ground voltage on the surface of a switchboard enclosure, determines an arc using a reference value for a characteristic value of the detected high-frequency induced-ground voltage, classifies the pattern of partial discharge using a short-time Fourier transform and a two-dimensional convolutional artificial neural network, and determines the type of partial discharge.

[0045] FIG. 1 is a configuration diagram of an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced-ground voltage signal according to the present invention applied to a switchboard, and FIG. 2 is a schematic configuration diagram of an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced-ground voltage signal according to the present invention.

[0046] Referring to the attached FIG. 1 and FIG. 2, the intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention includes a high-frequency induced ground voltage sensor (20) installed attached to the enclosure of a switchboard (10) and an analysis device (30) that determines and analyzes arcs and partial discharges based on signals detected from the high-frequency induced ground voltage sensor (20).

[0048] Figure 3 shows a schematic structure and equivalent circuit of a high-frequency induced ground voltage sensor applied to an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention.

[0049] Referring to the attached FIG. 3(a), the high-frequency induced ground voltage sensor (20) applied to the AI ​​pattern classification-based intelligent switchboard protection system using the high-frequency induced ground voltage signal according to the present invention may be configured to include a detection electrode (21) installed in contact with a detection surface (S) which may be composed of a circular copper plate and has a diameter of 105 mm and a thickness of 0.5 mm, a grounding electrode (22) which may be composed of a circular copper plate and has a diameter of 105 mm and a thickness of 0.5 mm, a dielectric (23) which is inserted between the detection electrode (21) and the grounding electrode (22) and is composed of epoxy resin with a thickness of 0.5 mm, and a BNC terminal (24) which is connected to the core wire of the detection electrode (21) and outputs a signal.

[0050] Here, the frequency band detected by the high-frequency induction ground voltage sensor (20) is 150 kHz to 16 MHz.

[0052] Referring to the equivalent circuit (b) of the structure of the high-frequency induction ground voltage sensor (20) above, it can be configured to have a composite impedance Av including a matching resistor (Rm) and a compensation resistor (Ri) of 50Ω, which is the same as the characteristic impedance of the coaxial cable, in order to eliminate signal distortion caused by capacitance.

[0053] Here, the combined impedance (Av) and impedance (Zs) can be expressed by the following Equation 1.

[0054] Formula 1)

[0055]

[0056]

[0057] Here, Av is the combined impedance, Zs is the impedance, Rm is the detection impedance, fc is the cutoff frequency, Ld is the inductive detection impedance, and Ri is the compensation resistance.

[0059] At this time, the voltage (V(t)) output from the high-frequency induction ground voltage sensor (20) is defined by the following Equation 2.

[0060] Formula 2)

[0061]

[0062] Here, V(t) is the output voltage, S is the area of ​​the detection electrode, Zs is the impedance, Rm is the detection impedance, and fc is the cutoff frequency.

[0063] The above cutoff frequency fc can be configured as a -3dB frequency of 100kHz.

[0064] In addition, the voltage (V(t)) output from the high-frequency induction ground voltage sensor (20) may be composed of a BNC coaxial cable having a characteristic impedance of 50Ω.

[0066] The signal detected by the high-frequency induction ground voltage sensor (20) is transmitted to the analysis device (30).

[0067] Referring to the attached FIG. 2, the analysis device (30) is configured to include a preprocessing unit (100), a signal processing unit (200), and an intelligent analysis unit (300).

[0069] Figure 4 shows a circuit diagram of a preprocessing unit applied to an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention.

[0070] Referring to the attached FIG. 4, the preprocessing unit (100) applied to the present invention performs the function of preprocessing a signal output from the high-frequency inductive ground voltage sensor (20), and is configured to include a high-pass filter (110, HPF), a low-pass filter (120, LPF), an RF detector (130), an amplifier (140), and a multiplexer (150).

[0071] The high-pass filter (110) removes frequency bands below a set lower limit frequency and is configured to include inductors (L2, L3) connected in parallel with respect to a capacitor (C42), and blocks 60Hz commercial power noise or unnecessary interference signals in the low-frequency band, thereby allowing only valid RF signals of 80MHz or higher to pass through.

[0072] The lower limit frequency set here can be configured to 80 MHz.

[0073] The capacitor (C42) acts as a barrier that cuts off the signal flow by acting as a high resistance (impedance) for low-frequency signals such as DC and 60 Hz. On the other hand, it allows high-frequency discharge signals of several MHz or higher to pass through.

[0074] Unlike the capacitor, the inductors (L7 and L8) connected in parallel have a lower resistance (impedance) as the frequency decreases. Therefore, the inductors (L7 and L8) completely divert low frequencies that cannot pass through the capacitor (C42) and linger to the ground wire.

[0075] That is, the above high-pass filter (110) adopts a pi-shaped structure with inductors placed on both sides, rather than using only a single simple series capacitor, thereby sending low-frequency noise to ground twice to block low-frequency noise and allow high-frequency noise to pass through.

[0077] The low-pass filter (120) removes frequency bands exceeding a set upper limit frequency and is configured to include capacitors (C5, C44) connected in parallel with respect to an inductor (L4), and blocks high frequency or harmonics and noise, thereby allowing only valid RF signals of the 1.2 GHz band or lower to pass through.

[0078] To elaborate, the low-pass filter (120) is composed of a CLC pi-type structure and consists of an inductor (L4) connected in series with a signal line, a capacitor (C9) connected in parallel between the signal line at the front end (input side) of the inductor (L4) and the ground line, and a capacitor (C44) connected in parallel between the signal line at the rear end (output side) of the inductor (L4) and the ground line.

[0079] The above inductor (L4) acts as a strong resistance to ultra-high frequency noise with extremely high frequencies, thereby preventing the high frequency from passing through.

[0080] Capacitors (C9 and C44) utilize the characteristic that resistance decreases as the frequency increases to discharge ultra-high frequency noise blocked by the inductor (L4) to the ground line at both the front and rear ends.

