Intelligent Distribution Board Having a Real-Time Integrated Safety Diagnosis System and Method Based on Partial Discharge

Abstract

The present invention relates to an intelligent switchboard device and a real-time diagnostic method for monitoring the internal state of power equipment connected to a power grid network, comprising: an internal environment monitoring module that continuously generates environmental state data by measuring changes in the internal physical environment of the intelligent switchboard device (100) in real time; a discharge signal detection module that acquires an analog discharge raw signal by capturing electromagnetic waves and acoustic radiation signals generated from internal power lines and insulators of the intelligent switchboard device; an edge inference computation unit that independently collects and processes the environmental state data and the analog discharge raw signal within the intelligent switchboard device without transmitting them to an external server to determine complex risk factors in real time; and an active control response unit that generates a response control signal instructing the operation of an external facility in response to at least one of the complex risk factors determined by the edge inference computation unit and an emergency environmental abnormal signal received directly from the internal environment monitoring module, and outputs the response to the external facility.

Description

Technology Field

[0001] The present invention relates to an intelligent switchboard equipped with a partial discharge-based real-time complex safety diagnosis system and a real-time diagnosis method using the same. More specifically, the invention relates to an intelligent switchboard equipped with a partial discharge-based real-time complex safety diagnosis system that measures changes in the physical environment and discharge signals inside the switchboard in real time, independently identifies complex risk factors through edge computing-based artificial intelligence inference, and performs active control responses, and a real-time diagnosis method using the same. Background Technology

[0002] In modern industrial society, a stable power supply has become an indispensable element of the national economy and industrial sites. Switchboards, which are core facilities of the power grid network, handle high voltage and high current, and various power devices such as circuit breakers, transformers, and relays are densely arranged inside. As these switchboards operate for long periods, they are continuously exposed to mechanical and electrical stresses as well as various environmental factors, such as insulation degradation, loosening of contacts, and the ingress of dust and moisture. Consequently, insulation performance deteriorates or partial discharge occurs, which carries serious potential risks that could ultimately lead to large-scale fires, explosions, and widespread power outages.

[0003] Conventional switchboard monitoring systems have primarily relied on threshold-based passive monitoring methods, equipped with basic sensors such as simple temperature and humidity sensors or smoke detectors, to sound an alarm only when a specific value exceeds a single preset threshold. However, defects occurring within a switchboard do not manifest merely as fragmentary phenomena such as a rise in temperature or the generation of smoke.

[0004] In the early stages of partial discharge or insulation degradation, various physical, chemical, and electromagnetic signs occur in combination, including minute ultraviolet corona, ultra-high frequency electromagnetic waves, and minute acoustic and vibration signals. Conventional fragmentary monitoring systems have structural limitations in that they dismiss these early signs of defect as individual noise or fail to detect them at all.

[0005] Looking at the prior art, registered patent No. 10-2079813, “25.8kV eco-friendly switchgear equipped with partial discharge diagnostic function and IoT technology,” is configured in a molded manner so that the UHF sensor detecting partial discharge is not exposed to air, thereby minimizing detection errors caused by noise. It also learns the partial discharge and temperature detected in the switchgear to obtain a normal operation pattern and, based on this, enables the diagnosis of the operating status of the switchgear.

[0006] Registered Patent No. 10-2295784, “Risk Management System and Method through Real-time Partial Discharge Tracking of Distribution Facilities,” distinguishes between PD and noise regarding partial discharge information provided by a real-time wide-area monitoring system, tracks facilities connected to underground lines where PD occurs and affects nearby facilities, and enables the management of relative risk for facilities under management by simultaneously considering the frequency of occurrence over a period and the impact of failure, in addition to the PD magnitude.

[0007] Registered Patent No. 10-2377936, “Intelligent Switchboard Partial Discharge Monitoring and Diagnosis System Using Ultra-High Frequency Electric Signals,” utilizes ultra-high frequency electric signals to rapidly detect partial discharges caused by insulation failure, connection failure, or open circuits in power equipment inside a switchboard using ultra-high frequency sensors, while monitoring the occurrence of partial discharge and enabling diagnosis through an intelligent algorithm.

[0008] Furthermore, some intelligent diagnostic systems recently introduced adopt a centralized data processing structure that transmits vast amounts of raw data collected from sensors to a remote central control server or cloud via wired or wireless communication networks for analysis. Due to limitations in network bandwidth, such centralized systems inevitably suffer from data transmission delays. In particular, during ultra-urgent situations where catastrophic consequences can occur in a split second—such as arc faults in power facilities—waiting for analysis results from the central server leads to the loss of the "golden time," which is a fatal flaw. Additionally, if the external communication infrastructure is paralyzed or severed due to a disaster, the switchboard possesses a critical vulnerability, being left helplessly without being able to perform any internal fault detection or circuit breaker control.

[0009] Furthermore, conventional partial discharge analysis technologies have been limited to simply filtering collected signals by frequency band or comparing only waveform amplitudes, raising concerns that they cannot accurately identify various external industrial noises or minute defect signals within the actual switchboard. Since powerful electromagnetic interference from large inverters and motors is ever-present in power plant sites, analyzing raw signals directly results in an extremely high probability of false alarms, increasing the fatigue of field managers and severely compromising system reliability. In particular, methods that quantitatively analyze only the frequency or amplitude of discharge signals without considering their phase characteristics cannot specifically identify the concrete cause of the defect—whether it is a void discharge, surface discharge, or corona discharge—thereby limiting the ability to establish appropriate preemptive maintenance measures.

[0010] To overcome these technical limitations, various attempts are being made to incorporate the latest artificial intelligence deep learning technology into diagnostic systems; however, neural network models inherently have limitations in that they require massive computing power and memory resources. In the environment of small edge devices with limited form factors that can be mounted inside switchboards, processing heavy AI models ranging from tens to hundreds of megabytes in real-time is nearly impossible due to hardware constraints. As a result, most diagnostic technologies remain limited to post-diagnosis analysis on a central server and have not evolved into true autonomous protection systems where field equipment independently performs real-time inference to decide on immediate circuit breaker opening or fire suppression system integration.

[0011] Therefore, at this current juncture where the complexity of power grids is increasing exponentially and unmanned facilities, such as unmanned smart substations where management personnel are not stationed, are gradually expanding, there is an urgent need to develop an advanced system capable of fusing and processing various types of physical sensing information within the switchboard itself and autonomously and immediately identifying complex risk factors through an ultra-lightweight artificial neural network that does not rely on communication networks, thereby enabling active and immediate control of external circuit breaker facilities upon confirmation of a dangerous situation. Prior art literature

[0012] Korean Patent Publication No. 10-2079813 (February 14, 2020) Korean Patent Publication No. 10-2295784 (August 25, 2021) US Patent Publication US 10,191,114 (January 29, 2019) Chinese Published Patent Publication CN 120686039 (September 23, 2025) US Patent Publication US 8,010,239 (August 30, 2011) The problem to be solved

[0013] The present invention has been devised to improve upon the aforementioned problems, and aims to provide an intelligent switchboard device that measures changes in the physical environment and electromagnetic discharge signals inside the switchboard in real time, and performs active control response by immediately identifying complex risk factors through an artificial neural network in an edge environment.

[0014] The present invention aims to provide an intelligent switchboard device that improves analysis reliability by comprehensively acquiring not only general environmental conditions such as temperature, humidity, gas, and vibration, but also ultraviolet corona, ultra-high frequency, acoustic vibration, and leakage high-frequency current using a heterogeneous sensor array and a composite sensor system equipped in the intelligent switchboard device. means of solving the problem

[0015] The present invention relates to an intelligent switchboard device for monitoring the internal state of power equipment connected to a power grid network, comprising: an internal environment monitoring module that continuously generates environmental state data by measuring changes in the internal physical environment of the intelligent switchboard device in real time; a discharge signal detection module that acquires an analog discharge raw signal by capturing electromagnetic waves and acoustic radiation signals generated from internal power lines and insulators of the intelligent switchboard device; an edge inference computation unit that independently collects and processes the environmental state data and the analog discharge raw signal within the intelligent switchboard device without transmitting them to an external server to determine complex risk factors in real time; and an active control response unit that generates a response control signal instructing the operation of external equipment in response to at least one of the complex risk factors determined by the edge inference computation unit and an emergency environmental anomaly signal directly received from the internal environment monitoring module, and outputs the response control signal to the external equipment. The edge inference computation unit detects the zero-crossing point of commercial power supplied to the intelligent switchboard device to generate a phase synchronization signal, amplifies the analog discharge raw signal applied from the discharge signal detection module through a low-noise amplifier provided within the phase synchronization and signal acquisition means, and then passes through commercial log scale conversion Phase synchronization and signal acquisition means for converting into digital discrete datasets, andA pattern mapping and normalization means for generating a two-dimensional phase-decomposed partial discharge pattern heatmap image by dividing the digital discrete dataset into multiple unit pieces based on time intervals corresponding to one period of the phase synchronization signal, superimposing them on a single plane according to the zero reference of the phase synchronization signal, dividing the single plane into a preset first designated number of unit grids composed of phase sections on the horizontal axis and signal intensity sections on the vertical axis, counting the occurrence frequency values ​​of signals mapped within each unit grid, and scaling the counted occurrence frequency values ​​to a preset second designated number of pixel values; and sequentially passing through a convolution operation layer for extracting multiple morphological feature maps from the two-dimensional phase-decomposed partial discharge pattern heatmap image, a pooling layer for reducing the spatial dimension of the multiple morphological feature maps, a fully connected layer for unfolding the multiple morphological feature maps whose spatial dimension has been reduced by the pooling layer into a one-dimensional array and merging the environmental state data received from the internal environment monitoring module to derive complex risk features, and an output layer for calculating the derived complex risk features into multiple risk factor type probability values, thereby the intelligent switchboard device An intelligent switchboard device is provided that includes an artificial neural network-based complex risk identification means for finally determining the aforementioned internal complex risk factors.

