A power distribution transformer low-voltage aging fault detection method and system

Abstract

This invention relates to the field of electrical variable measurement technology, and particularly to a method and system for detecting low-voltage aging faults in distribution transformers. The method includes constructing a dynamic multidimensional dataset, constructing a one-dimensional time-series health curve, generating a dynamic aging assessment model, identifying latent fault risks, and generating early warning information. The system includes a data acquisition module, a time-series health calculation module, an assessment module, an identification module, and an early warning module. This invention constructs a time-series health curve by fusing electrical parameters, thermal parameters, and insulation characteristic parameters, and uses a variable-weight neural network to achieve adaptive aging assessment. This enables accurate identification of latent faults and prediction of remaining service life before the fault becomes significant, significantly reducing false alarm and false negative rates.

Description

Technical Field

[0001] This invention relates to the field of electrical variable measurement technology, and in particular to a method and system for detecting low-voltage aging faults in distribution transformers. Background Technology

[0002] Distribution transformers are key node equipment in power distribution networks, and their operating status directly affects the reliability and quality of power supply. With the in-depth advancement of smart distribution network construction, online monitoring and fault early warning of distribution transformers have become necessary technical means to ensure the safe and stable operation of the power system. Among them, monitoring and assessing the aging status of low-voltage side winding insulation is an important link in preventing transformer failures and avoiding unplanned power outages.

[0003] In the practical operation and monitoring of distribution transformers, the evaluation of low-voltage side insulation performance mainly relies on a combination of periodic preventive testing and real-time threshold alarms. Periodic preventive testing includes offline detection methods such as insulation resistance measurement and dielectric loss factor testing, while real-time monitoring typically involves configuring current transformers or voltage transformers on the low-voltage side of the transformer. This collects basic parameters such as operating current, load rate, and temperature rise, and compares them with preset thresholds. When the measured parameters exceed the threshold range, an alarm signal is triggered. This technical solution can provide a certain degree of alarm function when a transformer experiences a sudden fault.

[0004] In the implementation of existing technologies, low-voltage side insulation aging exhibits typical chronic degradation characteristics. Its parameter changes often show a slow and gradual trend, and the early stages of degradation do not necessarily lead to operating parameters immediately exceeding conventional monitoring thresholds. Because the current technology's understanding of the insulation aging process is limited to judging the exceedance of a single parameter and lacks the ability to analyze the correlation between the evolution of multi-dimensional operating parameters over time, the system exhibits a significant blind spot when facing the latent fault mode of slow insulation degradation. This technical limitation, which lacks the ability to predict latent faults and identify interference signals, results in the false alarm rate and false negative rate of existing low-voltage aging detection methods for distribution transformers remaining at a high level for a long time, making it difficult to meet the requirements of smart distribution networks for the accuracy of equipment status perception. Summary of the Invention

[0005] The inventors discovered through research that when insulation aging causes leakage current to gradually increase and dielectric loss factor to rise slowly, if the change in these parameters has not yet reached the threshold trigger condition, the system will not be able to identify the degradation trend of the equipment's health status; conversely, when an occasional transient disturbance causes a parameter to momentarily exceed the limit, the system will misjudge it as an insulation fault and generate a false alarm.

[0006] The purpose of this invention is to provide a method and system for detecting low-voltage aging faults in distribution transformers, so as to effectively solve the technical problems mentioned in the background art.

[0007] One aspect of the present invention provides a method for detecting low-voltage aging faults in distribution transformers, comprising: S1. By deploying a multi-sensor array on the low-voltage side of the distribution transformer, electrical parameters, thermal parameters, and insulation characteristic parameters are collected simultaneously to construct a dynamic multidimensional dataset; S2. Using the dynamic multidimensional dataset, extract the steady-state mean, variance, slope, and peak factor of each parameter, determine the comprehensive weight through principal component analysis, and construct a one-dimensional time-series health curve; S3. Based on the one-dimensional time-series health curve, a variable weight neural network structure is adopted to dynamically adjust the weight coefficients according to recent operating data to generate a dynamic aging assessment model; S4. Based on the dynamic aging assessment model, calculate the second derivative of the health curve, capture the accelerating trend of aging rate, and identify hidden fault risks. S5. Input the current health status parameters into the prediction model, simulate the aging process within a preset period in the future, and generate early warning information.

