Power distribution transformer fault real-time early warning analysis method based on edge computing

Abstract

The application discloses a power distribution transformer fault real-time early warning analysis method based on edge computing and relates to the technical field of power distribution transformer fault early warning, which comprises the following steps: acquiring multi-dimensional electrical characteristic quantities such as winding hot spot temperature rise gradient and iron core leakage magnetic flux density fluctuation value collected by an edge sensing node of a power distribution transformer body, generating a standardized characteristic data stream through on-site preprocessing by an edge computing gateway; extracting a frequency domain attenuation coefficient reflecting an insulation aging trend based on the time sequence correlation of the data stream, fusing the coefficient and a real-time load current effective value to calculate an insulation state degradation evaluation index; and dynamically adjusting an early warning trigger threshold according to the distribution law of the index in a historical database and generating a graded alarm instruction when the index exceeds the threshold. The method realizes accurate and real-time early warning of power distribution transformer faults, adapts to different operating conditions and reduces early warning deviation.

Description

Technical Field

[0001] This invention belongs to the field of power distribution transformer fault early warning technology, specifically a real-time early warning and analysis method for power distribution transformer faults based on edge computing. Background Technology

[0002] Distribution transformers are core equipment in the power distribution chain of a power system, and their operating status directly affects the stability and security of power supply. Fault early warning is a key means to ensure their reliable operation. Currently, distribution transformer fault early warning mainly involves collecting electrical characteristic quantities through sensor nodes deployed on the equipment itself, transmitting the data to a cloud server for centralized processing, extracting basic time-domain characteristics such as voltage and current, and combining this with preset fixed warning thresholds to determine potential faults.

[0003] In existing technologies, the multi-dimensional electrical characteristics collected by sensor nodes lack targeted local preprocessing, and data transmission to the cloud is prone to delays and losses, affecting the timeliness of early warnings. Furthermore, the features extracted by existing technologies are mostly limited to the time domain, failing to capture subtle changes in electrical characteristics during insulation aging and making it difficult to accurately reflect insulation aging trends. In addition, fault warning trigger thresholds are set with fixed values, without dynamic adjustment based on real-time equipment load and historical operating data distribution patterns, making them unsuitable for different operating conditions and prone to warning deviations. These shortcomings need to be addressed through local preprocessing, accurate extraction of insulation aging-related features, and dynamic adjustment of warning thresholds. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art;

[0005] Therefore, this invention proposes a real-time early warning and analysis method for distribution transformer faults based on edge computing, including:

[0006] The multi-dimensional electrical characteristics collected by each edge sensing node deployed on the main body of the distribution transformer during a continuous operating cycle are obtained. The multi-dimensional electrical characteristics include the temperature rise gradient of the winding hot spot, the core leakage flux density fluctuation value, and the bushing dielectric loss tangent increment.

[0007] The multi-dimensional electrical characteristics are preprocessed locally by an edge computing gateway to generate a standardized feature data stream.

[0008] Based on the time-series correlation of the standardized feature data stream, a frequency domain attenuation coefficient reflecting the insulation aging trend is extracted.

[0009] By combining the frequency domain attenuation coefficient with the real-time effective value of the load current, the insulation condition deterioration assessment index is calculated.

[0010] Based on the distribution pattern of the insulation condition deterioration assessment index in the historical database, the fault warning trigger threshold is dynamically adjusted.

[0011] When the insulation condition deterioration assessment index monitored in real time exceeds the fault warning trigger threshold, a graded alarm command is generated and issued.

[0012] Furthermore, the acquisition of multi-dimensional electrical feature quantities collected by each edge sensing node deployed on the distribution transformer body during a continuous operating cycle specifically includes:

[0013] The fiber optic temperature sensor array integrated on the top of the transformer tank is controlled to scan the axial region of the winding at a preset spatial resolution to obtain temperature field distribution data at each temperature measurement point.

[0014] Based on the temperature field distribution data and the rate of temperature difference change along the winding height direction, the temperature rise gradient of the winding hot spot is calculated.

[0015] A magnetostrictive sensor attached to the transformer core clamp is driven to collect alternating magnetic flux signals, and Hilbert transform is performed on the alternating magnetic flux signals to obtain the leakage magnetic flux density fluctuation value of the core.

[0016] By using a digital dielectric loss bridge coupled in the high-voltage bushing end screen circuit of the transformer, the phase difference offset between dielectric loss voltage and dielectric loss current is measured, and the change in the tangent value of the phase difference offset is defined as the tangent increment of the bushing dielectric loss angle.

[0017] The synchronously collected winding hot spot temperature rise gradient, core leakage flux density fluctuation value, and bushing dielectric loss tangent increment are output after adding a unified time scale.

[0018] Furthermore, the multi-dimensional electrical characteristics are preprocessed locally via an edge computing gateway to generate a standardized feature data stream, specifically including:

[0019] The multi-dimensional electrical characteristic quantities are received and the temperature rise gradient of the winding hot spot is filtered by moving average to eliminate temperature jump interference caused by the start and stop of the cooling fan.

[0020] Wavelet packet energy entropy analysis was performed on the leakage flux density fluctuation value of the iron core to extract the characteristic frequency band energy ratio that characterizes the loosening or short circuit of the iron core.

[0021] Temperature normalization is performed on the incremental tangent of the bushing dielectric loss angle to compensate for the influence of ambient temperature changes on dielectric loss measurement.

[0022] The temperature rise gradient of the winding hot spot, the fluctuation value of the core leakage flux density, and the increment of the bushing dielectric loss tangent after filtering, feature extraction and normalization are mapped to a numerical range of zero to one hundred.

[0023] The mapped data is packaged and encapsulated according to a fixed time window to form the standardized feature data stream.

[0024] Furthermore, based on the time-series correlation of the standardized feature data stream, a frequency-domain attenuation coefficient reflecting the insulation aging trend is extracted, specifically including:

[0025] Perform a fast Fourier transform on the winding hotspot temperature rise gradient sequence in the standardized feature data stream to obtain its spectral amplitude distribution;

[0026] Identify the dominant frequency component with the largest amplitude in the spectral amplitude distribution, and record the amplitude intensity corresponding to the dominant frequency component;

[0027] Calculate the amplitude attenuation slope of the dominant frequency component over multiple consecutive time windows, and define the reciprocal of the amplitude attenuation slope as the initial attenuation factor;

[0028] Analyze the autocorrelation coefficient of the bushing dielectric loss tangent increment sequence in the standardized feature data stream. When the autocorrelation coefficient is less than a set threshold under a preset lag order, the initial attenuation factor is weighted and amplified.

[0029] The weighted and amplified initial attenuation factor is used as the frequency domain attenuation coefficient.

[0030] Furthermore, by integrating the frequency domain attenuation coefficient with the real-time effective value of the load current, an insulation condition degradation assessment index is calculated, specifically including:

[0031] The real-time three-phase current sampling value of the distribution transformer is retrieved from the substation monitoring system, and the root mean square value of the real-time three-phase current sampling value is calculated as the effective value of the real-time load current.

[0032] Construct a two-dimensional feature plane with the frequency domain attenuation coefficient as the horizontal axis and the real-time load current effective value as the vertical axis;

[0033] Divide the two-dimensional feature plane into several grid cells of equal area, and count the position index of the grid cell into which the data point at the current sampling time falls;

[0034] Query the pre-stored insulation state mapping table and find the corresponding state score value according to the location index;

[0035] The median filtering is applied to the state score values ​​at ten consecutive sampling times, and the filtered output is defined as the insulation state deterioration assessment index.

[0036] Furthermore, based on the distribution pattern of the insulation condition deterioration assessment index in the historical database, the fault early warning trigger threshold is dynamically adjusted, specifically including:

[0037] Access the historical database stored locally on the edge computing gateway and read all recorded insulation condition degradation assessment indices from the past thirty natural days;

[0038] Histogram statistics were performed on the read insulation condition deterioration assessment index to determine the peak position and distribution width of its probability density function;

[0039] The sum obtained by adding three times the distribution width to the peak position of the probability density function is set as the initial warning threshold;

[0040] Determine whether the rate of change of the current insulation condition deterioration assessment index exceeds a preset slope threshold;

[0041] If the slope threshold is exceeded, a correction amount proportional to the rate of change is added to the initial warning threshold to obtain the fault warning trigger threshold; if the slope threshold is not exceeded, the initial warning threshold is maintained as the fault warning trigger threshold.

