Power distribution transformer health state evaluation method and system fusing multi-source data

Abstract

The application relates to the technical field of power equipment monitoring, and discloses a power distribution transformer health state evaluation method and system fusing multi-source data. The method comprises the following steps: obtaining a multi-source monitoring data set containing a time stamp from a monitoring device; pre-processing the data set to remove noise; performing time correction processing on the cleaned data to unify the time reference; performing frequency domain transformation on the time-aligned data, extracting peak frequency and amplitude distribution to form a feature set; preliminarily classifying fault mechanisms according to the feature set; fusing the classification results with the time sequence trend and waveform features in the cleaned data to generate a comprehensive fault feature vector; and performing matching analysis on the comprehensive fault feature vector by using a pre-trained classifier to output a health state evaluation result. The application solves the problems of multi-source data fusion difficulty, inaccurate feature representation and unreliable critical state recognition, and improves the accuracy and early warning timeliness of state evaluation.

Description

Technical Field

[0001] This application relates to the field of power equipment monitoring technology, and in particular to a method and system for assessing the health status of distribution transformers by integrating multi-source data. Background Technology

[0002] Distribution transformers are key equipment in the power grid distribution chain, and their health status directly affects the regional power supply quality and electricity safety. With the advancement of smart grid construction and the digital transformation of distribution networks, real-time and accurate condition assessment of distribution transformers has become an important requirement for achieving predictive maintenance and lean asset management. Currently, online monitoring technology can simultaneously collect multi-source heterogeneous data such as dissolved gas concentration in the transformer oil, partial discharge pulse signals, and winding vibration response, providing a data foundation for comprehensively assessing their health status.

[0003] However, existing methods for health status assessment by fusing multi-source data have significant shortcomings, failing to meet the accuracy and timeliness requirements of distribution networks for condition assessment. First, due to differences in the physical mechanisms and sensing paths of different signal sources, there are significant spatiotemporal inconsistencies among multi-source data. For example, partial discharge pulse sequences reflect instantaneous insulation defect events, while cumulative changes in oil gas concentration typically lag by several hours, and winding vibration responses are simultaneously affected by electromagnetic forces and mechanical states, with propagation delays. Existing methods often ignore these differences, directly extracting and fusing features from the original time-series data. This results in the same health status deterioration event being represented fragmented across different signal dimensions, disrupting the correlation between physical quantities, and making it difficult for assessment models to capture early, weak, complex fault symptoms. Second, at the health status feature engineering level, existing methods typically employ a "feature stacking" strategy, that is, independently extracting time-domain, frequency-domain, or time-frequency-domain features from each signal and then simply splicing them together, failing to filter and weight based on the contribution of different features to the health status representation. This results in high-dimensionality and redundancy in the constructed health feature vectors, which contain a large number of invalid or even misleading features introduced by environmental noise, load fluctuations, and electromagnetic interference. This not only increases model complexity and computational burden but also reduces the robustness and generalization ability of health status classification. Finally, in the health status assessment decision-making stage, most methods rely on static classification models (such as support vector machines and neural networks) trained based on historical fault databases. These models are insufficient in recognizing the dynamic evolution process of health status from quantitative accumulation to sudden qualitative change. In particular, they are prone to low-confidence misjudgments for samples in "sub-healthy" or "critical" states, and lack closed-loop verification and adaptive optimization mechanisms for low-confidence results. This makes it difficult to guarantee the reliability of the assessment conclusions and fails to effectively support accurate decision-making for distribution network condition-based maintenance.

[0004] To address the above deficiencies, this application combines multi-source signal time synchronization correction, contribution-oriented feature fusion optimization, and confidence-driven dynamic evaluation and verification mechanisms to solve the problems of difficult multi-source heterogeneous data fusion in distribution transformers, inaccurate health status feature representation, and unreliable critical state identification, thereby improving the accuracy, sensitivity, and engineering practicality of health status assessment. Summary of the Invention

[0005] This application provides a method and system for assessing the health status of distribution transformers by integrating multi-source data. It solves the problems of difficulty in integrating multi-source heterogeneous data of distribution transformers, inaccurate characterization of health status features, and unreliable identification of critical states, thereby improving the accuracy, sensitivity, and engineering applicability of health status assessment.

[0006] Firstly, this application provides a method for assessing the health status of distribution transformers by integrating multi-source data, the method comprising:

[0007] Step S1: Obtain a multi-source monitoring dataset containing timestamps from the distribution transformer monitoring equipment. The multi-source monitoring dataset includes oil gas concentration data, partial discharge pulse sequence, and winding vibration response sequence.

[0008] Step S2: Preprocess the multi-source signal dataset to remove noise and interference, and obtain a cleaned signal dataset;

[0009] Step S3: Perform time correction processing on the cleaned signal dataset to obtain a time-aligned unified signal dataset;

[0010] Step S4: Perform frequency domain transformation on the time-aligned unified signal dataset, extract the peak frequency and amplitude distribution of each signal sequence, and form a multi-source frequency domain feature set;

[0011] Step S5: Perform preliminary fault mechanism classification based on the multi-source frequency domain feature set to obtain classification results;

[0012] Step S6: Combine the classification results with the time-series trend vector of the oil gas concentration data in the cleaned signal dataset and the time-domain waveform features of the winding vibration response sequence to generate a comprehensive fault feature vector;

[0013] Step S7: Perform matching analysis on the comprehensive fault feature vector using a pre-trained classifier to output the health status assessment result of the distribution transformer.

[0014] Secondly, this application provides a distribution transformer health status assessment system that integrates multi-source data, used to implement the distribution transformer health status assessment method that integrates multi-source data, the system comprising:

[0015] The data acquisition module is used to acquire a multi-source monitoring dataset containing timestamps from the distribution transformer monitoring equipment. The multi-source monitoring dataset includes oil gas concentration data, partial discharge pulse sequences, and winding vibration response sequences.

[0016] The data cleaning module is used to preprocess the multi-source signal dataset to remove noise and interference, and obtain a cleaned signal dataset.

[0017] The time correction module is used to perform time correction processing on the cleaned signal dataset to obtain a time-aligned unified signal dataset.

[0018] The feature extraction module is used to perform frequency domain transformation on the time-aligned unified signal dataset, extract the peak frequency and amplitude distribution of each signal sequence, and form a multi-source frequency domain feature set.

[0019] The fault classification module is used to perform preliminary fault mechanism classification based on the multi-source frequency domain feature set and obtain the classification result.

[0020] The feature module integrates the classification results with the time-series trend vector of the gas concentration data in the oil in the cleaned signal dataset and the time-domain waveform features of the winding vibration response sequence to generate a comprehensive fault feature vector.

[0021] The health assessment module is used to perform matching analysis on the comprehensive fault feature vector through a pre-trained classifier and output the health status assessment results of the distribution transformer.

[0022] This application proposes a method and system for assessing the health status of distribution transformers by fusing multi-source data. It solves the problems of difficulty in fusing heterogeneous multi-source data, inaccurate characterization of health status features, and unreliable identification of critical states in distribution transformers, thereby improving the accuracy, sensitivity, and engineering applicability of health status assessment. Compared with existing technologies, the beneficial effects of this application's technical solution are at least as follows:

[0023] First, by performing time delay calculation and correction based on cross-correlation function on oil gas concentration data, partial discharge pulse sequence and winding vibration response sequence, the precise alignment of multi-source heterogeneous signals on a unified time axis is achieved, effectively eliminating signal time delay caused by differences in physical mechanisms, and enabling the representation of the same health state event in different signals to be synchronously correlated.

[0024] Second, by extracting and fusing the time-series trend vector of gas concentration data in oil, the time-domain waveform features of winding vibration response sequence, and the quantization vector of preliminary fault classification results, an extended feature set is formed. Furthermore, feature selection and dimensionality reduction optimization are performed based on variance contribution rate to construct a comprehensive health feature vector with strong discriminative power and low redundancy. This overcomes the shortcomings of traditional "feature stacking" methods, which have high dimensionality and a lot of noise, and improves the accuracy of health status feature representation and model computation efficiency.

[0025] Third, by performing matching analysis on the comprehensive feature vector through a pre-trained classifier and triggering a secondary verification mechanism for low-confidence assessment results in conjunction with a confidence threshold, the original signal features are dynamically integrated for verification. This enables sensitive capture and reliable judgment of the critical process of health status from quantitative to qualitative change, enhances the assessment system's ability to identify boundary states and early latent defects, and improves the accuracy and reliability of status warnings. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a flowchart illustrating the health status assessment method for distribution transformers that integrates multi-source data in this application.

[0028] Figure 2 This is a graph showing the cross-correlation function between partial discharge and vibration signal in this application;

[0029] Figure 3 This is a comparison chart of the noise immunity performance in this application;

[0030] Figure 4 This is a radar chart comparing the overall performance in this application;

[0031] Figure 5 This is a schematic diagram of the structure of the distribution transformer health status assessment system that integrates multi-source data in this application. Detailed Implementation

[0032] This application provides a method and system for assessing the health status of distribution transformers by integrating multi-source data. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data used can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0033] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the method for assessing the health status of distribution transformers that integrates multi-source data in this application includes:

[0034] Step S1: Obtain a multi-source monitoring dataset containing timestamps from the distribution transformer monitoring equipment. The multi-source monitoring dataset includes oil gas concentration data, partial discharge pulse sequence, and winding vibration response sequence.