[0008] In some embodiments, in S1, the electrical parameters include operating current and terminal voltage, the thermal parameters include winding temperature and ambient temperature, and the insulation characteristic parameters include insulation dielectric loss factor and insulation leakage current.

[0009] In some embodiments, in step S1, the steady-state parameters and transient signals are acquired by capturing transient impact signals.

[0010] In some embodiments, in step S2, an adaptive weight allocation algorithm is used to dynamically adjust the weight coefficients based on the historical contribution rate and current rate of change of each parameter.

[0011] In some embodiments, in step S2, the correlation between parameters is analyzed to determine the dynamic evolution logic of the coupling effect between parameters during the aging process.

[0012] In some embodiments, in S3, the variable weight neural network integrates an anomaly detection submodule to identify and filter out occasional interference signals.

[0013] In some embodiments, in step S4, a confidence assessment mechanism based on historical statistical algorithms is introduced to give the probability of occurrence of latent faults by comparing the similarity between current features and historical fault cases.

[0014] In some embodiments, in step S5, a differential autoregressive moving average model combined with an exponential smoothing method is used to quantitatively predict the health decline trajectory and the time to reach the critical point.

[0015] In some embodiments, in step S5, multi-level early warning information is generated based on the prediction results, and the early warning levels include attention level, warning level and danger level.

[0016] Another aspect of the present invention provides a low-voltage aging fault detection system for distribution transformers, and implements a method for detecting low-voltage aging faults in distribution transformers, comprising: a data acquisition module, a time-series health calculation module, an evaluation module, an identification module, and an early warning module; The acquisition module is configured to construct a multi-parameter acquisition matrix on the low-voltage side of the distribution transformer, including a current sensor, a voltage sensor, a temperature sensor, a dielectric loss factor sensor, and a leakage current sensor. The time-series health calculation module is electrically connected to the acquisition module, and integrates a feature extraction unit and a weight allocation unit to generate a time-series health curve. The assessment module is connected to the time-series health calculation module and uses a variable weight neural network algorithm to perform adaptive aging status assessment. The identification module integrates curvature analysis algorithm and confidence assessment algorithm to identify early fault characteristic signals; The early warning module is logically linked to each functional module. Based on health data and assessment results, it executes a prediction algorithm and generates early warning information.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. By integrating and collecting multi-dimensional operating parameters, the traditional detection method's reliance on a single parameter is completely broken. Through the comprehensive collection and analysis of electrical parameters, thermal parameters, and insulation characteristic parameters, a comprehensive data profile reflecting the insulation status of the transformer's low-voltage side can be constructed, solving the problem of inaccurate evaluation caused by the single parameter dimension in existing technologies. 2. By constructing a time-series health index, the detection dimensions are extended in a time-series manner. By mapping multi-dimensional parameters to a unified time-series health index, the evolution of insulation status can be tracked in real time, and the gradual characteristics of the aging process can be accurately depicted. Compared with discrete single-point detection, this time-series health index can detect early signals of insulation degradation earlier and effectively avoid false alarms caused by misjudgment due to single-point over-limit. 3. The evaluation method is upgraded by using a dynamic aging assessment model. By introducing a variable weight neural network structure, the parameter weights can be adaptively adjusted according to recent operating data, so as to realize dynamic tracking of aging trends, avoid misjudgment caused by interference signals, and significantly reduce false alarm rate and false negative rate. 4. By accurately identifying and predicting latent faults, we can shift from post-event alarms to pre-event prevention. By analyzing subtle changes in the curvature of the health index and predicting future aging trajectories, we can issue early warnings before insulation faults become significant, providing maintenance personnel with sufficient response time, fundamentally reducing the probability of unplanned power outages, and ensuring stable power supply from the power grid. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the low-voltage aging fault detection method for distribution transformers according to the present invention; Figure 2 This is a diagram of the low-voltage aging fault detection system for distribution transformers according to the present invention. Detailed Implementation