[0042] Furthermore, when the insulation condition deterioration assessment index monitored in real time exceeds the fault warning trigger threshold, a graded alarm command is generated and issued, specifically including:

[0043] Compare the current insulation condition deterioration assessment index with the fault warning trigger threshold.

[0044] When the insulation condition deterioration assessment index exceeds the fault warning trigger threshold for the first time, a first-level early warning message is generated, which includes the current value of the winding hot spot temperature rise gradient.

[0045] When the insulation condition deterioration assessment index continues to exceed the fault warning trigger threshold for five minutes, and the core leakage flux density fluctuation value simultaneously exceeds the preset flux threshold, a secondary warning message is generated. The secondary warning message contains the spectral characteristics of the core leakage flux density fluctuation value.

[0046] When the insulation condition deterioration assessment index exceeds twice the fault warning trigger threshold, and the bushing dielectric loss tangent increment exceeds the preset dielectric loss threshold, a level three trip warning message is generated.

[0047] The generated Level 1 warning message, Level 2 warning message, or Level 3 trip warning message are sent to the remote master station through the communication interface of the edge computing gateway.

[0048] Furthermore, mapping the temperature rise gradient of the winding hotspot, the fluctuation value of the core leakage flux density, and the increment of the bushing dielectric loss tangent after filtering, feature extraction, and normalization to a numerical range of zero to one hundred specifically includes:

[0049] Find the maximum allowable temperature rise value of the winding hot spot temperature rise gradient under rated operating conditions, divide the currently collected winding hot spot temperature rise gradient by the maximum allowable temperature rise value and then multiply by one hundred to obtain the temperature rise mapping value.

[0050] Find the peak value of magnetic flux density corresponding to the core leakage flux density fluctuation value at the core saturation critical point, divide the currently collected core leakage flux density fluctuation value by the peak value of magnetic flux density and then multiply by one hundred to obtain the magnetic flux mapping value.

[0051] Find the range of difference between the tangential increment of the casing medium loss angle and the new casing and the severely aged casing. Subtract the current collected tangential increment of the casing medium loss angle from the dielectric loss value of the new casing, divide by the range of difference, and then multiply by one hundred to obtain the dielectric loss mapping value.

[0052] The temperature rise mapping value, the magnetic flux mapping value, and the dielectric loss mapping value are respectively subjected to upper and lower limit clamping to ensure that the output value is not lower than zero and does not exceed one hundred.

[0053] Furthermore, the autocorrelation coefficient of the bushing dielectric loss tangent increment sequence in the standardized feature data stream is analyzed. When the autocorrelation coefficient is less than a set threshold under a preset lag order, the initial attenuation factor is weighted and amplified, specifically including:

[0054] Extract the tangential increment of the bushing medium loss angle of continuous preset length sampling points from the standardized feature data stream to form the time series data to be analyzed;

[0055] Set the lag order for autocorrelation analysis, calculate the autocorrelation coefficient of the time series data at the lag order, and the autocorrelation coefficient reflects the degree of linear correlation of the series at different time points;

[0056] Obtain a pre-stored critical aging threshold, which is an empirical value obtained based on the statistics of historical failure samples.

[0057] The autocorrelation coefficient is compared with the aging critical judgment threshold to determine the degree of random fluctuation of the bushing medium loss tangent increment sequence;

[0058] When the autocorrelation coefficient is determined to be less than the aging critical threshold, it indicates that the dielectric loss shows a non-stationary accelerated aging trend. At this time, the preset amplification coefficient matrix is ​​called, and the corresponding weighted amplification coefficient is obtained by looking up the table according to the range of the initial attenuation factor.

[0059] The initial attenuation factor is multiplied by the weighting amplification factor to complete the weighting amplification operation of the initial attenuation factor;

[0060] When the autocorrelation coefficient is determined to be greater than or equal to the aging critical threshold, the initial decay factor is kept unchanged, and the process proceeds directly to the subsequent coefficient definition step.

[0061] Furthermore, the step of querying the pre-stored insulation state mapping table and looking up the corresponding state score value based on the location index specifically includes:

[0062] The two-dimensional feature plane is divided into a grid of 100 equal parts in both the horizontal and vertical directions, and each grid cell is assigned a unique combination of row and column numbers as the position index;

[0063] The insulation state mapping table is traversed. The insulation state mapping table stores the state score value corresponding to each grid cell. The state score value is in the range of integer zero to integer one hundred.

[0064] Based on the row and column number of the grid cell where the current data point is located, locate the corresponding entry in the insulation state mapping table;

[0065] Read the status score value stored in the located entry;

[0066] If the current data point falls at the boundary of two grid cells, the distance weights from the data point to the center of the two adjacent grid cells are calculated respectively. The final state score value is obtained by weighted averaging of the two state score values.

[0067] Compared with the prior art, the beneficial effects of the present invention are:

[0068] Based on the time-series correlation of standardized feature data streams, this method extracts the frequency-domain attenuation coefficient reflecting the insulation aging trend, captures the time-series correlation characteristics of feature data, and explores the frequency-domain variation patterns of electrical features during insulation aging. Compared to conventional methods that only extract time-domain electrical features, this method can accurately capture the subtle dynamics of insulation aging, avoid missing potential faults due to incomplete feature extraction, achieve precise perception of insulation aging status, make fault warnings more targeted, and solve the problem that conventional technologies cannot accurately reflect insulation aging trends.

[0069] This method calculates an insulation condition degradation assessment index by integrating the frequency domain attenuation coefficient with the real-time load current RMS value. Based on the distribution pattern of this index in a historical database, the fault warning trigger threshold is dynamically adjusted. By combining real-time equipment load conditions and historical operating data, the warning threshold can adapt to different load conditions and equipment operating status changes. Compared to the conventional method of using fixed warning thresholds, this avoids overly sensitive or delayed warnings caused by fixed thresholds, reduces false alarms and missed alarms, improves the reliability and adaptability of fault warnings, achieves real-time fault warnings, and solves the problems of conventional technologies being unable to adapt to different operating conditions and having large warning deviations. Attached Figure Description

[0070] Figure 1 This is a flowchart illustrating the steps of the real-time early warning and analysis method for distribution transformer faults based on edge computing described in this invention.

[0071] Figure 2 A flowchart for localized preprocessing and standardized data stream generation;

[0072] Figure 3 Box plot comparing the incremental sequence of the dielectric loss tangent of the bushing;

[0073] Figure 4 Heatmap of insulation state mapping table;

[0074] Figure 5 This is a sub-plot of the tangent increment of the dielectric loss angle in the bushing. Detailed Implementation

[0075] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0076] See Figure 1 This invention provides a real-time early warning and analysis method for distribution transformer faults based on edge computing. The specific method includes:

[0077] Various dedicated sensing nodes deployed on the distribution transformer simultaneously collect multi-dimensional electrical characteristics, such as winding hotspot temperature rise gradient, core leakage flux density fluctuation, and bushing dielectric loss tangent increment, within a continuous operating cycle. The edge computing gateway aggregates these raw characteristics via fieldbus or wireless means and performs a series of local preprocessing operations, including data filtering, interference cancellation, temperature compensation, and normalization mapping, ultimately generating a standardized characteristic data stream with uniform format and standardized dimensions. The time-series correlation of this standardized characteristic data stream is analyzed, and a frequency domain attenuation coefficient characterizing the aging trend of insulation materials is extracted using frequency domain analysis. This frequency domain attenuation coefficient is then fused with the real-time effective value of the load current synchronously obtained from the substation monitoring system to obtain a comprehensive insulation condition deterioration assessment index. The edge computing gateway maintains a historical database to store historical records of this insulation condition deterioration assessment index. By analyzing its distribution patterns, the trigger threshold for fault warnings can be dynamically adjusted. During real-time monitoring, the calculated insulation condition deterioration assessment index is continuously compared with the currently dynamically adjusted fault warning trigger threshold. Once the threshold is exceeded, different levels of graded alarm commands are generated based on the degree and duration of the exceedance, and then sent to the remote master station through the communication network.

