A method and system for real-time monitoring of transformer equipment status based on multi-source data fusion

Abstract

This invention discloses a method and system for real-time monitoring of transformer equipment status based on multi-source data fusion, specifically relating to the field of transformer equipment status monitoring technology. First, a unified time reference is obtained through multi-channel time alignment driven by event anchors, employing a strategy combining coarse and fine alignment. Based on this, channel alignment confidence is constructed using peak sharpness ratio, group delay consistency, phase-locked consistency, and geometric consistency based on temperature compensation. Then, acoustic emission spectrum energy and pulse count, vibration and shock characteristics, and cross-sensor consistency characteristics are extracted under the unified time reference, and dynamic threshold tuning is performed in conjunction with load and environmental context. Next, based on channel health factors, gated weighting, intra-group decorrelation, and probability calibration are performed on multi-source evidence to obtain calibrated fault probabilities and inconsistencies. Finally, graded results are output under multi-level judgment constraints controlled by error detection rate and physical consistency verification constraints.

Description

Technical Field

[0001] This invention relates to the field of transformer equipment condition monitoring technology, and more specifically, to a method and system for real-time monitoring of transformer equipment condition based on multi-source data fusion. Background Technology

[0002] Current transformer monitoring methods largely rely on single modes (such as AE / UHF or vibration) and fixed thresholds, making it difficult to simultaneously ensure sensitivity and stability under electromagnetic and mechanical noise, load and temperature fluctuations. While multi-sensor solutions are widespread, they generally lack a unified time base and alignment reliability quantification, and do not incorporate temperature-compensated geometric consistency checks, easily leading to misalignment and cross-channel "duplicate counting." At the feature and fusion level, there is a lack of channel health factor gating and intra-group decorrelation, and the output lacks probabilistic calibration, causing thresholds to drift with different scenarios. The judgment process is mostly Boolean alarms, lacking statistical constraints such as FDR and physical verification such as arrival time difference and coherence, and also failing to create event-level auditable traces. There is an urgent need for a real-time monitoring method and system that, under unified time base and alignment reliability constraints, integrates phase / delay consistency and geometric consistency, combines health factor-driven decorrelation fusion and probabilistic calibration, and uses an FDR + physical verification closed loop to ensure interpretability and traceability. Summary of the Invention

[0003] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a method and system for real-time monitoring of transformer equipment status based on multi-source data fusion to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A real-time monitoring method for transformer equipment status based on multi-source data fusion includes the following steps:

[0006] Event anchor point detection is performed on current, harmonic phase, vibration, and acoustic emission channels; mismatched anchor points are eliminated using a robust matching algorithm, and inter-channel time delay estimation and uncertainty assessment are completed by combining statistical estimation and random sampling methods to obtain a unified time reference; constrained dynamic time warping and state smoothing are implemented on coarse alignment segments to achieve refined alignment; channel alignment confidence is generated by fusing main peak sharpness ratio, group time delay consistency, phase lock consistency, and geometric consistency combined with temperature compensation.

[0007] Based on a unified time reference and channel alignment confidence, the spectral energy and pulse number of the acoustic emission channel, the impact characteristics of the vibration channel, and the consistency characteristics of cross-sensor coherence and amplitude ratio are extracted. The channel health factor is constructed based on the channel alignment confidence and noise ratio. Gated weighting and fusion are performed on the multi-channel fault evidence, and the probability calibration of the fused output is performed to obtain the fault probability and inconsistency index.

[0008] Under the constraints of failure probability and inconsistency index, a multi-level judgment process based on error detection rate control is adopted, and physical consistency is verified by combining the upper limit of arrival time difference, cross-sensor coherence and amplitude ratio, to generate graded results.

[0009] In a preferred embodiment, anchor point detection includes: for the current channel, using first-order difference and adaptive threshold to identify abrupt change points; for the harmonic phase channel, performing short-time discrete Fourier transform on the power frequency and several harmonics, and locating phase inflection points based on the sign reversal of the first derivative of the phase curve; for the acoustic emission channel, obtaining the envelope through Hilbert transform and extracting sudden peak values ​​using an adaptive threshold driven by the median absolute deviation; and for the vibration channel, locating the arrival time of the impact using bandpass filtering and envelope detection.

[0010] In a preferred embodiment, the construction and scoring of the main peak sharpness ratio are as follows: within each analysis window, a generalized cross-correlation phase transformation method is applied to the reference channel and the aligned channel to obtain a cross-correlation sequence, the main peak is identified and its amplitude is taken as the main peak amplitude; after removing the main lobe where the main peak is located, the side peak with the largest amplitude is selected from the remaining local maxima as a comparison; when the side peak amplitude is zero or close to zero, it is replaced with a preset minimum positive value to avoid anomalies; the main peak sharpness ratio is defined by the relative separation degree of the main peak and side peak amplitudes.