[0020] The following will be based on embodiments of the present invention. Figures 1-2 The technical solutions in the embodiments of the present invention will be clearly and completely described together. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0021] Example 1 This embodiment is applied to the online monitoring scenario of the low-voltage side insulation status of a 10kV distribution transformer in a power distribution network. In this scenario, the distribution transformer has a capacity of 400kVA, a rated voltage of 400V, and adopts a three-phase four-wire power supply method. The low-voltage side outgoing line is connected to the distribution cabinet of a residential community or commercial complex.

[0022] The low-voltage aging fault detection system for distribution transformers deployed in this scenario includes a data acquisition module, a data preprocessing module, a feature extraction module, a time-series health calculation module, and an evaluation module. The modules interact with each other via a data bus.

[0023] The data acquisition module is located in the low-voltage side outgoing line cabinet of the distribution transformer and includes at least a current acquisition unit, a voltage acquisition unit, a temperature acquisition unit, and a leakage current acquisition unit. The current acquisition unit uses a combination of a current transformer and a high-precision analog front-end chip. The primary side of the current transformer is connected in series in the outgoing line circuits of phases A, B, C, and N on the low-voltage side. The secondary side output signal is converted to I / V and conditioned by the analog front-end chip before being sent to an analog-to-digital converter for digital sampling. The sampling frequency is set to 80 points per cycle, i.e., 4kHz for a 50Hz power frequency signal. The voltage acquisition unit also uses a voltage transformer to acquire the voltage signals of phases A, B, C, and N. The primary side of the voltage transformer is connected in parallel to the low-voltage side bus. The secondary side signal is amplitude-adjusted by a voltage divider resistor network before being sent to the analog-to-digital converter shared with the current acquisition unit for synchronous sampling. The temperature acquisition unit uses a platinum resistance thermometer (RTD) sensor. The sensor probe is directly attached to the surface of the insulation material of the low-voltage winding of the transformer. The winding temperature is measured using a four-wire connection method, with an accuracy controlled within ±0.5℃. The leakage current acquisition unit uses a high-sensitivity micro-current sensor, positioned between the neutral point and the grounding wire on the low-voltage side of the transformer. It acquires the weak leakage current signal generated by insulation degradation, with a measurement range of 0.1mA to 100mA and a resolution of 0.01mA. The acquisition module synchronously acquires the above four parameters in each sampling cycle. The sampling cycle is set to 1 second, meaning that one set of raw data vectors containing three-phase current, three-phase voltage, winding temperature, and leakage current is generated every second.

[0024] The data preprocessing module is deployed in an intelligent monitoring terminal near the distribution transformer. This terminal uses an industrial-grade embedded processor with a main frequency of 800MHz, equipped with 512MB of RAM and 4GB of storage. After receiving the raw sampled data from the acquisition module, the data preprocessing module sequentially performs digital filtering and outlier removal. The digital filtering process uses an adaptive notch filter bank, which includes a 50Hz power frequency notch filter and a 2Hz low-pass filter. The power frequency notch filter is used to eliminate interference from the power frequency and its integer multiples of harmonics, while the low-pass filter is used to suppress high-frequency noise. The center frequency and Q value of the adaptive notch filter can be automatically adjusted according to the actual power grid frequency deviation, with an adjustment range of 49.5Hz to 50.5Hz. The outlier removal process uses the 3σ criterion based on statistical principles. That is, the mean and standard deviation of each parameter sequence are calculated, and data points exceeding the mean ± 3 times the standard deviation are identified as outliers and replaced with linear interpolation of adjacent data points.