[0081] The above low-pass filter (120) blocks unnecessary ultra-high frequency from entering the circuit after-end (RF detector or amplifier) ​​by placing capacitors in parallel.

[0083] The above high-pass filter (110, HPF) and low-pass filter (120, LPF) perform the function of a band-pass filter (BPF) that passes only frequencies in the 80 MHz to 1.2 GHz band.

[0085] The RF detector (130) converts the RF signal that has passed through the high-pass filter (110) and the low-pass filter (120) into a DC voltage proportional to the power amount (dBm), thereby compressing the high-frequency signal of a wide dynamic range into a low-frequency voltage signal.

[0086] At this time, the RF detector (130) may be a logarithmic amplifier-based detector.

[0087] That is, the RF detector (130) is converted into a signal suitable for processing by the MCU of the signal processing unit (200) described later and output.

[0088] The IC (U10, RF Log Detector IC) has multiple stages of amplifiers and detectors connected in series internally, compressing the amplitude of the input high-frequency signal into a direct current (DC) voltage that is linearly proportional to the decibel (dB) scale and outputting it.

[0089] The matching resistor (R53) connected between the signal line and ground suppresses the reflection of high-frequency signals by matching the characteristic impedance of the transmission line.

[0090] The series coupling capacitor (C18) blocks the DC component from flowing into the IC and allows only the pure AC high-frequency partial discharge signal to pass through to INHI (pin 2).

[0091] The bypass capacitor (C41) receives the signal via a single line. Therefore, the remaining input pin, INLO (pin 3), is grounded AC-wise through the bypass capacitor (C41) to operate in single-ended input mode.

[0092] 3.3V of power is supplied to the power pins of the IC (U10), VPOS1 (pin 1) and VPOS2 (pin 4).

[0093] The decoupling capacitors (C38, C40) are placed in parallel between the power line and ground, and act as energy tanks that supply pure power to the IC (U10) by diverting low-frequency and high-frequency power noise entering through the power line to ground, and immediately supply the instantaneous charge required when the IC (U10) operates at high speed.

[0095] The filter capacitor (C39) is a component connected between the CLPF (pin 14) of the IC (U10) and ground, and the capacitance value of the capacitor determines the bandwidth of whether the envelope of the partial discharge signal is extracted smoothly and roundly. That is, the filter capacitor (C39) ultimately filters the internal high-frequency vibrations and provides them as a single-shaped pulse.

[0096] Pin 11 (VSET) and Pin 12 (VOUT) are connected to each other in a feedback structure. This allows the IC (U10) to be used in a measurement mode that directly outputs a voltage proportional to the power of the input signal, rather than in a controller mode.

[0098] The amplifier (140) performs the function of scaling the low-frequency voltage signal output from the RF detector (130) to fit the input range of the ADC (Analog-to-digital converter) and outputting it.

[0099] The above amplifier (140) may be configured as a multi-stage amplifier and buffer circuit applying OP-Amps (U13.1, U13.2, U11.1, U11.2).

[0100] The OP-Amp (U13.1) is primarily stabilized and amplified using feedback resistors (R19, R72).

[0101] Partial discharge pulses can appear and disappear very briefly within microseconds. Consequently, when converting signals at the speed of a standard ADC (Analog-to-Digital Converter), there is a high probability that the partial discharge peaks will not be detected. To address this hardware-wise, a peak hold and pulse extension circuit (D2, C4, R2) is configured at the output of the OP-Amp (U13.1) to hold and extend the peak point for a certain period of time.

[0102] The diode (D2) conducts when the signal voltage passing through the OP-Amp (U13.1) rises, allowing current to pass through, but becomes reverse-directed when the voltage drops, preventing current from flowing backward.

[0103] The charging capacitor (C4) charges and holds the peak voltage of the partial discharge signal that has passed through the diode (D2). The discharge resistor (R2) gradually discharges the voltage charged in the charging capacitor (C4) to ground.

[0104] As a result, the peak hold and pulse extension circuit (D2, C4, R2) extends the pulse waveform of the instantaneous peak and transforms it into a smooth waveform.

[0106] The second amplification stage OP-Amp (U13.2) is a voltage follower structure in which the output terminal (pin 7) is directly connected to the inverting input terminal (pin 6), and it passes through without amplification, but the input impedance is close to infinity and the output impedance is close to 0.

[0107] This functions as an isolation barrier to prevent the peak voltage stored in the front charging capacitor (C4) from being unintentionally discharged quickly due to the rear circuit.

[0108] At this time, the voltage charged in the capacitor (C4) is supplied to the next stage OP-Amp (U11.2).

[0110] The output terminal (R89, R71, U11.2) is a resistor network that proportionally reduces the voltage level of the signal passing through the OP-Amp (U13.2) to the operating voltage range of the MCU (e.g., 3.3V).

[0111] Power consumption is high if the MCU operates the ADC every time to monitor the analog value. Therefore, a reference voltage (MCU_VPP, a threshold value provided by the MCU's DAC output, etc.) is applied to pin 2 of the OP-Amp (U11.2).

[0112] The moment the partial discharge signal (pin 3) exceeds this reference voltage, the OP-Amp (U11.2) detects this and generates a digital high signal through pin 1. This signal is transmitted to the interrupt pin and acts as a trigger.

[0114] The OP-Amp (U11.1) functions as an auxiliary output stage and is a buffer that branches the input signal of the OP-Amp (U13.1) to amplify the voltage. It functions as a diagnostic / auxiliary output port to send the pure analog envelope waveform itself, without passing through the peak hold (D2, C4), to external debugging equipment or another analog analysis circuit (NODE_OUT2).

[0116] The multiplexer (150) performs the function of selectively outputting a multi-channel signal output from the amplifier (140).