[0035] In one specific embodiment, performing step S1 includes the following steps:

[0036] Data on oil gas concentration, partial discharge pulse sequence, and winding vibration response sequence were collected using power distribution transformer monitoring equipment.

[0037] The data on gas concentration in oil, partial discharge pulse sequence, and winding vibration response sequence were integrated to form an initial monitoring dataset.

[0038] Add corresponding acquisition timestamp information to each signal sequence in the initial monitoring dataset;

[0039] The initial monitoring dataset is sorted along a timeline based on the timestamp information collected.

[0040] Record the acquisition source and acquisition time range of each signal sequence in the initial monitoring dataset;

[0041] The integrated and sorted signal sequences, corresponding acquisition timestamp information, acquisition sources, and acquisition time range records are used to generate a multi-source monitoring dataset containing timestamps.

[0042] Specifically, the monitoring equipment for the distribution transformer collects data on oil gas concentration, partial discharge pulse sequences, and winding vibration response sequences. The oil gas concentration data is collected by a gas sensor deployed inside the transformer tank, covering characteristic gases such as hydrogen, methane, and acetylene. The sampling frequency is set to once every 5 minutes, the acquisition range is 0~500ppm, and the data output format is digital signal. The partial discharge pulse sequence is collected by a partial discharge detector installed on the transformer tank, with a sampling frequency of 1MHz. The pulse amplitude (unit: pC), pulse width, and pulse occurrence time are recorded, and the data is stored in binary stream format. The winding vibration response sequence is collected by a piezoelectric vibration sensor fixed at the end of the winding, with a sampling frequency of 2kHz and a measurement range of 0-50m / s². The output signal is an analog quantity, which is transmitted in digital signal format after analog-to-digital conversion. The three types of raw data were distinguished according to signal type. The gas concentration data in the oil were organized into a one-dimensional numerical sequence according to the acquisition time. The partial discharge pulse sequence was sorted according to the pulse occurrence time to form a pulse event sequence. The winding vibration response sequence was composed of a vibration acceleration time sequence according to the sampling time. Then, the three sequences were integrated into the same data framework to form an initial monitoring dataset. This dataset uses row indexes to correspond to the logical order of data acquisition, and columns to correspond to the raw acquisition values ​​and signal type identifiers of the three types of signals, respectively.

[0043] For each signal sequence in the initial monitoring dataset, a corresponding acquisition timestamp is added. The timestamp uses UTC time format and is accurate to the millisecond level. The timestamp of the gas concentration data in the oil is marked as the moment when the sensor completes a single sampling. Each pulse event of the partial discharge pulse sequence is marked with the corresponding detection time. Each sampling point of the winding vibration response sequence is associated with the sampling time. The timestamp information corresponds one-to-one with the signal data to ensure that each data point has a clear time reference. The initial monitoring dataset is sorted along a timeline based on the collected timestamp information. The sorting uses an improved algorithm based on three-way fast sort, with the UTC millisecond value corresponding to the timestamp as the core sorting keyword. The algorithm sets the benchmark selection strategy to randomly sample three timestamps and take the median value, dividing the data into three partitions: less than, equal to, and greater than the benchmark value. The partition equal to the benchmark value retains the priority logic of the signal sequence. At the same time, the maximum recursion depth is set to 16 to avoid stack overflow, adapting to the scenario of mixed storage of multi-source data. During the sorting process, the integrity of the data within each signal sequence is maintained. If different signal sequences have the same timestamp, they are arranged in the order of oil gas concentration data, partial discharge pulse sequence, and winding vibration response sequence under the same time node. The sorted data is distributed in chronological order from early to late, forming a time-ordered dataset.

[0044] The initial monitoring dataset records the acquisition source of each signal sequence. For oil gas concentration data, the acquisition source includes the sensor installation location (e.g., top, middle, bottom of the tank) and sensor number. For partial discharge pulse sequences, the acquisition source records the detector channel number and installation location. For winding vibration response sequences, the acquisition source is marked with the sensor fixed point and number. The acquisition time range for each signal sequence is also recorded, i.e., the interval formed by the timestamp of the first data point and the timestamp of the last data point. The sorted signal sequences, corresponding acquisition timestamp information, acquisition source, and acquisition time range records are integrated to generate a multi-source monitoring dataset containing timestamps. This dataset is stored in a table structure, with fields including timestamp, oil gas concentration (hydrogen, methane, acetylene), partial discharge pulse amplitude, partial discharge pulse width, winding vibration acceleration, acquisition source code, and acquisition time range. The acquisition source code uses the format "signal type-location-number," for example, "GT-01" represents gas sensor number 1 at the top of the tank, "PD-CH-02" represents channel 2 of the partial discharge detector, and "VW-03" represents vibration sensor number 3 at the winding point.

[0045] Step S2: Preprocess the multi-source signal dataset to remove noise and interference, and obtain the cleaned signal dataset.

[0046] In one specific embodiment, performing step S2 includes the following steps:

[0047] The signal decomposition and reconstruction method is used to decompose each signal sequence in the multi-source monitoring dataset;

[0048] Based on the signal sequence decomposition results, high-frequency noise and electromagnetic interference components in the partial discharge pulse sequence and winding vibration response sequence are filtered out, and the gas concentration data in the oil are smoothed.

[0049] The signal sequences, after being filtered and smoothed to remove high-frequency noise and electromagnetic interference, are reassembled to form a preliminary processed dataset.

[0050] The effectiveness of the preliminary processed dataset is verified by comparing the quality of each signal in the preliminary processed dataset with the corresponding original signal.

[0051] Identify outliers in the initial processed dataset and perform interpolation on the outliers;

[0052] The preliminary processed dataset, after validity verification and outlier interpolation, is determined as the cleaned signal dataset.

[0053] Specifically, wavelet decomposition and reconstruction methods were used to decompose the oil gas concentration data, partial discharge pulse sequence, and winding vibration response sequence in the multi-source monitoring dataset. The db4 wavelet basis function was selected, and the decomposition level was set to 5 levels. The Mallat algorithm was used to decompose each signal sequence into 5 detail components and 1 approximate component. The detail components correspond to the high-frequency components of the signal, and the approximate component corresponds to the low-frequency trend components. Based on the decomposition results, a soft thresholding function was used to process the 4th and 5th level detail components of the partial discharge pulse sequence. The threshold calculation formula is as follows: ,in The noise standard deviation of the detail components is estimated using the median absolute deviation. To determine the data length of the partial discharge pulse sequence, the effective high-frequency components related to the discharge pulses are retained after processing, while high-frequency noise is filtered out. For the winding vibration response sequence, a hard thresholding function is used to process the 3rd to 5th level detail components, with a threshold value of [missing information]. , This represents the noise standard deviation of the detail components of the sequence. To determine the data length of the winding vibration response sequence, high-frequency components corresponding to electromagnetic interference are removed. For the gas concentration data in the oil, the decomposed approximate components are retained, while all detailed components are discarded to achieve smoothing and reduce data fluctuations. The smoothing of the gas concentration data is achieved by reconstructing the approximate components using inverse wavelet transform, ensuring that the trend characteristics of concentration changes are preserved. For example, the original data length of the partial discharge pulse sequence of a 10kV distribution transformer is... According to estimates Calculations yielded By using soft thresholding, detail components with an absolute value less than 3.43pC are set to zero, and components with an absolute value greater than this threshold are contracted, effectively separating the real discharge pulse signal with an amplitude of more than 5pC and filtering out false pulses caused by high-frequency interference from the power grid.

[0054] The processed signal sequences are reconstructed using inverse wavelet transform. The reconstructed partial discharge pulse sequence retains the valid discharge pulse signal, the reconstructed winding vibration response sequence highlights the mechanical vibration-related features, and the reconstructed oil gas concentration data forms a stable concentration time series. Subsequently, the three reconstructed signal sequences are reintegrated according to the timestamp correspondence of the original data to form a preliminary processed dataset. Each time node corresponds to the processed oil gas concentration value, partial discharge pulse parameters, and winding vibration response value, ensuring the temporal correspondence of the data. For example, at 15:00:00, the original data for a distribution transformer shows an oil hydrogen concentration of 85 ppm, a partial discharge pulse amplitude of 12 pC, and a winding vibration acceleration of 3.2 m / s². After processing, the corresponding data for that time is a hydrogen concentration of 83 ppm, a partial discharge pulse amplitude of 11.8 pC, and a winding vibration acceleration of 3.1 m / s². The timestamps of all three remain consistent, ensuring temporal correlation during subsequent time alignment and feature fusion.

