A transformer partial discharge detection method and system based on multi-source signal fusion

Abstract

The application relates to the technical field of fault detection, and discloses a transformer partial discharge detection method and system based on multi-source signal fusion, wherein the method comprises the following steps: clock deviation self-calibration and frequency self-adaptive oversampling are adopted to synchronously collect multi-channel waveform sequences; adaptive adjustable wavelet transform and local kurtosis spectrum dynamic parameter selection are adopted for noise removal; a directional enhancement energy operator and an adaptive energy ratio causal constraint are adopted to extract feature points; and transformer boundary element velocity field and layered adaptive grid filtering are combined for positioning. Compared with the prior art which depends on time stamp alignment, each sensor clock source is independent, and there is a millisecond-level time deviation, which leads to a meter-level error in positioning, and causes the technical problems of fault type misjudgment and position error. The application improves the accuracy of partial discharge source positioning and the reliability of fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of fault detection technology, and in particular to a method and system for detecting partial discharge in transformers based on multi-source signal fusion. Background Technology

[0002] Currently, in online monitoring of partial discharge in transformers with voltage levels of 110kV and above, a multi-physics sensing array is typically deployed, consisting of high-frequency current, ultra-high-frequency, and ultrasonic sensors. These three sensors simultaneously acquire partial discharge signals. The common processing path in existing technologies is as follows: each acquisition channel operates based on an independent local clock source, and software alignment is performed afterward using the timestamps carried in the data packets of each channel. Alternatively, a unified clock source such as GPS / IRIG-B is connected at the front end for hardware synchronization. The time-aligned signal is then sent to a positioning algorithm based on time difference of arrival or time of flight to calculate the discharge source coordinates.

[0003] For example, in substations, the electromagnetic environment is complex and temperature changes drastically. Differences in the crystal oscillator frequency scaling factor and initial clock bias offset among the various sensor acquisition units result in millisecond-level relative deviations between the clock planes of different channels. When purely based on timestamp alignment, this deviation is converted into an equivalent error in the discharge source location algorithm—the ultrasonic wave propagation time or the electromagnetic wave arrival time difference—leading to positioning errors exceeding meters. This can cause internal oil-paper insulation surface discharge to be misidentified as oil bubble discharge or inter-turn discharge to be misidentified as external interference, severely impacting fault diagnosis and maintenance decisions.

[0004] Therefore, there is an urgent need for a full-process detection method that achieves sub-millisecond-level cross-channel time alignment without requiring modifications to on-site sensor hardware or relying on an external unified clock source, utilizing only the inherent correlation of the waveform characteristics of multi-source signals. This method should be able to control the discharge power supply location error within 30 cm, even with independent operation of local clocks in each channel and the presence of frequency drift and initial clock difference, thereby improving the accuracy of fault location and the reliability of insulation diagnosis. Summary of the Invention

[0005] To address the aforementioned technical shortcomings, the purpose of this invention is to propose a transformer partial discharge detection method based on multi-source signal fusion. This method aims to solve the technical problem that existing technologies rely on timestamp alignment, but each sensor clock source is independent and has millisecond-level time deviations, resulting in positioning errors of more than meters and causing misjudgments of fault types and incorrect locations. In particular, it is impossible to achieve high-precision time alignment without modifying the hardware clock of the field sensors.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides a transformer partial discharge detection method based on multi-source signal fusion.

[0007] The transformer partial discharge detection method based on multi-source signal fusion includes: Step S10: Acquire multi-source raw signals in the transformer partial discharge detection scenario. Based on the multi-source raw signals, use a clock deviation self-calibration and frequency adaptive oversampling triggering mechanism to perform cross-channel synchronous acquisition tasks and output multi-channel raw waveform sequences with synchronization compensation. Step S20: Based on the original multi-channel waveform sequence with synchronization compensation, the multi-channel signal denoising task is performed using an adaptive adjustable Q wavelet transform and a local kurtosis spectrum dynamic parameter selection mechanism, and the denoised signal sequence is output. Step S30: Based on the denoised signal sequence, a joint detection mechanism of directional enhancement Teager-Kaiser energy operator and energy ratio factor constraint is used to perform cross-physics feature point extraction task and output a multi-channel feature point set; Step S40: Perform a heterogeneous signal global alignment task based on the multi-channel feature point set, and output the three-channel signal and the joint time offset estimate; Step S50: Perform the joint positioning task of the discharge source based on the three-channel signals and the joint time offset estimate, and output the three-dimensional coordinates of the discharge source.

[0008] Preferably, step S10, which involves acquiring multi-source raw signals in a transformer partial discharge detection scenario, performing cross-channel synchronous acquisition based on the multi-source raw signals using a clock deviation self-calibration and frequency adaptive oversampling triggering mechanism, and outputting a multi-channel raw waveform sequence with synchronization compensation, specifically includes: Step S101: Set a high-frequency current sensor at the transformer grounding circuit to collect high-frequency current signals, set an ultra-high frequency sensor at the transformer tank or detection window to collect ultra-high frequency signals, set an ultrasonic sensor on the outer wall of the transformer tank or at the acoustic coupling position to collect ultrasonic signals, and use the high-frequency current signal, the ultra-high frequency signal and the ultrasonic signal as the multi-source original signal; Step S102: Establish clock mapping relationships between local sampling times and unified reference times for the high-frequency current signal, the ultra-high frequency signal, and the ultrasonic signal respectively:

[0009] in, For the first Each channel is mapped to a sampling time after a unified reference time. It is the channel number, and Indicates a high-frequency current channel. Indicates UHF channel, Indicates the ultrasonic channel. To standardize reference times, For the first Frequency scaling factor of each channel For the first The initial clock bias offset of each channel relative to a unified reference time. For the first Random phase noise and sampling jitter terms for each channel; Step S103: Inject a detection pulse signal that is distinct from the partial discharge detection frequency band into the transformer grounding wire, obtain the detection arrival time of the detection pulse signal in the high-frequency current channel, ultra-high frequency channel and ultrasonic channel, and construct the residual sequence between the detection arrival time and the geometric propagation arrival time by combining the geometric propagation distance from the sensor of each channel to the detection pulse injection point. Step S104: Based on the residual sequence, the frequency scaling factor and initial clock bias of each channel are estimated using recursive least squares. When a partial discharge pulse exceeding the preset adaptive noise threshold is detected in any channel, the high-frequency current channel, ultra-high frequency channel and ultrasonic channel are controlled to record the waveforms of the pre-trigger segment and the post-trigger segment, respectively, to obtain the original waveform sequence of the multi-channel with synchronous compensation.