[0117] A0 (pin 11), A1 (pin 10), and A2 (pin 9) of the multiplexer (150) are input pins that receive signals, and these three pins form a binary address. That is, when digital signals from 000 to 111 are sent sequentially, a switch inside the multiplexer (150) selects and connects one of the eight input channels.

[0118] Pins 1, 2, 4, 5, 12, 13, 14, and 15 of the multiplexer (150) and protection resistors (R76, R78, R79, R77, R74, R81, R80, R73) connected in series thereto are input ports where partial discharge analog signals of different sensors wait. The connected resistors perform the function of suppressing noise that may occur during switching and protecting the multiplexer (150) from inrush current.

[0119] The output terminals (A0, A1, A2) output only a single channel signal selected by the control signal. The output signal is input to the analog-to-digital converter (ADC) channel.

[0121] The signal output from the multiplexer (150) of the preprocessing unit (100) is converted into a digital signal through an ADC and transmitted to the signal processing unit (200) at the rear.

[0122] Depending on the design conditions, the above ADC may be included in the signal processing unit (200).

[0124] FIG. 5 is a schematic diagram showing the configuration of a signal processing unit and an intelligent analysis unit applied to an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention.

[0126] The signal processing unit (200) performs the function of determining whether an arc has occurred by comparing a preset reference value according to the number of abnormal signal occurrences, the duration of abnormal signal occurrences, and sharpness (degree of sharpness of the signal) using the signal preprocessed by the preprocessing unit (100), and can be configured as a one-chip type MCU (Micro Controller Unit).

[0128] The intelligent analysis unit (300) receives data from the signal processing unit (200) and performs the function of determining the type of partial discharge by classifying the pattern of partial discharge using a Short-Time Fourier Transform (STFT) and a neural network algorithm, and may be composed of a small computer (Mini PC) or a graphics processor unit (GPU) that performs analysis / communication / storage / display, etc. by loading a program.

[0129] The output of the above intelligent analysis unit (300) can be transmitted to a remote user terminal or server, etc., through a Human Machine Interface (HMI).

[0131] First, the signal processing unit (200) will be explained.

[0133] The above signal processing unit (200) detects the number of occurrences of abnormal signals, the period time of occurrence of abnormal signals, and the sharpness of which can be determined as characteristic values ​​of an arc signal, and determines it as an arc when each of the detected characteristic values ​​is greater than or equal to a set reference value.

[0134] The signal transmitted and input from the above preprocessing unit (100) is defined as x[n] = [x1, x2, x3 … , xn], and the abnormal signal in the input signal data is defined by the following Equation 3.

[0135] Formula 3)

[0136]

[0137]

[0138] Here, I[n] is the abnormal signal of the n-th input signal, and T is the threshold value for determining whether it is abnormal.

[0139] In other words, if the input signal exceeds the reference value, it is defined as an abnormal signal, and if the input signal is below the reference value, it is defined as a normal signal.

[0141] From the above Equation 3, the number of occurrences of abnormal signals (N abn ) is calculated by the following formula 4.

[0142]

[0143] Here, N abn is the number of occurrences of abnormal signals, and I[n] is the abnormal signal of the nth input signal.

[0145] In addition, the generation period time of the abnormal signal (L max ) refers to the time difference from when an abnormal signal occurs until the time when the next abnormal signal occurs, and is the abnormal signal occurrence cycle time (L max ) is defined by the following formula 5.

[0146] Formula 5)

[0147]

[0148] Here, L max is the abnormal signal generation cycle.

[0150] In addition, sharpness (k) represents the height of the amplitude relative to the amplitude and is calculated by the following Equation 6.

[0151] Formula 6)

[0152]

[0153] Here, k is the vortex, E is the mean of the covariance, x[n] is the input signal, μ is the mean of the entire input signal, and σ is the standard deviation of the entire input signal.

[0155] Because arc discharge is accompanied by a very sharp and strong impulse instantaneously, it has a much higher sharpness than steady-state noise peaks. In other words, sharpness is an indicator that confirms the intensity of sudden peaks.

[0157] Through the above formulas 4 to 6, the conditions for determining an arc are that the number of occurrences and sharpness of abnormal signals are greater than or equal to a preset reference value, and the period of occurrence of abnormal signals is less than or equal to a reference value.

[0158] To elaborate, the number of occurrences of each calculated abnormal signal (N abn ), abnormal signal occurrence cycle time (L max ) and sharpness (k) and set reference value (N th , L th , k th The arc is determined by comparing each of the ).

[0159] In the above, N th is the reference value for the number of abnormal signal occurrences, L th is the reference value of the abnormal signal generation period time, and k th is the standard value of sharpness.

[0161] Here, the judgment of the arc is the number of abnormal signal occurrences (N abn ) is the reference value (N) of the number of abnormal signal occurrences th ) Anomaly, the occurrence period time of the abnormal signal (L max ) Reference value of the abnormal signal occurrence cycle time (L th ) or less, sharpness (k) is the reference value of sharpness (k th It can be determined as an arc if at least two of the above cases are satisfied.

[0162] Depending on the design conditions, the number of occurrences of the above abnormal signal (N abn ) is the reference value (N) of the number of abnormal signal occurrences th ) or more, and the generation period time (L) of the above abnormal signal max ) Reference value of the abnormal signal occurrence cycle time (L th) or less, and the above sharpness (k) is the reference value of sharpness (k th It can be configured to determine the arc in the case of ) or more.

[0163] This can drastically reduce false alarms and increase the reliability of the diagnosis.

[0164] When an arc is determined through Equations 3 to 6, the signal processing unit (200) is configured to output a Trip Signal and an Alarm Signal.

[0165] The above trip signal can be used as a signal to cut off power by operating a circuit breaker, etc., and the alarm signal can be used as a signal to generate a warning light or a warning sound.

[0166] Here, the signal processing unit (200) is configured as an Edge AI that processes the situation at the site independently.