[0078] In one embodiment of the invention, a fiber optic temperature sensor array integrated and mounted on the top of the transformer tank is controlled. This array scans along the winding axis with a preset spatial resolution to acquire temperature data from multiple temperature measurement points at different heights of the winding, thereby forming temperature field distribution data of the winding region. Based on this temperature field distribution data, the rate of change of temperature difference along the winding height direction is calculated, which is used as the temperature rise gradient of the winding hot spots. A magnetostrictive sensor attached to the surface of the transformer core clamp is driven to collect alternating magnetic flux signals near the core. The collected original alternating magnetic flux signals are subjected to Hilbert transform to resolve the signal envelope. The fluctuation amplitude of this envelope is defined as the core leakage flux density fluctuation value. Through a digital dielectric loss bridge coupled in the high-voltage bushing end screen circuit of the transformer, the dielectric loss voltage signal and dielectric loss current signal are simultaneously measured, and the phase difference offset between the two is calculated. The change of the tangent value of this phase difference offset compared to the reference value is calculated as the bushing dielectric loss angle tangent increment. Finally, a unified timestamp is added to the three characteristic quantities that are synchronously collected: the temperature rise gradient of the winding hot spot, the fluctuation value of the core leakage flux density, and the increment of the bushing dielectric loss tangent, and then packaged and output to the edge computing gateway.

[0079] In practical implementation, the fiber optic temperature sensor array integrated on the top of the distribution transformer tank scans the axial region of the winding at a preset spatial resolution, acquiring temperature data at each measuring point to form temperature field distribution data for the winding region. In one example scenario, the spatial resolution is set to one measuring point every 10 centimeters, acquiring temperature values ​​at 11 points along the axial direction of the winding at a height of 1 meter, thus forming the temperature field distribution data. The temperature gradient of the winding hotspots is calculated based on the rate of change of temperature difference along the winding height direction from the temperature field distribution data. The calculation process involves selecting the temperature values ​​of adjacent measuring points in the temperature field distribution data and their corresponding height coordinates. The rate of change of temperature difference can be obtained by calculating the ratio of the temperature difference between adjacent points to the height difference, expressed by the formula:

[0080]

[0081] in: and These represent the temperature values ​​at the i-th and (i+1)-th temperature measurement points, respectively. and This represents the corresponding winding axial height coordinate. This refers to the calculated winding hot spot temperature rise gradient. By traversing all adjacent points and selecting the maximum value as the final output winding hot spot temperature rise gradient, in data comparison, the winding hot spot temperature rise gradient of a normally stable transformer usually shows a gradual change, while in the early stage of a fault with local overheating, the winding hot spot temperature rise gradient will increase significantly.

[0082] In some embodiments, a magnetostrictive sensor attached to the core clamp of a distribution transformer is driven to collect the original alternating magnetic flux signal near the core. The original alternating magnetic flux signal is subjected to Hilbert transform to obtain its analytical signal, and then the signal envelope is obtained. The fluctuation amplitude of the envelope is defined as the core leakage flux density fluctuation value. In a specific implementation, the magnetostrictive sensor collects the analog signal in the form of voltage at a sampling rate of 10,000 times per second. After analog-to-digital conversion, a digital sequence is obtained. The analytical signal is calculated by applying Hilbert transform to the digital sequence. The magnitude sequence of the analytical signal constitutes the signal envelope. The peak-to-peak value of the envelope in one power frequency cycle is calculated, and this peak-to-peak value is output as the core leakage flux density fluctuation value. Compared with the flat envelope under normal operating conditions, when there is a loose core or a local short circuit hazard, the alternating magnetic flux signal envelope collected by the magnetostrictive sensor will show obvious periodic or non-periodic fluctuations, and the corresponding core leakage flux density fluctuation value will increase significantly.

[0083] Optionally, a digital dielectric loss bridge coupled in the high-voltage bushing end-screen circuit of the distribution transformer is used to synchronously measure the dielectric loss voltage and current signals. The high-precision lock-in amplifier inside the digital dielectric loss bridge measures the phase difference offset between the dielectric loss voltage and current signals. The tangent of the phase difference offset measured in the current sampling period is subtracted from a preset reference tangent value, and the difference is defined as the bushing dielectric loss angle tangent increment. In specific implementation, the reference tangent value is usually selected as the dielectric loss angle tangent value measured in the factory acceptance test or recent routine test of the bushing. The digital dielectric loss bridge calculates and outputs the phase difference at a rate of 100 times per second and calculates its tangent value in real time. The average value of the tangent value is calculated for each power frequency cycle, and then the average value of this cycle is subtracted from the reference tangent value to obtain the bushing dielectric loss angle tangent increment. In terms of data comparison, when the bushing insulation is in good condition, the bushing dielectric loss angle tangent increment fluctuates slightly near zero, while when the bushing insulation is damp or aged, the bushing dielectric loss angle tangent increment will show a continuous increasing trend.

[0084] It is understandable that the synchronously collected winding hotspot temperature rise gradient, core leakage flux density fluctuation value, and bushing dielectric loss tangent increment are output after adding a unified time stamp. In specific implementation, the edge computing gateway sends periodic synchronous acquisition instructions to each sensor node. After receiving the instruction, each sensor node immediately performs one-time data acquisition and packages the collected winding hotspot temperature rise gradient, core leakage flux density fluctuation value, and bushing dielectric loss tangent increment data together with the same timestamp issued from the edge computing gateway. This data is then transmitted to the data buffer of the edge computing gateway via fieldbus or wireless network. In some embodiments, the timestamp accuracy is at the millisecond level to ensure that different physical quantities are strictly aligned in the time dimension, thereby providing an accurate time reference for subsequent analysis of the time series correlation of multi-dimensional electrical characteristic quantities.

[0085] In one embodiment of the present invention, see [reference] Figure 2The edge computing gateway receives multi-dimensional electrical characteristics from various sensor nodes. A moving average filter is applied to the winding hotspot temperature rise gradient data to smooth out transient temperature fluctuations caused by the periodic start-stop of the cooling fan. Wavelet packet energy entropy analysis is performed on the core leakage flux density fluctuation sequence to decompose the signal into different frequency bands, and the proportion of energy in each characteristic frequency band to the total energy is calculated. This proportion can be used to characterize potential loosening or short-circuit defects in the core. The bushing dielectric loss tangent increment is normalized based on the real-time ambient temperature to compensate for the inherent influence of temperature changes on the dielectric loss tangent measurement. After the above preprocessing, the three characteristics are mapped to numerical ranges from zero to one hundred. The specific mapping process is as follows: First, find the maximum allowable temperature rise value of the winding hot spot temperature rise gradient under the transformer's rated operating conditions. Divide the currently collected winding hot spot temperature rise gradient by this maximum allowable temperature rise value and multiply by 100 to obtain the temperature rise mapping value. Second, find the peak value of the core leakage flux density fluctuation value corresponding to the core flux density reaching the saturation critical point. Divide the currently collected core leakage flux density fluctuation value by this peak value and multiply by 100 to obtain the flux mapping value. Third, find the range of differences between the typical value of the bushing dielectric loss tangent increment under new bushing conditions and the typical value under severely aged bushing conditions. Subtract the typical value of the new bushing dielectric loss from the currently collected bushing dielectric loss tangent increment, divide by this range, and multiply by 100 to obtain the dielectric loss mapping value. The calculated temperature rise mapping value, flux mapping value, and dielectric loss mapping value are then clamped to ensure that their output values ​​are neither lower than zero nor higher than 100. The edge computing gateway packages and encapsulates the mapped temperature rise, magnetic flux, and dielectric loss values ​​according to a fixed time window length, forming a standardized feature data stream for subsequent analysis.