[0025] The feature extraction module also runs within the intelligent monitoring terminal, extracting features from the preprocessed data. The extracted features fall into three main categories: time-domain statistical features, frequency-domain features, and time-series features. Time-domain statistical features include six statistics for each parameter type: mean, standard deviation, peak value, peak-to-peak value, skewness, and kurtosis. These statistics reflect the overall level and fluctuation of the parameters. Frequency-domain features utilize a Fast Fourier Transform algorithm to perform spectral analysis on current and voltage signals, extracting the amplitude and phase information of each harmonic, with a focus on the content of the 3rd, 5th, 7th, 9th, and 11th harmonics, as the presence of these higher harmonics may be related to partial discharge caused by insulation degradation. Time-series features include the rate of change between adjacent sampling points, the trend within the sliding window, and periodic fluctuation characteristics. The rate of change is obtained through the difference operation of adjacent sampling values, the trend is characterized by the slope of a linear regression fit, and the periodic fluctuation characteristics are determined by the position of the first valley of the autocorrelation function. The feature extraction module performs feature calculations in 1-minute cycles, which means that the 60 sets of sampled data collected within 60 seconds are summarized into an analysis window, and each window outputs a feature vector containing 54 feature values.

[0026] The time-series health assessment module is deployed in a cloud-based data processing platform. This platform adopts a distributed architecture, comprising two main parts: a data storage cluster and a computing cluster. The data storage cluster uses a time-series database to store historical feature data and health assessment results, with a data retention period of 3 years. The computing cluster is deployed in a containerized manner, including multiple parallel computing nodes, each configured with 8 CPU cores and 32GB of memory. After receiving the feature vectors uploaded by the feature extraction module, the time-series health assessment module first establishes a time-series health assessment model for each feature dimension.

[0027] Specifically, the time-series health calculation module maintains a one-dimensional time-series array of length 1008 for each of the 54 feature dimensions. Each array records the normalized health value of that feature over the past 1008 minutes (approximately 7 days). The normalized health is calculated using a membership function based on a Gaussian distribution, and the formula is as follows: Health = exp(-(|current value - reference value|) 2 ) / (2×threshold) 2 )) The reference value is the weighted moving average of the feature over the past 7 days, with recent data having a higher weight than older data. The weight decay factor is set to 0.95, and the threshold is twice the standard deviation of the feature.

[0028] After the normalized health scores for each feature dimension are calculated, the time-series health score calculation module further employs a weighted fusion algorithm to combine the 54 one-dimensional health scores into a single comprehensive health score. The fusion algorithm uses a combined weighting method that integrates the analytic hierarchy process (AHP) and entropy weighting. The AHP determines the basic weights for each feature category based on expert experience, the time-series health score calculation module dynamically adjusts the weight coefficients based on the contribution of each feature dimension to historical failure cases, and the entropy weighting method objectively assigns weights based on the degree of variation of each feature dimension. The final output comprehensive health score is a value between 0 and 100, where a health score of 100 indicates that the equipment is in perfect health, and a health score of 0 indicates that the equipment has experienced a serious failure.

[0029] The evaluation module is also deployed in the computing cluster of the cloud data processing platform. It receives the comprehensive health time series output by the time series health construction module and performs aging trend analysis and fault prediction based on this series. The core algorithm of the evaluation module adopts a combination model of variational autoencoder and long short-term memory network. The variational autoencoder is used to learn the low-dimensional manifold representation of normal operation from the high-dimensional feature space, and the long short-term memory network is used to model the temporal dependencies of the health time series.

[0030] In its specific operation, the evaluation module first reconstructs the current feature vector using a variational autoencoder, calculating the reconstruction error as an anomaly score. Then, it inputs the comprehensive health sequence of the past 24 hours into a long short-term memory network, which outputs predicted health values ​​for the next 1, 4, 12, and 24 hours. Finally, based on the predicted health trend and the current anomaly score, the comprehensive evaluation module calculates the equipment's aging acceleration index and failure probability. When the aging acceleration index exceeds a preset first threshold, the system outputs a warning signal; when the failure probability exceeds a preset second threshold, the system outputs an alarm signal.

