A transformer fault prediction method based on multi-source data fusion

Abstract

The present application relates to the field of power equipment state monitoring, and particularly relates to a transformer fault prediction method based on multi-source data fusion. The method comprises the following steps: synchronously collecting ultra-high frequency partial discharge and contact acoustic signature signals, extracting atlas and frequency band features, calculating low-frequency baseband energy derivatives to preliminarily screen transient impact states, comparing the acoustic signature library by using a dynamic time warping algorithm to confirm transient decoupling modes or output unknown early warnings, removing shallow interference and extracting phase atlas and high-frequency residuals as target features for diagnosis in the decoupling mode, and finally outputting prediction warnings or suppressing working condition artifacts according to power frequency phase aggregation degree and resonance energy peak value. The present application effectively improves the system false alarm problem caused by extreme transient working conditions, and realizes feature decoupling and early prediction of deep composite faults of the transformer under strong interference.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring, and in particular to a method for predicting transformer faults based on multi-source data fusion. Background Technology

[0002] Power transformers are the core hub of the power grid, and multi-source online monitoring combining ultra-high frequency partial discharge and acoustic vibration has become the mainstream technology for assessing their health status. In actual substation environments, transformers inevitably need to undergo normal transient dispatching operations such as on-load tap changer operation, circuit breaker opening and closing, and grid load ride-through. These routine operating processes will inevitably generate extremely strong broadband mechanical vibrations and spatial electromagnetic inrush currents.

[0003] When faced with such transient conditions, existing monitoring technologies often employ fixed threshold alarms or blind blocking and shielding strategies, which are difficult to effectively decouple and eliminate strong interference illusions. This can easily lead to frequent false alarms in the system and may also cause the system to miss deep insulation degradation and mechanical fault characteristics that are masked by noise during the impact. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a transformer fault prediction method based on multi-source data fusion, aiming to improve the problems of existing technologies being prone to false alarms under transient impact conditions and having difficulty decoupling the true fault characteristics.

[0005] This invention provides the following technical solution: a transformer fault prediction method based on multi-source data fusion, comprising: S1. Synchronously acquire the UHF partial discharge signal and contact acoustic pattern signal of the transformer, divide the contact acoustic pattern signal into low-frequency baseband signal and high-frequency residual band signal, and extract the UHF partial discharge signal into time-domain amplitude features and phase spectrum features. S2. Calculate the first derivative of the energy of the low-frequency baseband signal in real time. If the first derivative of the energy exceeds the preset derivative threshold, it is determined that the transformer has entered the transient impact screening state and the waveform comparison process is triggered. S3. In the transient impact screening state, the envelope features of the contact acoustic signal are extracted, and the envelope features are compared with the preset operation acoustic library using the dynamic time warping algorithm. If the matching similarity reaches the preset similarity threshold, it is confirmed that the transient decoupling mode has been entered; otherwise, it is determined that the transformer has experienced an unknown abnormal impact and an abnormal impact warning is directly output. S4. When in the transient decoupling mode, the low-frequency baseband signal and the time-domain amplitude feature are removed, and the phase spectrum feature and the high-frequency residual band signal are extracted as the target decoupling feature. S5. Diagnose the target decoupling characteristics. If the phase spectrum characteristics show power frequency phase aggregation and the energy peak of the high frequency residual band signal exceeds the preset resonance threshold, output a fault prediction alarm; otherwise, determine it as a working condition impact artifact and suppress the alarm.

[0006] Preferably, in step S1, the step of dividing the signal into a low-frequency baseband signal and a high-frequency residual band signal includes: The contact acoustic pattern signal is processed by low-pass filtering to extract the vibration component of a preset low-frequency band on the transformer surface to generate the low-frequency baseband signal. The contact acoustic signal is processed by bandpass filtering to extract the vibration component of a preset high-frequency band; A time delay compensation operation is performed on the extracted vibration components of the preset high-frequency band to generate the high-frequency residual band signal that is synchronously aligned with the low-frequency baseband signal on the time axis.

[0007] Preferably, in step S2, the step of calculating the first derivative of the energy of the low-frequency baseband signal in real time includes: Construct a sliding time window along the time axis and calculate the cumulative energy value of the low-frequency baseband signal within the current sliding time window; Obtain the difference sequence of the accumulated energy values ​​between adjacent sliding time windows, and calculate the time change rate of the difference sequence; The time rate of change is subjected to noise reduction by applying a smoothing filter operator, and the smoothed numerical sequence is output as the first derivative of the energy.

[0008] Preferably, in step S2, the step of determining that the transformer has entered the transient impact preliminary screening state and triggering the waveform comparison process includes: Extract the basis noise derivative sequence of the transformer during its historical steady-state operation, and calculate the preset derivative threshold based on the statistical distribution model; The first derivative of the energy is compared with the preset derivative threshold. When the first derivative of the energy is greater than the preset derivative threshold within a continuous preset sampling period, a preliminary screening trigger command is generated. In response to the initial screening trigger command, the working mode of the monitoring system is switched to the transient impact initial screening state, and the waveform comparison process is triggered.