[0016] The internal environment monitoring module includes a temperature detection sensor for measuring the internal temperature of the intelligent switchboard device, a humidity detection sensor for measuring internal humidity, a chemical gas detection sensor for measuring carbon monoxide concentration, a vibration detection sensor for measuring frequency, a door opening detection sensor for detecting door opening, and a flood detection sensor for detecting a flooded state; and the discharge signal detection module includes a corona light detection sensor for detecting corona in the ultraviolet region, an ultra-wideband antenna detection sensor for measuring electromagnetic radiation in the ultra-high frequency band, an acoustic vibration detection sensor for measuring acoustic and vibration acceleration in a preset frequency band, and a high-frequency current transformer sensor for detecting leakage high-frequency current through electromagnetic induction.

[0017] The above pattern mapping and normalization means generates a square unit grid matrix having a total of 1024 unit grids, which is the first designated number, by dividing the phase section of the horizontal axis of the single plane into 32 sections based on 360 degrees and dividing the signal strength section of the vertical axis into 32 sections based on the signal strength range after converting the analog discharge raw signal to a commercial log scale, and normalizes the accumulated occurrence frequency value for each of the 1024 unit grids to a numerical range of 0 to 255 to complete the two-dimensional phase decomposition partial discharge pattern heatmap image.

[0018] The active control response unit generates a first response control signal to operate external fire extinguisher equipment when the composite risk element determined by the edge inference operation unit is a fire risk element, generates a second response control signal to open a power circuit breaker when the composite risk element is an insulation damage risk element or a corona arc risk element, and generates a third response control signal to activate an internal alarm speaker when the composite risk element is additionally confirmed as an earthquake occurrence element or a flood risk element through measurement by the internal environment monitoring module. When the first, second, and third response control signals are generated in overlap, the first response control signal for the fire risk element is set as the highest priority, the second response control signal for the insulation damage or corona arc risk element is set as the next priority, and the third response control signal for the earthquake occurrence or flood risk element is assigned the next priority according to a preset ranking control logic, independently generates the signals, and outputs them to the external equipment respectively.

[0019] The present invention relates to a real-time diagnostic method performed by an intelligent switchboard device that monitors the internal state of power equipment connected to a power grid network, comprising: an internal environment monitoring step in which the intelligent switchboard device measures changes in the internal physical environment in real time and continuously generates environmental state data; a discharge signal detection step in which the intelligent switchboard device captures electromagnetic waves and acoustic radiation signals generated from internal power lines and insulators to acquire an analog discharge raw signal; an edge inference computation step in which the intelligent switchboard device independently collects and processes the environmental state data and the analog discharge raw signal internally without transmitting them to an external server to determine complex risk factors in real time; and an active control response step in which the intelligent switchboard device generates a response control signal that directs the operation of external equipment in response to at least one of the complex risk factors determined by the edge inference computation step and an emergency environmental anomaly signal directly received from the internal environment monitoring step, and outputs the response control signal to the external equipment. The edge inference computation step detects the zero-crossing point of the commercial power supply entering the intelligent switchboard device to generate a phase-locked signal, amplifies the analog discharge raw signal through an internally equipped low-noise amplifier, and then performs a commercial log-scale transformation to obtain a digital discrete dataset. A pattern that generates a two-dimensional phase-decomposed partial discharge pattern heatmap image by performing a process of: a phase synchronization and signal acquisition step for conversion; dividing the digital discrete dataset into a plurality of unit pieces based on a time interval corresponding to one period of the phase synchronization signal and superimposing them on a single plane according to the zero point reference of the phase synchronization signal; dividing the single plane into a preset first designated number of unit grids composed of a phase section on the horizontal axis and a signal intensity section on the vertical axis; counting the occurrence frequency values ​​of the signals mapped within each unit grid; and scaling the counted occurrence frequency values ​​to a preset second designated number of pixel values.A real-time diagnostic method is provided that includes a mapping and normalization step, a convolutional operation layer that extracts a plurality of morphological feature maps from a two-dimensional phase decomposition partial discharge pattern heatmap image, a pooling layer that reduces the spatial dimension of the plurality of morphological feature maps, a fully connected layer that expands the plurality of morphological feature maps whose spatial dimension has been reduced by the pooling layer into a one-dimensional array and merges the environmental state data to derive a complex risk feature, and an output layer that calculates the derived complex risk feature into a plurality of risk factor type probability values, thereby determining the complex risk factor inside the intelligent switchboard device.

[0020] The above internal environment monitoring step performs in parallel the process of acquiring carbon monoxide concentration data through a chemical gas detection sensor, the process of acquiring internal temperature data through a temperature detection sensor, the process of acquiring internal humidity data through a humidity detection sensor, the process of acquiring mechanical vibration frequency data through a vibration detection sensor, the process of acquiring door open status data through a door open detection sensor, and the process of acquiring flood status data through a flood detection sensor, and the above discharge signal detection step performs in parallel the process of acquiring a detection signal of corona light in the ultraviolet region, the process of acquiring an electromagnetic radiation signal in the ultra-high frequency band, the process of acquiring an acoustic vibration signal, and the process of acquiring an electromagnetic induction signal of leakage high-frequency current through a high-frequency current transformer sensor.

[0021] The pattern mapping and normalization steps include the process of generating a square unit grid matrix having a total of 1024 unit grids by dividing the phase section of the horizontal axis of the single plane into 32 sections based on 360 degrees and dividing the signal strength section of the vertical axis into 32 sections based on the signal strength range after commercial logarithmic scale conversion of the analog discharge raw signal, and the process of normalizing the accumulated occurrence frequency value for each of the 1024 unit grids to a numerical range of 0 to 255.

[0022] The above active control response step includes a process of independently generating and outputting each to the external equipment by assigning priorities according to a preset ranking control logic, wherein if the above-described composite risk element derived through the edge inference operation step is a fire risk element, a first response control signal is generated to operate external fire extinguisher equipment; if the above-described composite risk element is an insulation damage risk element or a corona arc risk element, a second response control signal is generated to open a power circuit breaker; and if the above-described composite risk element is additionally confirmed as an earthquake occurrence element or a flood risk element as a result of analyzing environmental status data obtained through the internal environment monitoring step, a third response control signal is generated to activate the internal alarm speaker of the intelligent switchboard device. In the event that the first, second, and third response control signals are generated in overlap, the first response control signal for the fire risk element is set as the highest priority, the second response control signal for the insulation damage or corona arc risk element is set as the next priority, and the third response control signal for the earthquake occurrence or flood risk element is assigned the next priority.

[0023] The method further includes a pre-model training step performed via an external server prior to performing the edge inference operation step, wherein the pre-model training step includes a labeling process that generates a training dataset for supervised learning by connecting a 2D phase decomposition partial discharge training pattern heatmap image collected in a simulated environment with a predefined risk class label one-to-one, and a weight optimization process in which the external server divides the training dataset for supervised learning according to a preset splitting ratio and then updates the weights of the convolution operation layer and the fully connected layer through a backpropagation algorithm to derive a trained neural network model.

[0024] The above-mentioned pre-model training step further includes a compatibility ensuring process of converting the trained neural network model into an Open Neural Network Exchange format file so that it can be run on different hardware platforms, and an edge device porting process of converting the Open Neural Network Exchange format file into C-language-based source code and loading it so that it can be run in the computational control environment inside the intelligent switchboard device.

[0025] The above internal environment monitoring step includes a process of instructing the active control response step to issue an immediate alarm by omitting the subsequent neural network operation of the edge inference operation step by directly applying a fire precursor state signal as an emergency environment abnormality signal, independently of the presence or absence of the analog discharge raw signal, when the carbon monoxide concentration value according to the acquired carbon monoxide concentration data exceeds a first threshold set for carbon monoxide and at the same time the internal temperature value according to the acquired internal temperature data exceeds a second threshold set for temperature.

[0026] The above phase synchronization and signal acquisition step includes a process of controlling an analog-to-digital converter to acquire the entire digital data continuously sampled during a preset acquisition time, and then dividing the entire digital data into periodic units corresponding to the frequency characteristics of the target commercial power supply to extract a plurality of preliminary datasets for separating unit pieces, and the above pattern mapping and normalization step performs a separation operation on the preliminary dataset according to a partitioning criterion to finally generate the plurality of unit pieces.

[0027] The output layer of the above artificial neural network-based complex risk identification step receives the derived complex risk features as input, calculates independent probability values ​​for each of a plurality of classes including fire risk, electric shock accident risk, insulation damage risk, dust foreign substance adhesion risk, and corona arc risk, and determines one or more classes exceeding a specified threshold as the complex risk factors, thereby including a process for simultaneously detecting complex defects.

[0028] The pooling layer of the above artificial neural network-based complex risk identification step includes a process of reducing spatial dimensions by applying a maximum pooling technique that extracts the maximum value within a preset operation window or an average pooling technique that extracts the average value to the plurality of morphological feature maps derived from the convolution operation layer.

[0029] The scaling process of the pattern mapping and normalization step comprises a process of subtracting the minimum frequency value among all grids from the occurrence frequency value of each counted unit grid, dividing the result by the difference between the maximum frequency value among all grids and the minimum frequency value, adding a preset exception constant to the denominator to prevent zero-point division errors caused by the maximum frequency value and the minimum frequency value being the same, and multiplying by the maximum pixel value of the image format to be normalized to replace it with pixel numeric data within the integer range of 0 to the maximum pixel value. Effects of the invention

[0030] According to one embodiment of the present invention, artificial intelligence inference that fuses environmental factors and discharge signals is independently performed at the distribution board terminal edge without being constrained by communication blind spots or network delays, thereby enabling the complete blocking of power accidents that occur in an instant.