[0031] Furthermore, the workflow of the distribution transformer low-voltage aging fault detection system deployed in this scenario includes, in chronological order, data synchronous acquisition, data preprocessing, feature batch calculation, time-series health update, and dynamic aging assessment.

[0032] During the data synchronization acquisition phase, the current acquisition unit, voltage acquisition unit, temperature acquisition unit, and leakage current acquisition unit sample synchronously according to a unified clock reference. The analog signals output from the current transformer and voltage transformer are conditioned by the analog front-end chip and then digitized at a sampling rate of 4kHz. The analog-to-digital converter (ADC) adopts a successive approximation architecture with a 16-bit resolution, covering a wide measurement range from 0.1% to 200% of the rated current. The thermistor sensor in the temperature acquisition unit measures the temperature every 100ms, and the measurement result is transmitted to the intelligent monitoring terminal after analog-to-digital conversion. The micro-current sensor in the leakage current acquisition unit uses an integrating measurement principle with a measurement period of 1 second, effectively suppressing white noise interference. At the end of each 1-second cycle, the acquisition module packages the sampling data from the nine channels into a single JSON message and sends it to the intelligent monitoring terminal via a 4G wireless network.

[0033] In the data preprocessing stage, after receiving the raw sampled data from the acquisition module, the data preprocessing module of the intelligent monitoring terminal first performs format parsing and integrity verification on the data. The format parsing process deserializes the JSON message into a structure containing timestamps and sampled values. The integrity verification process checks the frame header, frame trailer, and checksum of the data frame to ensure that no errors occur during data transmission. Digital filtering is implemented using a pipelined filtering algorithm. The data preprocessing module has two built-in filter chains for processing current / voltage signals and temperature / leakage current signals, respectively. The current / voltage signals are first passed through a 50Hz notch filter to eliminate power frequency interference. The notch filter has a bandwidth of 2Hz and a notch depth greater than 40dB. Then, high-frequency noise is eliminated by a low-pass filter with a cutoff frequency of 100Hz. The low-pass filter is designed as a fourth-order Butterworth filter. The temperature and leakage current signals are smoothed by a low-pass filter with a cutoff frequency of 1Hz. Outlier removal is performed after filtering. The data preprocessing module maintains a rolling queue of length 10 for each parameter. When a new data point exceeds three times the standard deviation of the queue mean, it is identified as an outlier and replaced with a valid value from the end of the queue. Normalization is performed after outlier removal. The data preprocessing module maintains a dynamically updated min-max calibration table for each parameter. The calibration table is updated every 24 hours to ensure that the calibration values ​​can be adaptively adjusted to follow seasonal load changes.

[0034] During the feature batch calculation stage, the feature extraction module of the intelligent monitoring terminal triggers the feature calculation task at a 1-minute interval. The feature calculation task first reads 60 sets of sampled data accumulated in the past 60 seconds from the local cache, and then calculates the time domain statistical features for each set of data.

[0035] Specifically, the mean calculation uses a recursive algorithm to avoid repeatedly traversing the data; the standard deviation calculation uses the Welford online algorithm to achieve numerical stability; and the skewness and kurtosis calculations use a grouping and aggregation method to reduce computational complexity. Frequency domain feature calculation uses the Cooley-Tukey Fast Fourier Transform algorithm, with a 1024-point sample sequence as input, and zero-padding used for any missing points. Time-series feature calculation uses a sliding window technique with a window length of 10 sample points; the rate of change is obtained by the difference between adjacent points, and the trend is obtained by linear fitting using the least squares method. The feature extraction module encapsulates the calculated 54-dimensional feature vector into a message and sends it to the cloud data processing platform via a message queue, with a message sending frequency of one per minute.

[0036] During the time-series health update phase, the time-series health calculation module of the cloud data processing platform receives the feature vector sent by the feature extraction module and triggers the health calculation pipeline.