[0009] Preferably, in step S3, the step of extracting the envelope features of the contact voiceprint signal includes: The contact acoustic signal is analyzed using an envelope extraction operator to obtain an initial envelope curve; The initial envelope curve is resampled using an interpolation algorithm to obtain a resampled curve; An amplitude normalization operation is performed on the resampled curve to generate the envelope feature with a unified data dimension.

[0010] Preferably, in step S3, the step of comparing the envelope feature with a preset operation voiceprint database includes: Retrieve the standard operation template from the preset operation voiceprint library and construct the cost matrix between the envelope feature and the standard operation template; Find the alignment path with the minimum cumulative distance in the cost matrix, and obtain the best matching distance between the envelope feature and the standard operation template on the time axis; Based on the optimal matching distance, the matching similarity between the envelope feature and the standard operation template is calculated using a mapping function.

[0011] Preferably, in step S4, the step of extracting the phase spectrum features and the high-frequency residual band signal as target decoupling features includes: The synchronous reference signal of the transformer is obtained, and the ultra-high frequency partial discharge signal is mapped to the phase interval corresponding to the synchronous reference signal to construct the two-dimensional distributed phase spectrum features. Perform time-frequency transformation processing on the high-frequency residual band signal to extract the spectral amplitude distribution data of the high-frequency residual band signal in a preset high-frequency band; The phase map features are fused and spliced ​​with the spectral amplitude distribution data to generate the target decoupling features.

[0012] Preferably, in step S5, the step of outputting a fault prediction alarm includes: The pulse distribution density of the phase spectrum features within a preset power frequency cycle is statistically analyzed, and the phase aggregation probability of the pulse distribution density within a preset phase interval is calculated. In the target decoupling features, determine the maximum energy peak value corresponding to the high-frequency residual band signal; When the phase convergence probability is greater than a preset density threshold and the maximum energy peak is greater than the preset resonance threshold, the fault prediction alarm is generated.

[0013] Preferably, in step S5, the step of otherwise determining it as an impact artifact and suppressing the alarm includes: When the phase spectrum features do not show power frequency phase clustering, or the energy peak value of the high frequency residual band signal is not greater than the preset resonance threshold, the ultra-high frequency partial discharge signal and the contact acoustic pattern signal are determined to be working condition impact artifacts. Intercept the corresponding over-limit alarm signals and generate artifact masking instructions for the monitoring alarm link; Assign working condition impact artifact labels to the current UHF partial discharge signal and the contact acoustic pattern signal, and update them to the local background noise dataset.

[0014] The present invention has the following beneficial effects: 1. In this invention, an adaptive feature decoupling mechanism is proposed. It captures transient impacts through dual verification of energy derivative and dynamic time warping algorithm, actively eliminates shallow signals affected by interference and extracts high-frequency residuals and phase spectrum, effectively improves the false alarm problem caused by extreme working conditions such as switching action, and realizes the separation of deep fault features under strong interference.

[0015] 2. In this invention, a multi-source diagnostic framework for partial discharge and acoustic signature is constructed. The phase convergence probability of the ultra-high frequency signal and the mechanical resonance peak of the high frequency residual band are spliced ​​together to overcome the limitation of single sensing technology in identifying complex intertwined anomalies. This enables early and accurate prediction of composite faults caused by insulation degradation induced by micro-mechanical deformation.

[0016] 3. In this invention, a closed-loop feedback artifact adaptive suppression strategy is introduced. Based on the historical steady-state noise distribution, a comparison threshold is generated in real time. When the signal is determined to be an artifact of the operating condition, the alarm output is actively intercepted, and the artifact feature is labeled and updated to the local background dataset. This enables the monitoring system to have autonomous evolution capability and enhances the stability of the equipment in long-term operation. Attached Figure Description