[0031] In addition, according to one embodiment of the present invention, by integrally operating a multi-directional environmental sensor and a multi-frequency band discharge sensor, the limitations of the existing method relying on a single sensor can be overcome, and initial defects can be diagnosed with high reliability.

[0032] In addition, according to one embodiment of the present invention, by aligning a discharge raw signal with high variability based on phase and normalizing it into a heatmap image of uniform scale, the efficiency of deep learning feature extraction can be increased exponentially.

[0033] In addition, according to one embodiment of the present invention, customized hardware shutdown and firefighting measures are automatically performed for each identified complex risk factor, and the entire switchboard can be prevented by prioritizing control of accidents with the greatest ripple effect in the event of multiple risks.

[0034] According to one embodiment of the present invention, a series of diagnostic algorithm processes leading to real-time data acquisition, edge inference, and active control are unified, thereby establishing the self-protection capability of the distribution panel even in the event of a central server failure.

[0035] In addition, according to one embodiment of the present invention, stable defect information collection is possible even in specific noise environments by operating parallel logic that simultaneously acquires gas, temperature, ultra-high frequency, sound, etc., in a multidimensional manner.

[0036] In addition, according to one embodiment of the present invention, complex discharge data can be standardized into a square image of fixed resolution, allowing artificial neural network computation processing to be performed quickly even within limited resources.

[0037] In addition, according to one embodiment of the present invention, control signal conflicts and system panic in emergency situations can be prevented at the source by independently designing control signals for each type of risk and using an algorithm that enforces a ranking when overlapping.

[0038] In addition, according to one embodiment of the present invention, the basic discrimination accuracy of the edge device model can be maximized by performing partitioned cross-validation and backpropagation weight updates in advance using the abundant resources of a cloud server.

[0039] In addition, according to one embodiment of the present invention, deep learning inference is enabled even in a microcontroller environment by lightweightly loading an artificial intelligence model with C language-based source code without a heavy language environment.

[0040] In addition, according to one embodiment of the present invention, when temperature and carbon monoxide concentration increase simultaneously, the neural network computation process is omitted and bypassed to an immediate response logic, thereby saving even a mere second of inference time and preventing the spread of a fatal fire.

[0041] In addition, according to one embodiment of the present invention, since extensive continuous sampling data is systematically machined in accordance with the power cycle, a long-term continuous diagnostic system can be maintained without exhaustion of computational resources or memory overflow.

[0042] In addition, according to one embodiment of the present invention, through multi-label classification, various types of risks occurring simultaneously can be detected early without omission, rather than simply identifying only a single failure.

[0043] In addition, according to one embodiment of the present invention, by applying a pooling technique to preserve key discharge features while smoothing noise and significantly reducing dimensions, real-time feature computation is performed without load even on low-spec edge processors.

[0044] In addition, according to one embodiment of the present invention, even in situations where no pattern occurs, the uninterrupted stability required of industrial control equipment can be guaranteed by preventing a fatal bug in which the system goes down due to an infinite error caused by a denominator of 0 through an exception constant. Brief explanation of the drawing

[0045] FIG. 1 is a block diagram schematically showing the overall hardware configuration of an intelligent switchboard device according to one embodiment of the present invention. FIG. 2 is a flowchart illustrating a real-time diagnosis and active control method using an intelligent switchboard device according to an embodiment of the present invention. FIG. 3 is a diagram illustrating a two-dimensional phase-decomposed partial discharge pattern heatmap image generated by processing in an edge inference operation unit according to an embodiment of the present invention. FIG. 4 is a conceptual diagram of an artificial neural network-based deep learning architecture that is driven in an edge inference operation unit according to an embodiment of the present invention to identify complex risk factors. FIG. 5 is a flowchart illustrating the entire pre-training process of training, optimizing, and porting an artificial neural network model to an edge device environment according to one embodiment of the present invention. FIG. 6 is a flowchart illustrating the preprocessing of analog discharge raw data and pattern data processing performed in an edge inference operation unit according to an embodiment of the present invention. FIG. 7 is a diagram showing the configuration of a simulated environment of a corona arc generator used for the purpose of collecting data for the preliminary learning of the present invention. FIG. 8 is a diagram showing the configuration of a simulated environment of a partial discharge generation test cell used for the purpose of collecting data for prior learning of the present invention. Specific details for implementing the invention

[0046] The present invention relates to an intelligent switchboard device (100) for monitoring the internal state of power equipment connected to a power grid network and a real-time diagnostic method using the same.

[0047] As illustrated in FIGS. 1 and 2, the intelligent switchboard device (100) includes an internal environment monitoring module (110) that measures changes in the internal physical environment in real time and continuously generates environmental state data.

[0048] The above internal environment monitoring module (110) performs the role of providing basic data for the complex risk analysis to be described later by continuously converting local environmental changes inside the distribution panel into physical and electrical signals.

[0049] Additionally, the intelligent switchboard device (100) includes a discharge signal detection module (130) that captures electromagnetic waves and acoustic radiation signals generated from internal power lines and insulators to acquire an analog discharge raw signal.

[0050] The above discharge signal detection module (130) captures abnormal electromagnetic waves caused by insulation degradation or minute contact failures, and this is used as a key indicator for diagnosing initial defects.

[0051] In addition, the intelligent switchboard device (100) includes an edge inference computation unit (150) that independently collects and processes the environmental status data and the analog discharge raw signal within the intelligent switchboard device (100) without transmitting them to an external server, and determines complex risk factors in real time.

[0052] The edge inference operation unit (150) is implemented as an independent operation processor integrated inside the field device to perform immediate fault determination without being affected by delays or disconnections of the external communication network.

[0053] Additionally, the intelligent switchboard device (100) includes an active control response unit (170) that generates a response control signal instructing the operation of an external facility in response to at least one of the composite risk factor determined by the edge inference operation unit (150) and an emergency environment abnormality signal directly received from the internal environment monitoring module (110), and outputs the signal to the external facility.

[0054] The edge inference operation unit (150) above is equipped with phase synchronization and signal acquisition means (151) for performing precise synchronization and conversion of the signal.

[0055] The above-mentioned phase synchronization and signal acquisition means (151) detects the zero-cross point of the commercial power supply entering the intelligent distribution panel device (100) and generates a phase synchronization signal.

[0056] The above zero-cross point refers to the moment when the voltage of the AC power source switches from positive to negative or from negative to positive and crosses 0 V, and the signal must be acquired based on this to accurately analyze the phase dependence of the partial discharge phenomenon.

[0057] In addition, the phase synchronization and signal acquisition means (151) amplifies the analog discharge raw signal applied from the discharge signal detection module (130) through a low-noise amplifier (151a) provided internally.

[0058] The above low-noise amplifier (151a) is a device that selectively amplifies only minute discharge signals while suppressing the amplification of background noise mixed in the raw signal.

[0059] Subsequently, the phase synchronization and signal acquisition means (151) converts the amplified signal into a digital discrete dataset by performing a commercial log scale conversion.

[0060] The purpose of commercial logarithmic scale conversion is to prevent data loss by compressing the amplitude of discharge signals with an extremely wide dynamic range into a processable numerical range.

[0061] The edge inference operation unit (150) includes a pattern mapping and normalization means (152) that processes the converted data into a visual pattern.

[0062] The pattern mapping and normalization means (152) divides the digital discrete dataset into a plurality of unit pieces based on time intervals corresponding to one period of the phase synchronization signal.

[0063] Next, the pattern mapping and normalization means (152) superimposes the divided unit pieces on a single plane according to the zero point reference of the phase synchronization signal.

[0064] Additionally, the pattern mapping and normalization means (152) divides the single plane into a first designated number of unit grids, each consisting of a phase section on the horizontal axis and a signal intensity section on the vertical axis, and counts the frequency values ​​of the signals mapped within each unit grid.

[0065] The above-mentioned counted occurrence frequency values ​​are scaled to a preset second designated number of pixel values, thereby generating a two-dimensional phase-decomposed partial discharge pattern heatmap image.

[0066] The edge inference operation unit (150) further includes an artificial neural network-based complex risk identification means (153) that analyzes the generated image.

[0067] The above artificial neural network-based composite risk identification means (153) passes through a convolutional operation layer that extracts a plurality of morphological feature maps from the above two-dimensional phase-decomposed partial discharge pattern heatmap image.

[0068] The above convolutional operation layer extracts spatial features such as edges, textures, and cluster shapes within the image through filter kernels.

[0069] Afterwards, the artificial neural network-based composite risk identification means (153) passes through a pooling layer that reduces the spatial dimension of the plurality of morphological feature maps.

[0070] And, the artificial neural network-based composite risk identification means (153) expands the plurality of morphological feature maps, whose spatial dimension is reduced by the pooling layer, into a one-dimensional array.

[0071] In this process, the artificial neural network-based composite risk identification means (153) passes the environment state data received from the internal environment monitoring module (110) through a fully connected layer that derives composite risk features by merging the data.

[0072] Finally, the artificial neural network-based complex risk identification means (153) sequentially passes the derived complex risk features into a plurality of risk element type probability values ​​to finally determine the complex risk elements inside the intelligent switchboard device (100).

[0073] The above internal environment monitoring module (110) is composed of a heterogeneous sensor array for measuring various physical limit values.

[0074] The internal environment monitoring module (110) includes a temperature detection sensor that measures the internal temperature of the intelligent distribution panel device (100).