[0167] That is, the preprocessing unit (100) removes unnecessary noise from the complex signal detected by the high-frequency inductive ground voltage sensor (20), significantly reduces the amount of computation to be performed, and converts it into a signal pattern that the signal processing unit (200) can read immediately, and provides it.

[0168] Accordingly, the signal processing unit (200) is installed at the site (Edge) where the signal is detected, i.e., at the distribution panel, and is configured to immediately determine whether the signal transmitted from the preprocessing unit (100) is an arc or not.

[0169] The arc progresses into a fire within a few milliseconds (ms). Consequently, there is no time to send the signal detected by the high-frequency induction ground voltage sensor (20) to a remote server.

[0170] Accordingly, by processing the signal detected by the high-frequency inductive ground voltage sensor (20) on-site, sensitive power equipment data is processed internally without being sent to an external device. As a result, there is an advantage of excellent security, reduced load on the communication network (or network, etc.) since there is no need to transmit the entire detected signal, and the ability to perform intelligent protection functions through the judgment of the signal processing unit (200) itself even in an offline state where communication cannot be performed through the communication network (or network).

[0172] Next, signals other than arcs, that is, signals with an amplitude smaller than that determined to be arcs, are transmitted from the signal processing unit (200) to the intelligent analysis unit (300).

[0174] The intelligent analysis unit (300) receives data from the signal processing unit (200), classifies the pattern of partial discharge using a Short-Time Fourier Transform (STFT) and a neural network algorithm, and performs the function of determining the type of partial discharge based on the classified pattern.

[0176] The Short-Time Fourier Transform (STFT) is an algorithm that simultaneously analyzes the time-frequency information of abnormal signals whose frequency components change over time by dividing a long signal into short time intervals (frames) and performing a Fourier Transform (FFT) on each. While the general Fourier Transform (FT) provides only frequency information for the entire time domain, the Short-Time Fourier Transform is characterized by including the point in time when the frequency occurred.

[0177] That is, the short-time Fourier transform above divides the signal into short, overlapping intervals, applies a window function to each divided interval to perform an FFT on each interval, and combines the results of each interval in a time-frequency representation, thereby generating a spectrogram that shows how the frequency changes over time.

[0178] The above Short-Time Fourier Transform is used when the signal changes over time, transient events occur, and time-frequency characteristics are important.

[0180] The process of converting a signal detected by a high-frequency induction ground voltage sensor (20) through the short-time Fourier transform in the intelligent analysis unit (300) is explained.

[0181] Continuous signal samples collected from the high-frequency inductive ground voltage sensor (20) are defined as S1, S2, …, Sn, and each signal is divided into frames of the minimum analysis unit, and the nth sample value of the kth frame is defined by the following formula 7.

[0182] Formula 7)

[0183]

[0184] Here, S k is the signal of the k-th frame, s k [n] is the k-th frame of the n-th sample value.

[0185] For example, if the number of samples is 600 and one sample is divided into 128 frames, a total of 76,800 frames are generated.

[0186] For each generated frame, the Short-Time Fourier Transform (STFT) is defined by the following Equation 8.

[0187] Formula 8)

[0188]

[0189] Here, Xk(m,ω) is the spectrogram at angular frequency (ω) for time position (m), n is the number of samples, k is the number of frames, and s k [n] is the k-th frame of the n-th sample value, ω[nR] is the window function, e -jωm is the complex exponential function of the Fourier transform.

[0190] In the above, the window function (ω[nR]) is a function for selecting only a specific section within the frame, where R is the position where the window moves within the frame.

[0192] Figure 6 shows a spectrogram of an embodiment generated through a short-time Fourier transform in an AI pattern classification-based intelligent switchboard protection system using a high-frequency induced ground voltage signal according to the present invention.

[0193] Referring to the attached Fig. 6, the spectrogram represents three-dimensional information on a two-dimensional plane, where the horizontal axis represents time and the vertical axis represents frequency, and the color / intensity represents the energy intensity (magnitude / power) at a specific time and frequency.

[0194] Accordingly, a two-dimensional spectrogram is generated through a short-time Fourier transform in the intelligent analysis unit (300). The spectrogram is a two-dimensional image, and by analyzing the two-dimensional image having time-frequency, the pattern and type (classification) of the partial discharge are determined.

[0195] That is, the intelligent analysis unit (300) uses a spectrogram generated through a short-time Fourier transform as data to classify the pattern of partial discharge through a neural network algorithm, and determines the type of partial discharge based on the classified pattern.

[0196] In the present invention, a 2D Convolutional Neural Network (2D-CNN) is applied as the neural network algorithm.

[0197] That is, since the spectrogram generated through the Short-Time Fourier Transform (STFT) is two-dimensional data (image) with intensity of time and frequency, the neural network algorithm infers partial discharge using a two-dimensional convolutional artificial neural network (2D-CNN), classifies the pattern of partial discharge, and determines the type of partial discharge based on the classified pattern.

[0199] Figure 7 is a diagram showing the structure of a two-dimensional convolutional artificial neural network in an intelligent switchboard protection system based on AI pattern classification using high-frequency induced ground voltage signals according to the present invention.

[0200] In addition, FIG. 7a is a structural diagram of a one-dimensional convolutional neural network explaining the change in data form in an intelligent switchboard protection system based on AI pattern classification using high-frequency induced ground voltage signals according to the present invention.

[0201] A 2D convolutional neural network (2D-CNN) is a deep learning optimized for extracting spatial features from 2D data of spectrograms.

[0202] Referring to the attached Fig. 7, the two-dimensional convolutional artificial neural network consists of an input layer, a convolution layer, a pooling layer, a fully connected layer corresponding to the hidden layer of a general Neural Network (NN), and an output layer. Among these layers, the convolution layer and the pooling layer, which perform convolution operations, play important roles in the CNN.

[0204] The input layer receives data representing amplitude changes over time, and the convolutional layer extracts local features of the data by sliding multiple filters (kernels) over the signal; using K kernels (filters) generates K feature maps, through which specific patterns or frequency characteristics within the signal are detected.