[0037] The first processing node in the health calculation pipeline is the feature grouping node. This node divides the 54-dimensional feature vector into three groups according to electrical quantity features, temperature features, and insulation features. The electrical quantity features include 6 current / voltage related features, the insulation features include 3 leakage current related features, and the remaining 45 features are assigned to the temperature feature group.

[0038] The second processing node is the reference value update node. This node maintains an exponentially weighted moving average as a reference value for each feature. The update formula for the mean is: Reference value = α × current value + (1-α) × historical reference value Where α is the learning rate, set to 0.05.

[0039] The third processing node is the health calculation node, which calculates the normalized health score for each feature using the aforementioned Gaussian membership function.

[0040] The fourth processing node is the weight fusion node, which merges 54 one-dimensional health scores into a comprehensive health score using a combined weighting algorithm. The combined weights are calculated at the beginning of each calculation cycle. First, the baseline weight ratio of the electrical quantity feature group, temperature feature group, and insulation feature group is determined to be 3:2:5 using the analytic hierarchy process (AHP). Then, the weights within each group are adjusted using the entropy weighting method based on the coefficient of variation of each feature over the past 24 hours. The final comprehensive health score, together with the historical health scores previously output by the time-series health score construction module, forms a comprehensive health score time series of length 1008, which is transmitted to the evaluation module via a message queue.

[0041] In the dynamic aging assessment phase, the assessment module of the cloud data processing platform receives the comprehensive health time series output by the time-series health calculation module and sequentially performs anomaly detection, trend prediction, and risk assessment. In anomaly detection, the assessment module inputs the current 54-dimensional feature vector into a pre-trained variational autoencoder model. The reconstruction error of the variational autoencoder model is calculated using the root mean square error; a larger reconstruction error indicates that the current device operating state deviates more from the normal mode. In trend prediction, the assessment module inputs the comprehensive health sequence of the past 24 hours into a long short-term memory network model. The model is a four-layer LSTM structure, with each layer containing 128 hidden units. The model's output is the predicted health values ​​for the next four time points. In risk assessment, the assessment module calculates anomaly scores based on the anomaly detection results and an aging acceleration index based on the trend prediction results. The aging acceleration index is calculated as follows: The rate of change of the predicted health status over the next 12 hours relative to the current health status is calculated. If the rate of change is negative and the absolute value is greater than 0.5 / hour, it is considered an accelerated aging state, and the aging acceleration index is the ratio of the absolute value of the rate of change to 0.5. If the rate of change is positive or the absolute value is less than 0.5, it is considered a stable state, and the aging acceleration index is 0. The failure probability is calculated using a Bayesian inference method, comprehensively considering the anomaly score, the aging acceleration index, and historical failure prior probabilities, outputting a failure probability value between 0 and 1. When the aging acceleration index is greater than 1.2, the evaluation module generates an early warning message, which includes the current overall health status value, anomaly score, aging acceleration index, and the predicted health status value for the next 12 hours, and is pushed to maintenance personnel via SMS and mobile APP. When the failure probability is greater than 0.7, the evaluation module generates an alarm message, which includes all the content of the early warning message, as well as a failure type prediction and handling suggestions, and is simultaneously pushed to maintenance personnel and the dispatch center via SMS, telephone, and mobile APP.

[0042] Example 2 Based on Example 1, this example further illustrates the adaptive adjustment scheme of the present invention under extreme operating environments. This example is applied to the scenario of monitoring the insulation status of the low-voltage side of an outdoor pole-mounted distribution transformer. In this scenario, the distribution transformer has a capacity of 200kVA, a rated voltage of 400V, and is installed on a pole in a coastal area with severe salt spray corrosion. The ambient temperature range is -20℃ to +50℃, and the humidity range is 30% to 95%.