[0055] The quality of each signal in the pre-processed dataset is compared with that of the corresponding original signals. The effectiveness is verified by calculating the signal-to-noise ratio (SNR) and correlation coefficient. The SNR calculation formula is as follows: ,in The effective power of the processed signal. The noise power is represented by the Pearson correlation coefficient between the processed signal and the original signal. Set the signal-to-noise ratio improvement to be no less than 15% and the correlation coefficient to be... The validity criterion is that if this criterion is not met, the process is repeated after readjusting the wavelet decomposition level or threshold parameters. For example, the original signal-to-noise ratio (SNR) of the winding vibration response signal of a distribution transformer was 16 dB. After processing, the SNR increased to 20 dB, an improvement of 25%, and the correlation coefficient... One set of data met the validity criteria; however, the signal-to-noise ratio of the other set of data only improved by 12% after processing. Therefore, the wavelet decomposition level was readjusted to 6, and the threshold was adjusted accordingly. Adjusted to After reprocessing, the improvement reached 18%, and the correlation coefficient... The set standards have been met.

[0056] Identify outliers in the initial processing dataset, using The criteria are as follows: For oil gas concentration data, calculate the mean and standard deviation of the sequence; values ​​exceeding the range of the mean plus or minus three standard deviations are considered outliers. For partial discharge pulse sequences, calculate the mean and standard deviation of the pulse amplitude; pulse amplitudes exceeding the range of the mean plus or minus three standard deviations are considered outliers. For winding vibration response sequences, calculate the mean and standard deviation of the vibration acceleration; values ​​exceeding the range of the mean plus or minus three standard deviations are considered outliers. Linear interpolation is used to process the identified outliers. Assuming the location of the outlier is... Its previous valid data point is The next valid data point is The interpolation formula is: This ensures the continuity and integrity of the data sequence. For example, in the methane concentration data of a distribution transformer oil, the value at a certain moment is 210 ppm, while the mean of the sequence is 50 ppm, the standard deviation is 30 ppm, and the corresponding interval is [-40 ppm, 140 ppm]. 210 ppm is outside this interval and is judged as an outlier. The value at the previous moment 10 minutes ago is 48 ppm, and the value at the next moment 20 minutes ago is 52 ppm. Through linear interpolation, the corresponding value at 15 minutes ago is 50 ppm. The supplemented data sequence maintains continuity.

[0057] The preliminary processed dataset, after validity verification and outlier interpolation, is the cleaned signal dataset. This dataset eliminates noise and interference in multi-source signals, corrects outlier data, and provides a reliable data foundation for subsequent time correction and feature extraction. It solves the technical problem that fault features are masked due to strong noise interference, ensures the accuracy of subsequent cross-physical quantity feature fusion, and avoids interference from invalid data on health status assessment results.

[0058] Step S3: Perform time correction processing on the cleaned signal dataset to obtain a time-aligned unified signal dataset.

[0059] In one specific embodiment, performing step S3 includes the following steps:

[0060] The cross-correlation function was used to calculate the gas concentration data in the oil, the time delay offset between the winding vibration response sequence and the partial discharge pulse sequence in the signal dataset after cleaning.

[0061] The hysteresis adjustment of the oil gas concentration data and the winding vibration response sequence in the cleaned signal data is determined based on the time delay offset.

[0062] Time correction is performed on the oil gas concentration data and winding vibration response sequence in the cleaned signal dataset based on the hysteresis adjustment.

[0063] The time-corrected oil gas concentration data, partial discharge pulse sequence, and winding vibration response sequence are integrated to form a candidate unified dataset.

[0064] The accuracy of time alignment is verified by comparing the timestamps of each signal sequence in the candidate unified dataset.

[0065] The candidate unified datasets, verified for accuracy, were identified as time-aligned unified signal datasets.

[0066] Specifically, the time delay offset between the oil gas concentration data, winding vibration response sequence, and partial discharge pulse sequence in the cleaned signal dataset is calculated using the cross-correlation function. Correlation is measured by calculating the product of the two signal sequences under different time delays. This requires unifying the time granularity of the multi-source signals before performing the cross-correlation operation. In the specific calculation, the partial discharge pulse sequence is used as a reference. First, window statistical aggregation processing is performed on its 1MHz high-frequency sampling characteristics. The time window is divided according to the sampling period of the oil gas concentration data (5 minutes = 300 seconds). Three core features are statistically analyzed within each window: total pulse energy, average peak amplitude, and pulse frequency. This transforms the high-frequency discrete partial discharge pulse sequence into a partial discharge aggregated feature sequence with the same granularity as the gas data, preserving key information about the discharge event while significantly reducing the data volume. For the 2kHz sampled winding vibration response sequence, the average vibration peak value and the proportion of the main frequency energy are also statistically analyzed within a 300-second window to generate a vibration aggregated feature sequence, ensuring that all three types of signals are unified into a 300-second interval time-series feature sequence. A sliding window length for cross-correlation calculation is set, which is represented by the number of data points in the aggregated sequence, corresponding to a time span of 200 times the sampling period of the original partial discharge pulse sequence. For example, if the original partial discharge sampling period is 1µs, the time span corresponding to the sliding window is 200µs. In the aggregated sequence (with a time interval of 5 minutes), the corresponding number of data points is calculated to ensure that the window can cover the typical time delay range, avoiding both excessively short windows that miss delay information and excessively long windows that increase the computational burden. Subsequently, the time delay range is set to [-T, T] (where T is taken as 500 times the sampling period of the gas concentration data in the oil), which fully covers the typical delay range of gas diffusion (several hours) and vibration propagation (milliseconds), while avoiding excessively large time delay ranges that increase the computational burden. For the cross-correlation calculation of oil gas concentration data and partial discharge aggregation feature sequence, there is no need to interpolate and resample the gas data. Instead, the original 300-second interval sequence and the partial discharge aggregation feature sequence are directly subjected to sliding window product summation. The correlation values ​​are solved window by window to generate the cross-correlation sequence. The time delay corresponding to the peak value in this sequence is the time delay offset of the oil gas concentration data relative to the partial discharge pulse sequence. The cross-correlation calculation logic of the winding vibration aggregation feature sequence and the partial discharge aggregation feature sequence is the same. Based on the 300-second interval feature sequence of the same granularity, the sliding window product summation is directly performed. The time delay offset between the two is obtained by locating the peak value of the cross-correlation sequence.For example, the partial discharge pulse sequence of a 10kV distribution transformer is aggregated through a 300-second window to form a partial discharge aggregated feature sequence. The oil gas concentration data is sampled for 5 minutes. The peak time delay is calculated by cross-correlation between the two, which is 3600 seconds. That is, the time delay of the oil gas concentration data relative to the partial discharge pulse sequence is 3600 seconds. The winding vibration response sequence of the transformer is aggregated through a 300-second window and cross-correlated with the partial discharge aggregated feature sequence to obtain a peak time delay of 0.02 seconds. That is, the time delay of the winding vibration response sequence relative to the partial discharge pulse sequence is 0.02 seconds.

[0067] The hysteresis adjustment of the oil gas concentration data and winding vibration response sequence in the cleaned signal dataset is determined based on the time delay offset. The value of the hysteresis adjustment is equal to the corresponding time delay offset, and the sign is determined according to the signal propagation direction. If the oil gas concentration data or winding vibration response sequence lags behind the partial discharge pulse sequence, the hysteresis adjustment is positive; if it leads, it is negative. For cases with multi-peak cross-correlation, the time delay offset corresponding to the main peak is selected as the core hysteresis adjustment. If there are multiple secondary peaks with amplitudes exceeding 80% of the main peak, the weighted average of the time delay offsets corresponding to these secondary peaks is calculated as the hysteresis adjustment, with the weight proportional to the peak amplitude. For example, if the cross-correlation sequence of the oil gas concentration data and the partial discharge pulse sequence of a distribution transformer shows two peaks, with the main peak amplitude of 0.92 (corresponding to a time delay of 3600 seconds) and the secondary peak amplitude of 0.85 (corresponding to a time delay of 3500 seconds), the hysteresis adjustment is calculated as follows: Second.

[0068] Based on the hysteresis adjustment, time correction is performed on the concentrated oil gas concentration data and winding vibration response sequence of the cleaned signal dataset. For the oil gas concentration data, linear interpolation is used to shift the entire time axis forward by the duration corresponding to the hysteresis adjustment, i.e., subtracting the hysteresis adjustment from the timestamp of each data point. If overlapping or missing timestamps occur after the shift, missing data is supplemented or overlapping data is merged using linear interpolation. For the winding vibration response sequence, the same linear interpolation method is used for time axis shift correction to ensure consistency with the timestamps of the partial discharge pulse sequence. For example, if the hysteresis adjustment for the oil gas concentration data is 3600 seconds, and the original timestamp of a certain data point is 1620000000 seconds, the corrected timestamp becomes 16200000000 - 3600 = 1619996400 seconds. If data points are missing within a certain time period after correction, linear interpolation is used to supplement them based on the valid data points before and after that time period.