[0075] In addition, the internal environment monitoring module (110) includes a humidity detection sensor that measures humidity, which causes internal condensation or insulation breakdown.

[0076] And the internal environment monitoring module (110) includes a chemical gas detection sensor that measures the concentration of carbon monoxide generated when the cable sheath or insulating resin deteriorates.

[0077] In addition, the internal environment monitoring module (110) includes a vibration detection sensor that measures the frequency to detect mechanical loosening or resonance.

[0078] In addition, the internal environment monitoring module (110) includes a door opening detection sensor that detects unauthorized access or the destruction of the sealed state.

[0079] Finally, the internal environment monitoring module (110) includes a flood detection sensor that detects a flooded state to prevent a short circuit accident caused by external moisture inflow.

[0080] The above discharge signal detection module (130) also has a composite sensor system for multi-faceted analysis of the discharge phenomenon.

[0081] The above discharge signal detection module (130) includes a corona light detection sensor that detects the corona in the ultraviolet region that occurs during the initial stage of discharge.

[0082] Additionally, the discharge signal detection module (130) includes an ultra-wideband antenna detection sensor that measures electromagnetic radiation in the ultra-high frequency band radiated when an internal arc occurs.

[0083] And the discharge signal detection module (130) includes an acoustic vibration detection sensor that measures acoustic and vibration acceleration in a preset frequency band that occurs when an internal bubble of the insulator is destroyed.

[0084] Additionally, the discharge signal detection module (130) includes a high-frequency current transformer sensor that detects leakage high-frequency current flowing through a ground wire, etc., as an electromagnetic induction phenomenon.

[0085] The above pattern mapping and normalization means (152) precisely processes data to match the input specifications of the neural network model.

[0086] The above pattern mapping and normalization means (152) divides the phase section of the horizontal axis of the single plane into 32 sections based on 360 degrees.

[0087] At the same time, the pattern mapping and normalization means (152) divides the signal strength section of the vertical axis into 32 sections based on the signal strength range (dBm) after the commercial logarithmic scale conversion of the analog discharge raw signal.

[0088] Through this partitioning operation, the pattern mapping and normalization means (152) generates a square unit grid matrix having a total of 1024 unit grids, which is the first designated number.

[0089] Afterwards, the pattern mapping and normalization means (152) normalizes the accumulated occurrence frequency value for each of the 1024 unit grids into a numerical range of 0 to 255.

[0090] Through the above normalization processing, the deviation of the data is reduced and visual contrast is maximized to increase the feature extraction efficiency of the deep learning model, thereby completing the above 2D phase decomposition partial discharge pattern heatmap image.

[0091] The active control response unit (170) performs immediate and physical facility protection measures according to the type and level of risk determined by the edge inference calculation unit (150).

[0092] The above active control response unit (170) generates a first response control signal that operates an external fire extinguisher equipment when the composite risk factor determined by the edge inference operation unit (150) is a fire risk factor.

[0093] In addition, the active control response unit (170) generates a second response control signal to open the power breaker when the composite risk factor is an insulation damage risk factor or a corona arc risk factor.

[0094] Opening the above power circuit breaker is an important measure to immediately cut off the power applied to the faulty area, thereby preventing secondary explosions or cascading accidents.

[0095] And the active control response unit (170) generates a third response control signal that activates an internal alarm speaker when the composite risk factor is additionally identified as an earthquake occurrence factor or a flood risk factor through measurement of the internal environment monitoring module (110).

[0096] In particular, the active control response unit (170) activates a preset ranking control logic when multiple accidents occur simultaneously and the first, second, and third response control signals are superimposed.

[0097] The above-mentioned preset ranking control logic assigns the first response control signal to the fire risk factor with the greatest impact as the highest priority.

[0098] Next, the above-mentioned preset ranking control logic assigns the second corresponding control signal to the insulation damage or corona arc risk factor that may cause mechanical destruction of the power grid as a next priority.

[0099] Finally, the above-mentioned preset ranking control logic assigns the above-mentioned third response control signal to the next priority for earthquake occurrence or flood risk factors with a strong purpose of on-site evacuation and calling maintenance personnel.

[0100] The active control response unit (170) above independently generates control signals according to assigned priority and outputs them to the external equipment, thereby preventing system overload and ensuring an optimal protection sequence.

[0101] The above intelligent switchboard device (100) performs a real-time diagnostic method that operates as an organic process in response to the hardware configuration.

[0102] First, the intelligent switchboard device (100) performs an internal environment monitoring step (S110) that measures changes in the internal physical environment in real time and continuously generates environmental state data.

[0103] Simultaneously or sequentially, the intelligent switchboard device (100) performs a discharge signal detection step (S130) to acquire an analog discharge raw signal by capturing electromagnetic waves and acoustic radiation signals generated from internal power lines and insulators.

[0104] Subsequently, the intelligent switchboard device (100) performs an edge inference operation step (S150) in which it independently collects and processes the environmental status data and the analog discharge raw signal internally without transmitting them to an external server, thereby determining complex risk factors in real time.

[0105] Finally, the intelligent switchboard device (100) performs an active control response step (S170) in which it generates a response control signal instructing the operation of an external facility and outputs it to the external facility in response to at least one of the composite risk factor determined by the edge inference operation step (S150) and an emergency environment abnormal signal directly received from the internal environment monitoring step (S110).

[0106] Specifically, the edge inference operation step (S150) is configured to include sub-detailed processes.

[0107] The edge inference operation step (S150) includes a phase synchronization and signal acquisition step (S151) that detects the zero-crossing point of the commercial power supply entering the intelligent distribution panel device (100) and generates a phase synchronization signal.

[0108] In the phase synchronization and signal acquisition step (S151), the analog discharge raw signal is amplified through an internally provided low-noise amplifier and then converted into a digital discrete dataset through a commercial logarithmic scale conversion.

[0109] Additionally, the edge inference operation step (S150) includes a pattern mapping and normalization step (S152) for visualizing the transformed data.

[0110] The pattern mapping and normalization step (S152) divides the digital discrete dataset into a plurality of unit pieces based on time intervals corresponding to one period of the phase synchronization signal and superimposes them on a single plane according to the zero point reference of the phase synchronization signal.

[0111] The pattern mapping and normalization step (S152) then divides the single plane into a first designated number of pre-set unit grids, each consisting of a phase section on the horizontal axis and a signal intensity section on the vertical axis, and counts the frequency values ​​of the signals mapped within each unit grid.

[0112] The above-mentioned counted occurrence frequency values ​​are scaled to a preset second designated number of pixel values ​​and are ultimately used to generate a two-dimensional phase-decomposed partial discharge pattern heatmap image.

[0113] The above edge inference operation step (S150) goes through an artificial neural network-based complex risk identification step (S153) that performs intelligent inference based on the generated image.

[0114] The above artificial neural network-based complex risk identification step (S153) first applies a convolutional operation layer that extracts a plurality of morphological feature maps from the above two-dimensional phase-decomposed partial discharge pattern heatmap image.

[0115] Subsequently, the artificial neural network-based complex risk identification step (S153) applies a pooling layer that reduces the spatial dimension of the plurality of morphological feature maps.

[0116] Next, the artificial neural network-based complex risk identification step (S153) applies a fully connected layer that expands the plurality of morphological feature maps, whose spatial dimensions are reduced by the pooling layer, into a one-dimensional array and merges the environment state data to derive complex risk features.

[0117] Finally, the artificial neural network-based complex risk identification step (S153) sequentially applies an output layer that calculates the derived complex risk features into multiple risk factor type probability values ​​to determine the complex risk factors within the intelligent switchboard device (100).

[0118] The above internal environment monitoring step (S110) simultaneously performs the process of acquiring various types of sensor data.

[0119] The above internal environment monitoring step (S110) performs in parallel the process of acquiring carbon monoxide concentration data through a chemical gas detection sensor, the process of acquiring internal temperature data through a temperature detection sensor, the process of acquiring internal humidity data through a humidity detection sensor, the process of acquiring mechanical vibration frequency data through a vibration detection sensor, the process of acquiring door open state data through a door open detection sensor, and the process of acquiring flood state data through a flood detection sensor.

[0120] Likewise, the discharge signal detection step (S130) ensures the collection of three-dimensional defect information by performing in parallel the process of acquiring a detection signal of corona light in the ultraviolet region, the process of acquiring an electromagnetic radiation signal in the ultra-high frequency band, the process of acquiring an acoustic vibration signal, and the process of acquiring an electromagnetic induction signal of leakage high-frequency current through a high-frequency current transformer sensor.

[0121] The pattern mapping and normalization step (S152) performs a detailed algorithm for pixelation and numerical normalization of the image.

[0122] The pattern mapping and normalization step (S152) includes the process of dividing the phase section of the horizontal axis of the single plane into 32 sections based on 360 degrees.

[0123] At the same time, the signal strength section of the vertical axis is divided into 32 sections based on the signal strength range after the commercial logarithmic scale conversion of the analog discharge raw signal.

[0124] This completes the process of generating a square unit cell matrix with a total of 1024 unit cells.

[0125] Subsequently, the pattern mapping and normalization step (S152) performs a process of normalizing the accumulated occurrence frequency values ​​for each of the 1024 unit grids into a numerical range of 0 to 255, thereby converting the data into an 8-bit black-and-white or color map image format optimized for a deep learning model.

[0126] The above active control response step (S170) systematizes the response procedure of the actual device to the judgment result.

[0127] The active control response step (S170) generates a first response control signal that operates an external fire extinguisher equipment when the composite risk factor derived through the edge inference operation step (S150) is a fire risk factor.