[0205] The pooling layer performs downsampling to reduce the size of detected feature maps, compressing the amount of data to decrease computational load and filtering out noise from the model by retaining only important features.

[0206] The flattening layer generates a single long 1D vector from data arranged in a 2D form (number of feature maps × time axis). The flattening layer serves as a preparatory step for feeding all extracted features into a classifier, where N represents the total length of this vector.

[0207] The fully connected layer connects flattened features to all nodes of the neural network for comprehensive analysis. It performs high-level decision-making by combining high-dimensional extracted features to ultimately determine which class a component belongs to.

[0208] The output layer outputs the final classification result as a probability value. The probability for a specific result appears highest, and through this, the type of partial discharge for the input signal is finally determined.

[0210] 1. Input layer

[0211] The input layer is the layer that receives the spectrogram.

[0212] The above spectrogram is an image generated by the short-time Fourier transform of the intelligent analysis unit (300), and the size (width, height) and number of channels (color) of the spectrogram are defined.

[0214] 2. Convolution layer

[0215] A convolution layer generates a feature map by moving multiple kernels of a predetermined size through the input data, multiplying the kernel and pixel value at each position, and summing them.

[0216] FIG. 8 is a diagram illustrating the process of extracting a feature map from a convolutional layer according to one embodiment in a one-dimensional convolutional neural network in an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention.

[0217] Referring to the attached Fig. 8, the input data in the input layer is superimposed with a fixed kernel (or filter), and after multiplying the input data and the kernel by a corresponding weight, the total sum is calculated and the calculated result is mapped to the first component of the feature map.

[0218] The kernel is moved one step to the right according to the set stride (Stride, set to 1 in Fig. 8) and overlaps with the data in the next area. A summation operation is performed again using the same kernel weight as the data in the moved area, and the resulting value is mapped to the second component of the feature map.

[0219] The above operations are performed sequentially in the row and column directions over the entire area of ​​the input data to complete the final feature map.

[0220] In the process of performing operations in the above convolutional layer, padding and stride must be considered.

[0221] Padding is the process of expanding input data by filling the surroundings with specific values ​​before performing a convolution operation, and zero padding is generally used.

[0222] The spatial size decreases with each operation performed in the convolutional layer, and the sizes of the input and output feature maps change.

[0223] Consequently, a problem arises where information at the edges is lost, so padding is used to prevent this.

[0224] For example, if zero padding is added to 4×4 input data, it becomes 6×6 data, and if a 3×3 kernel (or filter) is used, a 4×4 feature map is generated. If padding of size 2 is added to 4×4 input data, it becomes 8×8 data, and if padding of size 3 is added, it becomes 10×10 data.

[0225] In other words, padding is primarily used to adjust the output size.

[0226] Stride refers to the interval at which the kernel moves during the process of applying a kernel (filter) to input data.

[0227] That is, by adjusting the stride, the size of the feature map is reduced relative to the input data, and by adjusting the stride, the amount of computation can be significantly reduced, and the diversity of feature extraction can be derived using the stride value.

[0228] This operation method compresses information while preserving the spatial relationship of the input data.

[0229] Furthermore, computational efficiency is maximized by connecting only local regions the size of the kernel instead of connecting the entire input data at once; a higher computational result indicates that the corresponding region has a high degree of similarity to specific features (Edge, Texture, etc.) defined by the kernel.

[0230] The above kernel can be composed of 3×3 or 5×5.

[0231] The image, converted into input data in the form of a multidimensional array in the above input layer, is shifted by a specified number of cells and performs a convolution operation with a kernel to extract features of the input data while simultaneously reducing the dimensionality of the data to generate a feature map. Here, the size of the feature map depends on the size of the input data, the size of the filter, the presence or absence of padding, and the shift interval. This is expressed by the following Equation 9.

[0232] Formula 9)

[0233]

[0234] Here, O is the size of the feature map, I is the size of the input image, F is the size of the filter, S is the shift interval of the filter, and P is the size of the padding.

[0235] The padding mentioned above is a method of creating additional extra space in the input data to prevent data information loss.

[0236] However, as the number of filters increases, a relatively large number of feature maps accumulate. This requires high-dimensional parameters, which can lead to a decrease in learning speed and overfitting problems. Pooling layers exist to solve this.

[0238] 3. Pooling layer

[0239] The pooling layer outputs multidimensional feature information by selecting the maximum or average value of a local region for the feature map generated by the convolutional layer, thereby reducing the size and computational load of the feature map and lowering the sensitivity to changes in spatial position.

[0241] Max pooling is applied to extract key information by selecting the largest value among the values. This method significantly reduces computational load while preserving the strongest peak energy in partial discharge, enabling the model to stably recognize even minute changes in the temporal position of the signal.

[0243] Since the pooling layer performs simple operations of extracting the maximum value from a specific range or calculating the average, rather than multiplying weights, there are no parameters to learn.

[0244] Since the above pooling layer is performed independently for each channel, the number of channels in the input data does not change even after the operation, and even if an object within the input data moves or rotates slightly, it can be recognized as the same feature by abstracting the location of the feature through pooling.

[0245] Max pooling and average pooling are generally used in pooling layers.

[0246] Max pooling is a method that selects only the largest value from the target area.

[0247] Max pooling is the most widely used method for removing noise and extracting image boundaries or textures, as it has the characteristic of preserving strong signals (key features).

[0248] Average pooling is a method of calculating the average of all values ​​within a target area.

[0249] The above average pooling evenly reflects the overall information of the input data, but it is used less frequently than maximum pooling because strong features can be diluted.

[0251] FIG. 9 is a diagram illustrating the process of extracting a feature vector using maximum pooling in a pooling layer according to an embodiment of a one-dimensional convolutional neural network in an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention.