[0017] Figure 1 This is a flowchart of a transformer fault prediction method based on multi-source data fusion proposed in this invention; Figure 2 This is a flowchart of the multi-source signal synchronous acquisition and group delay adaptive alignment proposed in this invention; Figure 3 This is a flowchart of the baseband energy differential differentiation and statistical adaptive initial screening proposed in this invention; Figure 4 This is a flowchart of the voiceprint pattern recognition and intelligent triage based on the DTW algorithm proposed in this invention. Figure 5 This is a flowchart of the deep multidimensional feature decoupling and multimodal feature fusion proposed in this invention; Figure 6 This is a flowchart of the electroacoustic combined dual threshold diagnosis and artifact adaptive suppression closed-loop control proposed in this invention. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 In a first embodiment of the present invention, the present invention provides a transformer fault prediction method based on multi-source data fusion, such as... Figure 1 As shown, it includes: S1. Synchronously acquire the UHF partial discharge signal and contact acoustic pattern signal of the transformer, divide the contact acoustic pattern signal into low frequency baseband signal and high frequency residual band signal, and extract the UHF partial discharge signal into time domain amplitude features and phase spectrum features. Further, step S1, the step of dividing the signal into a low-frequency baseband signal and a high-frequency residual band signal, includes: Low-pass filtering is applied to the contact acoustic pattern signal to extract the vibration component of the transformer surface in a preset low-frequency band to generate a low-frequency baseband signal. Bandpass filtering is applied to the contact acoustic signal to extract the vibration components of a preset high-frequency band. A time delay compensation operation is performed on the extracted vibration components of the preset high-frequency band to generate a high-frequency residual band signal that is synchronously aligned with the low-frequency baseband signal on the time axis.

[0020] Specifically, the system first acquires underlying signals using a UHF partial discharge sensor and a contact-type acoustic fingerprint sensor installed on the transformer body. The system's main control unit uses a unified hardware clock source to trigger the analog-to-digital conversion module, ensuring that both the original UHF partial discharge signal and the original contact acoustic fingerprint signal are acquired simultaneously, thus guaranteeing strict timestamp alignment of the multi-source data at the hardware physical level.

[0021] After acquiring continuous raw contact acoustic signals, the system uses a finite impulse response (FIR) digital filter to divide the frequency band. The cutoff frequency of the low-pass filter is set to cover the fundamental frequency of normal mechanical vibration of the transformer core and windings, while the upper and lower limit frequencies of the band-pass filter are set to lock in the high-frequency excitation band generated by local deformation or abnormal structural friction. Let the raw contact acoustic signal be... The unit impulse response of the low-pass filter is The unit impulse response of the bandpass filter is ,in For discrete-time series indexing. Low-frequency baseband signal. With the initially extracted high-frequency vibration components The extraction process satisfies the following discrete convolution formula: ; ; In the above formula, and These represent the orders of the low-pass and band-pass filters, respectively. This is the summation and integration variable for the convolution operation. Through the above operations, the physical characteristics of the complex broadband vibration signal on the transformer surface are initially decoupled.

[0022] Due to the dispersion effect of the transformer tank material on sound waves of different frequencies and the group delay differences caused by digital filters of different orders, the initially extracted high-frequency vibration components will be out of sync with the low-frequency baseband signal in terms of timing. The system determines a fixed discrete time delay compensation amount based on the pre-set difference in the propagation velocity of the sound wave in the transformer structure and the phase response parameters of the digital filter. Delay compensation is performed through signal sequence shifting operations to generate the final high-frequency residual band signal. Its formula is: ; This operation eliminates the time misalignment of heterogeneous frequency band signals during transmission and computation, ensuring that the high-frequency residual band signal and the low-frequency baseband signal are strictly synchronized and aligned on the time axis.

[0023] For the synchronously acquired UHF partial discharge raw signal, the system divides it into fixed time windows to extract time-domain amplitude features. Let the length of the time window be... The original signal sequence of the UHF partial discharge is Then the temporal amplitude characteristics within the current window Defined as the maximum absolute value of the electromagnetic pulse signal within the time window, the extraction formula is: ; in the formula This serves as the starting sampling index for the current time window. Simultaneously, the system acquires the transformer's power frequency AC voltage reference signal in real time and extracts the corresponding power frequency phase angle at the current moment via a phase-locked loop. The temporal amplitude features extracted within the aforementioned time window. Mapped to the corresponding power frequency phase angle Within the AC phase interval, after continuous two-dimensional scattering over multiple power frequency cycles, a two-dimensional distribution matrix reflecting the correlation between the partial discharge intensity inside the transformer and the power frequency voltage phase is constructed, thus completing the extraction of phase spectrum features.

[0024] This step, through rigorous clock synchronization acquisition and digital filtering delay compensation calculations, achieves precise alignment and feature stripping of multi-source acoustic and electrical signals of the transformer in the spatiotemporal dimension, providing a high-fidelity data foundation for subsequent diagnosis.

[0025] S2. Calculate the first derivative of the energy of the low-frequency baseband signal in real time. If the first derivative of the energy exceeds the preset derivative threshold, the transformer is determined to enter the transient impact screening state and the waveform comparison process is triggered. Further, step S2, the step of calculating the first derivative of the energy of the low-frequency baseband signal in real time, includes: Construct a sliding time window along the time axis and calculate the cumulative energy of the low-frequency baseband signal within the current sliding time window; Obtain the difference sequence of energy accumulation values ​​between adjacent sliding time windows, and calculate the time rate of change of the difference sequence; A smoothing filter operator is applied to the rate of change of time for noise reduction, and the smoothed numerical sequence is output as the first derivative of energy.