[0128] Additionally, the active control response step (S170) generates a second response control signal to open the power breaker if the composite risk factor is an insulation damage risk factor or a corona arc risk factor.

[0129] And the active control response step (S170) generates a third response control signal that activates the internal alarm speaker of the intelligent switchboard device (100) when the complex risk factor is additionally identified as an earthquake occurrence factor or a flood risk factor as a result of analyzing the environmental state data obtained through the internal environment monitoring step (S110).

[0130] Here, when the first, second, and third corresponding control signals are generated in overlap, the active control response step (S170) calls a preset ranking control logic that sets the first corresponding control signal for the fire risk element as the highest priority.

[0131] The above-mentioned preset ranking control logic assigns the second response control signal for insulation damage or corona arc risk factors as the next priority, and the third response control signal for earthquake occurrence or flood risk factors as the next priority.

[0132] The above active control response step (S170) performs the process of independently generating control signals according to this priority and outputting them to the external equipment respectively.

[0133] The present invention further includes a pre-model training step performed via an external server prior to performing the edge inference operation step (S150) for high-accuracy inference.

[0134] Since the computational resources of the intelligent distribution panel device (100) itself are limited, weight updates using vast amounts of data are performed on an external server with abundant computing power.

[0135] The above-mentioned pre-model training step involves a labeling process that generates a training dataset for supervised learning by connecting pattern heatmap images for 2D phase decomposition partial discharge training collected in a simulated environment with predefined risk class labels one-to-one.

[0136] The above-mentioned labeled data is prepared with noise removed after undergoing expert verification.

[0137] Subsequently, the aforementioned pre-model training step performs a weight optimization process in which the external server divides the training dataset for supervised learning according to a preset splitting ratio.

[0138] In the above weight optimization process, the external server cross-validates the training data and validation data, and continuously updates the weights of the convolutional layer and the fully connected layer through the backpropagation algorithm to derive a trained neural network model that achieves the target performance.

[0139] Models trained on an external server must undergo optimization and conversion procedures to be deployed on field equipment.

[0140] To this end, the aforementioned pre-model training step undergoes a compatibility assurance process that converts the trained neural network model into an ONNX format file, an open neural network exchange format, so that it can be run on different hardware platforms.

[0141] The above ONNX format file has a universal graph structure that is not dependent on a framework, making it easy to port to various edge devices.

[0142] Next, the above-mentioned pre-model learning step performs an edge device porting process in which the ONNX format file is converted into C language-based source code and loaded so that it can be operated in a low-power computation control environment inside the intelligent distribution panel device (100).

[0143] Through this porting process, neural network inference becomes possible with lightweight pure C language binaries without a heavy deep learning engine.

[0144] The above internal environment monitoring step (S110) incorporates hardware bypass logic to prepare for ultra-urgent situations where even a delay in inference operations is not allowed.

[0145] The above internal environment monitoring step (S110) continuously monitors whether the carbon monoxide concentration value according to the acquired carbon monoxide concentration data exceeds a preset first threshold for carbon monoxide.

[0146] At the same time, the internal environment monitoring step (S110) determines, using an AND condition, whether the internal temperature value according to the acquired internal temperature data exceeds a second threshold preset for the temperature.

[0147] If both conditions are met, there is a sign that a fire is already in progress or imminent, so a fire precursor state signal is directly applied to the active control response step (S170) as an emergency environment abnormality signal, independently of the presence or absence of the analog discharge raw signal or the inference result of the neural network.

[0148] Through this direct authorization process, the subsequent neural network operation of the edge inference operation step (S150) is omitted, and an immediate alarm is issued, thereby minimizing equipment loss.

[0149] The above phase synchronization and signal acquisition step (S151) presents a sampling procedure optimized to capture high-speed electromagnetic signals without omission.

[0150] The above phase synchronization and signal acquisition step (S151) includes a process of controlling an analog-to-digital converter to acquire the entire digital data that has been continuously sampled for a preset acquisition time.

[0151] The above analog-to-digital converter operates with a sampling rate of tens of megahertz or more to resolve high-frequency discharge pulses.

[0152] Subsequently, the phase synchronization and signal acquisition step (S151) extracts a plurality of preliminary datasets for separating unit pieces by dividing the entire acquired digital data into periodic units corresponding to the frequency characteristics of the target commercial power supply.

[0153] And the pattern mapping and normalization step (S152) performs a separation operation on the preliminary dataset according to the partitioning criteria to finally generate the plurality of unit pieces.

[0154] This fundamentally prevents data buffer overflow that may occur during long-term continuous measurements.

[0155] The output layer of the above artificial neural network-based complex risk identification step (S153) diagnoses complex defects by adopting a multi-label classification technique rather than simple binary classification.

[0156] The above output layer receives the derived composite risk features as input and calculates independent probability values ​​for each of a plurality of classes, including fire risk, electric shock accident risk, insulation damage risk, dust foreign substance adhesion risk, and corona arc risk.

[0157] The above probability values ​​are derived as real values ​​between 0 and 1 through a softmax or sigmoid activation function.

[0158] Subsequently, the output layer completes the process of simultaneously detecting two or more composite defects rather than a single cause by finally determining one or more classes exceeding a specified threshold as the composite risk factors.

[0159] Therefore, it is possible to accurately predict and diagnose the occurrence of multiple and serial accidents.

[0160] The pooling layer of the above artificial neural network-based complex risk identification step (S153) performs compression of feature data to maximize computational efficiency in a limited memory environment.

[0161] The pooling layer above may apply a maximum pooling technique to the plurality of morphological feature maps derived from the convolution operation layer, which extracts the maximum value within a preset operation window.

[0162] The above maximum pooling technique is advantageous for preserving the most prominent features, such as the peak point of the discharge pattern.

[0163] Alternatively, the pooling layer may include a process of reducing spatial dimensions by applying an average pooling technique that extracts the average value within the operation window.

[0164] The above average pooling technique is effective for smoothing background noise and identifying the overall dispersion pattern of the data.

[0165] Depending on the designer's purpose, one or a combination of these two techniques is applied to exponentially reduce the number of parameters passed to the next layer.

[0166] The scaling process of the pattern mapping and normalization step (S152) applies a normalization formula designed to ensure the mathematical stability of the data.

[0167] The above scaling process first performs an operation to subtract the minimum frequency value among all grids from the occurrence frequency value of each counted unit grid.

[0168] Subsequently, the result is divided by the difference between the maximum frequency value among the entire grid and the aforementioned minimum frequency value.

[0169] At this time, due to the characteristics of the data, when the above maximum frequency value and the above minimum frequency value are the same, that is, in a zero state where no pattern occurs, the denominator becomes 0, and a zero division error may occur in which an infinite error occurs.

[0170] To prevent this, the above scaling process includes a safety technique that performs the operation by adding a preset exception constant to the denominator.

[0171] Finally, the process of multiplying the calculated decimal-point result value by the maximum pixel value of the image format to be normalized and replacing it with pixel numeric data within the integer range of 0 to the maximum pixel value is completed.

[0172] This enables stable and error-free image processing to be continuously performed in the switchboard edge environment.

[0173] As illustrated in FIG. 3, the edge inference operation unit (150) generates a two-dimensional phase-decomposed partial discharge pattern heatmap image to process the collected partial discharge data into an optimal visual pattern that can be recognized by an artificial neural network.

[0174] The pattern mapping and normalization means provided within the edge inference operation unit (150) superimposes the acquired digital discrete dataset onto a single plane according to the zero point reference of the phase synchronization signal.

[0175] Specifically, the horizontal axis of the above single plane represents a 360-degree commercial power phase section, and this is divided into 32 sections at uniform intervals.

[0176] At the same time, the vertical axis of the single plane represents the signal strength range after the commercial logarithmic scale conversion of the analog discharge raw signal, and this is also divided into 32 sections.

[0177] Through this orthogonal partitioning process, a square unit cell matrix having a total of 1024 unit cells is derived by intersecting 32 horizontal axes and 32 vertical axes.

[0178] The above pattern mapping and normalization means performs a process of precisely counting the occurrence frequency values ​​of the mapped signal for each of the 1024 unit grids.

[0179] Afterwards, the above-mentioned frequency values ​​per unit grid are normalized into a numerical range of 0 to 255 to conform to the 8-bit image format, which is the input specification of the deep learning model.

[0180] In the above normalization process, scaling is performed using the difference between the maximum and minimum frequency values ​​among the entire grid as the denominator; however, to fundamentally prevent zero-division errors caused by the maximum and minimum frequency values ​​being identical, a pre-set exception constant is added to the denominator to ensure computational stability.

[0181] Through this series of data processing, a heatmap diagram with 32 times 32 resolution is completed, which intuitively expresses the frequency of discharge occurrence according to signal strength and power phase as shades or brightness of color, as shown in Figure 3 above.

[0182] The edge inference operation unit (150) transmits the generated heatmap data with a resolution of 1024 pixels to an artificial neural network-based composite risk identification means to perform learning and inference operations.

[0183] The above artificial neural network-based complex risk identification means extracts the signal intensity distribution, discharge occurrence frequency, and concentration in a specific phase band within the imaged heatmap diagram as a comprehensive morphological feature map.

[0184] The extracted morphological feature map passes through a convolutional layer and a pooling layer to reduce its spatial dimension, and is merged with other environmental state data obtained from the internal environment monitoring module in a fully connected layer.

[0185] As a result, the edge inference operation unit (150) can independently and accurately identify and classify specific risk factors, such as insulation degradation, corona discharge, or physical defects inside the switchboard, through artificial neural network inference based on the heatmap image.