[0252] Figure 9 (a) shows a maximum pooling method, which extracts the maximum value from a selected area of ​​a feature map, and Figure 9 (b) shows an average pooling method, which calculates the average of all values ​​within a target area.

[0253] As a result, the pooling layer compresses the data and outputs it.

[0254] In both the max pooling method and the average pooling method, the size of the feature map is reduced by half, which is expressed by the following Equation 10.

[0255] Formula 10)

[0256]

[0257] Here, O is the size of the feature map, I is the size of the input image, Fp is the size of the pooling filter, and S is the filter shift interval.

[0258] In this case, Fp is expressed as [Fp×Fp], and the sizes of Fp and S are set to be the same.

[0260] In this invention, the Max Pooling method is applied in the Pooling layer.

[0261] In this way, the feature map, with its dimensionality reduced through the processes of the Convolutional Layer and Pooling Layer, is finally passed to the Fully Connected Layer.

[0263] 4. Fully Connected Layer

[0264] The fully connected layer converts the multidimensional feature information output from the pooling layer into a one-dimensional vector, and classifies the pattern and severity of partial discharge through linear combination and activation functions based on the converted one-dimensional vector.

[0265] At this time, during the training process of the convolutional neural network (CNN), the fully connected layer is configured to deactivate a randomly selected neuron among the neurons composed of the linear combination and activation function to reduce coexistence dependency between the neurons and prevent overfitting.

[0266] The pattern and severity of the partial discharge classified in the above fully connected layer are transmitted to the output layer along with the activation function.

[0268] 5. Output Layer

[0269] The output layer calculates a probability calculation of the partial discharge type for the partial discharge pattern classified in the fully connected layer using a softmax function, and determines one type of partial discharge selected from noise, corona, floating, insulator void, and insulator surface based on the calculated probability calculation.

[0270] As described above, the softmax function is a function that normalizes all input values ​​to values ​​between 0 and 1, and has the characteristic that the sum of the normalized values ​​is always 1.

[0271] The softmax function is expressed by the following Equation 11.

[0272] Formula 11)

[0273]

[0274] Here, is the softmax function of the k-th node, n is the number of nodes in the output layer, z k is the input of the k-th node of the output layer, z i is the input of the i-th node of the output layer.

[0275] In the above Equation 11, the Cross-Entropy function is used as the loss function, and the above Cross-Entropy function is expressed by the following Equation 12.

[0276] Formula 12)

[0277]

[0278] Here, CE is the loss function, S is the total number of samples, k is the node, and t ik is the output of the actual class, is the output of the pattern classifier.

[0279] In the present invention, backpropagation is used for CNN training.

[0281] In this way, the probability that an image belongs to each label is output for each label through the softmax function of the output layer, and the label with the highest probability among them is selected as the final prediction.

[0283] In the present invention, the prediction of partial discharge based on inference is classified into four classes, such as noise, defective insulator discharge, floating electrode discharge, and corona discharge, and the type of partial discharge is determined and output.

[0285] Noise is a signal caused by external environmental factors rather than actual discharge within the equipment. It is generated by surrounding communication lines, broadcast waves, switching noise from motor drive circuits, fluorescent lights, or electrical interference from the measuring instruments themselves.

[0286] The aforementioned noise appears randomly regardless of the power system's phase or occurs continuously in a specific frequency band.

[0288] Defective Insulator Discharge is a discharge that occurs in voids within an insulator, at the interface between different insulators, or on the surface of an insulator.

[0289] The above-mentioned insulator discharges include internal discharge (Internal PD) occurring within voids generated during the manufacturing process of epoxy or solid insulators, and surface discharge (Surface PD) occurring along the creepage distance when contaminants or moisture are present on the surface of the insulator.

[0290] The aforementioned insulator discharge is a dangerous type with the highest potential to lead to equipment accidents; as the discharge progresses, it forms a treeing phenomenon within the insulator and ultimately causes insulation breakdown.

[0292] Floating Electrode Discharge is a discharge that occurs when a conductor is not grounded or is not fully connected to other conductors, and the electric potential within the electric field is in an unfixed (floating) state.

[0293] The above-mentioned floating discharge occurs when metal parts (bolts, nuts, spacers, etc.) inside the equipment become separated due to vibration or when contact is unstable due to poor assembly.

[0294] The above floating discharge has a constant discharge magnitude and a very regular pattern. In addition, it tends to occur symmetrically in both positive (+) and negative (-) cycles of the applied voltage, and since it is a discharge between metals, the electromagnetic wave signal is detected very strongly.

[0296] Corona discharge is a discharge that occurs when an electric field is concentrated at the sharp ends (pointed protrusions) or regions of high curvature of a conductor within a gaseous insulator.

[0297] The above corona discharge occurs when an electric field exceeding the dielectric breakdown strength of air is formed near scratches on the surface of the wire, sharp metal protrusions, or insulators of overhead transmission lines.

[0298] The above corona discharge mainly occurs near the peak of the positive (+) or negative (-) half-cycle of the applied voltage phase angle, and becomes more active when humidity is high or atmospheric pressure is low. Although the corona discharge may not be fatal to the equipment in the early stages, if it persists, it generates ozone, which causes oxidation and corrosion of surrounding insulation materials.

[0300] FIG. 10 is a schematic diagram showing the configuration of a signal processing unit and an intelligent analysis unit of another embodiment in an AI pattern classification-based intelligent switchboard protection system using a high-frequency induced ground voltage signal according to the present invention.

[0301] Referring to the attached FIG. 10, the signal preprocessed in the preprocessing unit (100) can be transmitted to the signal processing unit (200) or to the intelligent analysis unit (300) to determine an arc or diagnose a partial discharge, respectively.

[0302] That is, the signal processing unit (200) and the intelligent analysis unit (300) are configured in parallel to detect arcs and partial discharges, respectively.

[0304] The arc detected by the signal processing unit (200) is an accident that occurs after partial discharge has occurred. Partial discharge is a cause or precursor phenomenon that triggers the arc, and the arc can be seen as the final result of insulation breakdown caused by partial discharge being left unchecked.