[0069] The time-corrected oil gas concentration data, partial discharge pulse sequence, and winding vibration response sequence are integrated to form a candidate unified dataset. During integration, the timestamp of the partial discharge pulse sequence is used as a benchmark, and the corrected oil gas concentration data and winding vibration response sequence are correlated one-to-one according to their timestamps. Each timestamp contains the corresponding oil gas concentration value, partial discharge pulse parameter, and winding vibration response value, ensuring synchronous correlation of the three types of data in the time dimension. The timestamps of each signal sequence in the candidate unified dataset are compared to verify the accuracy of time alignment. A timestamp deviation threshold of 0.5 times the sampling period of the partial discharge pulse sequence is set. The differences between the timestamps of the oil gas concentration data, the winding vibration response sequence, and the partial discharge pulse sequence at each timestamp in the candidate unified dataset are calculated. If the absolute value of all differences is less than the set threshold, the time alignment is considered accurate. If the absolute value of any difference exceeds the threshold, the corresponding time delay offset and hysteresis adjustment are recalculated, and time correction and integration are performed again until the accuracy requirements are met. For example, the sampling period of the partial discharge pulse sequence is 1 microsecond, and the set timestamp deviation threshold is 0.5 microseconds. If the difference between the timestamp of the gas concentration data in the oil and the timestamp of the partial discharge pulse sequence at a certain timestamp is 0.3 microseconds, and the difference between the timestamp of the winding vibration response sequence and the timestamp of the partial discharge pulse sequence is 0.2 microseconds, both are less than the threshold, and the alignment is determined to be accurate. If the difference is 0.6 microseconds at a certain timestamp, then the adjustment is performed.

[0070] The candidate unified dataset, verified for accuracy, was identified as the time-aligned unified signal dataset. This dataset effectively addresses the technical problem of spatiotemporal inconsistency in multi-source data. By calculating and correcting time delays based on cross-correlation functions, it eliminates time lags caused by differences in the physical mechanisms of different signal sources. This allows the representation of the same health status event in oil gas concentration data, partial discharge pulse sequences, and winding vibration response sequences to be synchronously correlated. This provides a unified time reference for subsequent frequency domain transformation and feature fusion, avoiding fault feature fragmentation and misjudgment caused by signal timing misalignment, and improving the accuracy and reliability of health status assessment of distribution transformers.

[0071] Step S4: Perform frequency domain transformation on the time-aligned unified signal dataset, extract the peak frequency and amplitude distribution of each signal sequence, and form a multi-source frequency domain feature set.

[0072] In one specific embodiment, performing step S4 includes the following steps:

[0073] Perform frequency domain transformation on each signal sequence in a time-aligned unified signal dataset;

[0074] Based on the frequency domain transformation results, the peak frequency of each signal sequence is extracted, and the amplitude distribution characteristics of each signal sequence in different frequency ranges are statistically analyzed.

[0075] The peak frequency and corresponding amplitude distribution characteristics of a single signal sequence are combined to generate a frequency domain feature vector;

[0076] Integrate the frequency domain feature vectors of all signal sequences to form an initial multi-source frequency domain feature set;

[0077] Identify anomalous frequency domain features in the initial multi-source frequency domain feature set, and perform secondary verification on the anomalous frequency domain features;

[0078] The effective frequency domain feature vectors after secondary verification are retained and re-integrated to form a multi-source frequency domain feature set.

[0079] Specifically, frequency domain transformation is performed on each signal sequence in the time-aligned unified signal dataset. Fast Fourier Transform (FFT) is selected as the frequency domain transformation algorithm. Transform parameters are set separately for the oil gas concentration data, partial discharge pulse sequence, and winding vibration response sequence. The number of sampling points for the Fourier transform of the oil gas concentration data is set to 1024, the number of sampling points for the partial discharge pulse sequence is set to 4096, and the number of sampling points for the winding vibration response sequence is set to 2048. The number of sampling points is always a power of 2 to improve transformation efficiency. During the transformation, window functions are used to suppress spectral leakage. Hanning windows are used for the oil gas concentration data, Hamming windows for the partial discharge pulse sequence, and Blackman windows for the winding vibration response sequence. The selection of the window function is determined based on the spectral characteristics of each signal sequence to ensure the integrity of the signal characteristics after frequency domain transformation. For example, the winding vibration response sequence of a 10kV distribution transformer has a length of 2000 points. During FFT, zeros are automatically padded to 2048 points. After applying the Blackman window, the spectral leakage is reduced by 30%, effectively preserving the true frequency domain characteristics of the vibration signal.

[0080] Based on the frequency domain transformation results, the peak frequency of each signal sequence is extracted, and the amplitude distribution characteristics of each signal sequence in different frequency intervals are statistically analyzed. Peak frequency extraction is achieved by scanning the amplitude spectrum after frequency domain transformation, and the frequency point with the largest amplitude is selected as the peak frequency. If there are multiple peak frequencies with similar amplitudes, the first three frequency points with the largest amplitudes are extracted as the multi-peak frequency feature. The frequency interval division is determined according to the typical frequency band characteristics of the distribution transformer signal. For example, the frequency intervals of the gas concentration data in oil are divided into 0~1Hz, 1~5Hz, and 5~10Hz; the frequency intervals of the partial discharge pulse sequence are divided into 0~100kHz, 100~500kHz, and 500kHz~1MHz; and the frequency intervals of the winding vibration response sequence are divided into 0~50Hz, 50~200Hz, and 200~500Hz. The proportion of the sum of the amplitudes in each frequency interval to the total amplitude is statistically analyzed as the amplitude distribution feature. For example, after frequency domain transformation, the amplitude of the partial discharge pulse sequence of a certain distribution transformer is the largest at 150kHz, which is determined as the peak frequency. The amplitude accounts for 25% in the 0~100kHz range, 60% in the 100~500kHz range, and 15% in the 500kHz~1MHz range, forming the amplitude distribution characteristics of the sequence.

[0081] The peak frequency and corresponding amplitude distribution characteristics of a single signal sequence are combined to generate a frequency domain feature vector. The frequency domain feature vector of the gas concentration data in oil consists of the peak frequency and the amplitude proportion of the three frequency intervals, with a dimension of 4. The frequency domain feature vector of the partial discharge pulse sequence consists of the first three peak frequencies and the amplitude proportion of the three frequency intervals, with a dimension of 6. The structure of the frequency domain feature vector of the winding vibration response sequence is consistent with that of the partial discharge pulse sequence, with a dimension of 6. The elements in the vector are arranged in the order of "peak frequency - amplitude proportion of frequency interval 1 - amplitude proportion of frequency interval 2 - amplitude proportion of frequency interval 3" to ensure the consistency and interpretability of the feature vector. For example, the peak frequency of the gas concentration data in the oil is 3Hz, and the amplitude proportions of the three frequency intervals are 30%, 50%, and 20%, respectively. Its frequency domain feature vector is [3, 0.3, 0.5, 0.2]. The three peak frequencies of the partial discharge pulse sequence are 150kHz, 200kHz, and 300kHz, and the amplitude proportions of the three frequency intervals are 25%, 60%, and 15%, respectively. Its frequency domain feature vector is [150000, 200000, 300000, 0.25, 0.6, 0.15].

[0082] The frequency domain feature vectors of all signal sequences are integrated to form an initial multi-source frequency domain feature set. During the integration process, the timestamps of the time-aligned unified signal dataset are used as indices to associate and store the frequency domain feature vectors of oil gas concentration data, partial discharge pulse sequences, and winding vibration response sequences corresponding to the same timestamp. Each index corresponds to three frequency domain feature vectors, which constitute the basic unit of the initial multi-source frequency domain feature set. The set is stored using a three-dimensional array structure. The first dimension is the time index, the second dimension is the signal type identifier, and the third dimension is the frequency domain feature vector element, which facilitates rapid retrieval and processing later.

[0083] Abnormal frequency domain features in the initial multi-source frequency domain feature set are identified. Abnormal feature identification is based on the statistical threshold method. For each signal type's frequency domain feature vector, the mean and standard deviation of each feature element are calculated. An abnormality threshold is set as the mean plus or minus three times the standard deviation. Feature elements exceeding this range are judged as abnormal, and the corresponding frequency domain feature vector is marked as an abnormal frequency domain feature. A secondary verification of abnormal frequency domain features is performed, using a combination of neighborhood comparison verification and historical data comparison verification. Neighborhood comparison verification compares the frequency domain feature vectors of the five indices before and after the time index where the abnormal feature is located. If the similar features of adjacent indices are all within the normal range, further historical data comparison verification is performed. Similar frequency domain feature data of the distribution transformer over the past 30 days are retrieved, and the similarity between the abnormal feature and the historical data is calculated. A similarity below 0.7 confirms an abnormality, a similarity above 0.8 is judged as normal fluctuation, and a similarity between 0.7 and 0.8 requires manual review in conjunction with the original signal data. For example, in the frequency domain feature vector of the winding vibration response sequence of a distribution transformer, the peak frequency is 800Hz, which is far beyond the normal range [-100Hz, 500Hz] corresponding to the mean of 200Hz and the standard deviation of 100Hz. It is marked as an abnormal frequency domain feature. Neighbor comparison shows that the peak frequencies of the five indices before and after are all between 150 and 250Hz. Historical data comparison shows a similarity of 0.65, confirming that the feature is abnormal.