[0086] In practical implementation, the edge computing gateway receives multi-dimensional electrical characteristic quantities from the sensing nodes and performs moving average filtering on the winding hot spot temperature rise gradient to eliminate temperature jump interference caused by the start and stop of the cooling fan. The window length of the moving average filter is set to 30 seconds, which means that each output data point is the arithmetic mean of the winding hot spot temperature rise gradient at the current moment and the previous 29 sampling points (1 point per second). In the example scenario, at the moment the cooling fan starts, the original winding hot spot temperature rise gradient may jump from 5.0K / m to 7.5K / m instantaneously. After moving average filtering, the output value will transition smoothly, avoiding sharp pulse interference. In terms of data comparison, the waveform of the filtered winding hot spot temperature rise gradient sequence is smoother and more conducive to observing its true trend change. Wavelet packet energy entropy analysis is performed on the core leakage flux density fluctuation value to extract the characteristic frequency band energy proportion representing core loosening or short circuit. In specific implementation, four-level wavelet packet decomposition is used to decompose the core leakage flux density fluctuation value signal into 16 frequency bands, and the energy of each frequency band is calculated. The characteristic frequency band is usually predefined as a specific frequency range related to core mechanical vibration or local saturation. The energy of this frequency band divided by the total signal energy gives the characteristic frequency band energy proportion, expressed by the formula:

[0087]

[0088] in: The proportion of energy in the representative characteristic frequency band. The sum of the energies of all sub-bands within the characteristic frequency band. This represents the sum of the energies of all 16 sub-bands. Under normal conditions, the core is firmly secured, and the energy proportion of the characteristic band is relatively low. When the core becomes loose, the vibrational energy related to the loosening frequency increases, leading to a higher proportion of the characteristic band energy. Significant increase.

[0089] In some embodiments, temperature normalization is performed on the bushing dielectric loss tangent increment to compensate for the influence of ambient temperature changes on dielectric loss measurement. In a specific implementation, a temperature-dielectric loss reference curve is built into the edge computing gateway. This curve describes the typical relationship between the bushing dielectric loss tangent increment and ambient temperature changes. During processing, the values ​​of the ambient temperature sensor installed on the transformer tank wall are read in real time. Based on the temperature-dielectric loss reference curve, the expected baseline value of the bushing dielectric loss tangent increment under the current ambient temperature is found. The actual measured bushing dielectric loss tangent increment is subtracted from this expected baseline value to obtain the temperature-normalized bushing dielectric loss tangent increment. Optionally, the temperature-dielectric loss reference curve is obtained by fitting transformer factory test or historical long-term monitoring data. In data comparison, the unnormalized bushing dielectric loss tangent increment will exhibit periodic fluctuations in environments with large diurnal temperature differences. After normalization, this fluctuation caused by pure temperature changes is effectively suppressed, making the data more realistically reflect the changes in insulation state.

[0090] It is understandable that the winding hotspot temperature rise gradient, core leakage flux density fluctuation value, and bushing dielectric loss tangent increment, after filtering, feature extraction, and normalization, are mapped to a numerical range of zero to one hundred. The maximum allowable temperature rise value of the winding hotspot temperature rise gradient under rated operating conditions is then found. The currently collected winding hotspot temperature rise gradient is divided by the maximum allowable temperature rise value and multiplied by one hundred to obtain the temperature rise mapping value. For example, if the maximum allowable value of the winding hotspot temperature rise gradient is 15 K / m, and the current filtered value is 6 K / m, then the temperature rise mapping value is (6 / 15)*100=40. The peak flux density value corresponding to the core leakage flux density fluctuation value at the core saturation critical point is then found. The current core leakage flux density fluctuation value after characteristic frequency band energy ratio analysis is divided by the peak flux density value and multiplied by one hundred to obtain the flux mapping value. In specific implementations, the peak flux density value can be obtained from transformer design parameters or type test reports and used as a fixed threshold. Find the range of differences between the tangential increment of the bushing dielectric loss angle between the new bushing and the severely aged bushing. Subtract the dielectric loss value of the new bushing from the tangential increment of the bushing dielectric loss angle after normalization at the current temperature, divide by the range of differences, and then multiply by 100 to obtain the dielectric loss mapping value. In some embodiments, the dielectric loss value of the new bushing is 0.002, and the dielectric loss value of the severely aged bushing is 0.020, so the range of differences is 0.018. If the current normalized value is 0.005, then the dielectric loss mapping value is ((0.005-0.002) / 0.018)*100≈16.7.

[0091] In practice, the temperature rise mapping value, magnetic flux mapping value, and dielectric loss mapping value are clamped to ensure that the output value is not lower than zero and does not exceed one hundred. This means that if the calculated mapping value is less than zero, the output is forced to be zero; if the mapping value is greater than one hundred, the output is forced to be one hundred. This step ensures that all feature quantities are within a uniform, dimensionless scaling range in subsequent processing. The mapped data is packaged and encapsulated according to a fixed time window to form a standardized feature data stream. Optionally, the fixed time window is set to 5 minutes. Every 5 minutes, the edge computing gateway packages all temperature rise mapping values, magnetic flux mapping values, and dielectric loss mapping values ​​generated during this period into a data packet. In addition to containing the three mapping value sequences, the data packet also contains the start and end timestamps of the time window. This data packet is a standardized feature data stream unit and is sent to the buffer of the subsequent analysis module.

[0092] In one embodiment of the present invention, a Fast Fourier Transform (FFT) is performed on the winding hotspot temperature rise gradient mapping value sequence in the standardized feature data stream to transform it from the time domain to the frequency domain, thereby obtaining its spectral amplitude distribution. In this spectral amplitude distribution, the frequency component with the largest amplitude is identified and determined as the dominant frequency component, and the amplitude intensity corresponding to this dominant frequency component is recorded. The attenuation slope of the amplitude intensity of this dominant frequency component over multiple consecutive time windows is calculated, and the reciprocal of this attenuation slope is defined as the initial attenuation factor. Simultaneously, the autocorrelation coefficient of the bushing dielectric loss tangent increment mapping value sequence in the standardized feature data stream is analyzed. A lag order for the autocorrelation analysis is set, and the autocorrelation coefficient at this lag order is calculated. This coefficient reflects the degree of linear correlation of the sequence at different time points. A pre-stored aging critical judgment threshold is obtained, and the calculated autocorrelation coefficient is compared with this aging critical judgment threshold. When the autocorrelation coefficient is less than the critical aging threshold, it indicates that the randomness of the incremental sequence of the bushing dielectric loss tangent is enhanced, showing a non-stationary accelerated aging trend. At this time, a preset amplification coefficient matrix is ​​invoked, and the corresponding weighted amplification coefficient is obtained from the matrix according to the range of the initial attenuation factor. The initial attenuation factor is multiplied by this weighted amplification coefficient to complete the weighted amplification operation of the initial attenuation factor. When the autocorrelation coefficient is greater than or equal to the critical aging threshold, the initial attenuation factor remains unchanged. The value obtained after judgment and possible weighted amplification operations is finally defined as the frequency domain attenuation coefficient.

[0093] In practical implementation, a Fast Fourier Transform (FFT) is performed on the winding hotspot temperature rise gradient mapping value sequence in the standardized feature data stream to obtain its spectral amplitude distribution. The input of the FFT is the winding hotspot temperature rise gradient mapping value sequence containing 1024 consecutive sampling points, and the output is 512 complex frequency components. The magnitudes of these components are used to construct the spectral amplitude distribution. The frequency component with the largest amplitude is identified in the spectral amplitude distribution and determined as the dominant frequency component. The amplitude intensity corresponding to the dominant frequency component is recorded. The amplitude attenuation slope of the dominant frequency component within multiple consecutive time windows is calculated. The reciprocal of the amplitude attenuation slope is defined as the initial attenuation factor. Specifically, the amplitude intensity of the dominant frequency component in the current time window is compared with the amplitude intensity of the dominant frequency component in the previous ninth time window. The formula is expressed as:

[0094]

[0095] in: The amplitude intensity of the dominant frequency component represents the current time window. This represents the amplitude intensity of the dominant frequency component in the previous ninth time window. It is the duration of a single time window. The calculated attenuation slope is the initial attenuation factor. In the data comparison, as the insulation aging process accelerates, the amplitude intensity of the dominant frequency component decays more rapidly, leading to a faster attenuation slope. The absolute value of increases, thereby increasing the initial attenuation factor. Decrease.