[0010] Preferably, step S20, which involves performing multi-channel signal denoising based on the original multi-channel waveform sequence with synchronization compensation using an adaptive adjustable Q-wavelet transform and a local kurtosis spectrum dynamic parameter selection mechanism, and outputting a denoised signal sequence, specifically includes: Step S201: Perform a short-time Fourier transform on each channel of the original waveform sequence in the multi-channel original waveform sequence with synchronization compensation to obtain the time-frequency amplitude matrix of the corresponding channel, and calculate the local kurtosis value of each frequency bin according to the following formula:

[0011] in, For the first The first channel The local kurtosis value of each frequency bin. For the first The first channel in the The frequency compartment and the first Time-frequency amplitude within a time window For the first The first channel The average time-frequency amplitude of each frequency compartment For the first The first channel Standard deviation of time-frequency amplitude for each frequency compartment The number of time windows included in the statistics. The sampling point number, A small positive constant used to avoid a denominator of zero; Step S202: Search for significant peaks in the local kurtosis value sequence of each channel, obtain the center frequency and half-peak bandwidth corresponding to the significant peaks, and determine the optimal quality factor of the corresponding channel according to the following formula:

[0012] in, For the first The optimal quality factor for each channel. For the first The center frequency corresponding to the local kurtosis peak in each channel For the first The half-peak bandwidth corresponding to the significant peak value in each channel; Step S203: Configure the adjustable Q-wavelet transform decomposition parameters for each channel according to the optimal quality factor, perform adjustable Q-wavelet decomposition on the original waveform sequence of each channel to obtain multiple sub-band signals for the corresponding channel, and apply Bayesian thresholding criteria to adaptively shrink the sub-band signals:

[0013] in, For the first The first channel Bayesian threshold for each sub-signal, For sub-sequence numbers, For the first The first channel The noise variance of the individual signal. For the first The first channel The standard deviation of the signal in the individual. A small positive constant is used to avoid a denominator of zero; the sub-band signal after threshold shrinkage is reconstructed to obtain the denoised signal sequence.

[0014] Preferably, in step S202, the optimal quality factor is determined independently for the high-frequency current channel, the ultra-high frequency channel, and the ultrasonic channel respectively; wherein, the optimal quality factor corresponding to the ultra-high frequency channel has a value range of 1.0 to 2.5, the optimal quality factor corresponding to the high-frequency current channel has a value range of 3.0 to 5.0, and the optimal quality factor corresponding to the ultrasonic channel has a value range of 6.0 to 10.0.

[0015] Preferably, in step S30, the step of performing cross-physics feature point extraction based on the denoised signal sequence using a joint detection mechanism of directional enhancement Teager-Kaiser energy operator and energy ratio factor constraint, and outputting a multi-channel feature point set, specifically includes: Step S301: For the denoised signal sequence, the first... Denoising discrete signal of each channel The directionality-enhancing Teager-Kaiser energy value is calculated according to the following formula:

[0016] in, For the first The first channel The directional enhancement of the Teager-Kaiser energy value at each sampling point For the first The first channel The amplitude of the denoised signal at each sampling point For the first The amplitude of the denoised signal from the previous sampling point in each channel. For the first The amplitude of the denoised signal at the last sampling point in each channel. For the first Positive enhancement factor for each channel, For the first Gradient enhancement factor for each channel; Step S302: Set the dynamic energy ratio at each sampling point according to the directionality enhancement Teager-Kaiser energy value. ; Step S303: When the first Dynamic energy ratio of each channel When the energy ratio threshold of the channel is exceeded for the first time, the local maximum point of the directional enhancement Teager-Kaiser energy near the corresponding sampling point in the channel is determined as a candidate feature arrival time marker. The absolute refractory period constraint and the causal consistency test between the high-frequency current channel, the ultra-high frequency channel and the ultrasonic channel are performed on the candidate feature arrival time marker to eliminate false markers caused by multipath reflection or electromagnetic oscillation, thus forming the multi-channel feature point set.

[0017] Preferably, step S40, which involves performing a global alignment task of heterogeneous signals based on the multi-channel feature point set and outputting the three-channel signal and the joint time offset estimate, specifically includes: Step S401: Using the UHF channel feature point set in the multi-channel feature point set as a reference, extract the cross-scale envelope trajectory of the denoised signals of the high-frequency current channel, UHF channel and ultrasonic channel respectively, and construct the cross-scale envelope trajectory matrix of each channel. Step S402: Construct a weighted local cost matrix based on the cross-scale envelope trajectory matrix; Step S403: Based on the weighted local cost matrix, use regularized dynamic time warp to solve for the estimated joint time offset of the high-frequency current channel and the ultrasonic channel relative to the ultra-high frequency channel, and perform time compensation on the three-channel signals according to the estimated joint time offset to obtain the time-synchronized three-channel signals.