[0084] Valid frequency domain feature vectors after secondary verification are retained and re-integrated into a multi-source frequency domain feature set. Frequency domain feature vectors confirmed to be abnormal are removed. For feature vectors that cannot be determined after manual review, they are retained and marked as to be observed. A three-dimensional array structure is reconstructed to ensure that the frequency domain feature vectors under each time index in the set are valid or clearly marked. This process effectively solves the problem of redundant and invalid features caused by "feature stacking" at the feature engineering level. By extracting targeted features through frequency domain transformation and combining anomaly identification and secondary verification, effective features related to the health status of distribution transformers are screened out, reducing interference from environmental noise and load fluctuations. This provides high-quality feature input for subsequent fault mechanism classification and improves the robustness and generalization ability of health status assessment.

[0085] Step S5: Perform preliminary fault mechanism classification based on the multi-source frequency domain feature set to obtain the classification results.

[0086] In one specific embodiment, performing step S5 includes the following steps:

[0087] If the peak frequency of the partial discharge pulse sequence in the multi-source frequency domain feature set exceeds a preset frequency threshold, then a potential discharge-dominated fault is identified; otherwise, a non-discharge-dominated potential fault is identified.

[0088] Calculate the matching degree between the amplitude distribution characteristics of the winding vibration response sequence in the multi-source frequency domain feature set and the potential discharge-dominated fault.

[0089] Determine if the matching degree is higher than the preset matching threshold. If yes, determine if there is a discharge-dominant potential fault; otherwise, determine if there is a discharge-suspected potential fault.

[0090] The distribution of discharge-dominant potential faults, suspected discharge-dominant potential faults, and non-discharge-dominant potential faults was statistically analyzed to obtain preliminary classification results.

[0091] For potential discharge-related faults identified in the preliminary classification results, the weights of the corresponding signal sequence frequency domain features in the multi-source frequency domain feature set are adjusted.

[0092] Based on the adjusted feature weights, the matching degree is re-evaluated, the preliminary classification results are optimized, and the optimized classification of various potential faults is integrated to obtain the final classification result.

[0093] Specifically, the system determines whether the peak frequency of a partial discharge pulse sequence in the multi-source frequency domain feature set exceeds a preset frequency threshold. The peak frequency refers to the frequency value corresponding to the data point with the largest amplitude in the amplitude spectrum after frequency domain transformation of the partial discharge pulse sequence. The preset frequency threshold is determined based on the voltage level, insulation structure, and operating conditions of the distribution transformer. For example, the preset frequency threshold is set to 160kHz for a 10kV distribution transformer, 180kHz for a 35kV distribution transformer, and 200kHz for 110kV and above distribution transformers. This threshold is obtained through statistical fitting of a large amount of field fault measurement data and can effectively distinguish the frequency characteristics of partial discharge pulses corresponding to normal operation and potential discharge faults. When extracting the peak frequency of a partial discharge pulse sequence from a multi-source frequency domain feature set, the amplitude spectrum after frequency domain transformation is first filtered and denoised using a 5-point moving average to eliminate high-frequency glitches. Then, the frequency value corresponding to the maximum amplitude is traversed and selected. If there are multiple extreme points where the ratio of amplitude to maximum amplitude is ≥0.85, these frequencies are retained as multi-peak features. The core judgment is still based on the frequency corresponding to the maximum amplitude. After extraction, it is compared with the preset frequency threshold for the corresponding voltage level. If the peak frequency is greater than or equal to the preset frequency threshold, a potential discharge-dominated fault is determined to exist; if the peak frequency is less than the preset frequency threshold, a non-discharge-dominated potential fault is determined to exist. For example, after denoising, the peak frequency of the partial discharge pulse sequence of a 35kV distribution transformer is 192kHz, which exceeds the preset threshold of 180kHz, and is determined to be a potential discharge-dominated fault; the peak frequency of the partial discharge pulse sequence of a 10kV distribution transformer is 148kHz, which is lower than the preset threshold of 160kHz, and is determined to be a non-discharge-dominated potential fault.

[0094] The matching degree between the amplitude distribution characteristics of the potential discharge-dominated fault and the winding vibration response sequence in the multi-source frequency domain feature set is calculated. A cosine similarity algorithm is used to calculate the matching degree. The standard amplitude distribution template corresponding to the potential discharge-dominated fault and the actual amplitude distribution characteristics of the winding vibration response sequence are constructed as two vectors of the same dimension. The degree of matching is quantified by calculating the cosine angle between the vectors. The matching degree calculation formula is as follows: Where C is the standard amplitude distribution template vector, D is the actual amplitude distribution feature vector, and "·" indicates the vector dot product operation. and Let C and D represent the Euclidean norms of vectors C and D, respectively. The standard amplitude distribution template is set according to the winding vibration characteristics of distribution transformers with discharge faults at different voltage levels. For example, the standard amplitude distribution template vector for a 10kV distribution transformer discharge fault is [0.22, 0.43, 0.35], corresponding to the amplitude proportions in the three frequency ranges of 0~50Hz, 50~200Hz, and 200~500Hz; the standard amplitude distribution template vector for a 35kV distribution transformer is [0.18, 0.41, 0.41]; and the standard amplitude distribution template vector for 110kV and above distribution transformers is [0.15, 0.38, 0.47]. For example, the actual amplitude distribution feature vector of the winding vibration response sequence of a certain 10kV distribution transformer is [0.23, 0.42, 0.35], and its dot product with the standard template vector is... The norm of vector C is The norm of vector D is Final matching degree .

[0095] The matching degree is determined to be higher than a preset matching threshold. Preferably, the preset matching threshold is set to 0.8. If the matching degree is greater than 0.8, a discharge-dominant potential fault is identified; if the matching degree is less than or equal to 0.8, a discharge-suspected potential fault is identified. For example, the matching degree of the aforementioned 10kV distribution transformer is 0.9997, which is higher than the preset threshold of 0.8, and is therefore identified as a discharge-dominant potential fault; the matching degree of a certain 35kV distribution transformer is 0.76, which is lower than the preset threshold of 0.8, and is therefore identified as a discharge-suspected potential fault. The distribution of discharge-dominant potential faults, discharge-suspected potential faults, and non-discharge-dominant potential faults is statistically analyzed. A 24-hour statistical period is used as a statistical cycle to record the occurrence frequency, duration, and corresponding timestamp interval of the three types of faults within each statistical cycle, forming a distribution statistical table. Based on this table, preliminary classification results are obtained, which clearly define the proportion, occurrence time, and evolution trend of the three types of faults within each statistical cycle. For example, within a 24-hour statistical period, a certain distribution transformer experienced 4 discharge-dominant potential faults, lasting a total of 150 minutes; 2 discharge-suspected potential faults, lasting a total of 80 minutes; and 1 non-discharge-dominant potential fault, lasting 40 minutes. The preliminary classification results clearly show the distribution characteristics of various types of faults.

[0096] For potential discharge-related faults identified in the preliminary classification results, the weights of the corresponding signal sequence frequency domain features in the multi-source frequency domain feature set are adjusted. This weight adjustment employs a dynamic weighting strategy based on information gain, specifically calculated using the information entropy difference. The steps are as follows: Taking the discharge-related sample set as the object, the initial information entropy E is first calculated. The sample set contains three categories of samples: discharge-dominant, discharge-related, and non-discharge-dominant, with a total sample size of M and a sample size for each category of M. The formula is The sample set is then divided into two subsets based on the features to be calculated (e.g., peak frequency is divided into "higher / lower" based on a threshold, and amplitude distribution features are divided into "higher / lower" based on matching degree), and the conditional information entropy of each subset is calculated. The total conditional information entropy is obtained by weighted summation. ( (Number of samples in the subset). Information gain value. This value directly quantifies the contribution of features to fault classification; the larger the value, the stronger the discriminative ability and the higher the corresponding weight. For example, a sample set of suspected discharge types... initial entropy After dividing the samples according to peak frequency, the entropy of subset 1 (60 samples) is 1.491, and the entropy of subset 2 (40 samples) is 1.322. This can be calculated... Gain value Similarly, , The total gain is 0.257, and the normalized weights are 0.58, 0.24, and 0.18, respectively.

[0097] Based on the adjusted feature weights, the matching degree is re-evaluated. The adjusted frequency domain feature weights are incorporated into the matching degree calculation process. The matching degree between potential discharge-dominant faults and the amplitude distribution characteristics of winding vibration response sequences is recalculated. The preliminary classification results are optimized based on the comparison between the new matching degree and the preset matching threshold. If the recalculated matching degree is higher than the preset matching threshold, the original suspected discharge-dominant potential fault is adjusted to a discharge-dominant potential fault. If it is still lower than the preset matching threshold, the suspected discharge-dominant potential fault classification is maintained. If the matching degree is significantly reduced, a comprehensive judgment is made based on the frequency domain characteristics of oil gas concentration data to determine whether to adjust it to a non-discharge-dominant potential fault. The optimized classifications of various potential faults are integrated to form the final classification result, which clearly defines the current potential fault types, distribution patterns, and key characteristics of the distribution transformer.

[0098] Step S6: Combine the classification results with the time-series trend vector of the gas concentration data in the oil and the time-domain waveform features of the winding vibration response sequence in the cleaned signal dataset to generate a comprehensive fault feature vector.