[0043] The system deployed in this scenario features the following targeted hardware design: The sensor housing of the data acquisition module is made of 316L stainless steel, achieving an IP67 protection rating. The connection between the sensor and the low-voltage side lead of the transformer is double-sealed with heat-shrink tubing and silicone rubber sealing rings for waterproofing. The intelligent monitoring terminal adopts a wide-temperature industrial-grade design, with key components operating within a temperature range of -40℃ to +85℃, and includes a built-in temperature compensation circuit to ensure the sampling accuracy of the analog-to-digital converter under extreme temperature conditions. The communication module uses an industrial-grade 4G DTU, supporting BeiDou positioning and timing functions to ensure reliable data transmission in remote areas and environments with poor communication signals.

[0044] In this scenario, the system undergoes the following adaptive adjustments at the software level: The digital filtering algorithm in the data preprocessing module incorporates adaptive temperature compensation, dynamically adjusting filter parameters based on data collected from an ambient temperature sensor to ensure consistent filtering performance under varying temperature conditions. The feature extraction module adds a salt spray corrosion feature recognition algorithm, analyzing the high-frequency components of leakage current signals to detect the degree of salt spray deposition on the insulation surface; a sustained increase in these high-frequency components indicates intensified corrosion. The time-series health calculation module introduces an environmental correction factor into its health calculation model. This factor dynamically adjusts based on the impact of temperature and humidity on the insulation aging rate, ensuring the health assessment results are not subject to spurious interference from environmental factors. The prediction model in the evaluation module employs transfer learning technology, using the model parameters trained in the scenario of Example 1 as initial parameters and fine-tuning them with historical data from the current scenario, significantly shortening the model training cycle and improving prediction accuracy.

[0045] Example 3 Based on Example 1, this invention further describes the parameter optimization configuration scheme under various transformer operating conditions. This example is applied to the monitoring scenario of distribution transformer groups in an urban complex. In this scenario, a total of 10 distribution transformers are managed, with the capacity of each transformer ranging from 315kVA to 1250kVA. The low-voltage side is connected to a variety of loads such as central air conditioning, elevators, lighting and power equipment, and the load characteristics are complex and fluctuate drastically.

[0046] This scenario employs a centralized architecture design, deploying a data acquisition master station and an edge computing gateway in the power distribution room of the complex. The acquisition modules from 10 transformers are wired to the data acquisition master station. Data exchange between the master station and the edge computing gateway is conducted via Gigabit Ethernet, and the edge computing gateway uploads data to the cloud data processing platform via a dedicated fiber optic line. The data acquisition master station's hardware architecture uses a PXIe bus structure, equipped with 8 synchronous sampling channels at a sampling rate of 200kHz per channel, capable of simultaneously acquiring 90 electrical quantity channels from 10 transformers. The edge computing gateway uses a high-performance multi-core processor, configured with a 16-core CPU and 64GB of memory, enabling it to complete all feature extraction calculations locally, effectively reducing the computational load and network bandwidth consumption of the cloud platform.

[0047] In this scenario, the following optimizations are made to the multi-condition recognition: The feature extraction module adds a working condition identification submodule. This submodule automatically divides historical operating data into several working condition categories based on an unsupervised clustering algorithm. The clustering algorithm uses the DBSCAN density clustering method, eliminating the need to pre-set the number of categories. During system operation, the working condition identification submodule continuously analyzes the distance between the current feature vector and each working condition category. When the distance exceeds a threshold, a new working condition category is automatically created. The feature extraction module dynamically adjusts the feature weights based on the identified working condition categories. For steady-state working conditions, it focuses on changes in time-domain statistical characteristics; for impact load working conditions, it focuses on instantaneous peak values ​​and rate of change characteristics; and for periodic load working conditions, it focuses on frequency domain characteristics and periodic fluctuation characteristics.

[0048] The following optimizations were made to the health assessment in this scenario: The time-series health calculation module adopts a multi-condition adaptive health model, which establishes an independent health assessment baseline for each type of operating condition.