[0096] In some embodiments, the autocorrelation coefficient of the bushing medium loss angle tangent incremental mapping value sequence in the standardized feature data stream is analyzed. The bushing medium loss angle tangent incremental mapping values ​​of 1024 consecutive sampling points are extracted from the standardized feature data stream to form the time series data to be analyzed. The lag order of the autocorrelation analysis is set to 100, and the autocorrelation coefficient of the time series data at a lag order of 100 is calculated. The autocorrelation coefficient reflects the degree of linear correlation of the sequence at different time points. A pre-stored aging critical judgment threshold is obtained; this threshold is an empirical value obtained based on historical fault sample statistics. The calculated autocorrelation coefficient is compared with the aging critical judgment threshold to determine the degree of random fluctuation in the bushing medium loss angle tangent incremental mapping value sequence. When the autocorrelation coefficient is determined to be less than the aging critical judgment threshold, it indicates that the medium loss exhibits a non-stationary accelerated aging trend. At this time, a preset amplification coefficient matrix is ​​invoked, and the corresponding weighted amplification coefficient is obtained by looking up the table according to the range of the initial attenuation factor values. See Table 1 for an example of the weighted amplification coefficient matrix.

[0097] Table 1: Weighted Amplification Coefficient Matrix

[0098] Initial attenuation factor range Weighted amplification factor 0.0≤ <0.5 2.5 0.5≤ <1.0 2.0 1.0≤ <2.0 1.5 2.0≤ 1.0 Initial attenuation factor interval Weighted amplification factor

[0099] It can be understood that multiplying the initial attenuation factor by the weighting amplification factor completes the weighted amplification operation of the initial attenuation factor. For example, if the calculated initial attenuation factor... If the autocorrelation coefficient is 0.8 and less than the critical aging threshold, then it belongs to the interval "0.5≤" from the matrix. <1.0” corresponds to the weighted amplification factor The initial autocorrelation factor is 2.0, and the weighted amplification result is 0.8 * 2.0 = 1.6. When the autocorrelation coefficient is determined to be greater than or equal to the aging critical threshold, the initial attenuation factor remains unchanged, and the process proceeds directly to the subsequent coefficient definition step. In practice, the value obtained after the judgment and possible weighted amplification operations is ultimately defined as the frequency domain attenuation coefficient. At the data comparison level, a series of incremental changes in the tangent of the bushing dielectric loss angle that exhibits a stable and slow change has a high autocorrelation coefficient and typically does not trigger weighted amplification. The frequency domain attenuation coefficient... equal to the initial decay factor However, a sequence exhibiting violent random fluctuations has a low autocorrelation coefficient, which triggers weighted amplification, resulting in a decrease in the frequency domain attenuation coefficient of the final output. Significantly greater than the initial decay factor This amplifies the weight of insulation aging trends in subsequent assessments.

[0100] See Figure 3In the real-time fault early warning analysis method for distribution transformers based on edge computing, the bushing dielectric loss tangent increment is a core multi-dimensional electrical characteristic quantity characterizing the aging state of the high-voltage bushing insulation of the transformer. Its sequence distribution characteristics are directly used for insulation condition deterioration assessment and fault early warning triggering logic. The figure shows a comparison of the bushing dielectric loss tangent increment sequence under normal and accelerated aging states after standardization (mapped to the 0-100 value range), used to intuitively quantify the statistical distribution differences of dielectric loss increment under the two states. In the specific analysis, the upper and lower boundaries of the box in the box plot represent the upper quartile (Q3) and lower quartile (Q1) of the data, respectively. The horizontal line inside the box is the median of the data, and the upper and lower whiskers correspond to the maximum and minimum values ​​of the data (after removing outliers), respectively. As shown in the figure: In the normal state group, the median of the increase in the dielectric loss tangent of the bushing is approximately 20, with the distribution concentrated in the 15-25 range. The overall data fluctuation range is small (the beard line covers 12-28), indicating that the bushing insulation performance is stable under normal operating conditions, and the dielectric loss increment exhibits a stable, low-fluctuation time series characteristic, meeting the criteria of "high autocorrelation coefficient and no accelerated aging trend." In the accelerated aging group, the median of the increase in the dielectric loss tangent of the bushing significantly increases to 25, and the distribution of the beard line expands significantly to the 17-33 range. The data fluctuation range significantly increases (the beard line covers 5-45), indicating that the bushing insulation performance deteriorates under accelerated aging conditions, and the dielectric loss increment exhibits a non-stationary, high-random-fluctuation time series characteristic, corresponding to the accelerated aging judgment condition of "autocorrelation coefficient less than the critical aging judgment threshold, triggering weighted amplification." The comparison results provide data support for "weighted correction of frequency domain attenuation coefficient based on autocorrelation coefficient of bushing dielectric loss increment sequence": the high volatility of dielectric loss increment sequence under accelerated aging state will lead to a decrease in autocorrelation coefficient, thereby triggering the weighted amplification of the initial attenuation factor and increasing the weight of insulation aging trend in subsequent evaluation; at the same time, it provides a quantitative basis for setting dielectric loss threshold for three-level trip warning, and can set the aging critical judgment threshold of dielectric loss increment based on the distribution difference between the two states to realize graded early warning of fault.

[0101] In one embodiment of the present invention, real-time three-phase current sampling values ​​of the distribution transformer are synchronously retrieved from the substation monitoring system, and the root mean square value of these sampling values ​​is calculated as the real-time effective load current value. A two-dimensional feature plane is constructed with the frequency domain attenuation coefficient as the abscissa and the real-time effective load current value as the ordinate. Within this two-dimensional feature plane, the abscissa and ordinate are each uniformly divided into ten parts, thereby forming one hundred grid cells of equal area. Each grid cell is assigned a unique combination of row and column numbers, which is the location index. The insulation state mapping table pre-stored in the edge computing gateway is queried. This table stores the state score value corresponding to each grid cell determined by the row and column numbers. The state score value is an integer between zero and one hundred. Based on the data point formed by the frequency domain attenuation coefficient and the real-time effective load current value at the current sampling time on the two-dimensional feature plane, the row and column numbers of the grid cell to which it falls are determined, i.e., the location index. Based on this location index, the corresponding entry in the insulation state mapping table is located, and the state score value stored in that entry is read. If the current data point falls near the boundaries of two or more grid cells (i.e., at the intersection), the distance from the data point to the center of each of its adjacent grid cells is calculated. Using the reciprocal of the distance as the weight, a weighted average is calculated for the state score values ​​corresponding to each grid cell. This weighted average is then used as the final state score value. The state score value sequence obtained from ten consecutive sampling times is then subjected to median filtering. The filtered output value is defined as the insulation state degradation assessment index for that time moment.

[0102] In practical implementation, real-time three-phase current sampling values ​​of the distribution transformer are retrieved from the substation monitoring system. The root mean square (RMS) value of the real-time three-phase current sampling values ​​is calculated as the real-time effective load current value. In the example scenario, the substation monitoring system collects instantaneous current values ​​of phases A, B, and C at a rate of 1600 points per second. For each phase, data for one power frequency cycle (20 milliseconds, containing 32 sampling points) is collected. The square root of the average of the sum of squares of these 32 sampling values ​​is calculated to obtain the effective current value of that phase in that cycle. Then, the effective current values ​​of the three phases are averaged to obtain the real-time effective load current value at the current sampling time. In the data comparison, the effective value of the real-time load current is relatively low when the transformer is unloaded, but increases proportionally with the increase of load. A frequency domain attenuation coefficient is constructed... The x-axis represents the real-time effective value of the load current. The vertical axis represents a two-dimensional characteristic plane, and the horizontal axis range of this two-dimensional characteristic plane is defined as the frequency domain attenuation coefficient. The possible value range is [0, 10], and the range of the vertical axis is defined as the effective value of the real-time load current. The possible value range is [0, 1.5 times the rated current]. The two-dimensional feature plane is divided into several equal-area grid cells. In specific implementation, the horizontal coordinate range [0, 10] is evenly divided into 10 equal parts, and the vertical coordinate range is evenly divided into 10 equal parts, thus forming 100 equal-area rectangular grid cells. Each grid cell is assigned a unique row and column number combination as a position index, with row numbers from top to bottom (1 to 10) and column numbers from left to right (1 to 10). The data points at the current sampling time are statistically analyzed (based on the frequency domain attenuation coefficient). and real-time load current RMS value The location index of the grid cell into which the data point falls is determined. The positioning method is to determine the row and column to which it belongs by comparing the coordinate values ​​of the data point with the grid boundary values. For example, if the frequency domain attenuation coefficient... Real-time load current RMS value If the x-axis value of 3.2 times the rated current falls within the interval [3,4), corresponding to column number 4; and the y-axis value of 0.8 times the rated current falls within a preset interval, corresponding to row number 5, then the position index is (row 5, column 4).