[0026] Further, in step S2, the step of determining that the transformer has entered the transient impact initial screening state and triggering the waveform comparison process includes: Extract the basis noise derivative sequence of the transformer during its historical steady-state operation, and calculate the preset derivative threshold based on the statistical distribution model; The first derivative of energy is compared with a preset derivative threshold. When the first derivative of energy is greater than the preset derivative threshold within a continuous preset sampling period, a preliminary screening trigger command is generated. In response to the initial screening trigger command, the monitoring system's operating mode is switched to transient impact initial screening state, and the waveform comparison process is triggered.

[0027] Specifically, after acquiring the low-frequency baseband signal, the system constructs a sliding time window with a fixed step size along the time axis to calculate the signal energy in real time. Let the number of sampling points included in the sliding time window be... The discrete sequence of the low-frequency baseband signal is Then the current number Energy accumulation value within a sliding time window The calculation formula is: ; in the formula Let be the step size of the adjacent sliding time windows along the time axis. As the time windows continue to slide, the system acquires the difference sequence of the accumulated energy values ​​between adjacent sliding time windows in real time and calculates the rate of change of this difference sequence with respect to time. Let the time interval corresponding to adjacent sliding time windows be . Then the first Rate of change over time corresponding to each window The calculation formula is: ; Because random mechanical vibration spikes exist in the transformer operating environment, directly using this time rate of change would lead to drastic numerical fluctuations. Therefore, the system employs a moving average smoothing filter operator to denoise the time rate of change sequence. Let the smoothing window length be... The numerical sequence output after smoothing and filtering is used as the first derivative of the energy of the low-frequency baseband signal. The calculation formula is as follows: ; To achieve adaptive identification of abnormal transient impacts on transformers, the system pre-extracts a sequence of base noise derivatives from the transformer's historical steady-state operation under conditions free from external interference, and calculates a preset derivative threshold based on a Gaussian statistical distribution model. Let the total number of energy first-order derivative samples extracted during the historical steady-state operation be... , No. The historical sample value is The system calculates the mathematical expectation of the historical sequence in sequence. with standard deviation The calculation formula is: ; ; The system then generates a preset derivative threshold for real-time comparison based on the aforementioned statistical parameters. The calculation formula is: ; in the formula This is a tolerance coefficient pre-set based on the anti-interference requirements of the transformer's operating environment.

[0028] In the real-time monitoring stream, the system compares the currently calculated first derivative of the energy frame by frame. With preset derivative threshold The magnitude of the energy is relative to the frequency of electromagnetic interference. To prevent false triggering caused by isolated electromagnetic interference or random impacts, the system's logic judgment unit will only generate a preliminary screening trigger command when the first derivative of the energy is continuously greater than a preset derivative threshold for multiple consecutive preset sampling periods. In response to this preliminary screening trigger command, the system directly switches the working mode of the underlying monitoring process from the normal feature accumulation state to the transient impact preliminary screening state, and simultaneously triggers the subsequent waveform comparison process targeting the acoustic signature features.

[0029] This step, through the joint determination of energy differential and statistical adaptive threshold, effectively overcomes the interference of transformer steady-state mechanical noise and realizes rapid and accurate perception of high-energy transient impact conditions in complex field environments.

[0030] S3. Under the transient impact screening state, extract the envelope features of the contact acoustic signal, and use the dynamic time warping algorithm to compare the envelope features with the preset operation acoustic library. If the matching similarity reaches the preset similarity threshold, it is confirmed to enter the transient decoupling mode; otherwise, it is determined that the transformer has experienced an unknown abnormal impact and an abnormal impact warning is directly output.

[0031] Further, step S3, the step of extracting the envelope features of the contact voiceprint signal, includes: The initial envelope curve is obtained by analyzing the contact acoustic signal using an envelope extraction operator. An interpolation algorithm is used to resample the initial envelope curve to obtain a resampled curve. Amplitude normalization is performed on the resampled curve to generate envelope features with uniform data dimensions.

[0032] Further, step S3, the step of comparing the envelope features with a preset operation voiceprint database, includes: Retrieve the standard operation template from the preset operation voiceprint library and construct the cost matrix between the envelope features and the standard operation template; Find the alignment path with the minimum cumulative distance in the cost matrix, and obtain the best matching distance between the envelope feature and the standard operation template on the time axis; Based on the optimal matching distance, the similarity between the envelope features and the standard operating template is calculated using a mapping function.