[0186] As illustrated in FIG. 4, the edge inference operation unit (150) processes multidimensional data collected from the intelligent switchboard device (100) through an artificial neural network to predict the type of complex risk factor that may occur and drives a neural network processing structure that independently determines the type of risk factor that has currently occurred.

[0187] The edge inference operation unit (150) described above is configured to include a deep learning architecture that sequentially executes a convolutional layer, a pooling layer, a fully connected layer, and an output layer to comprehensively analyze the 2D phase-decomposed partial discharge pattern heatmap image with 32 x 32 resolution described above and environmental state data obtained from various sensors.

[0188] The convolutional layer of the edge inference operation unit (150) extracts spatial features of the pattern by performing a convolutional operation while sliding a filter kernel of a preset size on the pattern heatmap image input data consisting of 1024 pixels.

[0189] In this process, the edge inference operation unit (150) converts local two-dimensional visual patterns, such as the phase concentration and signal strength distribution of the discharge signal, into multiple feature map forms and uses them as key indicators for risk factor classification to be performed later.

[0190] Next, the pooling layer of the edge inference operation unit (150) spatially reduces the size of the extracted feature map, thereby significantly reducing the absolute amount of computation of parameters to be processed.

[0191] At the same time, the pooling layer ensures invariance that reduces sensitivity to minute noise changes or small pixel shifts, thereby supporting robust feature maintenance even in environments with severe external industrial noise.

[0192] High-dimensional discharge feature data summarized while passing through the pooling layer is flattened and then merged with physical environment state data of an input layer, including a temperature sensor, humidity sensor, chemical gas sensor, vibration sensor, door opening detection sensor, or flood detection sensor inside the intelligent switchboard device (100), and transmitted to the fully connected layer.

[0193] The fully connected layer of the edge inference operation unit (150) performs a key role in connecting the high-dimensional partial discharge features summarized by the preceding convolutional layer and pooling layer, and the environment state data input as a single numerical type, into meaningful judgment logic.

[0194] The above fully connected layer passes through multiple hidden layer nodes and performs cross-analysis of discharge data features and environmental data through a non-linear activation function, and infers in depth complex correlations between variables rather than simply whether a threshold is exceeded.

[0195] Finally, the output layer of the edge inference operation unit (150) applies a softmax function to normalize the operation result transmitted from the fully connected layer into an independent occurrence probability value for each predefined class and outputs it.

[0196] The edge inference operation unit (150) specifically determines whether the currently input pattern is a corona discharge type or an internal insulation breakdown discharge based on the output value of the softmax function, and comprehensively analyzes the extracted feature combinations to finally recognize a multidimensional complex risk pattern.

[0197] The types of risk factors identified by the intelligent switchboard device (100) according to the present invention are classified as fire risk, electric shock accident risk, earthquake occurrence, insulation damage risk, dust foreign matter risk, corona arc risk, or flooding risk.

[0198] In order for the above edge inference operation unit (150) to accurately predict and determine these specific risk factors, a prior learning process using high-quality data that precisely simulates each accident situation must be performed beforehand.

[0199] Specifically, the edge inference operation unit (150) separately produces a standardized fire model capable of reproducing an actual fire situation and performs pre-learning to predict the fire risk or determine the occurrence of a fire by collecting training data that measures environmental changes resulting from burning the model.

[0200] In addition, the edge inference operation unit (150) collects learning data of abnormal patterns obtained by artificially manipulating the door opening detection sensor and the vibration sensor to simulate abnormal situations caused by worker inattention or unauthorized access, and performs prior learning to predict the risk of electric shock accidents or determine the occurrence of electric shock accidents through this.

[0201] In addition, the edge inference operation unit (150) firmly fixes the vibration sensor to a precision-controlled exciter and mechanically generates seismic waves of various frequency bands to simulate an actual earthquake situation, then collects learning data based on the acquired vibration acceleration criteria, and performs pre-learning to predict an earthquake situation or determine whether an earthquake has occurred.

[0202] In addition, the edge inference operation unit (150) generates artificial insulation breakdown signs using a corona arc generator to which high voltage is applied, and performs prior learning to predict the risk of insulation damage or determine the occurrence of insulation damage by collecting learning data including the discharge signal characteristics radiated at that time.

[0203] In addition, the edge inference operation unit (150) configures a foreign substance partial discharge generation test cell to implement a situation where dust accumulates inside the distribution board and causes tracking, and collects learning data including the characteristics of the discharge pulses generated therefrom to predict the risk of the dust foreign substance or to determine the occurrence of a defect caused by the dust foreign substance.

[0204] In addition, the edge inference operation unit (150) collects discharge learning data concentrated on a specific phase using a corona partial discharge generation test cell to simulate the air insulation breakdown phenomenon occurring at a sharp conductor surface or a spaced terminal, and performs prior learning to predict the corona arc risk or determine the occurrence of a corona arc through this.

[0205] Finally, the edge inference operation unit (150) simulates an actual flooding situation by bringing moisture into contact with the flood detection sensor located at the bottom inside the intelligent distribution panel device (100), thereby collecting learning data based on the water level rise pattern, and performs pre-learning to predict the flood risk or determine the flood occurrence situation based on this, thereby maximizing the accuracy of determination when applied in the field.

[0206] Accordingly, the intelligent switchboard device (100) can perfectly classify subtle complex accident signs that could not be detected by conventional simple threshold comparison methods in real time at the edge and direct preemptive protective measures by organically combining a vast amount of pre-training data specialized for each type of risk and a multi-layered neural network structure.

[0207] As illustrated in FIG. 5, the entire pre-training process is performed to train and optimize an artificial neural network model in advance and implant it into an edge device so that the edge inference operation unit (150) can independently and accurately determine risk factors inside the distribution board.

[0208] The aforementioned pre-training process first performs a data labeling process to apply the vast amount of raw data acquired from the field to a supervised learning-based algorithm.

[0209] In the above data labeling process, a class label file serving as a ground truth is precisely created by perfectly mapping the previously processed 2D phase-resolved partial discharge pattern heatmap image dataset with a resolution of 32 x 32 to the actual risk factor type at the time the discharge pattern occurred.

[0210] The class label file created above and the corresponding data of tens of thousands of original heatmap images are uploaded to an external, cloud-based high-performance computing server environment, such as Google Colab, to smoothly perform weight update operations of the artificial neural network, which require massive computing power and memory resources of the graphics processing unit.

[0211] The aforementioned external high-performance computing server undergoes a data splitting process that subdivides the entire collected dataset according to pre-set splitting ratios for the systematic training of the artificial neural network model and objective performance verification. (Divided into training (70%), validation (20%), and test (10%).)

[0212] Specifically, the aforementioned external high-performance computing server allocates 70 percent of the total dataset as training data used to directly update the weights and bias parameters of the neural network, and utilizes it as basic data for the backpropagation algorithm.

[0213] Next, the aforementioned external high-performance computing server allocates 20 percent of the total dataset as validation data and uses it to monitor in real time overfitting, where the model is excessively fitted to the training data during the training cycle, and to fine-tune hyperparameters.

[0214] The remaining 10 percent of the data is stored in complete isolation as test data and is used to objectively evaluate the general generalization performance of the neural network model after it has been trained, specifically how accurately it performs inference on new discharge patterns it has never encountered.

[0215] Once the above data partitioning is completed, the external high-performance computing server designs a lightweight convolutional neural network architecture with a minimized number of parameters, taking into account the hardware operating environment of the edge inference computing unit (150) inside the distribution board where memory and power supply are extremely limited.

[0216] The convolutional neural network model designed above starts with an input layer that accommodates visual information of 1024 pixels, passes through a convolutional layer that extracts local spatial features of the image in parallel, and a pooling layer that significantly reduces computational load and provides robustness against noise.

[0217] Subsequently, the aforementioned convolutional neural network model is tightly composed of a fully connected layer that expands high-dimensional morphological features summarized through pooling into a logical decision domain, and a softmax output layer that finally calculates the occurrence probability values ​​of individual risk factors as continuous real values ​​between 0 and 1. (Input→Convolution→Pooling→Dense→Softmax)

[0218] The above external high-performance computing server runs a general-purpose deep learning framework such as PyTorch or TensorFlow to sequentially inject the divided training data into the designed convolutional neural network model.

[0219] The above general-purpose deep learning framework intensively executes a learning training process that converges all numerical weights within the convolutional neural network model to an optimal state by calculating the error between the input image features and the actual correct label as a loss function and repeatedly performing gradient descent and backpropagation algorithms in the direction of minimizing it.

[0220] The optimal weight matrix and the internal computation graph structure of the neural network model derived through the above learning training process initially take the form of a file that is strongly dependent on a specific deep learning framework.

[0221] Therefore, a format conversion process is essential to provide universality, enabling it to run on microcontrollers from various manufacturers or in independent hardware environments without compatibility issues.

[0222] The above-mentioned external high-performance computing server converts the dependent model, whose training has been completely finished, into an ONNX format file, which is a standardized open neural network exchange format that perfectly guarantees structural compatibility between different hardware platforms, and outputs and downloads it.

[0223] Next, the edge inference operation unit (150) performs a high-level optimization and C language source code conversion process so that the ONNX format file can be executed independently without the help of a heavy Python interpreter or deep learning engine in a low-power microcontroller environment.

[0224] The aforementioned source code conversion process refers to the core task of converting the complex graph operation structures and vast floating-point weight matrices of the ONNX format into a serialized source code array based on a pure C language, utilizing embedded-specific artificial neural network optimization software tools such as STM32Cube.AI.

[0225] In this conversion process, the artificial neural network optimization software tool actively applies quantization techniques to minimize dynamic memory allocation by reconfiguring static memory buffers and converting unnecessary 32-bit floating-point operations into 8-bit or 16-bit integer operations.