[0099] In one specific embodiment, performing step S6 includes the following steps:

[0100] From the cleaned dataset, the time-series trend vector of gas concentration data in oil is extracted, and the time-domain waveform features of the winding vibration response sequence are extracted. The time-domain waveform features include vibration peak amplitude and waveform period.

[0101] By fusing time-series trend vectors and time-domain waveform features, an initial fused feature set is obtained;

[0102] The classification results are converted into quantized feature vectors, added to the initial fusion feature set, and then the expanded fusion feature set is obtained.

[0103] Calculate the contribution rate of each feature in the extended fusion feature set, select core features based on the contribution rate and remove redundant features to complete feature dimensionality reduction optimization;

[0104] The extended fusion feature set, after dimensionality reduction optimization, is normalized to generate a comprehensive fault feature vector.

[0105] Specifically, a time-series trend vector of oil gas concentration data was extracted from the cleaned dataset. A sliding window linear regression algorithm was selected, with a window length of 24 hours (corresponding to 288 sampling points of oil gas concentration data) and a sliding step size of 1 hour (corresponding to 12 sampling points). Linear fitting was performed on the gas concentration data (covering hydrogen, methane, and acetylene) within each window to obtain the slope and intercept of each window. The slopes and intercepts of all windows were arranged in chronological order to form a time-series trend vector with dimensions of 2×3×number of windows, where 2 represents the slope and intercept, and 3 represents the three characteristic gases. Simultaneously, the time-domain waveform features of the winding vibration response sequence were extracted. The cleaned vibration response sequence was used as the object, with a sampling frequency of 2kHz (corresponding to a sampling interval of 0.5 milliseconds). The data was divided into segments with 1 second as one analysis cycle, and each cycle contained 2000 sampling points. Vibration peak amplitude extraction: Vibration acceleration data within each cycle is traversed. Outliers exceeding the mean plus or minus three standard deviations are removed. The maximum value of the remaining data is selected as the peak amplitude for that cycle. This process is repeated for 24 cycles to form a peak amplitude sequence. Waveform period extraction: For the vibration signal within each cycle, a threshold method (threshold set at 30% of the peak amplitude for that cycle) is used to identify the positions of two adjacent peaks corresponding to the rising edge crossing the threshold point. The time interval between the two points is calculated. The average of five adjacent peak intervals for each cycle is taken as the waveform period for that cycle. This process also forms a waveform period sequence of 24 cycles. Finally, these are integrated into a two-dimensional time-domain waveform feature vector (dimension 2×24, corresponding to the peak amplitude sequence and waveform period sequence, respectively). For example, after fitting the hydrogen concentration in the oil of a distribution transformer over 24 hours using a sliding window, 24 slope values ​​and 24 intercept values ​​are obtained. Methane and acetylene each have 24 slope values ​​and 24 intercept values, forming a 2×3×24 time-series trend vector. The peak amplitude of each period of the winding vibration response sequence is between 3.2 and 4.5 m / s², and the average waveform period is 0.02 seconds, forming a corresponding two-dimensional time-domain waveform feature vector.

[0106] The time-series trend vector and time-domain waveform features are fused using a vector concatenation method. The time-series trend vector is expanded into a one-dimensional vector (dimension 2×3×number of windows), and concatenated with the time-domain waveform feature vector (dimension 2) to form an initial fused feature set with a dimension of 2×3×number of windows + 2. For example, the 2×3×24 time-series trend vector expanded to a dimension of 144, concatenated with the time-domain waveform feature vector (dimension 2), results in an initial fused feature set with a dimension of 146.

[0107] The classification results are converted into quantized feature vectors using one-hot encoding. If the classification result is a discharge-dominant potential fault, the quantized feature vector is [1,0,0]; if it is a discharge-suspected potential fault, the quantized feature vector is [0,1,0]; and if it is a non-discharge-dominant potential fault, the quantized feature vector is [0,0,1]. This three-dimensional quantized feature vector is added to the initial fusion feature set to obtain an expanded fusion feature set. The expanded set dimension is 146+3=149.

[0108] The contribution rate of each feature in the extended fusion feature set is calculated using a variance contribution rate algorithm. First, the covariance matrix of the extended fusion feature set is calculated. Then, the eigenvalues ​​of each feature are obtained through eigenvalue decomposition. The contribution rate is calculated as: Contribution Rate = (Single Feature Value / Sum of All Feature Values) × 100%. A contribution rate threshold of 1% is set. Core features with a contribution rate higher than 1% are selected, and redundant features are removed, thus completing feature dimensionality reduction optimization. For example, after variance decomposition, the contribution rates of the first 35 features in the aforementioned 149-dimensional extended fusion feature set are all higher than 1%, with a cumulative contribution rate of 92%. These 35 core features are retained, and the remaining 114 redundant features are removed.

[0109] The expanded fusion feature set, after dimensionality reduction optimization, is normalized using a min-max normalization algorithm. For each core feature, the normalization is performed according to the formula... Perform the transformation, where x is the original feature value. This is the minimum value of the feature. The maximum value of this feature is used. After normalization, all feature values ​​are mapped to the interval [0,1] to generate a comprehensive fault feature vector.

[0110] Step S7: Perform matching analysis on the comprehensive fault feature vector using a pre-trained classifier to output the health status assessment results of the distribution transformer.

[0111] In one specific embodiment, performing step S7 includes the following steps:

[0112] The similarity between the comprehensive fault feature vector and the standard feature vector in the historical fault mode library is calculated by using a pre-trained classification model.

[0113] The matching similarity is used to determine whether the critical process of the health status of the distribution transformer from quantitative to qualitative change has been captured;

[0114] Extract the fault state category with the highest matching similarity and the corresponding fault type probability distribution, and combine it with the critical process judgment result to generate the initial health status assessment result;

[0115] The system identifies results in the initial health status assessment that have a confidence level lower than a preset confidence threshold. Key matching features are extracted from the comprehensive feature vector of the identified results and then combined with the cleaned signal dataset for secondary verification.

[0116] Based on the secondary verification results, the probability distribution of fault types is adjusted and the initial health status assessment results are optimized to generate and output the health status assessment results of the distribution transformer.

[0117] Specifically, based on a pre-built historical fault mode library, which stores standard feature vectors of distribution transformers under different health states, the library covers feature data corresponding to various typical states such as discharge-dominant faults, suspected discharge faults, and non-discharge-dominant faults. The standard feature vector dimension is consistent with the comprehensive fault feature vector, both being 35-dimensional vectors after dimensionality reduction optimization. The classification model is constructed using the random forest algorithm, containing 100 decision trees. The maximum depth of each decision tree is set to 15, the minimum number of sample splits is 2, and the minimum number of leaf nodes is 1. The model training process uses the comprehensive fault feature vector from historical fault cases as input and the corresponding actual health state as the label. The model parameters are optimized through 5-fold cross-validation. The ratio of the training set to the test set is 7:3. The training iterations are performed until the model's classification accuracy on the test set stabilizes above 90%. The process of outputting matching similarity in the model requires the collaboration of an internal decision-making mechanism and a cosine similarity algorithm: First, the comprehensive fault feature vector is input into the trained random forest model. Each decision tree in the model determines the correlation between the vector and various standard feature vectors in the historical fault pattern library based on the feature vector's dimensions and node splitting rules, outputting a preliminary correlation score for each type of fault. The correlation scores of all decision trees are weighted and summed to obtain a comprehensive correlation score matrix. Subsequently, based on the comprehensive correlation score matrix, the model achieves preliminary correlation positioning between the comprehensive fault feature vector and the standard feature vectors in the historical fault pattern library. Focusing on the selected high-correlation standard feature vector subset, the model uses a cosine similarity algorithm to calculate the matching similarity between the comprehensive fault feature vector and each standard feature vector in the historical fault pattern library. The model automatically traverses all standard feature vectors in this subset of the historical fault pattern library, performs cosine similarity calculations one by one, and finally outputs a set of similarity values, forming a similarity vector. Each element in the vector corresponds to the degree of matching between the comprehensive fault feature vector and a certain standard feature vector. The calculated similarity value ranges between [0,1], with a larger value indicating a higher degree of matching.

[0118] Based on the matching similarity results, it is determined whether the health status of the distribution transformer is in a critical process of quantitative change leading to qualitative change. A similarity change rate threshold of 0.15 is set. The similarity change rate between the comprehensive fault feature vector and the same standard feature vector is calculated over five consecutive monitoring periods. The formula for calculating the change rate is as follows: ,in For the current period's similarity, The similarity is calculated based on the similarity over four cycles. If the rate of change of similarity exceeds a set threshold, and the core features related to fault development in the comprehensive fault feature vector (such as the temporal trend slope of acetylene concentration in oil and the peak frequency of winding vibration) consistently approach the feature threshold corresponding to the historical critical state, then the critical process of a healthy state is captured. If the rate of change of similarity is less than or equal to the threshold, and the core feature values ​​remain stable within the normal fluctuation range, then the critical process has not been entered. For example, the similarity of a distribution transformer with the standard feature vector of a discharge-dominant fault over five consecutive cycles is 0.62, 0.68, 0.75, 0.83, and 0.91, respectively. The calculated rate of change is... The concentration of acetylene in the oil was greater than the threshold of 0.15, and the slope of the time trend of acetylene concentration in the oil increased from 0.02 ppm / h to 0.08 ppm / h, which is close to the historical critical state threshold of 0.1 ppm / h. Therefore, it was determined that a critical process was captured.