[0018] Preferably, step S50, which involves performing a joint positioning task for the discharge source based on the three-channel signals and the joint time offset estimate, and outputting the three-dimensional coordinates of the discharge source, specifically includes: Step S501: Obtain the boundary element model of the transformer structure, and configure the equivalent propagation velocity of electromagnetic waves and the equivalent propagation velocity of sound waves based on the boundary element model of the transformer structure to form the equivalent propagation velocity field of the boundary element of the transformer structure. Step S502: Based on the three-channel signals after time synchronization, determine the arrival time of the high-frequency current signal, the arrival time of the ultra-high frequency signal, and the arrival time of the ultrasonic signal, and use the joint time offset estimate to perform time compensation on the arrival time of the high-frequency current signal and the arrival time of the ultrasonic signal to obtain the compensated multi-channel arrival time. Step S503: Based on the compensated multi-channel arrival time and the equivalent propagation velocity field of the transformer structural boundary element, determine the candidate discharge region, and use a weighted least squares Kalman filter recursive fusion within the candidate discharge region to output the three-dimensional coordinates of the discharge source.

[0019] This invention also provides a transformer partial discharge detection system based on multi-source signal fusion, comprising: The synchronous acquisition module is used to acquire multi-source raw signals in the transformer partial discharge detection scenario. Based on the multi-source raw signals, a clock deviation self-calibration and frequency adaptive oversampling triggering mechanism is used to perform cross-channel synchronous acquisition tasks and output multi-channel raw waveform sequences with synchronous compensation. An adaptive denoising module is used to perform multi-channel signal denoising based on the multi-channel original waveform sequence with synchronization compensation, using adaptive adjustable Q wavelet transform and local kurtosis spectrum dynamic parameter selection mechanism, and outputting a denoised signal sequence. The feature point extraction module is used to perform cross-physics field feature point extraction based on the denoised signal sequence by using a joint detection mechanism of directional enhancement Teager-Kaiser energy operator and energy ratio factor constraint, and output a multi-channel feature point set. The heterogeneous signal alignment module is used to perform a global alignment task of heterogeneous signals based on the multi-channel feature point set, and output the three-channel signal and the joint time offset estimate. The discharge source positioning module is used to perform a joint positioning task of the discharge source based on the three-channel signals and the joint time offset estimate, and output the three-dimensional coordinates of the discharge source.

[0020] The present invention also provides a transformer partial discharge detection device based on multi-source signal fusion. The transformer partial discharge detection device based on multi-source signal fusion includes: a memory, a processor, and a transformer partial discharge detection program based on multi-source signal fusion stored in the memory and executable on the processor. When the transformer partial discharge detection program based on multi-source signal fusion is executed by the processor, it implements the above-mentioned method.

[0021] The present invention also provides a computer program product, the computer program product including a transformer partial discharge detection program based on multi-source signal fusion, wherein the transformer partial discharge detection program based on multi-source signal fusion implements the above method when executed by a processor.

[0022] The beneficial effects of this invention are as follows: This invention achieves high-precision synchronous acquisition of multi-channel raw waveform sequences through clock deviation self-calibration and frequency adaptive oversampling triggering mechanism, without relying on an external unified clock source or modifying the hardware circuit of the field sensors. It suppresses cross-channel time alignment errors caused by clock deviation and provides a reliable premise for subsequent signal processing.

[0023] This invention achieves adaptive adjustment Wavelet transform and local kurtosis spectrum dynamic parameter selection mechanism achieve differentiated denoising for the time-frequency characteristics of signals from different physical channels, effectively protecting the transient characteristics of partial discharge pulses. Then, multi-scale envelope feature pyramid and structural geometric constraint dynamic time warp technology are used to align heterogeneous signals, solving the serious impact of millisecond-level synchronization deviation on positioning. Finally, the combination of transformer structure boundary element equivalent propagation velocity field and hierarchical grid search-Kalman filter recursive positioning improves the accuracy and robustness of partial discharge source positioning and reduces the risk of fault misjudgment. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the first embodiment of a transformer partial discharge detection method based on multi-source signal fusion according to the present invention. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0026] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Example 1: As Figure 1 The diagram shown is a flowchart of the first embodiment of a transformer partial discharge detection method based on multi-source signal fusion according to the present invention. The first embodiment of the transformer partial discharge detection method based on multi-source signal fusion according to the present invention is presented.

[0028] In the first embodiment, the transformer partial discharge detection method based on multi-source signal fusion includes: Step S10: Acquire multi-source raw signals in the transformer partial discharge detection scenario. Based on the multi-source raw signals, use a clock deviation self-calibration and frequency adaptive oversampling triggering mechanism to perform cross-channel synchronous acquisition tasks and output multi-channel raw waveform sequences with synchronization compensation. The multi-source raw signals in this step refer to the partial discharge raw response waveforms sensed by the UHF sensor, high-frequency current sensor, and ultrasonic sensor deployed on the transformer. The clock skew self-calibration mechanism injects a self-test pulse with a known pulse width and defined propagation geometry into the transformer grounding wire. This pulse is simultaneously received by the three channels or recorded via electrical coupling. The characteristic arrival time of each channel for this pulse is extracted. The residual between the measured arrival time and the theoretical arrival time calculated based on the transformer size and signal propagation speed is used as input. A recursive least squares algorithm is then used to estimate the frequency scaling factor of the local clock for each channel online. and initial clock bias offset This completes the clock parameter calibration. The frequency adaptive oversampling trigger mechanism immediately initiates an oversampling process under an independent clock when the signal amplitude detected in any channel exceeds the adaptive threshold calculated in real-time based on background noise. This is done at an integer multiple of the nominal sampling rate of each channel, and simultaneously triggers waveform recording across all three channels, including pre-trigger and post-trigger segments. The pre-trigger segment ensures complete preservation of the pulse leading edge information. The resulting multi-channel original waveform sequence with synchronization compensation refers to the sequence composed of the discrete sampled values ​​of the three channels after the aforementioned clock parameter compensation, and its time reference has been initially unified during acquisition.