[0033] Specifically, after the monitoring system enters the transient impact initial screening state, it needs to perform pattern recognition on the transformer contact acoustic signal that causes energy mutations. The system first uses the Hilbert transform as an envelope extraction operator to analyze the contact acoustic signal. Let the discretized contact acoustic signal sequence be... The corresponding Hilbert transform result is The system constructs an analytical signal and calculates its instantaneous complex modulus to obtain the initial envelope curve. The calculation formula is as follows: ; Because the durations of different operations or impact events vary, in order to align with the template dimensions in the preset voiceprint library, the system uses a cubic spline interpolation algorithm to resample the initial envelope curve along the time axis, uniformly mapping its length to a preset standard number of points to obtain the resampled curve. Subsequently, extreme value amplitude normalization is performed on this resampled curve to generate envelope features with uniform data dimensions. The formula is: ; in the formula This is a resampling curve sequence. and These represent the global maximum and global minimum values ​​in the resampled curve sequence, respectively. This is the index of the normalized data points. This envelope feature removes the interference of absolute amplitude and purely characterizes the macroscopic fluctuations of the transient mechanical vibration energy of the transformer over time.

[0034] After extracting the envelope features, the system retrieves a standard operation template from a preset operation voiceprint library for comparison. This preset operation voiceprint library pre-stores standard energy envelope sequences for compliant operation procedures such as transformer on-load tap changer actions and circuit breaker opening and closing. Let the retrieved standard operation template sequence be... The system constructs envelope features With standard operating template Cost matrix between Local differences are measured using Euclidean distance, and the calculation formula is as follows: ; Based on the constructed cost matrix, the system employs a dynamic time warping algorithm to find the alignment path with the minimum cumulative distance. Let the cumulative cost matrix be... Its state transition programming formula is: ; Matrix boundary conditions are set as follows The system fills the entire matrix through the above iterative calculation, taking the bottom right corner element of the matrix. The value is used as the best matching distance between the envelope feature and the standard operating template on the time axis. , in the formula and These represent the total data length of the envelope feature and the standard operation template, respectively.

[0035] After obtaining the optimal matching distance, the system uses an exponential decay mapping function to convert it into a quantified matching similarity. The calculation formula is: ; in the formula A preset attenuation coefficient is used to adjust the sensitivity of distance mapping. The system judges the calculated matching similarity. Is the similarity greater than or equal to a preset similarity threshold? If the condition is met, it indicates that the signal causing the energy surge originates from a known structural switching operation of the transformer. The system confirms entry into transient decoupling mode to further investigate whether the operation is accompanied by deep electrical defects. If the matching similarity is strictly less than the preset similarity threshold, it indicates that the current signal belongs to an unexpected destructive event such as a severe external impact or a fracture of internal support components not recorded in the voiceprint database. The system directly determines that the transformer has experienced an unknown abnormal impact and outputs an abnormal impact warning signal on the main interface of the monitoring terminal.

[0036] This step utilizes a dynamic time warping algorithm to overcome the natural fluctuations in the execution time of similar mechanical actions in transformers, achieving accurate classification and intelligent diversion of the true physical properties of transient impact events.

[0037] S4. When in transient decoupling mode, remove low-frequency baseband signals and time-domain amplitude features, and extract phase spectrum features and high-frequency residual band signals as target decoupling features. Further, step S4, which involves extracting phase spectrum features and high-frequency residual band signals as target decoupling features, includes: The synchronous reference signal of the transformer is obtained, and the ultra-high frequency partial discharge signal is mapped to the phase interval corresponding to the synchronous reference signal to construct a two-dimensional distributed phase spectrum feature. Perform time-frequency transformation processing on the high-frequency residual band signal to extract the spectral amplitude distribution data of the high-frequency residual band signal within a preset high-frequency band; The phase map features and the spectral amplitude distribution data are fused and stitched together to generate the target decoupling features.

[0038] Specifically, when the system confirms entry into transient decoupling mode, the normal on-load tap changer operation or circuit breaker opening and closing of the transformer generates significant low-frequency mechanical vibrations and spatial electromagnetic inrush currents, masking subtle fault characteristics. The system's main control logic unit actively eliminates low-frequency baseband signals and time-domain amplitude characteristics from the buffer. After eliminating macroscopic interference data, the system acquires the AC voltage signal output from the secondary side of the transformer voltage transformer as a synchronization reference signal, and uses a zero-crossing comparison algorithm to extract the instantaneous power frequency phase angle of this reference signal. Assume the AC voltage reference period is uniformly divided into... The amplitude range of the UHF partial discharge signal is divided into discrete phase intervals as follows: The system extracts UHF partial discharge pulse sequences into discrete amplitude intervals, maps them to corresponding phase and amplitude intervals, and counts the frequency of pulse occurrence within each interval grid to construct a grid of size [value missing]. The phase map feature matrix of the two-dimensional distribution The row indices of this matrix correspond to specific power frequency phases, the column indices correspond to discharge intensities, and the matrix element values ​​represent the probability density of partial discharge under specific operating conditions.