[0103] In some embodiments, a pre-stored insulation state mapping table is queried, and the corresponding state score value is found based on the location index. The insulation state mapping table is a two-dimensional lookup table stored in the edge computing gateway. Its rows and columns correspond one-to-one with the grids divided by the two-dimensional feature plane. Each entry stores the state score value of the corresponding grid cell. The state score value is pre-set based on historical operating data, simulation data, or expert experience, and ranges from integer 0 to integer 100. The higher the value, the higher the risk of insulation state degradation. See Table 2.

[0104] Table 2: Insulation State Mapping Table

[0105] Rows and Columns Column 1 Column 2 Column 3 Column 4 Column 5 Line 1 5 5 8 10 15 Line 2 8 10 12 15 20 Line 3 10 12 15 20 30 line 4 12 15 20 30 45 Line 5 15 20 30 45 60

[0106] Based on the row and column number of the grid cell where the current data point is located, locate the corresponding entry in the insulation state mapping table. For example, if the location index is (row 5, column 4), the state score value obtained by looking up the table is 45. Read the state score value stored in the located entry. If the current data point falls at the boundary of two grid cells, it is necessary to calculate the distance weights from the current data point to the centers of the two adjacent grid cells separately, and then perform a weighted average of the two state score values ​​to obtain the final state score value. The formula is expressed as:

[0107]

[0108] in: It is the final status score. and These are the state score values ​​corresponding to two adjacent grid cells A and B, respectively. and It is a distance weight, calculated as follows: , ,here It is the Euclidean distance from the current data point to the center of grid cell A. This is the Euclidean distance from the current data point to the center of grid cell B. It can be understood that median filtering is applied to the state score values ​​over ten consecutive sampling times, and the filtered output is defined as the insulation condition deterioration assessment index. In practice, the edge computing gateway maintains a state score value buffer of length 10. Each time a new state score value is calculated, it is placed in the buffer, and the oldest value is removed. Then, the 10 values ​​in the buffer are sorted, and the average of the 5th and 6th values ​​after sorting is taken as the median filter output. This output value is the insulation condition deterioration assessment index at the current moment. In terms of data comparison, when the transformer's operating state is stable, the continuously calculated state score values ​​fluctuate little, and the median-filtered insulation condition deterioration assessment index is stable. When the insulation condition begins to change rapidly, the state score values ​​show a trend of increasing, causing the median-filtered insulation condition deterioration assessment index to also show an upward trend. Optionally, the grid division can also be dynamic and non-uniform. In some embodiments, a finer grid division can be used in areas with dense data points and a coarser division can be used in sparse areas, based on the distribution density of historical data points, but the total number of grid cells remains at 100.

[0109] See Figure 4In the quantitative characterization stage of insulation condition deterioration assessment for distribution transformers, a two-dimensional feature plane was constructed with the frequency domain attenuation coefficient as the abscissa and the real-time load current effective value as the ordinate. A 10×10 equal-area grid cell division and insulation condition mapping table scoring assignment were completed. Specifically, the abscissa was divided into 10 equal parts according to the frequency domain attenuation coefficient value range [0, 10], and the ordinate corresponding to the real-time load current effective value range [0, 1.5 times the rated current] was divided into 10 equal parts, forming 100 grid cells and assigning unique row and column indices. The figure shows the thermal distribution of the mapping table scoring in the 5×5 core area. The scoring is based on historical operating data, simulation analysis, and expert experience. The status score covers the integer range of 0-100, and the higher the value, the higher the risk level of insulation degradation. The grid cell score shows a significant gradient distribution. As the column number of the horizontal axis increases from row 1 to row 5, the score gradually increases from 5 to 15, 20, 30, 45, and 60. In the same column, as the row number increases, the score also shows an upward trend. Among them, row 5 and column 5 correspond to the highest score of 60, which intuitively represents the extreme insulation degradation risk under high load and high frequency domain attenuation coefficient. The color mapping of the heat map is highly matched with the score value. The low score (5-15) is presented with a light yellow / light white background, which corresponds to a stable operating state with extremely low risk. The medium score (20-45) is presented with orange-yellow and orange-red tones, which represents a medium degradation risk and requires continuous monitoring. The high score (60) is marked with a dark red color scheme, which corresponds to a warning state of severe insulation degradation. The visualization results intuitively present the insulation status classification under different combinations of load and frequency domain attenuation coefficient in the two-dimensional feature plane, providing a quantitative basis for subsequent querying of the mapping table to obtain the status score value and generating the insulation status degradation assessment index through median filtering. It also lays a visual representation foundation for dynamically adjusting the fault warning trigger threshold and generating graded alarm commands.

[0110] In one embodiment of the present invention, the edge computing gateway accesses its locally stored historical database to read all calculated and recorded insulation condition deterioration assessment indices from the past thirty natural days. Histogram statistics are performed on these historical insulation condition deterioration assessment indices to fit their probability density distribution, determining the peak position and distribution width of the probability density function. The peak position of the probability density function is added to three times the distribution width, and the sum is set as the initial warning threshold. Simultaneously, the rate of change of the current insulation condition deterioration assessment index relative to the previous moment is calculated, and it is determined whether the rate of change exceeds a preset slope threshold. If the current rate of change exceeds the preset slope threshold, a correction amount proportional to the magnitude of the rate of change is superimposed on the initial warning threshold, and the superimposed result is used as the final fault warning trigger threshold; if the current rate of change does not exceed the preset slope threshold, the initial warning threshold is directly used as the fault warning trigger threshold. In the real-time monitoring loop, the relationship between the insulation condition deterioration assessment index calculated at the current moment and the dynamically adjusted fault warning trigger threshold is compared. When the insulation condition deterioration assessment index exceeds the fault warning trigger threshold for the first time, a Level 1 warning message is generated, which includes the current value of the winding hot spot temperature rise gradient. When the insulation condition deterioration assessment index continues to exceed the fault warning trigger threshold for five minutes, and the core leakage flux density fluctuation value simultaneously exceeds its preset flux threshold, a Level 2 warning message is generated, which includes detailed spectral characteristic data of the core leakage flux density fluctuation value. When the insulation condition deterioration assessment index exceeds twice the fault warning trigger threshold value, and the bushing dielectric loss tangent increment simultaneously exceeds its preset dielectric loss threshold, a Level 3 trip warning message is generated. The edge computing gateway sends the generated Level 1, Level 2, or Level 3 trip warning messages to the remote master station system through its configured communication interface.

[0111] In practical implementation, the historical database stored locally on the edge computing gateway is accessed to retrieve all recorded insulation condition degradation assessment indices from the past thirty natural days. The historical database is stored in time-series format, with each insulation condition degradation assessment index bearing a precise timestamp. Histogram statistics are performed on the retrieved insulation condition degradation assessment indices to determine the peak position and distribution width of their probability density function. Specifically, the range of values ​​for the insulation condition degradation assessment indices is divided into 50 equal intervals. The frequency of each insulation condition degradation assessment index falling into each interval over the past thirty days is counted, forming a frequency histogram. The interval with the highest frequency in the histogram is found, and the median of that interval is used as the peak position of the probability density function. Approximation, distribution width This can be estimated by calculating the standard deviation of these insulation condition degradation assessment indices. The initial warning threshold is set by adding three times the distribution width to the peak position of the probability density function. The formula is expressed as: In the data comparison, for a transformer with extremely stable operating conditions, the historical data of its insulation condition deterioration assessment index is very concentrated and widely distributed. The calculated initial warning threshold is very small. Near peak position The early warning is relatively sensitive; however, for a transformer with large load fluctuations and inherently large fluctuations in insulation condition assessment values, its distribution width is... The calculated initial warning threshold is relatively large. The corresponding level will also be higher, avoiding too many false alarms caused by normal fluctuations.