[0029] After adopting a clock deviation self-calibration and frequency adaptive oversampling triggering mechanism, the millisecond-level random clock deviation between the three channels is estimated online and compensated according to the frequency scaling factor and clock offset of each channel. After compensation, the relative time error between the waveforms of each channel is compressed to the subsampling interval level. Originally, due to the independent operation of the crystal oscillators of each sensor acquisition unit and factors such as aging and temperature drift, the overall waveform of one channel was offset by several sampling points or even more relative to another channel. After synchronous compensation, these offsets are suppressed, and the ultra-high frequency, high frequency current, and ultrasonic signals obtain a unified time reference in the time domain. This allows the time domain segmentation of the short-time Fourier transform window or wavelet transform used in subsequent denoising processing to be accurately aligned with the effective range of the same discharge event, without mixing background noise segments into the signal segments due to time misalignment. It also avoids the signal transient characteristics being fragmented or blurred during window truncation, providing a reliable premise for cross-channel joint feature analysis and localization.

[0030] Traditional multi-source signal acquisition generally relies on two approaches: First, adding a hardware synchronization line or a unified GPS / IRIG-B clock source at the front end to lock each acquisition channel to the same time reference. However, the long distance between the transformer's field sensors and the acquisition unit, the complex electromagnetic environment, and the difficulty and high cost of wiring, coupled with the introduction of transmission jitter and electromagnetic interference by long-distance clock distribution networks, make it difficult to guarantee implementation integrity. Second, after independent acquisition by each channel, post-processing software alignment is performed using the nominal sampling clock and the local timestamp. However, the frequency scaling factor and initial phase differences of the crystal oscillators in each channel result in millisecond-level clock plane deviations, which change slowly over time. Post-processing linear interpolation cannot eliminate high-order random phase noise, and the residual deviation is converted into meter-level spatial errors in power supply positioning. This step injects a self-test pulse into the transformer's own grounding circuit, eliminating the need for additional synchronization cables. It utilizes recursive least squares to track the clock parameters of each channel online and perform real-time compensation, effectively suppressing initial positioning deviations caused by asynchrony and controlling time alignment errors from the acquisition source.

[0031] For example, in a 110kV transformer online monitoring system, an ultra-high frequency (UHF) sensor is installed at the drain valve, a high-frequency current sensor is mounted on the core grounding lead, and an ultrasonic sensor is attached to the outer wall of the tank. The three acquisition units operate independently. Upon startup, the system first injects a 5-nanosecond self-test pulse through the grounding wire. This pulse propagates at near-light speed along the metal conductor and electromagnetic path. Theoretically, the arrival time difference between the UHF channel and the high-frequency current channel relative to the pulse injection point is 0 nanoseconds and approximately 2 nanoseconds, respectively. While the ultrasonic sensor cannot directly receive this pulse, its transmission cable generates an extremely short electrical pulse through coupling. The residual is obtained by comparing the measured time with the theoretical time. For example, in one measurement, the residual was 0.8 nanoseconds for UHF, 3.3 nanoseconds for high-frequency current, and 1.2 nanoseconds for ultrasonic characterization. Based on this, an observation equation is constructed, and the frequency scaling factor and initial clock bias are estimated using recursive least squares, completing the calibration. Subsequently, during continuous monitoring, the system automatically injects self-test pulses for recalibration at regular intervals or when the ambient temperature changes significantly, ensuring that the synchronization deviation between the three-channel waveform sequences remains within less than one sampling interval during long-term operation, thus providing well-time-aligned raw data for subsequent steps.

[0032] Step S20: Based on the original multi-channel waveform sequence with synchronization compensation, the multi-channel signal denoising task is performed using an adaptive adjustable Q wavelet transform and a local kurtosis spectrum dynamic parameter selection mechanism, and the denoised signal sequence is output. By using optimal quality factor selection guided by local kurtosis spectrum and independent channel denoising, the effective partial discharge pulse components and background noise of each channel signal are more finely distinguished. The steep transient pulse rise edges of the UHF channel are not overly smoothed due to the use of a lower quality factor, thus preserving rise time details. The oscillation period and attenuation characteristics of the damped oscillation waveform in the high-frequency current channel are not smoothed out. The noise baseline of the ultrasonic channel, affected by power frequency vibration and low-frequency environmental interference, is effectively filtered out, while the pulse packet profile remains intact after noise removal. The signal-to-noise ratio of each channel signal is improved after denoising, resulting in a significant reduction in noise-induced false energy spikes when applying the directional enhancement Teager-Kaiser energy operator in step S30. This reduces the false alarm rate of pulse arrival time detection and helps improve the accuracy and consistency of feature point extraction.

[0033] Traditional denoising methods typically employ fixed wavelet basis functions and uniform quality factors, or apply the same threshold rule to all channels. This fails to account for the fundamentally different time-frequency characteristics of ultra-high frequency (UHF), high-frequency current, and ultrasonic sensors, which are physically distinct. UHF signals have extremely short pulse durations and very steep rise edges, requiring high-time-resolution wavelet transforms, i.e., low... Factor, if a conventional fixed high Value wavelet decomposition blurs the steep rising edge of the signal, causing nanosecond-level shifts in the localization of subsequent feature points; and since ultrasonic signals are narrowband mechanical waves, if low-frequency wavelet decomposition is used... Factor-based denoising often fails to separate interference components generated by transformer vibration and core magnetostriction that overlap with the partial discharge pulse frequency band, resulting in incomplete denoising. This step utilizes the local kurtosis spectrum to dynamically determine the independent optimal quality factor for each channel, enabling the wavelet transform's oscillation characteristics to adaptively match the time-frequency structure of the signal to be analyzed. This overcomes the problems of pulse feature loss or insufficient noise separation caused by fixed parameters, preserving the original physical characteristics of heterogeneous signals while ensuring denoising effectiveness.