[0039] Simultaneously, the system performs Discrete Fourier Transform processing on the retained high-frequency residual band signal to extract high-frequency resonance features. Let the high-frequency residual band signal sequence be... The number of sampling points used for the transformation is The system converts the signal sequence from the time domain to the frequency domain to obtain the complex spectrum. The calculation formula is: ; in the formula This is a frequency domain discrete spectral line index. The imaginary unit is used. The system further calculates the complex spectrum. The absolute value is used to obtain the true physical amplitude of each frequency component. Based on the typical high-frequency distribution range excited by micro-loosening of the transformer core or windings, the system extracts effective frequency points within the corresponding preset upper and lower limits of the high-frequency band, generating a spectral amplitude distribution data vector composed of the amplitudes of specific high-frequency components. , in the formula This is a frequency index located within the preset high-frequency band boundary, and this data corresponds to the mechanical response index of minute deformation of the transformer structure.

[0040] After obtaining the above two sets of features, the system uses the tensor tiling algorithm to transform the two-dimensional distributed phase map feature matrix. The data is then flattened and reconstructed into a one-dimensional electromagnetic eigenvector. The system then combines this one-dimensional electromagnetic eigenvector with the spectral amplitude distribution data vector. Feature fusion and splicing are performed along the column direction. To eliminate the physical dimensional barrier between electrical and acoustic signals, the system performs zero-mean normalization on the spliced ​​joint sequence to generate target decoupled features that integrate electrical insulation and mechanical structure states.

[0041] This implementation process effectively eliminates the illusion of macroscopic interference caused by routine operations, and accurately extracts the real partial discharge and microscopic mechanical degradation parameters hidden in the strong transient background.

[0042] S5. Diagnose the target decoupling characteristics. If the phase spectrum features show power frequency phase aggregation and the energy peak of the high-frequency residual band signal exceeds the preset resonance threshold, output a fault prediction alarm. Otherwise, it is judged as a working condition impact artifact and the alarm is suppressed.

[0043] Furthermore, step S5, the step of outputting the fault prediction alarm, includes: Statistical analysis of the pulse distribution density of the phase spectrum characteristics within a preset power frequency cycle, and calculation of the phase clustering probability of the pulse distribution density within a preset phase interval; In the target decoupling characteristics, determine the maximum energy peak corresponding to the high-frequency residual band signal; When the phase convergence probability is greater than the preset density threshold and the maximum energy peak is greater than the preset resonance threshold, a fault prediction alarm is generated.

[0044] Furthermore, in step S5, the step of otherwise determining it as an operational impact artifact and suppressing the alarm includes: When the phase spectrum features do not show power frequency phase clustering, or the energy peak of the high frequency residual band signal is not greater than the preset resonance threshold, the ultra-high frequency partial discharge signal and the contact acoustic pattern signal are determined to be working condition impact artifacts. Intercept the corresponding over-limit alarm signals and generate artifact masking instructions for the monitoring alarm link; Assign working condition impact artifact labels to the current UHF partial discharge signal and contact acoustic pattern signal, and update them to the local background noise dataset.

[0045] Specifically, after acquiring the target decoupling features, the system first extracts the phase spectrum feature matrix contained therein and then calculates the pulse distribution density of this feature within a preset power frequency cycle. Let the two-dimensional distributed phase spectrum feature matrix be... It contains Each discrete phase interval and The system calculates the pulse summation within all amplitude intervals to obtain the nth discrete amplitude interval. Pulse distribution density in each phase interval The calculation formula is: ; in the formula This serves as the amplitude interval index. The system then calculates the pulse aggregation probability within the preset phase interval to determine whether typical insulation defect discharge characteristics exist. Let the preset phase interval set be... This set typically corresponds to the rising edges of the peaks in the positive and negative half-cycles of the AC voltage waveform, which are prone to partial discharge. The system calculates the phase clustering probability within the preset phase interval. The formula is: ; While calculating the phase convergence probability, the system iterates through the spectral amplitude distribution data vector of the high-frequency residual band signal contained in the target decoupling characteristics. Let this data vector be... The system uses a comparison and sorting algorithm to determine the maximum energy peak within this frequency band. The calculation formula is: ; in the formula This is the index of effective frequency points within the preset high-frequency band. After obtaining the phase convergence probability and the maximum energy peak, the system compares them numerically with the system's preset density threshold and resonance threshold, respectively.

[0046] When the judgment logic determines that the phase convergence probability is greater than the preset density threshold and the maximum energy peak is greater than the preset resonance threshold, it indicates that the transformer is generating significant high-frequency mechanical deformation resonance, accompanied by strong power frequency-correlated electrical discharge. Based on this, the system generates a fault prediction alarm containing equipment fault correlations and uploads it to the substation remote monitoring platform via the industrial communication bus.