[0226] Through such high-intensity quantization and optimization operations, the overall parameter capacity of the artificial neural network model and the RAM memory occupancy required for inference are exponentially reduced to precisely fit the limited flash memory specifications of edge devices.

[0227] Finally, the inference source code and quantized weight binary files, which have been fully converted and optimized based on the aforementioned language, are integrated and compiled into executable firmware.

[0228] The integrated compiled firmware undergoes a deployment step in which it is directly implanted and permanently loaded into the central computing unit and non-volatile flash memory area of ​​the edge inference computing unit (150) installed inside the intelligent distribution board device via a communication cable or debugging equipment.

[0229] By successfully completing the above distribution step, the edge inference computation unit (150) does not rely at all on the connection of an external cloud communication network or computational support from an upper central control server.

[0230] As a result, the edge inference computation unit (150) ultimately obtains a powerful and autonomous edge intelligence judgment capability that can accurately infer critical complex risk factors, such as fire or insulation damage, within an extremely short delay time of milliseconds by immediately performing self-computation on a pattern heatmap image of 32 x 32 resolution that is continuously input in real time at a physical distribution board site.

[0231] As illustrated in FIG. 6, the edge inference operation unit (150) performs a series of systematic preprocessing processes to extract and process the collected complex analog discharge raw data into normalized two-dimensional phase-decomposed partial discharge pattern data.

[0232] The edge inference operation unit (150) first performs a phase synchronization signal extraction process to precisely detect zero-crossing points where the voltage level crosses zero by monitoring the voltage waveform of the AC commercial power supply entering the intelligent distribution panel device.

[0233] The edge inference operation unit (150) internally generates a phase synchronization signal of a preset frequency band synchronized with the operating frequency of the target power system, using the detected zero-cross point as a starting point.

[0234] The generated phase synchronization signal is applied to a group of multiple sensors including a corona light sensing sensor, an ultra-wideband antenna sensing sensor, an acoustic vibration sensing sensor, or a high-frequency current transformer sensor, and is used as a synchronization measure to provide an absolute reference phase of 0 to 360 degrees to the individual signals acquired by each sensor.

[0235] Next, the edge inference operation unit (150) introduces a fine discharge pulse signal physically captured through the multi-type sensor group into an internal circuit and passes it through a low-noise amplifier to perform a high-frequency signal acquisition process that selectively amplifies only the desired high-frequency signal while suppressing the increase of thermal noise or background noise as much as possible.

[0236] After the signal-to-noise ratio is improved through the above high-frequency signal acquisition process, the edge inference operation unit (150) executes a signal conversion process that compresses the analog signal power to a commercial logarithmic scale by considering the physical characteristics of the discharge pulse having an extremely wide amplitude fluctuation range.

[0237] The edge inference operation unit (150) above can safely extract the envelope of a signal without data overflow or loss, ranging from minute discharge signs to strong arc waveforms through this commercial log scale conversion.

[0238] Next, the edge inference operation unit (150) performs a digitization process that converts the analog waveform into a digital dataset capable of machine calculation by controlling a built-in high-speed analog-to-digital converter and continuously performing sampling for a predetermined acquisition time.

[0239] The edge inference operation unit (150) undergoes a periodic division process to precisely divide the acquired long-length digital dataset into multiple pre-set division units in accordance with one cycle of the previously generated phase synchronization signal, in order to prevent memory saturation caused by the accumulation of continuous data.

[0240] When the above periodic division operation is completed, the edge inference operation unit (150) performs a superposition process in which the separated multiple unit pieces are collectively accumulated and diagrammed on a single two-dimensional plane according to the zero point reference of the phase synchronization signal.

[0241] For example, superposition involves superimposing 60 cycles of data collected over 1 second in a 60Hz environment, which is the domestic commercial power frequency, to match the phase synchronization signal. Finally, for zone-specific frequency extraction, the horizontal axis is divided into 32 sections with a 360-degree phase interval of 11.25 degrees (0.52ms) and the vertical axis is divided into 32 sections with a 2dB interval for signal intensity. The number of signals corresponding to each section is counted from 60 data points, and this count value is used as the frequency value of that section.

[0242] The edge inference operation unit (150) sets the horizontal axis of the single plane to a phase range of 360 degrees and divides it equally into predetermined angle intervals to set a plurality of phase sections in order to replace the superimposed data with a matrix that can be interpreted by an artificial neural network.

[0243] At the same time, the edge inference operation unit (150) designates the vertical axis of the single plane as a signal strength range and divides it equally into predetermined decibel intervals to set a plurality of signal strength sections.

[0244] The edge inference operation unit (150) performs a zone-specific frequency extraction process for each of the plurality of unit grids derived through the orthogonal division of the horizontal axis and the vertical axis, by summing the number of hits of discharge signals mapped within the corresponding grid area from the overlapping plurality of data pieces.

[0245] The edge inference operation unit (150) determines the final accumulated number of signals counted in each unit grid as the absolute frequency value of the corresponding section and completes normalization processing to scale it within a preset pixel value range to match the image input specifications of the deep learning model.

[0246] As a result, the edge inference operation unit (150) perfectly converts complex time-series discharge raw data into a two-dimensional pattern image form in which the phase, signal strength, and frequency of occurrence are replaced with intuitive numerical values, thereby maximizing the computational efficiency and discrimination reliability of the subsequent complex risk judgment algorithm.

[0247] As illustrated in FIG. 7, a corona arc generator that artificially simulates a fatal corona arc situation inside a switchboard is utilized to support precise pre-learning of the edge inference operation unit (150).

[0248] The above corona arc generator is configured to include a pointed bed planar electrode or a sphere-to-sphere electrode structure and a mechanical calibration means for adjusting the minute separation distance between electrodes with precision in micrometers.

[0249] The above corona arc generator physically and perfectly reproduces the corona discharge phenomenon at the stage just before air insulation breakdown that occurs when insulation distance is not secured at the busbar or terminal block of an actual distribution panel by applying high-voltage AC power while maintaining the above-mentioned separation distance.

[0250] The discharge signal detection module (130) equipped in the above intelligent distribution panel device (100) precisely captures minute ultraviolet corona light and electromagnetic wave signals in the ultra-high frequency band that are artificially radiated from the corona arc generator through a corona light detection sensor and an ultra-wideband antenna detection sensor, and acquires them as analog discharge raw signals.

[0251] The analog discharge raw signal acquired through the discharge signal detection module (130) is input to the phase synchronization and signal acquisition means (151) and undergoes a process of being synchronized with the phase of the commercial power supply.

[0252] Specific discharge characteristic data collected through these simulation experiments is transmitted to the edge inference operation unit (150) and processed into core training data that enables the artificial neural network-based composite risk identification means (153) to preemptively and accurately identify initial corona arc risk factors that may occur at the site.

[0253] As shown in FIG. 8, a partial discharge occurrence test cell that comprehensively simulates various partial discharge occurrence situations is separately constructed and utilized so that the edge inference operation unit (150) can learn not only single defects but also various types of complex defects.

[0254] The above partial discharge generation test cell is designed as a sealed insulating structure that simulates void discharge occurring in fine voids within epoxy resin or silicone insulators, or precisely replicates surface discharge conditions flowing along the surface of the insulator in a laboratory environment.

[0255] In addition, the above partial discharge generation test cell artificially induces a complex environmental defect situation, such as tracking discharge, by forcibly injecting dust or metallic foreign matter into the test cell or by increasing humidity to an extreme level in order to derive a correlation with the physical environmental factors measured by the internal environment monitoring module (110).

[0256] When a physical defect under various conditions occurs through the partial discharge generation test cell, the discharge signal detection module (130) immediately detects it and applies the generated signal to the phase synchronization and signal acquisition means (151).

[0257] The above-described phase synchronization and signal acquisition means (151) passes the input signal through the internally provided low-noise amplifier (151a) to suppress background noise of the external laboratory environment and maximize the signal-to-noise ratio of the pure defect discharge signal to extract a high-quality digital discrete dataset.

[0258] Afterwards, the pattern mapping and normalization means (152) performs processing to shape the extracted dataset into a two-dimensional phase-decomposed partial discharge pattern heatmap image having a resolution of 32 x 32.

[0259] The heatmap image formed by the pattern mapping and normalization means (152) is combined with label data of a specific defect assigned by the test cell and provided as an input node of the artificial neural network-based composite risk identification means (153).

[0260] The artificial neural network-based composite risk identification means (153) that has completed learning based on such vast simulation training data will acquire an advanced identification capability that can perfectly distinguish between fine pattern shape differences of void discharge, corona discharge, and surface tracking discharge when subsequently deployed to the actual field.

[0261] Finally, the active control response unit (170) is able to perform preemptive and active response measures without malfunction, such as opening the power breaker or sending a maintenance alarm to the manager at the most appropriate time before a fire occurs, based on the accurate diagnosis results derived through the advanced identification ability.

[0262] In addition, an intelligent switchboard device (100) according to one embodiment of the present invention may further include a data logging unit (not shown) to preserve the basis for autonomous judgment of the active control response unit (170) and to precisely analyze the cause of the accident after the fact.