[0034] Step S30: Based on the denoised signal sequence, a joint detection mechanism of directional enhancement Teager-Kaiser energy operator and energy ratio factor constraint is used to perform cross-physics feature point extraction task and output a multi-channel feature point set; Traditional feature point extraction methods often employ fixed threshold crossing or the standard Teager-Kaiser energy operator. Fixed thresholds are highly susceptible to signal amplitude fluctuations and changes in background noise levels, leading to missed detections at low signal-to-noise ratios and false detections during brief increases in interference. The standard TK energy operator is inherently symmetric; for asymmetric partial discharge pulses, its maximum output often occurs near the peak of the pulse, rather than at the physical discharge initiation moment. This introduces a rise-time-related systematic bias for different types of discharges, directly affecting the accuracy of time difference localization. Furthermore, traditional methods rarely impose physical causal constraints on the extracted feature points. Marker points extracted from multiple channels may exhibit temporal disorder, causing subsequent dynamic time warping to converge to an unreasonable warped path due to incorrect matching point relationships. This improved energy operator intentionally enhances sensitivity to the rise edge through a positive enhancement factor and gradient cube term, bringing feature points closer to the discharge initiation moment. Combined with an adaptive energy ratio threshold, refractory period, and causality check, false markers violating physical laws are eliminated, improving the robustness of subsequent alignment algorithms.

[0035] Step S40: Perform a heterogeneous signal global alignment task based on the multi-channel feature point set, and output the three-channel signal and the joint time offset estimate; This step takes the multi-channel feature point set obtained in step S30 and the denoised three-channel signal as input, and outputs the time-synchronized three-channel signal and the joint time offset estimate. The construction of the multi-scale Gaussian Laplace envelope feature pyramid uses the feature point set of the UHF channel as a reference. First, the absolute value of the denoised signal is taken to obtain the signal amplitude envelope. Then, a series of exponentially progressively scaled Gaussian kernels are applied for smoothing. The Laplace difference between adjacent smoothing scales is calculated to obtain a set of multi-scale feature vectors characterizing the change of the signal envelope shape at different scales. Then, a local window is extracted centered on each feature point and arranged to form a feature trajectory matrix. In the weighted regularization dynamic time warp process, a geometric penalty term is incorporated into the local cost matrix. This penalty term is constructed based on the allowable range of the equivalent propagation distance and physical size inside the transformer. High penalties or hard constraints are applied to pairing relationships in the warp path that exceed the reasonable propagation time upper limit, forcing the alignment path to remain within the physically feasible region. By solving the weighted regularization objective function, the joint time offset estimate of each channel relative to the UHF channel is obtained, and it is decomposed into an integer delay part and a fractional delay part. For the fractional delay part, a method based on... The piecewise parabolic interpolation filter of the polynomial is used for resampling compensation to finally obtain the three-channel signal after time synchronization. The joint time offset estimate is then directly used for the propagation time calculation in the positioning target function in step S50.

[0036] The multi-scale Gaussian Laplace envelope feature pyramid transforms the original time-domain waveform into a robust shape feature representation, insensitive to fluctuations in instantaneous signal amplitude and differences in inter-channel sensitivity. It extracts the structural commonality of the discharge pulse envelope profile in multi-scale space, thus overcoming the interference of significant waveform morphology differences between UHF, high-frequency current, and ultrasonic signals caused by different sensing mechanisms on time alignment. The geometric penalty term added to the dynamic time warp effectively limits the path search range, preventing the ultrasonic channel pulse from being incorrectly aligned to subsequent reflected waves or the next discharge event of the UHF channel, ensuring the alignment results have physical interpretability. The decomposition of integer and fractional delays and... Interpolation compensation enables the three-channel signals to achieve a high degree of overlap at the sub-sampling interval level near the feature points corresponding to the same discharge event, obtaining accurate cross-channel time delay, and providing high-precision time difference and time-of-flight information that can be directly substituted for joint positioning.

[0037] Existing multi-channel signal alignment methods mostly employ cross-correlation or standard dynamic time warping algorithms. Cross-correlation methods assume high similarity between signal waveforms, but UHF and high-frequency currents, as well as ultrasonic signals, differ greatly in frequency components, envelope shape, and duration. Direct cross-correlation often produces spurious peaks, leading to incorrect time delay estimations. While standard dynamic time warping allows for local time stretching and compression, it doesn't incorporate the physical constraints of the transformer structure. During alignment, it may generate warped paths spanning unreasonable time intervals, such as aligning the pulse corresponding to the actual discharge in an ultrasonic signal to the second reflected pulse of the UHF channel instead of the direct wave, resulting in an offset containing hundreds of microseconds of systematic error, leading to completely inaccurate subsequent positioning. This step utilizes an envelope feature pyramid for feature space mapping, improving the comparability between the three dissimilar signals. The geometric constraint penalty term incorporates domain knowledge of the actual transformer size and equivalent propagation speed into the path search, ensuring the physical validity of the path. This makes the final time offset estimate closer to the actual propagation delay and less affected by waveform distortion and noise.

[0038] For example, in a certain detection instance, a feature point of the UHF channel was set to a time reference of zero microseconds. Based on the geometric relationship between the power supply inside the transformer and the sensor, the actual propagation delay of the high-frequency current channel relative to this feature point was approximately 1.2 microseconds, and the ultrasonic channel delay was approximately 240 microseconds. When constructing the envelope feature pyramid, a Gaussian smoothing scale set of {2, 4, 8, 16} sampling points was used. The four feature dimensions generated reflected the envelope curvature information of the signal under different smoothing levels. During the dynamic time warp alignment process, if the offset of the ultrasonic channel was set to zero, i.e., directly aligned with the UHF reference point, the corresponding geometric penalty term would calculate a maximum allowable delay of approximately 429 microseconds based on the maximum external dimension of the transformer (60 cm) and the ultrasonic velocity in oil (approximately 1400 m / s), imposing a certain penalty value. However, the distance of this candidate path in the feature space was still extremely large, prompting the optimization process to shift towards the actual delay. In the final joint time offset estimate, the high-frequency current channel offset is 1.3 sampling points (0.3 fractional part), and the ultrasonic channel offset is 192.2 sampling points (corresponding to 240.25 microseconds at a sampling rate of 80 MHz, with a fractional part of 0.25).