[0047] When the logical judgment result indicates that the phase convergence probability is not greater than the preset density threshold, or the maximum energy peak value is not greater than the preset resonance threshold, the system determines that the currently captured drastic waveform change is not caused by a real internal transformer fault, but by a condition impact artifact caused by normal dispatching operations or external grid inrush current. The system's underlying logic controller immediately intercepts the corresponding over-limit alarm signal to the hardware relay and generates an artifact masking command for the monitoring alarm link to prevent false alarm signal output. At the same time, the system assigns the current UHF partial discharge signal and contact acoustic pattern signal that triggered this judgment to condition impact artifact data tags, saves them, and updates them to the local background noise dataset for continuous correction of the benchmark parameters for subsequent condition identification.

[0048] This step enables accurate identification of genuine and false faults under strong interference environments, effectively intercepts false alarms caused by complex operating conditions, and improves the system's adaptive anti-interference capability by utilizing accumulated artifact data.

[0049] Example 2 Taking a 500kV main transformer at a key substation as an example, the transformer was operating under high load. At 2 PM, the dispatch center issued an instruction to switch the on-load tap changer on the transformer. During this process, a loud mechanical impact sound and an accompanying electromagnetic inrush current were generated inside the transformer.

[0050] The monitoring system simultaneously acquired intense acoustic signatures and ultra-high frequency signals. Calculations revealed a sharp increase in the first derivative of the low-frequency baseband signal energy, far exceeding the preset threshold generated based on historical statistical models. The system determined that it had encountered a strong disturbance and immediately entered the transient impact screening state.

[0051] The initial envelope curve of the voiceprint is extracted and normalized, and then compared with the standard template of voltage regulation action in the database using the DTW algorithm. Due to the extremely high matching similarity exceeding the safety threshold, it is confirmed as a compliant operation and enters transient decoupling mode.

[0052] Faced with a significant acoustic-electric masking effect, the system actively eliminates macroscopic low-frequency and amplitude extreme interference. Instead, it extracts the high-frequency residual band signal for time-frequency transformation and maps the partial discharge pulse to the power frequency voltage period to construct a two-dimensional PRPD phase spectrum feature.

[0053] Quantitative analysis of the decoupling characteristics revealed that the PRPD spectrum exhibited a significant power frequency phase clustering phenomenon; simultaneously, the high-frequency residual spectrum also detected the maximum energy peak exceeding the safety resonance threshold.

[0054] Final decision: The logic unit determined that this fluctuation was not a simple operational artifact, but rather a voltage regulation oscillation that exposed a latent winding looseness, accompanied by actual partial insulation discharge. The system immediately reported a composite fault prediction alarm.

[0055] Traditional methods, under the same operating conditions, were compared using static threshold interception and manual analysis of error logs. The specific performance data is shown in the table below: Table 1. Comparison of the effects of traditional monitoring methods and the method of this invention.

[0056] As shown in the table above, addressing the limitations of traditional single-threshold methods, which are prone to false alarms and false negatives, this invention successfully achieves effective stripping of high-fidelity features under strong interference environments through an adaptive multi-source feature decoupling and joint diagnosis mechanism. This not only reduces the operational transient false alarm rate from 68.5% to below 1.2%, but also increases the detection rate of deep-seated hidden dangers from 35.0% to 96.4%, further enhancing the advance warning capability for composite faults. The day was extended to This series of indicators demonstrates the engineering application value of this invention in the very early prediction of composite faults in transformers.

[0057] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A transformer fault prediction method based on multi-source data fusion, characterized in that, include: S1. Synchronously acquire the UHF partial discharge signal and contact acoustic pattern signal of the transformer, divide the contact acoustic pattern signal into low-frequency baseband signal and high-frequency residual band signal, and extract the UHF partial discharge signal into time-domain amplitude features and phase spectrum features. S2. Calculate the first derivative of the energy of the low-frequency baseband signal in real time. If the first derivative of the energy exceeds the preset derivative threshold, it is determined that the transformer has entered the transient impact screening state and the waveform comparison process is triggered. S3. In the transient impact screening state, the envelope features of the contact acoustic signal are extracted, and the envelope features are compared with the preset operation acoustic library using the dynamic time warping algorithm. If the matching similarity reaches the preset similarity threshold, it is confirmed that the transient decoupling mode has been entered; otherwise, it is determined that the transformer has experienced an unknown abnormal impact and an abnormal impact warning is directly output. S4. When in the transient decoupling mode, the low-frequency baseband signal and the time-domain amplitude feature are removed, and the phase spectrum feature and the high-frequency residual band signal are extracted as the target decoupling feature. S5. Diagnose the target decoupling characteristics. If the phase spectrum characteristics show power frequency phase aggregation and the energy peak of the high frequency residual band signal exceeds the preset resonance threshold, output a fault prediction alarm; otherwise, determine it as a working condition impact artifact and suppress the alarm.