[0263] The data logging unit stores the two-dimensional phase-decomposed partial discharge pattern heatmap image and the environmental state data for a preset time (e.g., 5 seconds before and after operation) in an internal non-volatile flash memory for a short period based on the time of operation when the active control response unit (170) outputs the first, second, or third response control signal to an external facility. The stored data acts as a kind of black box and is transmitted to an external server when the communication network is restored after the disaster situation ends or when there is a request from a field manager terminal, and is used as an analysis indicator of explainable AI to visually verify the basis for the judgment of the deep learning model. Explanation of the symbols

[0264] 100: Intelligent Switchgear Device 110: Internal Environment Monitoring Module 130: Discharge signal detection module 150: Edge Inference Operation Unit 151: Phase synchronization and signal acquisition means 151a: Low-noise amplifier 152: Pattern Mapping and Normalization Methods 153: Artificial Neural Network-Based Complex Risk Identification Means 170: Active control corresponding unit

Claims

Claim 1 An intelligent switchboard device (100) for monitoring the internal state of power equipment connected to a power grid network, comprising: an internal environment monitoring module (110) that continuously generates environmental state data by measuring changes in the internal physical environment of the intelligent switchboard device (100) in real time; a discharge signal detection module (130) that acquires an analog discharge raw signal by capturing electromagnetic waves and acoustic radiation signals generated from internal power lines and insulators of the intelligent switchboard device (100); and an edge inference computation unit (150) that determines complex risk factors in real time by collecting and processing data internally within the intelligent switchboard device (100) independently of the communication network connection status with an external server based on the environmental state data and the analog discharge raw signal. The system includes an active control response unit (170) that generates a response control signal instructing the operation of an external facility in response to at least one of the composite risk factor determined by the edge inference operation unit (150) and an emergency environment abnormal signal directly received from the internal environment monitoring module (110), and outputs the response control signal to the external facility; wherein the edge inference operation unit (150) includes a phase synchronization and signal acquisition means (151) that detects a zero-cross point of commercial power supplied to the intelligent distribution panel device (100) to generate a phase synchronization signal, amplifies the analog discharge raw signal applied from the discharge signal detection module (130) through a low-noise amplifier (151a), and then converts it into a digital discrete dataset through a commercial log (log10) scale conversion;A pattern mapping and normalization means (152) for dividing the digital discrete dataset into a plurality of unit pieces based on a time interval corresponding to one period of the phase synchronization signal, superimposing them on a single plane according to the zero point reference of the phase synchronization signal, dividing the single plane into a first designated number of unit grids configured with a phase section on the horizontal axis and a signal intensity section on the vertical axis, counting the occurrence frequency values ​​of the signal mapped within each unit grid, and scaling the counted occurrence frequency values ​​to a second designated number of pixel values ​​to generate a two-dimensional phase-resolved partial discharge (PRPD) pattern heatmap image; An intelligent switchboard equipped with a partial discharge-based real-time complex safety diagnosis system, characterized by having: a convolutional operation layer that extracts a plurality of morphological feature maps from the above two-dimensional phase-decomposed partial discharge pattern heatmap image; a pooling layer that reduces the spatial dimension of the plurality of morphological feature maps; a fully connected layer that expands the plurality of morphological feature maps, whose spatial dimension has been reduced by the pooling layer, into a one-dimensional array and merges the environmental state data received from the internal environment monitoring module (110) to derive complex risk features; and an output layer that calculates the derived complex risk features into a plurality of risk element type probability values, thereby sequentially passing the results through the intelligent switchboard device (100) to finally determine the complex risk elements within the intelligent switchboard device (100). Claim 2 An intelligent switchboard equipped with a partial discharge-based real-time complex safety diagnosis system, wherein the internal environment monitoring module (110) includes a temperature detection sensor for measuring the internal temperature of the intelligent switchboard device (100), a humidity detection sensor for measuring internal humidity, a chemical gas detection sensor for measuring carbon monoxide concentration, a vibration detection sensor for measuring vibration frequency, a door opening detection sensor for detecting door opening, and a flood detection sensor for detecting a flood state, and the discharge signal detection module (130) includes a corona light detection sensor for detecting corona in the ultraviolet region and a high-frequency current transformer sensor for detecting leakage high-frequency current through electromagnetic induction. Claim 3 delete Claim 4 An intelligent switchboard equipped with a partial discharge-based real-time complex safety diagnosis system according to claim 1, wherein the active control response unit (170) generates a first response control signal to operate an external fire extinguisher equipment when the complex risk element determined by the edge inference operation unit (150) is a fire risk element, generates a second response control signal to open a power circuit breaker when the complex risk element is an insulation damage risk element or a corona arc risk element, and generates a third response control signal to activate an internal alarm speaker when the complex risk element is additionally confirmed as an earthquake occurrence element or a flood risk element through measurement by the internal environment monitoring module (110), wherein when the first, second, and third response control signals are superimposed, the first response control signal for the fire risk element is given the highest priority, the second response control signal for the insulation damage or corona arc risk element is given the next priority, and the third response control signal for the earthquake occurrence or flood risk element is given the next priority, and respectively generates and outputs to the external equipment by assigning priority according to a preset ranking control logic. Claim 5 A real-time diagnostic method performed by an intelligent switchboard device (100) that monitors the internal state of power equipment connected to a power grid network, wherein the intelligent switchboard device (100) comprises: an internal environment monitoring step (S110) in which the intelligent switchboard device (100) measures changes in the internal physical environment in real time and continuously generates environmental state data; a discharge signal detection step (S130) in which the intelligent switchboard device (100) captures electromagnetic waves and acoustic radiation signals generated from internal power lines and insulators to acquire an analog discharge raw signal; and an edge inference operation step (S150) in which the intelligent switchboard device (100) independently collects and processes the environmental state data and the analog discharge raw signal internally without transmitting them to an external server to determine complex risk factors in real time. The intelligent switchboard device (100) comprises an active control response step (S170) that generates a response control signal instructing the operation of an external facility in response to at least one of the composite risk factor determined by the edge inference operation step (S150) and an emergency environment abnormal signal directly received from the internal environment monitoring step (S110), and outputs the response control signal to the external facility; wherein the edge inference operation step (S150) comprises a phase synchronization and signal acquisition step (S151) that detects a zero-cross point of the commercial power supply entering the intelligent switchboard device (100) to generate a phase synchronization signal, amplifies the analog discharge raw signal through a low-noise amplifier provided internally, and then converts it into a digital discrete dataset through a commercial log (log10) scale conversion;A pattern mapping and normalization step (S152) for generating a two-dimensional Phase Resolved Partial Discharge (PRPD) pattern heatmap image by performing a process of dividing the digital discrete dataset into a plurality of unit pieces based on time intervals corresponding to one period of the phase synchronization signal and superimposing them on a single plane according to the zero point reference of the phase synchronization signal, dividing the single plane into a first designated number of unit grids composed of a phase section on the horizontal axis and a signal intensity section on the vertical axis, counting the occurrence frequency values ​​of the signals mapped within each unit grid, and scaling the counted occurrence frequency values ​​to a second designated number of pixel values. A real-time diagnostic method using an intelligent switchboard equipped with a partial discharge-based real-time complex safety diagnostic system, characterized by including: a convolutional operation layer that extracts a plurality of morphological feature maps from the 2D phase decomposition partial discharge pattern heatmap image; a pooling layer that reduces the spatial dimension of the plurality of morphological feature maps; a fully connected layer that expands the plurality of morphological feature maps, whose spatial dimension has been reduced by the pooling layer, into a 1D array and merges the environmental state data to derive complex risk features; and an output layer that calculates the derived complex risk features into a plurality of risk element type probability values, wherein the artificial neural network-based complex risk identification step (S153) is applied sequentially to determine the complex risk elements inside the intelligent switchboard device (100). Claim 6 In claim 5, the internal environment monitoring step (S110) performs in parallel the process of acquiring carbon monoxide concentration data through a chemical gas detection sensor, the process of acquiring internal temperature data through a temperature detection sensor, the process of acquiring internal humidity data through a humidity detection sensor, the process of acquiring mechanical vibration frequency data through a vibration detection sensor, the process of acquiring door open state data through a door open detection sensor, and the process of acquiring flood state data through a flood detection sensor; and the discharge signal detection step (S130) performs in parallel the process of acquiring a detection signal of corona light in the ultraviolet region, the process of acquiring an electromagnetic radiation signal in the ultra-high frequency band, the process of acquiring an acoustic vibration signal, and the process of acquiring an electromagnetic induction signal of leakage high-frequency current through a high-frequency current transformer sensor, thereby providing a real-time diagnosis method using an intelligent switchboard equipped with a partial discharge-based real-time complex safety diagnosis system. Claim 7 In claim 5, the pattern mapping and normalization step (S152) comprises: a process of generating a square unit grid matrix having a total of 1024 unit grids by dividing the phase section of the horizontal axis of the single plane into 32 sections based on 360 degrees, and dividing the signal strength section of the vertical axis into 32 sections based on the signal strength (dBm) range after commercial logarithmic scale conversion of the analog discharge raw signal; and a process of normalizing the accumulated occurrence frequency value for each of the 1024 unit grids to a numerical range of 0 to 255. Claim 8 In claim 5, the active control response step (S170) is characterized by including a process of generating a first response control signal to operate an external fire extinguisher equipment when the composite risk element derived through the edge inference operation step (S150) is a fire risk element, generating a second response control signal to open a power breaker when the composite risk element is an insulation damage risk element or a corona arc risk element, and generating a third response control signal to activate an internal alarm speaker of the intelligent switchboard device (100) when the composite risk element is additionally confirmed as an earthquake occurrence element or a flood risk element as a result of analyzing environmental state data obtained through the internal environment monitoring step (S110), wherein when the first, second, and third response control signals are superimposed, the first response control signal for the fire risk element is assigned the highest priority, the second response control signal for the insulation damage or corona arc risk element is assigned the next priority, and the third response control signal for the earthquake occurrence or flood risk element is assigned the next priority according to a preset ranking control logic, and generating them independently and outputting them to the external equipment respectively. Real-time diagnostic method using an intelligent switchboard equipped with a partial discharge-based real-time complex safety diagnostic system.

Images