[0112] In some embodiments, it is determined whether the rate of change of the current insulation condition deterioration assessment index exceeds a preset slope threshold. When calculating the rate of change, the insulation condition deterioration assessment index at the current moment is used. Compared with the insulation condition deterioration assessment index five minutes ago The difference is then divided by the time interval (5 minutes) to obtain the rate of change. Preset slope threshold It is a fixed value set based on experience, such as 0.5 exponential units per minute. If the rate of change... Exceeding the preset slope threshold Then at the initial warning threshold Based on this, a correction amount proportional to the rate of change is added to obtain the fault warning trigger threshold. The correction amount is calculated as follows: ,in If it is a proportional coefficient, then the fault warning trigger threshold is... If the rate of change The slope threshold was not exceeded. Then maintain the initial warning threshold. As a fault warning trigger threshold It is understandable that the warning is triggered by comparing the current insulation condition deterioration assessment index with the fault warning trigger threshold. When the insulation condition deterioration assessment index first exceeds the fault warning trigger threshold, a Level 1 warning message is generated. The Level 1 warning message includes the current value of the winding hotspot temperature rise gradient. In the example scenario, the insulation condition deterioration assessment index slowly rises from 45, reaching the fault warning trigger threshold of 48 at time t. At this point, a Level 1 warning message is immediately generated, and the message content clearly records the trigger time and the winding hotspot temperature rise gradient as 6.8 K / m. From a data comparison perspective, a Level 1 warning indicates that the insulation condition may have entered an early abnormal stage, but it is not yet accompanied by other obvious characteristics.

[0113] When the insulation condition deterioration assessment index continuously exceeds the fault warning trigger threshold for five minutes, and the core leakage flux density fluctuation value simultaneously exceeds the preset flux threshold, a level two warning message is generated. The preset flux threshold is a fixed value set based on the upper limit of the normal flux fluctuation range of the core. The level two warning message includes the spectral characteristics of the core leakage flux density fluctuation value, which may include detailed information such as the energy proportion of characteristic frequency bands and the main frequency. In specific implementation, the edge computing gateway has a timer that starts timing when the insulation condition deterioration assessment index first exceeds the fault warning trigger threshold and continuously monitors it for the next five minutes. If the insulation condition deterioration assessment index remains above the fault warning trigger threshold, and the core leakage flux density fluctuation value also exceeds the preset flux threshold at the five-minute mark, then a level two warning is triggered. When the insulation condition deterioration assessment index exceeds twice the fault warning trigger threshold, and the bushing dielectric loss tangent increment exceeds the preset dielectric loss threshold, a level 3 trip warning message is generated. The preset dielectric loss threshold is set based on the allowable limit of bushing dielectric loss. The level 3 trip warning message is the highest level alarm, indicating a potential risk of an emergency fault. Optionally, the triggering conditions for level 2 and level 3 warnings are based on an AND logic relationship, requiring simultaneous fulfillment of the insulation condition deterioration assessment index condition and the corresponding characteristic quantity (core leakage flux density fluctuation value or bushing dielectric loss tangent increment) condition. The generated level 1, level 2, or level 3 trip warning messages are sent to the remote master station via the communication interface of the edge computing gateway. The communication interface can be a wireless communication module (such as 4G / 5G) or a wired Ethernet interface. The messages are encapsulated and transmitted using the standard IEC61850 or 104 protocol to ensure that the remote master station can correctly receive and parse them.

[0114] See Figure 5In the edge computing-based real-time fault early warning analysis method for distribution transformers, the time series variation trend of the bushing dielectric loss angle tangent increment is one of the core characteristic quantities characterizing the aging state of transformer bushing insulation. Specifically, this time series curve is acquired by a digital dielectric loss bridge deployed in the high-voltage bushing end-screen circuit of the transformer. It is generated by measuring the phase difference offset between dielectric loss voltage and dielectric loss current, calculating the change in its tangent value, and adding a unified time scale. In the localized preprocessing stage of the edge computing gateway, temperature normalization is performed on this time series data to compensate for the influence of ambient temperature changes on dielectric loss measurement, and it is mapped to a standardized numerical range of 0-100, providing standardized feature input for subsequent insulation status assessment. In the frequency domain feature extraction stage, by analyzing the autocorrelation coefficient of the bushing dielectric loss angle tangent increment sequence, when the autocorrelation coefficient is less than a set threshold under a preset lag order, the initial attenuation factor extracted from the winding hot spot temperature rise gradient sequence is weighted and amplified, ultimately generating a frequency domain attenuation coefficient reflecting the insulation aging trend. In the hierarchical alarm logic, this characteristic quantity is one of the core triggering conditions for the Level 3 trip warning message: when the insulation condition deterioration assessment index exceeds twice the fault warning trigger threshold, and the bushing dielectric loss tangent increment exceeds the preset dielectric loss threshold, the system will generate the highest-level Level 3 trip warning message, which will be sent to the remote master station through the edge computing gateway to achieve real-time early warning and control of severe insulation faults in transformer bushings. As can be seen from the time-series curve in the figure, the bushing dielectric loss tangent increment exhibits non-stationary fluctuation characteristics, with multiple peak surges and declines. This fluctuation characteristic can be used to quantify the aging trend and deterioration rate of bushing insulation, providing data support for dynamically adjusting the fault warning trigger threshold and realizing hierarchical alarms.

[0115] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A real-time early warning and analysis method for distribution transformer faults based on edge computing, characterized in that, The method includes the following steps: The multi-dimensional electrical characteristics collected by each edge sensing node deployed on the main body of the distribution transformer during a continuous operating cycle are obtained. The multi-dimensional electrical characteristics include the temperature rise gradient of the winding hot spot, the core leakage flux density fluctuation value, and the bushing dielectric loss tangent increment. The multi-dimensional electrical characteristics are preprocessed locally by an edge computing gateway to generate a standardized feature data stream. Based on the time-series correlation of the standardized feature data stream, a frequency domain attenuation coefficient reflecting the insulation aging trend is extracted. By combining the frequency domain attenuation coefficient with the real-time effective value of the load current, the insulation condition deterioration assessment index is calculated. Based on the distribution pattern of the insulation condition deterioration assessment index in the historical database, the fault warning trigger threshold is dynamically adjusted. When the insulation condition deterioration assessment index monitored in real time exceeds the fault warning trigger threshold, a graded alarm command is generated and issued.

2. The real-time early warning and analysis method for distribution transformer faults based on edge computing as described in claim 1, characterized in that, The acquisition of multi-dimensional electrical feature quantities collected by each edge sensing node deployed on the distribution transformer body during a continuous operating cycle specifically includes: The fiber optic temperature sensor array integrated on the top of the transformer tank is controlled to scan the axial region of the winding at a preset spatial resolution to obtain temperature field distribution data at each temperature measurement point. Based on the temperature field distribution data and the rate of temperature difference change along the winding height direction, the temperature rise gradient of the winding hot spot is calculated. A magnetostrictive sensor attached to the transformer core clamp is driven to collect alternating magnetic flux signals, and Hilbert transform is performed on the alternating magnetic flux signals to obtain the leakage magnetic flux density fluctuation value of the core. By using a digital dielectric loss bridge coupled in the high-voltage bushing end screen circuit of the transformer, the phase difference offset between dielectric loss voltage and dielectric loss current is measured, and the change in the tangent value of the phase difference offset is defined as the tangent increment of the bushing dielectric loss angle. The synchronously collected winding hot spot temperature rise gradient, core leakage flux density fluctuation value, and bushing dielectric loss tangent increment are output after adding a unified time scale.