[0305] Specifically, partial discharge is a local discharge phenomenon occurring in bubbles, impurities, or sharp parts of electrodes within an insulator, and can be considered an initial stage of dielectric breakdown that does not completely bridge between electrodes.

[0306] If such micro-discharges are repeated, insulation performance gradually deteriorates and can eventually lead to an arc, a massive discharge phenomenon.

[0307] An arc is a high-energy discharge that occurs when dielectric breakdown progresses and the insulating performance between electrodes is completely lost.

[0308] While partial discharge occurs with very low energy, an arc is accompanied by heat and light and possesses powerful energy that causes direct damage (melting, burning, etc.) to equipment.

[0309] If partial discharge accumulates within power equipment such as switchboards and forms a carbonization track, an arc occurs along it, causing a failure of the entire system.

[0311] In the present invention, the signal processing unit (200) detects an arc using the signal preprocessed in the preprocessing unit (100), and if it is determined that it is not an arc, the intelligent analysis unit (300) detects a partial discharge.

[0312] In this case, arc accidents can be prevented if partial discharge is detected and measures are taken in advance.

[0313] Therefore, the signal processing unit (200) may be omitted depending on the design conditions.

[0314] That is, the arc and partial discharge detection system for a switchboard using a high-frequency induction-to-ground voltage sensor according to the present invention comprises a high-frequency induction-to-ground voltage sensor (20) installed attached to the enclosure of a switchboard (10) and an analysis device (30) that determines and analyzes arcs and partial discharges based on signals detected from the high-frequency induction-to-ground voltage sensor (20). The analysis device (30) may be configured to include a preprocessing unit (100) that preprocesses signals output from the high-frequency induction-to-ground voltage sensor (20); and an intelligent analysis unit (300) that receives data from the preprocessing unit (100) and determines the occurrence and type of partial discharge using wavelet transform and neural network algorithms.

[0316] Next, the learning process regarding the performance of an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention will be described.

[0318] FIG. 11 is a diagram showing the configuration of a simulation for an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention.

[0319] Referring to the attached Fig. 11, an enclosure (E) and a PD generator (PG) that generates partial discharge upon the application of voltage are arranged inside the enclosure (E), and a power supply unit that supplies voltage to the PD generator (PG) is configured.

[0320] A high-frequency induction ground voltage sensor (20) applied to the present invention is attached to the outer surface of the above-mentioned housing (E), and an analysis device (30) is configured to analyze the signal output from the high-frequency induction ground voltage sensor (20).

[0321] The above enclosure (E) used the company's 22.9KV enclosed switchgear.

[0322] The above PD generator (PG) conducted experiments by replacing each electrode to generate noise, insulator discharge, floating discharge, and corona discharge.

[0323] The power supply unit applied voltage to the PD generator (PG) using a high-voltage transformer with a maximum output voltage (50kV) and current (30mA) of 220V AC power, and adjusted the applied voltage through a voltage regulator to apply a voltage between 4.1 kV and 7 kV. Each data sample was stored at a time interval of 30 seconds so that the analysis device (30) could diagnose partial discharge every 30 seconds.

[0324] Learning using the above analysis device (30) was performed through repetition for each type of partial discharge.

[0326] Figure 12 is a verification graph of the accuracy in the process of learning partial discharge applied to an intelligent switchboard protection system based on AI pattern classification using high-frequency induced ground voltage signals according to the present invention.

[0327] Figure 12 attached is a graph showing how the accuracy of 2D-CNN changed over 2,000 epochs while learning partial discharge.

[0328] Training converged in about 500 epochs, and the comprehensive classification accuracy was 99.8%, accurately classifying 1,551 out of a total of 1,554 data points.

[0329] In the early stages of training (between 0 and 100 epochs), the accuracy increased rapidly and converged to 1.0 (100%) very quickly.

[0330] This means that the neural network algorithm of the present invention distinguishes whether or not there is partial discharge or the type of partial discharge very accurately through 2D data (spectrogram).

[0331] In addition, the learning line (blue line) and the experiment line (orange line) appear to overlap, and since both indicators converge to 1.0, it can be seen that the model has very high reliability even with new data.

[0332] The fact that the graph reached 1.0 smoothly and quickly, as shown in the attached drawing, means that the neural network algorithm used in the experiment has very clear and easy-to-distinguish characteristics to explain each partial discharge state.

[0333] Accordingly, from the perspective of partial discharge learning, this is interpreted to mean that noise and partial discharge are clearly distinguished in the data collected by the high-frequency inductive ground voltage sensor.

[0335] Figure 13 is a verification graph of the standard deviation in the process of learning partial discharge applied to an intelligent switchboard protection system based on AI pattern classification using high-frequency induced ground voltage signals according to the present invention.

[0336] Figure 13 attached shows how the Standard Deviation (STD) changes as learning progresses, and as a graph, it is possible to see how confidently the model predicts during the learning of the neural network algorithm of the present invention.

[0337] Referring to the attached Fig. 13, the STD (standard deviation) value is very high and fluctuates during the initial stage (0 to 250 epochs). This indicates an unstable state of learning.

[0338] However, after 1,000 epochs, the standard deviation (STD) converges to a very low value close to zero. This means that as the model repeats the learning process, the variability in the prediction results has almost disappeared, indicating that it produces consistent and stable performance regardless of the input data.

[0339] The fact that the quality (STD) of the partial discharge experimental data converges to such a low level suggests that the characteristics within the experimental data are highly consistent.

[0341] In this way, experiments on each partial discharge were conducted using the 2D-CNN, which is the neural network algorithm of the present invention learned.

[0342] For each partial discharge, the number of experiments was 192 for noise, 605 for insulator discharge, 281 for corona discharge, and 474 for floating discharge in a random order.