[0039] Step S50: Perform the joint positioning task of the discharge source based on the three-channel signals and the joint time offset estimate, and output the three-dimensional coordinates of the discharge source.

[0040] This step utilizes the time-synchronized three-channel signals output from step S40 and their joint time offset estimate, combined with the transformer structure boundary element equivalent propagation velocity field and a hierarchical adaptive grid search and Kalman filter recursive mechanism to perform a joint localization task. The equivalent propagation velocity field lookup table is constructed offline based on a precise 3D CAD model of the transformer windings, core, insulation components, and tank. It uses the boundary element method to solve for the equivalent propagation velocity field of sound waves considering reflection, refraction, and blocking effects at the dielectric interface. The lookup table covers all spatial grid nodes within the transformer, with each node storing its equivalent propagation velocity and equivalent propagation distance to each ultrasonic sensor. The joint localization objective function consists of a weighted sum of the electromagnetic wave arrival time difference term and the sound wave flight time term. The propagation time in the acoustic term is calculated by dividing the equivalent propagation distance by the node equivalent sound velocity obtained from the lookup table. The fusion weights are adaptively adjusted based on the variance of the dynamic time warp path cost from step S40 to balance the uncertainties of the two types of measurements. The hierarchical adaptive grid search first traverses the entire solution space with a coarser grid step size to quickly filter out multiple candidate regions with lower objective function values. Then, it performs a fine search within these candidate regions with a smaller grid step size to obtain the optimal coordinate solution, which is used as the initial value of the weighted least squares Kalman filter recursive. The filter recursive integrates the arrival time series information of multiple partial discharge events and gradually converges through time updates and measurement updates, outputting the final three-dimensional coordinates of the discharge source.

[0041] After achieving time synchronization at the sampling interval level for multi-source heterogeneous signals, the joint time offset estimate directly provides the localization model with a reliable time difference of arrival between the UHF and high-frequency current channels and the absolute time of flight of the ultrasonic channel. The combined use of electromagnetic and acoustic physical information enhances the observability of the localization problem, reducing the dependence of the 3D coordinate solution on a single measurement. The equivalent propagation velocity field lookup table corrects for modeling errors introduced by the bending of the propagation path and local sound velocity variations caused by the transformer's internal winding core obstruction, insulation plate reflection, and the composite medium of the paperboard, making the physical model of the ultrasonic localization term more realistic. The hierarchical grid search strategy avoids the huge computational overhead of ultra-fine grid search in the 3D global space, while ensuring localization accuracy through a combination of coarse and fine searches. The Kalman filter recursive, by fusing measurements of multiple discharge events, suppresses the influence of single arrival time picking errors and random interference, enabling the localization results to converge stably during continuous monitoring. The deviation between the final output discharge source coordinates and the actual location is controlled within a small range.

[0042] Traditional transformer partial discharge localization methods typically rely on the assumption of straight-line propagation in a uniform medium, using a single equivalent sound velocity for spherical or hyperboloidal intersection localization. However, inside actual transformers, ultrasonic waves experience significant path bending as they traverse different filling media and pass through winding wires and core sidewalls, with equivalent sound velocities varying by over 10% in different directions. Ignoring this difference and directly using a uniform sound velocity introduces systematic localization errors. Furthermore, common methods often rely solely on the arrival time data of a single discharge event, making the results extremely sensitive to sensor geometry and latency acquisition errors. The localization point often fluctuates significantly around its actual location, lacking stability. This step calibrates the sound propagation path and velocity point-by-point using a boundary element equivalent propagation velocity field lookup table, reducing fixed errors introduced by model simplification. The combination of hierarchical search and Kalman filtering balances computational feasibility with robustness in the localization process. By recursively fusing multiple measurements, the actual discharge source location is gradually approximated, overcoming the divergent and unreliable results of single-shot localization in traditional methods.

[0043] Example 2: Furthermore, the present invention provides a transformer partial discharge detection system based on multi-source signal fusion, which employs a transformer partial discharge detection method based on multi-source signal fusion as described in the above embodiments, and can solve the technical problem of transformer partial discharge detection based on multi-source signal fusion. The beneficial effects of the transformer partial discharge detection system based on multi-source signal fusion provided by the present invention are the same as those of the transformer partial discharge detection method based on multi-source signal fusion provided in the above embodiments, and other technical features of the transformer partial discharge detection system based on multi-source signal fusion are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0044] Example 3: This invention provides a transformer partial discharge detection device based on multi-source signal fusion. The device includes at least one processor and a memory communicatively connected to the processor. The memory stores instructions executable by the processor, which are then executed to enable the processor to perform the transformer partial discharge detection method based on multi-source signal fusion described in Example 1. This transformer partial discharge detection device based on multi-source signal fusion, as described in this invention, can include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. This transformer partial discharge detection device based on multi-source signal fusion is merely an example and should not be construed as limiting the functionality or scope of this invention. A transformer partial discharge detection device based on multi-source signal fusion may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory or a program loaded from a storage device into a random access memory. The random access memory also stores various programs and data required for the operation of the transformer partial discharge detection device based on multi-source signal fusion. The processing unit, read-only memory, and random access memory are interconnected via a bus. An I / O interface is also connected to the bus. Typically, the following systems can be connected to the I / O interface: input devices including touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices including liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices including magnetic tapes, hard disks, etc.; and communication devices. The communication device allows the transformer partial discharge detection device based on multi-source signal fusion to communicate wirelessly or wiredly with other devices to exchange data. Although a transformer partial discharge detection device based on multi-source signal fusion with various systems has been described, it should be understood that implementation or possession of all the described systems is not required. It can be implemented alternatively or with more or fewer systems.

[0045] Example 4: This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the transformer partial discharge detection method based on multi-source signal fusion as described above. The computer program product provided by this invention can solve the technical problem of transformer partial discharge detection based on multi-source signal fusion. Compared with the prior art, the beneficial effects of the computer program product provided by this invention are the same as those of the transformer partial discharge detection method based on multi-source signal fusion provided in the above embodiments, and will not be repeated here.