3. The real-time early warning and analysis method for distribution transformer faults based on edge computing as described in claim 2, characterized in that, The multi-dimensional electrical characteristics are preprocessed locally via an edge computing gateway to generate a standardized feature data stream, specifically including: The multi-dimensional electrical characteristic quantities are received and the temperature rise gradient of the winding hot spot is filtered by moving average to eliminate temperature jump interference caused by the start and stop of the cooling fan. Wavelet packet energy entropy analysis was performed on the leakage flux density fluctuation value of the iron core to extract the characteristic frequency band energy ratio that characterizes the loosening or short circuit of the iron core. Temperature normalization is performed on the incremental tangent of the bushing dielectric loss angle to compensate for the influence of ambient temperature changes on dielectric loss measurement. The temperature rise gradient of the winding hot spot, the fluctuation value of the core leakage flux density, and the increment of the bushing dielectric loss tangent after filtering, feature extraction and normalization are mapped to a numerical range of zero to one hundred. The mapped data is packaged and encapsulated according to a fixed time window to form the standardized feature data stream.

4. The real-time early warning and analysis method for distribution transformer faults based on edge computing as described in claim 3, characterized in that, Based on the time-series correlation of the standardized feature data stream, a frequency-domain attenuation coefficient reflecting the insulation aging trend is extracted, specifically including: Perform a fast Fourier transform on the winding hotspot temperature rise gradient sequence in the standardized feature data stream to obtain its spectral amplitude distribution; Identify the dominant frequency component with the largest amplitude in the spectral amplitude distribution, and record the amplitude intensity corresponding to the dominant frequency component; Calculate the amplitude attenuation slope of the dominant frequency component over multiple consecutive time windows, and define the reciprocal of the amplitude attenuation slope as the initial attenuation factor; Analyze the autocorrelation coefficient of the bushing dielectric loss tangent increment sequence in the standardized feature data stream. When the autocorrelation coefficient is less than a set threshold under a preset lag order, the initial attenuation factor is weighted and amplified. The weighted and amplified initial attenuation factor is used as the frequency domain attenuation coefficient.

5. The real-time early warning and analysis method for distribution transformer faults based on edge computing as described in claim 4, characterized in that, By combining the frequency domain attenuation coefficient with the real-time effective value of the load current, the insulation condition degradation assessment index is calculated, specifically including: The real-time three-phase current sampling value of the distribution transformer is retrieved from the substation monitoring system, and the root mean square value of the real-time three-phase current sampling value is calculated as the effective value of the real-time load current. Construct a two-dimensional feature plane with the frequency domain attenuation coefficient as the horizontal axis and the real-time load current effective value as the vertical axis; Divide the two-dimensional feature plane into several grid cells of equal area, and count the position index of the grid cell into which the data point at the current sampling time falls; Query the pre-stored insulation state mapping table and find the corresponding state score value according to the location index; The median filtering is applied to the state score values ​​at ten consecutive sampling times, and the filtered output is defined as the insulation state deterioration assessment index.

6. The real-time early warning and analysis method for distribution transformer faults based on edge computing as described in claim 5, characterized in that, Based on the distribution pattern of the insulation condition deterioration assessment index in the historical database, the fault early warning trigger threshold is dynamically adjusted, specifically including: Access the historical database stored locally on the edge computing gateway and read all recorded insulation condition degradation assessment indices from the past thirty natural days; Histogram statistics were performed on the read insulation condition deterioration assessment index to determine the peak position and distribution width of its probability density function; The sum obtained by adding three times the distribution width to the peak position of the probability density function is set as the initial warning threshold; Determine whether the rate of change of the current insulation condition deterioration assessment index exceeds a preset slope threshold; If the slope threshold is exceeded, a correction amount proportional to the rate of change is added to the initial warning threshold to obtain the fault warning trigger threshold; if the slope threshold is not exceeded, the initial warning threshold is maintained as the fault warning trigger threshold.

7. The real-time early warning and analysis method for distribution transformer faults based on edge computing as described in claim 6, characterized in that, When the insulation condition deterioration assessment index monitored in real time exceeds the fault warning trigger threshold, a graded alarm command is generated and issued, specifically including: Compare the current insulation condition deterioration assessment index with the fault warning trigger threshold. When the insulation condition deterioration assessment index exceeds the fault warning trigger threshold for the first time, a first-level early warning message is generated, which includes the current value of the winding hot spot temperature rise gradient. When the insulation condition deterioration assessment index continues to exceed the fault warning trigger threshold for five minutes, and the core leakage flux density fluctuation value simultaneously exceeds the preset flux threshold, a secondary warning message is generated. The secondary warning message contains the spectral characteristics of the core leakage flux density fluctuation value. When the insulation condition deterioration assessment index exceeds twice the fault warning trigger threshold, and the bushing dielectric loss tangent increment exceeds the preset dielectric loss threshold, a level three trip warning message is generated. The generated Level 1 warning message, Level 2 warning message, or Level 3 trip warning message are sent to the remote master station through the communication interface of the edge computing gateway.

8. The real-time early warning and analysis method for distribution transformer faults based on edge computing as described in claim 7, characterized in that, The process of mapping the temperature rise gradient of the winding hotspot, the fluctuation value of the core leakage flux density, and the increment of the bushing dielectric loss tangent after filtering, feature extraction, and normalization to a numerical range of zero to one hundred specifically includes: Find the maximum allowable temperature rise value of the winding hot spot temperature rise gradient under rated operating conditions, divide the currently collected winding hot spot temperature rise gradient by the maximum allowable temperature rise value and then multiply by one hundred to obtain the temperature rise mapping value. Find the peak value of magnetic flux density corresponding to the core leakage flux density fluctuation value at the core saturation critical point, divide the currently collected core leakage flux density fluctuation value by the peak value of magnetic flux density and then multiply by one hundred to obtain the magnetic flux mapping value. Find the range of difference between the tangential increment of the casing medium loss angle and the new casing and the severely aged casing. Subtract the current collected tangential increment of the casing medium loss angle from the dielectric loss value of the new casing, divide by the range of difference, and then multiply by one hundred to obtain the dielectric loss mapping value. The temperature rise mapping value, the magnetic flux mapping value, and the dielectric loss mapping value are respectively subjected to upper and lower limit clamping to ensure that the output value is not lower than zero and does not exceed one hundred.

9. The real-time early warning and analysis method for distribution transformer faults based on edge computing as described in claim 8, characterized in that, Analyze the autocorrelation coefficient of the bushing dielectric loss tangent increment sequence in the standardized feature data stream. When the autocorrelation coefficient is less than a set threshold at a preset lag order, the initial attenuation factor is weighted and amplified, specifically including: Extract the tangential increment of the bushing medium loss angle of continuous preset length sampling points from the standardized feature data stream to form the time series data to be analyzed; Set the lag order for autocorrelation analysis, calculate the autocorrelation coefficient of the time series data at the lag order, and the autocorrelation coefficient reflects the degree of linear correlation of the series at different time points; Obtain a pre-stored critical aging threshold, which is an empirical value obtained based on the statistics of historical failure samples. The autocorrelation coefficient is compared with the aging critical judgment threshold to determine the degree of random fluctuation of the bushing medium loss tangent increment sequence; When the autocorrelation coefficient is determined to be less than the aging critical threshold, it indicates that the dielectric loss shows a non-stationary accelerated aging trend. At this time, the preset amplification coefficient matrix is ​​called, and the corresponding weighted amplification coefficient is obtained by looking up the table according to the range of the initial attenuation factor. The initial attenuation factor is multiplied by the weighting amplification factor to complete the weighting amplification operation of the initial attenuation factor; When the autocorrelation coefficient is determined to be greater than or equal to the aging critical threshold, the initial decay factor is kept unchanged, and the process proceeds directly to the subsequent coefficient definition step.

10. The real-time early warning and analysis method for distribution transformer faults based on edge computing as described in claim 9, characterized in that, The process of querying the pre-stored insulation state mapping table and finding the corresponding state score value based on the location index specifically includes: The two-dimensional feature plane is divided into a grid of 100 equal parts in both the horizontal and vertical directions, and each grid cell is assigned a unique combination of row and column numbers as the position index; The insulation state mapping table is traversed. The insulation state mapping table stores the state score value corresponding to each grid cell. The state score value is in the range of integer zero to integer one hundred. Based on the row and column number of the grid cell where the current data point is located, locate the corresponding entry in the insulation state mapping table; Read the status score value stored in the located entry; If the current data point falls at the boundary of two grid cells, the distance weights from the data point to the center of the two adjacent grid cells are calculated respectively. The final state score value is obtained by weighted averaging of the two state score values.

Images