[0343] In each experiment, partial discharge was detected by a high-frequency induction ground voltage sensor (20), a spectrogram was generated through a short-time Fourier transform, and the generated spectrogram was applied to a 2D Convolutional Neural Network (2D-CNN) to infer partial discharge through probabilistic calculation.

[0345] Figure 14 is a table showing the learning results of a simulation for an intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to the present invention.

[0346] In the attached Fig. 14, class 0.0 means noise, 1.0 means insulator discharge, 2.0 means corona discharge, and 3.0 means floating discharge.

[0347] Referring to the attached Fig. 11, when the spectrogram generated through the short-time Fourier transform applied in the present invention is trained with a 2D convolutional artificial neural network (2D-CNN), noise is detected with an accuracy of 99.5%, insulator discharge with 100.0%, corona discharge with 99.3%, and floating discharge with 100.0%, it can be confirmed that the classification rate performance is high.

[0349] According to the present invention, an arc is quickly determined from the detected high-frequency induced ground voltage, and a trip signal is output immediately when it is determined to be an arc to prevent secondary accidents caused by the arc. In the case of non-arc, the occurrence of partial discharge can be diagnosed, so there is an advantage of being able to selectively respond to arcs and partial discharges and improve reliability.

[0350] In addition, since the type and risk level of partial discharge can be accurately determined from the initial stages, there is an advantage in being able to take appropriate measures before it progresses into a major accident such as an arc.

[0352] Although preferred embodiments of an intelligent switchboard protection system based on AI pattern classification using high-frequency induced ground voltage signals according to the present invention have been described above, the present invention is not limited thereto and may be implemented with various modifications within the scope of the claims, the description of the invention, and the attached drawings, and such modifications also fall within the scope of the present invention. Explanation of the symbols

[0354] 10: Switchboard 20: High-frequency induction ground voltage sensor 30: Analysis device 100: Preprocessing unit 110: High-pass filter 120: Low-pass filter 130: RF detector 140: Amplifier 150: Multiplexer 200: Signal processing unit 300: Intelligent Analysis Unit

Claims

Claim 1 The apparatus includes a high-frequency induction ground voltage sensor (20) attached to the enclosure of a distribution board (10) and an analysis device (30) that determines and analyzes arcs and partial discharges based on a signal detected from the high-frequency induction ground voltage sensor (20). The analysis device (30) comprises: a preprocessing unit (100) that preprocesses a signal output from the high-frequency induction ground voltage sensor (20); and a signal processing unit (200) that determines whether an arc has occurred by comparing a preset reference value according to the number of abnormal signal occurrences, the duration of abnormal signal occurrences, and sharpness using the signal preprocessed by the preprocessing unit (100). An intelligent distribution board protection system based on AI pattern classification using high-frequency induced ground voltage signals, comprising: an intelligent analysis unit (300) that receives data from the signal processing unit (200), classifies patterns of partial discharge using a Short-Time Fourier Transform (STFT) and a neural network algorithm, and determines the type of partial discharge based on the classified patterns; wherein the neural network algorithm of the intelligent analysis unit (300) applies a 2D Convolutional Neural Network (2D-CNN) to infer partial discharge through probabilistic calculation. Claim 2 An intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal, wherein, in claim 1, the high-frequency induced ground voltage sensor (20) comprises: a detection electrode (21) installed in contact with a detection surface; a ground electrode (22); a dielectric (23) inserted between the detection electrode (21) and the ground electrode (22); and a BNC terminal (24) to which the detection electrode (21) is connected to a core wire and outputs a signal. Claim 3 An intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to claim 1, wherein the preprocessing unit (100) comprises: a high-pass filter (110) that removes a frequency band below a set lower limit frequency; a low-pass filter (120) that removes a frequency band exceeding a set upper limit frequency; an RF detector (130) that converts a signal passing through the high-pass filter (110) and the low-pass filter (120) into a DC voltage proportional to the power amount (dBm); an amplifier (140) that scales and outputs a low-frequency voltage signal output from the RF detector (130) to fit the input range of an ADC (Analog-to-digital converter); and a multiplexer (150) that selectively outputs a multi-channel signal output from the amplifier (140). Claim 4 An intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to claim 1, wherein the signal processing unit (200) determines an arc when at least two of the following conditions are satisfied: the number of occurrences of the abnormal signal is greater than or equal to a reference value for the number of occurrences of the abnormal signal, the period time of occurrence of the abnormal signal is less than or equal to a reference value for the period time of occurrence of the abnormal signal, and the sharpness is greater than or equal to a reference value for the sharpness. Claim 5 An intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to claim 1, wherein the signal processing unit (200) determines an arc when the number of occurrences of the abnormal signal is greater than or equal to a reference value for the number of occurrences of the abnormal signal, the period time of occurrence of the abnormal signal is less than or equal to a reference value for the period time of occurrence of the abnormal signal, and the sharpness is greater than or equal to a reference value for the sharpness. Claim 6 An intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal, characterized in that, in claim 1, the signal processing unit (200) is configured to output a trip signal and an alarm signal when an arc is determined. Claim 7 An intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to claim 1, wherein the intelligent analysis unit (300) generates a spectrogram through the short-time Fourier transform in which the horizontal axis represents time, the vertical axis represents frequency, and the color / intensity represents the energy intensity (magnitude / power) at a specific time and frequency. Claim 8 delete Claim 9 An intelligent switchboard protection system based on AI pattern classification using a high-frequency induced ground voltage signal according to claim 1, wherein the probability calculation infers partial discharge through the following formula. Formula) Here, is the softmax function of the k-th node, n is the number of nodes in the output layer, z k is the input of the k-th node of the output layer, z i is the input of the i-th node of the output layer. Claim 10 An intelligent switchboard protection system based on AI pattern classification using high-frequency induced ground voltage signals, characterized in that, in claim 9, the prediction based on the inference of the partial discharge is noise, defective insulator discharge, floating electrode discharge, and corona discharge. Claim 11 delete

Images