[0119] The fault state category corresponding to the standard feature vector with the highest matching similarity is extracted as the initial fault type. The probability distribution of this fault type is then output by the classification model. The probability distribution is calculated using the softmax function, and the formula is as follows:

[0120]

[0121] in The score for the j-th type of fault in the output layer of the classification model is given, where m is the total number of fault categories (m=3 here). The probability values ​​for discharge-dominant, discharge-suspected, and non-discharge-dominant faults are obtained, and the sum of these three is 1. Based on the previous critical process assessment results, if a critical process has been captured, this state is highlighted in the initial health status assessment results, indicating that the fault is in a rapid development stage. If no critical process has been captured, the fault is marked as being in a stable or slow development stage, generating the initial health status assessment results, which include the fault type, probability distribution, critical process determination conclusion, and core feature values.

[0122] A preset confidence threshold of 0.8 is set, with the confidence level defined as the probability of the fault type corresponding to the highest probability value in the initial assessment results. If this probability value is lower than 0.8, the result is considered low confidence. For low confidence results, the top 10 key matching features with the highest contribution rate are extracted from the comprehensive fault feature vector. These features include the slope of the time-series trend vector of gas concentration in oil, the peak amplitude of winding vibration response, and the peak frequency of partial discharge pulse sequence. The key matching features are then validated a second time using the cleaned signal dataset. For example, for features related to gas concentration in oil, the trend of this feature over the past 10 monitoring periods in the cleaned dataset is compared to verify whether it matches the change pattern reflected by the key matching features. For features related to winding vibration, the time-domain waveform parameters corresponding to these features are recalculated and compared with the values ​​in the key matching features. For example, in an initial assessment, the probability of a suspected discharge fault was 0.76, which is lower than the confidence threshold of 0.8. The extracted key matching features included an acetylene concentration trend slope of 0.05 ppm / h and a vibration peak amplitude of 3.8 m / s². During the secondary verification, it was found that the acetylene concentration trend slope in the nearly 10 cycles of the cleaned dataset was between 0.04 and 0.06 ppm / h, and the vibration peak amplitude was stable at 3.7 to 3.9 m / s², which was consistent with the values ​​of the key matching features. Moreover, the peak frequency of the partial discharge pulse sequence was always between 150 and 160 kHz, which is close to the preset frequency threshold of 160 kHz.

[0123] The probability distribution of fault types is adjusted based on the secondary verification results. If the secondary verification confirms that the key matching features are true and effective, and the supplemented cleaned signal data supports the fault type, the probability of the corresponding fault type is increased by (0.8 - original probability) × 0.5, while the probabilities of other fault types are reduced proportionally. If the secondary verification finds anomalies in the key matching features that conflict with the patterns reflected in the cleaned signal data, the similarity between the comprehensive fault feature vector and other standard feature vectors in the historical fault mode library is recalculated, and the probability distribution of fault types is updated. The initial health status assessment results are optimized, clarifying the final fault type, adjusted probability distribution, critical process judgment conclusion, and secondary verification status. The health status assessment results of the distribution transformer are generated and output. The output results are presented in structured data form, including assessment timestamp, fault type, probability of various faults, critical process status, key feature values, and secondary verification explanation, providing accurate basis for distribution network condition-based maintenance.

[0124] Please see Figure 2 , Figure 2The cross-correlation function curves of partial discharge and vibration signals show the distribution of cross-correlation function values ​​between the partial discharge pulse sequence and the winding vibration response sequence under different time delays. The peak point corresponds to the time delay position where the correlation between the two is strongest. The curve in the figure fluctuates symmetrically around the zero point, and the amplitude of the fluctuations on both sides reflects the degree of interference between the two types of signals. The closer the absolute value of the peak is to 1, the stronger the temporal correlation between the signals. This illustrates that there is a clear temporal correlation between the partial discharge signal and the vibration signal. By locating this peak point, the time delay offset between the two can be accurately obtained, providing a core quantitative basis for multi-source signal time correction. This directly supports the accurate alignment of multi-source heterogeneous signals on a unified time axis and solves the problem of spatiotemporal inconsistency caused by the differences in the physical mechanisms of different signal sources.

[0125] Please see Figure 3 , Figure 3 The noise immunity performance comparison chart shows the fault detection accuracy of the proposed method and the traditional method under different signal-to-noise ratio (SNR) conditions. The curve of the proposed method consistently remains above that of the traditional method, and the accuracy decrease of the proposed method is significantly less than that of the traditional method as the SNR decreases. This demonstrates that the proposed method effectively improves the fault feature identification in low SNR scenarios by specifically filtering out high-frequency noise and electromagnetic interference from partial discharge and vibration signals, and by smoothing gas concentration data. It solves the technical problems of traditional methods, such as fault features being masked by noise interference and weak anti-interference capability.

[0126] Please see Figure 4 , Figure 4 The radar chart comparing the overall performance shows the performance of our method compared to traditional methods in five core dimensions: real-time performance, robustness, accuracy, feature quality, and computational efficiency. Our method scores higher than traditional methods in all dimensions, with the most significant advantages in feature quality and accuracy, exceeding 0.9. This demonstrates that our method achieves comprehensive performance improvement through end-to-end technical optimization: the advantage in feature quality stems from frequency domain feature selection in step S4 and feature fusion and dimensionality reduction in step S6; the improvement in accuracy benefits from time correction, feature optimization, and the secondary verification mechanism in step S7; the improvement in real-time performance and computational efficiency comes from reducing the computational burden on the model due to feature dimensionality reduction; and the enhanced robustness relies on the end-to-end anti-interference design of data cleaning, abnormal feature verification, and dynamic weight adjustment, solving problems such as inaccurate feature representation, low evaluation reliability, and insufficient computational efficiency inherent in traditional methods.

[0127] Please see Figure 5 The following describes a distribution transformer health status assessment system that integrates multi-source data according to an embodiment of this application. The distribution transformer health status assessment system that integrates multi-source data includes:

[0128] The data acquisition module is used to acquire multi-source monitoring datasets containing timestamps from the distribution transformer monitoring equipment. The multi-source monitoring datasets include oil gas concentration data, partial discharge pulse sequences, and winding vibration response sequences.

[0129] The data cleaning module is used to preprocess multi-source signal datasets to remove noise and interference, and obtain cleaned signal datasets.

[0130] The time correction module is used to perform time correction processing on the cleaned signal dataset to obtain a time-aligned unified signal dataset.

[0131] The feature extraction module is used to perform frequency domain transformation on the time-aligned unified signal dataset, extract the peak frequency and amplitude distribution of each signal sequence, and form a multi-source frequency domain feature set.

[0132] The fault classification module is used to perform preliminary fault mechanism classification based on a multi-source frequency domain feature set and obtain the classification results.

[0133] The feature module integrates the classification results with the time-series trend vector of the gas concentration data in the oil in the cleaned signal dataset and the time-domain waveform features of the winding vibration response sequence to generate a comprehensive fault feature vector.

[0134] The health assessment module is used to perform matching analysis on the comprehensive fault feature vector through a pre-trained classifier and output the health status assessment results of the distribution transformer.

[0135] Through the collaborative efforts of the aforementioned components, this system constructs a complete health status assessment technology system for distribution transformers, encompassing data acquisition, cleaning and correction, feature extraction and fusion, and intelligent evaluation. It achieves automated processing and in-depth analysis of multi-source heterogeneous monitoring data, improving the automation level and decision-making accuracy of status assessment. Specifically:

[0136] The data acquisition module integrates multiple sensors and a unified timing mechanism to acquire time-stamped raw data of multi-source signals, providing a time-correlated foundational dataset for subsequent processing. The data cleaning module utilizes signal decomposition and reconstruction techniques to perform differentiated filtering and smoothing of noise characteristics for different types of signals, effectively removing noise introduced by electromagnetic interference, load fluctuations, etc., ensuring the quality of data used for analysis, and solving the technical problems of low signal-to-noise ratio and masked fault features due to complex on-site environments. The time correction module calculates and compensates for the inherent time delay between different physical signals through cross-correlation functions, achieving precise alignment of oil gas concentration data, partial discharge pulse sequences, and winding vibration response sequences on a unified time axis. This module directly addresses the core technical deficiency of spatiotemporal inconsistency in multi-source signals, enabling the synchronous correlation and analysis of the representation of the same fault event across different physical quantities. The feature extraction module performs frequency domain transformation on the time-aligned signals, extracting peak frequencies and spectral energy distributions that reflect fault characteristics, constructing an initial multi-source frequency domain feature set, and filtering out effective features strongly correlated with equipment status through anomaly identification and secondary verification mechanisms, providing highly discriminative input for preliminary fault classification. The fault classification module, based on frequency domain features, uses threshold judgment and matching degree calculation to initially identify potential faults of discharge-dominant, discharge-suspected, and non-discharge-dominant types. A dynamic weight adjustment strategy based on information gain is employed to optimize the classification results, improving the ability to identify fuzzy fault states. The feature fusion module integrates the preliminary classification results with features extracted from post-cleaning data, such as oil gas concentration trends and winding vibration time-domain waveforms, in a multi-dimensional manner. Variance contribution rate analysis is used for feature dimensionality reduction and optimization, generating a feature vector that comprehensively and compactly represents the overall state of the transformer, overcoming the high dimensionality and redundancy problems of traditional "feature stacking" methods. The health assessment module uses a pre-trained random forest classification model to match the generated comprehensive feature vector with a historical fault pattern library, introducing a secondary verification mechanism based on confidence thresholds. This module not only outputs fault types and probability distributions but also, by analyzing similarity change trends, sensitively captures the critical process of equipment health status shifting from quantitative accumulation to sudden qualitative change, achieving high-precision early warning of latent defects and sub-health states, providing a reliable decision-making basis for predictive maintenance of distribution transformers.