[0049] Specifically, based on the operational data from the past 30 days, five typical operating conditions are first identified: light-load steady-state condition, full-load steady-state condition, air conditioning start-stop impact condition, elevator periodic load condition, and power equipment start-up condition. Each operating condition corresponds to an independent health assessment model. The structure of the model is the same as in Example 1, but the reference values ​​and threshold parameters are independently calibrated based on the historical data of that operating condition. During operation, the time-series health calculation module first determines the current operating condition category through the operating condition identification submodule, then calls the corresponding operating condition's health assessment model to calculate the one-dimensional health score, and finally merges the one-dimensional health scores of multiple operating conditions to obtain the comprehensive health score. Through this multi-operating-condition adaptive mechanism, the difference between parameter fluctuations caused by load changes and parameter anomalies caused by insulation aging can be accurately identified, effectively reducing the false alarm rate.

[0050] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0051] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for detecting low-voltage aging faults in distribution transformers, characterized in that, include: S1. By deploying a multi-sensor array on the low-voltage side of the distribution transformer, electrical parameters, thermal parameters, and insulation characteristic parameters are collected simultaneously to construct a dynamic multidimensional dataset; S2. Using the dynamic multidimensional dataset, extract the steady-state mean, variance, slope, and peak factor of each parameter, determine the comprehensive weight through principal component analysis, and construct a one-dimensional time-series health curve; S3. Based on the one-dimensional time-series health curve, a variable weight neural network structure is adopted to dynamically adjust the weight coefficients according to recent operating data to generate a dynamic aging assessment model; S4. Based on the dynamic aging assessment model, calculate the second derivative of the health curve, capture the accelerating trend of aging rate, and identify hidden fault risks. S5. Input the current health status parameters into the prediction model, simulate the aging process within a preset period in the future, and generate early warning information.

2. The method according to claim 1, characterized in that, In S1, the electrical parameters include operating current and terminal voltage, the thermal parameters include winding temperature and ambient temperature, and the insulation characteristic parameters include insulation dielectric loss factor and insulation leakage current.

3. The method according to claim 1, characterized in that, In S1, by capturing transient impact signals, dual acquisition of steady-state parameters and transient signals is achieved.

4. The method according to claim 1, characterized in that, In S2, an adaptive weight allocation algorithm is used to dynamically adjust the weight coefficients based on the historical contribution rate and current rate of change of each parameter.

5. The method according to claim 1, characterized in that, In S2, the correlation between various parameters is analyzed, and the dynamic evolution logic of the coupling effect between parameters in the aging process is determined.

6. The method according to claim 1, characterized in that, In S3, the variable weight neural network integrates an anomaly detection submodule to identify and filter out occasional interference signals.

7. The method according to claim 1, characterized in that, In step S4, a confidence assessment mechanism based on historical statistical algorithms is introduced, which gives the probability of occurrence of latent faults by comparing the similarity between current features and historical fault cases.

8. The method according to claim 1, characterized in that, In S5, a differential autoregressive moving average model combined with exponential smoothing is used to quantitatively predict the health decline trajectory and the time to reach the critical point.

9. The method according to claim 1, characterized in that, In step S5, multi-level early warning information is generated based on the prediction results. The early warning levels include attention level, warning level, and danger level.

10. A low-voltage aging fault detection system for distribution transformers, implementing the method described in any one of claims 1 to 9, characterized in that, include: The system includes a data acquisition module, a time-series health status calculation module, an assessment module, an identification module, and an early warning module. The acquisition module is configured to construct a multi-parameter acquisition matrix on the low-voltage side of the distribution transformer, including a current sensor, a voltage sensor, a temperature sensor, a dielectric loss factor sensor, and a leakage current sensor. The time-series health calculation module is electrically connected to the acquisition module, and integrates a feature extraction unit and a weight allocation unit to generate a time-series health curve. The assessment module is connected to the time-series health calculation module and uses a variable weight neural network algorithm to perform adaptive aging status assessment. The identification module integrates curvature analysis algorithm and confidence assessment algorithm to identify early fault characteristic signals; The early warning module is logically linked to each functional module. Based on health data and assessment results, it executes a prediction algorithm and generates early warning information.

Images