[0046] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a read-only memory. When the computer program is executed by a processing device, it performs the functions defined in the methods of the embodiments disclosed in this invention.

[0047] It should be understood that the various parts disclosed in this invention can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0048] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the present invention and its equivalents, the present invention also intends to include these modifications and variations.

Claims

1. A method for detecting partial discharge in transformers based on multi-source signal fusion, characterized in that, The methods include: Step S10: Acquire multi-source raw signals in the transformer partial discharge detection scenario. Based on the multi-source raw signals, use a clock deviation self-calibration and frequency adaptive oversampling triggering mechanism to perform cross-channel synchronous acquisition tasks and output multi-channel raw waveform sequences with synchronization compensation. Step S20: Based on the original multi-channel waveform sequence with synchronization compensation, the multi-channel signal denoising task is performed using an adaptive adjustable Q wavelet transform and a local kurtosis spectrum dynamic parameter selection mechanism, and the denoised signal sequence is output. Step S30: Based on the denoised signal sequence, a joint detection mechanism of directional enhancement Teager-Kaiser energy operator and energy ratio factor constraint is used to perform cross-physics feature point extraction task and output a multi-channel feature point set; Step S40: Perform a heterogeneous signal global alignment task based on the multi-channel feature point set, and output the three-channel signal and the joint time offset estimate; Step S50: Perform the joint positioning task of the discharge source based on the three-channel signals and the joint time offset estimate, and output the three-dimensional coordinates of the discharge source.

2. The transformer partial discharge detection method based on multi-source signal fusion as described in claim 1, characterized in that, Step S10 involves acquiring multi-source raw signals in a transformer partial discharge detection scenario, performing cross-channel synchronous acquisition based on the multi-source raw signals using a clock deviation self-calibration and frequency adaptive oversampling triggering mechanism, and outputting a multi-channel raw waveform sequence with synchronization compensation. Specifically, this includes: Step S101: Set a high-frequency current sensor at the transformer grounding circuit to collect high-frequency current signals, set an ultra-high frequency sensor at the transformer tank or detection window to collect ultra-high frequency signals, set an ultrasonic sensor on the outer wall of the transformer tank or at the acoustic coupling position to collect ultrasonic signals, and use the high-frequency current signal, the ultra-high frequency signal and the ultrasonic signal as multi-source raw signals; Step S102: Establish clock mapping relationships between local sampling times and unified reference times for the high-frequency current signal, the ultra-high frequency signal, and the ultrasonic signal respectively: in, For the first Each channel is mapped to a sampling time after a unified reference time. It is the channel number, and Indicates a high-frequency current channel. Indicates UHF channel, Indicates the ultrasonic channel. To standardize reference times, For the first Frequency scaling factor of each channel For the first The initial clock bias offset of each channel relative to a unified reference time. For the first Random phase noise and sampling jitter terms for each channel; Step S103: Inject a detection pulse signal that is distinct from the partial discharge detection frequency band into the transformer grounding wire, obtain the detection arrival time of the detection pulse signal in the high-frequency current channel, ultra-high frequency channel and ultrasonic channel, and construct the residual sequence between the detection arrival time and the geometric propagation arrival time by combining the geometric propagation distance from the sensor of each channel to the detection pulse injection point. Step S104: Based on the residual sequence, the frequency scaling factor and initial clock bias of each channel are estimated by recursive least squares. When a partial discharge pulse exceeding the preset adaptive noise threshold is detected in any channel, the high-frequency current channel, ultra-high frequency channel and ultrasonic channel are controlled to record the waveforms of the pre-trigger segment and the post-trigger segment respectively to obtain the original waveform sequence of the multi-channel channel with synchronous compensation.

3. The transformer partial discharge detection method based on multi-source signal fusion as described in claim 1, characterized in that, Step S20, based on the multi-channel original waveform sequence with synchronization compensation, employs an adaptive adjustable Q-wavelet transform and a local kurtosis spectrum dynamic parameter selection mechanism to perform multi-channel signal denoising and outputs a denoised signal sequence. This step specifically includes: Step S201: Perform a short-time Fourier transform on each channel of the original waveform sequence in the multi-channel original waveform sequence with synchronization compensation to obtain the time-frequency amplitude matrix of the corresponding channel, and calculate the local kurtosis value of each frequency bin according to the following formula: in, For the first The first channel The local kurtosis value of each frequency bin. For the first The first channel in the The frequency compartment and the first Time-frequency amplitude within a time window For the first The first channel The average time-frequency amplitude of each frequency compartment For the first The first channel Standard deviation of time-frequency amplitude for each frequency compartment The number of time windows included in the statistics. The sampling point number, A small positive constant used to avoid a denominator of zero; Step S202: Search for significant peaks in the local kurtosis value sequence of each channel, obtain the center frequency and half-peak bandwidth corresponding to the significant peaks, and determine the optimal quality factor of the corresponding channel according to the following formula: in, For the first The optimal quality factor for each channel. For the first The center frequency corresponding to the local kurtosis peak in each channel For the first The half-peak bandwidth corresponding to the significant peak value in each channel; Step S203: Configure the adjustable Q-wavelet transform decomposition parameters for each channel according to the optimal quality factor, perform adjustable Q-wavelet decomposition on the original waveform sequence of each channel to obtain multiple sub-band signals for the corresponding channel, and apply Bayesian thresholding criteria to adaptively shrink the sub-band signals: in, For the first The first channel Bayesian threshold for each sub-signal, For sub-sequence numbers, For the first The first channel The noise variance of the individual signal. For the first The first channel The standard deviation of the signal in the individual. A small positive constant is used to avoid the denominator being zero; the sub-band signal after threshold shrinkage is reconstructed to obtain the denoised signal sequence.