[0137] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the methods and systems described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0138] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for assessing the health status of distribution transformers by integrating multi-source data, characterized in that, Includes the following steps: Step S1: Obtain a multi-source monitoring dataset containing timestamps from the distribution transformer monitoring equipment. The multi-source monitoring dataset includes oil gas concentration data, partial discharge pulse sequence, and winding vibration response sequence. Step S2: Preprocess the multi-source monitoring dataset to remove noise and interference, and obtain a cleaned signal dataset; Step S3: Perform time correction processing on the cleaned signal dataset to obtain a time-aligned unified signal dataset; Step S4: Perform frequency domain transformation on the time-aligned unified signal dataset, extract the peak frequency and amplitude distribution of each signal sequence, and form a multi-source frequency domain feature set; Step S5: Determine whether the peak frequency of the partial discharge pulse sequence in the multi-source frequency domain feature set exceeds a preset frequency threshold. If yes, determine that there is a potential discharge-dominant fault; otherwise, determine that there is a non-discharge-dominant potential fault. Calculate the matching degree between the potential discharge-dominant fault and the amplitude distribution characteristics of the winding vibration response sequence in the multi-source frequency domain feature set. Determine whether the matching degree is higher than a preset matching threshold. If yes, determine that there is a discharge-dominant potential fault; otherwise, determine that there is a suspected discharge-dominant potential fault. Statistically analyze the distribution of discharge-dominant potential faults, suspected discharge-dominant potential faults, and non-discharge-dominant potential faults to obtain preliminary classification results. For suspected discharge-dominant potential faults in the preliminary classification results, adjust the weights of the frequency domain features of the corresponding signal sequence in the multi-source frequency domain feature set. Based on the adjusted feature weights, the matching degree is re-evaluated, the preliminary classification results are optimized, and the optimized classification results of various potential faults are integrated to obtain the final classification result. Step S6: Combine the classification results with the time-series trend vector of the oil gas concentration data in the cleaned signal dataset and the time-domain waveform features of the winding vibration response sequence to generate a comprehensive fault feature vector; Step S7: Calculate the matching similarity between the comprehensive fault feature vector and the standard feature vector in the historical fault mode library using a pre-trained classification model; determine whether the critical process of the distribution transformer's health status changing from quantitative to qualitative change has been captured based on the rate of change of the matching similarity over N consecutive monitoring periods, where N is a positive integer. Extract the fault state category with the highest matching similarity and the corresponding fault type probability distribution. Combine the critical process judgment results to mark the development stage of the fault and generate the initial health status assessment results. Identify the results in the initial health status assessment results with a confidence level lower than the preset confidence threshold. Extract the key matching features from the comprehensive feature vector from the identified results and perform secondary verification in combination with the cleaned signal dataset. Based on the secondary verification results, the probability distribution of fault types is adjusted and the initial health status assessment results are optimized to generate and output the health status assessment results of the distribution transformer.

2. The method according to claim 1, characterized in that, Step S1 includes: Data on oil gas concentration, partial discharge pulse sequence, and winding vibration response sequence were collected using power distribution transformer monitoring equipment. The oil gas concentration data, the partial discharge pulse sequence, and the winding vibration response sequence are integrated to form an initial monitoring dataset. Add corresponding acquisition timestamp information to each signal sequence in the initial monitoring dataset; The initial monitoring dataset is sorted along a timeline based on the collected timestamp information. Record the acquisition source and acquisition time range of each signal sequence in the initial monitoring dataset; The integrated and sorted signal sequences, corresponding acquisition timestamp information, acquisition sources, and acquisition time range records are used to generate a multi-source monitoring dataset containing timestamps.

3. The method according to claim 2, characterized in that, Step S2 includes: Each signal sequence in the multi-source monitoring dataset is decomposed using a signal decomposition and reconstruction method; Based on the signal sequence decomposition results, high-frequency noise and electromagnetic interference components in the partial discharge pulse sequence and the winding vibration response sequence are filtered out, and the gas concentration data in the oil is smoothed. The signal sequences, after being filtered and smoothed to remove high-frequency noise and electromagnetic interference, are reassembled to form a preliminary processed dataset. The effectiveness of the preliminary processed dataset is verified by comparing the quality of each signal in the preliminary processed dataset with the corresponding original signal. Identify outliers in the preliminary processed dataset and perform interpolation on the outliers; The preliminary processed dataset, after validity verification and outlier interpolation, is determined as the cleaned signal dataset.

4. The method according to claim 1, characterized in that, Step S3 includes: The time delay offset between the oil gas concentration data, the winding vibration response sequence, and the partial discharge pulse sequence in the cleaned signal dataset is calculated using the cross-correlation function. The hysteresis adjustment amount of the oil gas concentration data and the winding vibration response sequence in the cleaned signal data is determined based on the time delay offset. The time correction is performed on the oil gas concentration data and winding vibration response sequence in the cleaned signal dataset based on the hysteresis adjustment amount. The time-corrected oil gas concentration data, partial discharge pulse sequence, and winding vibration response sequence are integrated to form a candidate unified dataset. The accuracy of time alignment is verified by comparing the timestamps of each signal sequence in the candidate unified dataset. The candidate unified datasets, verified for accuracy, were identified as time-aligned unified signal datasets.

5. The method according to claim 1, characterized in that, Step S4 includes: Perform frequency domain transformation on each signal sequence in the time-aligned unified signal dataset; Based on the frequency domain transformation results, the peak frequency of each signal sequence is extracted, and the amplitude distribution characteristics of each signal sequence in different frequency ranges are statistically analyzed. The peak frequency and corresponding amplitude distribution characteristics of a single signal sequence are combined to generate a frequency domain feature vector; Integrate the frequency domain feature vectors of all signal sequences to form an initial multi-source frequency domain feature set; Identify anomalous frequency domain features in the initial multi-source frequency domain feature set, and perform secondary verification on the anomalous frequency domain features; The effective frequency domain feature vectors after secondary verification are retained and re-integrated to form a multi-source frequency domain feature set.

6. The method according to claim 1, characterized in that, Step S6 includes: From the cleaned signal dataset, the time-series trend vector of the gas concentration data in the oil is extracted, and the time-domain waveform features of the winding vibration response sequence are extracted. The time-domain waveform features include the vibration peak amplitude and the waveform period. By fusing the time-series trend vector with the time-domain waveform features, an initial fused feature set is obtained; The classification result is converted into a quantized feature vector, which is then added to the initial fusion feature set to obtain an expanded fusion feature set. Calculate the contribution rate of each feature in the extended fusion feature set, select core features based on the contribution rate and remove redundant features to complete feature dimensionality reduction optimization; The extended fusion feature set, after dimensionality reduction optimization, is normalized to generate a comprehensive fault feature vector.

7. A distribution transformer health status assessment system integrating multi-source data, used to implement the distribution transformer health status assessment method integrating multi-source data as described in any one of claims 1 to 6, characterized in that, The aforementioned distribution transformer health status assessment system that integrates multi-source data includes: The data acquisition module is used to acquire a multi-source monitoring dataset containing timestamps from the distribution transformer monitoring equipment. The multi-source monitoring dataset includes oil gas concentration data, partial discharge pulse sequences, and winding vibration response sequences. The data cleaning module is used to preprocess the multi-source monitoring dataset to remove noise and interference, and obtain a cleaned signal dataset. The time correction module is used to perform time correction processing on the cleaned signal dataset to obtain a time-aligned unified signal dataset. The feature extraction module is used to perform frequency domain transformation on the time-aligned unified signal dataset, extract the peak frequency and amplitude distribution of each signal sequence, and form a multi-source frequency domain feature set. The fault classification module is used to perform preliminary fault mechanism classification based on the multi-source frequency domain feature set and obtain the classification result. The feature module integrates the classification results with the time-series trend vector of the gas concentration data in the oil in the cleaned signal dataset and the time-domain waveform features of the winding vibration response sequence to generate a comprehensive fault feature vector. The health assessment module is used to perform matching analysis on the comprehensive fault feature vector through a pre-trained classification model and output the health status assessment results of the distribution transformer.

Images