2. The transformer fault prediction method based on multi-source data fusion according to claim 1, characterized in that, Step S1, the step of dividing the signal into a low-frequency baseband signal and a high-frequency residual band signal, includes: The contact acoustic pattern signal is processed by low-pass filtering to extract the vibration component of a preset low-frequency band on the transformer surface to generate the low-frequency baseband signal. The contact acoustic signal is processed by bandpass filtering to extract the vibration component of a preset high-frequency band; A time delay compensation operation is performed on the extracted vibration components of the preset high-frequency band to generate the high-frequency residual band signal that is synchronously aligned with the low-frequency baseband signal on the time axis.

3. The transformer fault prediction method based on multi-source data fusion according to claim 1, characterized in that, Step S2, the step of calculating the first derivative of the energy of the low-frequency baseband signal in real time, includes: Construct a sliding time window along the time axis and calculate the cumulative energy value of the low-frequency baseband signal within the current sliding time window; Obtain the difference sequence of the accumulated energy values ​​between adjacent sliding time windows, and calculate the time change rate of the difference sequence; The time rate of change is subjected to noise reduction by applying a smoothing filter operator, and the smoothed numerical sequence is output as the first derivative of the energy.

4. The transformer fault prediction method based on multi-source data fusion according to claim 1, characterized in that, Step S2, the step of determining that the transformer has entered the transient impact initial screening state and triggering the waveform comparison process, includes: Extract the basis noise derivative sequence of the transformer during its historical steady-state operation, and calculate the preset derivative threshold based on the statistical distribution model; The first derivative of the energy is compared with the preset derivative threshold. When the first derivative of the energy is greater than the preset derivative threshold within a continuous preset sampling period, a preliminary screening trigger command is generated. In response to the initial screening trigger command, the working mode of the monitoring system is switched to the transient impact initial screening state, and the waveform comparison process is triggered.

5. The transformer fault prediction method based on multi-source data fusion according to claim 1, characterized in that, Step S3, the step of extracting the envelope features of the contact acoustic signature signal, includes: The contact acoustic signal is analyzed using an envelope extraction operator to obtain an initial envelope curve; The initial envelope curve is resampled using an interpolation algorithm to obtain a resampled curve; An amplitude normalization operation is performed on the resampled curve to generate the envelope feature with a unified data dimension.

6. The transformer fault prediction method based on multi-source data fusion according to claim 1, characterized in that, Step S3, the step of comparing the envelope feature with a preset operation voiceprint database, includes: Retrieve the standard operation template from the preset operation voiceprint library and construct the cost matrix between the envelope feature and the standard operation template; Find the alignment path with the minimum cumulative distance in the cost matrix, and obtain the best matching distance between the envelope feature and the standard operation template on the time axis; Based on the optimal matching distance, the matching similarity between the envelope feature and the standard operation template is calculated using a mapping function.

7. The transformer fault prediction method based on multi-source data fusion according to claim 1, characterized in that, Step S4, the step of extracting the phase spectrum features and the high-frequency residual band signal as target decoupling features, includes: The transformer's synchronization reference signal is obtained, and the ultra-high frequency partial discharge signal is mapped to the phase interval corresponding to the synchronization reference signal to construct the two-dimensional distributed phase spectrum features. Perform time-frequency transformation processing on the high-frequency residual band signal to extract the spectral amplitude distribution data of the high-frequency residual band signal in a preset high-frequency band; The phase map features are fused and spliced ​​with the spectral amplitude distribution data to generate the target decoupling features.

8. The transformer fault prediction method based on multi-source data fusion according to claim 1, characterized in that, Step S5, the step of outputting a fault prediction alarm, includes: The pulse distribution density of the phase spectrum features within a preset power frequency period is statistically analyzed, and the phase aggregation probability of the pulse distribution density within a preset phase interval is calculated. In the target decoupling features, determine the maximum energy peak value corresponding to the high-frequency residual band signal; When the phase convergence probability is greater than a preset density threshold and the maximum energy peak is greater than the preset resonance threshold, the fault prediction alarm is generated.

9. The transformer fault prediction method based on multi-source data fusion according to claim 1, characterized in that, In step S5, the step of otherwise determining it as an impact artifact and suppressing the alarm includes: When the phase spectrum features do not show power frequency phase clustering, or the energy peak value of the high frequency residual band signal is not greater than the preset resonance threshold, the ultra-high frequency partial discharge signal and the contact acoustic pattern signal are determined to be working condition impact artifacts. Intercept the corresponding over-limit alarm signals and generate artifact masking instructions for the monitoring alarm link; Assign working condition impact artifact labels to the current UHF partial discharge signal and the contact acoustic pattern signal, and update them to the local background noise dataset.

Images