4. The transformer partial discharge detection method based on multi-source signal fusion as described in claim 3, characterized in that, In step S202, the optimal quality factor is determined independently for the high-frequency current channel, the ultra-high frequency channel, and the ultrasonic channel respectively; wherein, the optimal quality factor for the ultra-high frequency channel has a value range of 1.0 to 2.5, the optimal quality factor for the high-frequency current channel has a value range of 3.0 to 5.0, and the optimal quality factor for the ultrasonic channel has a value range of 6.0 to 10.

0.

5. The transformer partial discharge detection method based on multi-source signal fusion as described in claim 1, characterized in that, In step S30, based on the denoised signal sequence, a joint detection mechanism using the directional enhancement Teager-Kaiser energy operator and energy ratio factor constraint is employed to perform a cross-physics feature point extraction task and output a multi-channel feature point set. This step specifically includes: Step S301: For the denoised signal sequence, the first... Denoising discrete signal of each channel The directionality-enhancing Teager-Kaiser energy value is calculated according to the following formula: in, For the first The first channel The directional enhancement of the Teager-Kaiser energy value at each sampling point For the first The first channel The amplitude of the denoised signal at each sampling point For the first The amplitude of the denoised signal from the previous sampling point in each channel. For the first The amplitude of the denoised signal at the last sampling point in each channel. For the first Positive enhancement factor for each channel, For the first Gradient enhancement factor for each channel; Step S302: Set the dynamic energy ratio at each sampling point according to the directionality enhancement Teager-Kaiser energy value. ; Step S303: When the first Dynamic energy ratio of each channel When the energy ratio threshold of the channel is exceeded for the first time, the local maximum point of the directional enhancement Teager-Kaiser energy near the corresponding sampling point in the channel is determined as a candidate feature arrival time marker. The absolute refractory period constraint and the causal consistency test between the high-frequency current channel, the ultra-high frequency channel and the ultrasonic channel are performed on the candidate feature arrival time marker to eliminate false markers caused by multipath reflection or electromagnetic oscillation, forming a multi-channel feature point set.

6. The transformer partial discharge detection method based on multi-source signal fusion as described in claim 1, characterized in that, Step S40, which involves performing a global alignment task for heterogeneous signals based on the multi-channel feature point set and outputting the three-channel signal and joint time offset estimate, specifically includes: Step S401: Using the UHF channel feature point set in the multi-channel feature point set as a reference, extract the cross-scale envelope trajectory of the denoised signals of the high-frequency current channel, UHF channel and ultrasonic channel respectively, and construct the cross-scale envelope trajectory matrix of each channel. Step S402: Construct a weighted local cost matrix based on the cross-scale envelope trajectory matrix; Step S403: Based on the weighted local cost matrix, use regularized dynamic time warp to solve for the estimated joint time offset of the high-frequency current channel and the ultrasonic channel relative to the ultra-high frequency channel, and perform time compensation on the three-channel signals according to the estimated joint time offset to obtain the time-synchronized three-channel signals.

7. The transformer partial discharge detection method based on multi-source signal fusion as described in claim 6, characterized in that, Step S50, which involves performing a joint localization task for the discharge source based on the three-channel signals and the joint time offset estimate, and outputting the three-dimensional coordinates of the discharge source, specifically includes: Step S501: Obtain the boundary element model of the transformer structure, and configure the equivalent propagation velocity of electromagnetic waves and the equivalent propagation velocity of sound waves based on the boundary element model of the transformer structure to form the equivalent propagation velocity field of the boundary element of the transformer structure. Step S502: Determine the arrival time of the high-frequency current signal, the arrival time of the ultra-high frequency signal, and the arrival time of the ultrasonic signal based on the three-channel signals, and use the joint time offset estimate to perform time compensation on the arrival time of the high-frequency current signal and the arrival time of the ultrasonic signal to obtain the compensated multi-channel arrival time. Step S503: Based on the compensated multi-channel arrival time and the equivalent propagation velocity field of the transformer structural boundary element, determine the candidate discharge region, and use a weighted least squares Kalman filter recursive fusion within the candidate discharge region to output the three-dimensional coordinates of the discharge source.

8. A transformer partial discharge detection system based on multi-source signal fusion, applied to the transformer partial discharge detection method based on multi-source signal fusion as described in any one of claims 1 to 7, characterized in that, The system includes: The synchronous acquisition module is used to acquire multi-source raw signals in the transformer partial discharge detection scenario. Based on the multi-source raw signals, a clock deviation self-calibration and frequency adaptive oversampling triggering mechanism is used to perform cross-channel synchronous acquisition tasks and output multi-channel raw waveform sequences with synchronous compensation. An adaptive denoising module is used to perform multi-channel signal denoising based on the multi-channel original waveform sequence with synchronization compensation, using adaptive adjustable Q wavelet transform and local kurtosis spectrum dynamic parameter selection mechanism, and outputting a denoised signal sequence. The feature point extraction module is used to perform cross-physics field feature point extraction based on the denoised signal sequence by using a joint detection mechanism of directional enhancement Teager-Kaiser energy operator and energy ratio factor constraint, and output a multi-channel feature point set. The heterogeneous signal alignment module is used to perform a global alignment task of heterogeneous signals based on the multi-channel feature point set, and output the three-channel signal and the joint time offset estimate. The discharge source positioning module is used to perform a joint positioning task of the discharge source based on the three-channel signals and the joint time offset estimate, and output the three-dimensional coordinates of the discharge source.

9. A transformer partial discharge detection device based on multi-source signal fusion, characterized in that, The transformer partial discharge detection device based on multi-source signal fusion includes: a memory, a processor, and a transformer partial discharge detection program based on multi-source signal fusion stored in the memory and executable on the processor. When the transformer partial discharge detection program based on multi-source signal fusion is executed by the processor, it implements a transformer partial discharge detection method based on multi-source signal fusion as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a transformer partial discharge detection program based on multi-source signal fusion, which, when executed by a processor, implements a transformer partial discharge detection method based on multi-source signal fusion as described in any one of claims 1 to 7.

Images