[0011] In a preferred embodiment, group delay consistency is specifically obtained as follows: a group delay sequence is formed by phase expansion of the cross-spectral phase of the reference channel and the aligned channel; group delay samples are obtained at discrete frequency points within the service-focused frequency band, with the interquartile range representing the intra-band dispersion; the reference delay scale is taken as the representative value of the healthy period group delay fluctuation or the maximum acceptable fluctuation for the service; a three-segment linear mapping inversely proportional to the dispersion is used for scoring: full marks are given when the score is below the "excellent" threshold, zero marks are given when the score is above the "poor" threshold, and the score decreases linearly when the score is between the two thresholds. The "poor / excellent" threshold is preferably set at the 80th and 95th percentiles of the healthy period sample distribution; to suppress the influence of isolated frequency points, the group delay sequence is preferably smoothed using three to five points of median.

[0012] In a preferred embodiment, phase-locked consistency is measured by using the phase of power frequency voltage or current as a reference to measure the convergence of the reference phase of registered acoustic emission or ultra-high frequency pulses within the same event.

[0013] In a preferred embodiment, the geometric consistency of temperature compensation is based on the sound velocity-temperature relationship calibrated at the station and the sensor geometry. The positioning residual of the cross-channel arrival time difference is evaluated and normalized to a monotonic score where the smaller the residual, the higher the score.

[0014] In a preferred embodiment, the original alignment confidence is obtained by weighted fusion based on the main peak sharpness ratio, group delay consistency, phase lock consistency, and geometric consistency combined with temperature compensation according to preset weights; the original alignment confidence is then subjected to distribution-preserving quantile mapping so that the mapping result is consistent with the empirical distribution of the old version of the alignment confidence within the interval, and this result is output as the channel alignment confidence.

[0015] In a preferred embodiment, the specific steps of gated weighted fusion and probability calibration include:

[0016] Based on channel alignment confidence and noise ratio, a channel health factor ranging from 0 to 1 is generated using a monotonic mapping function. The channel health factor is multiplied by the original fault evidence to obtain discounted channel-level evidence. For channel groups with statistical correlation, intragroup whitening or shrinkage estimation operations are performed to eliminate duplicate information. All discounted evidence is weighted and fused using a log-likelihood function. The fused output probability is calibrated using temperature scaling or Platt scaling techniques, and the quantile output confidence interval is calculated based on a rolling time window.

[0017] In a preferred embodiment, the following modules are included:

[0018] Time alignment and evaluation module: performs multi-channel event anchor detection, coarse / fine two-level time alignment, and outputs channel-level alignment confidence;

[0019] Features and threshold module: Extracts acoustic emission spectrum energy, vibration and shock features, and cross-sensor consistency features, and dynamically adjusts the threshold based on the load environment context;

[0020] Fusion and calibration module: Utilizes channel health factors to gating and weight multi-source evidence, and outputs the failure probability through probability stacking and calibration;

[0021] Judgment and Verification Module: Implements a multi-level judgment process to control the error detection rate, performs physical verification by combining sound speed propagation constraints and sensor consistency, and generates hierarchical event records.

[0022] The technical effects and advantages of this invention are as follows:

[0023] This invention establishes a unified time reference and quantifies time delay uncertainty by using anchor point detection, RANSAC to remove mismatches, coarse alignment, and refined alignment using constrained dynamic time warping (DTW) + Kalman smoothing. Based on this, a four-dimensional consistency quantity is introduced to form a channel alignment confidence level that is "more credible the larger it is," thereby suppressing the systematic bias caused by misalignment from the source and significantly improving the comparability and reproducibility of cross-channel evidence.

[0024] On the fusion side, evidence is gating and discounted using channel health factors, and within-group whitening or shrinkage estimation is performed on statistically correlated channel groups to avoid "double counting." Linear-nonlinear synthesis is achieved using differentiable stacking functions such as log-likelihood, and probability calibration is implemented through temperature scaling or Platt scaling to match the output alarm probability with the actual occurrence frequency, thus maintaining robust discrimination thresholds and transferability across different scenarios. Compared to solutions relying solely on hard thresholds, this invention achieves a higher detection rate at the same false alarm level and can directly provide confidence intervals to support operational decisions.

[0025] On the judgment and verification side, the BH program is used to implement controllable multi-level statistical judgment of FDR, and physical thresholds such as the upper limit of arrival time difference, cross-sensor coherence and amplitude ratio are superimposed for verification, which significantly reduces sporadic noise and false source triggering. The judgment results are locally incremented and solidified in the form of event-level records (ELR), and metadata such as probability, threshold, physical verification, health factors and version are completely saved, which facilitates auditing, playback and continuous improvement. This mechanism ensures real-time performance while achieving interpretability and traceability of the monitoring link. Attached Figure Description

[0026] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings;

[0027] Figure 1 This is a flowchart illustrating the real-time monitoring method for transformer equipment status based on multi-source data fusion according to the present invention.

[0028] Figure 2 This is a schematic diagram of the structure of the real-time monitoring system for transformer equipment status based on multi-source data fusion according to the present invention. Detailed Implementation

[0029] The technical solutions of 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.

[0030] Example 1

[0031] This invention relates to a real-time monitoring method for transformer equipment status based on multi-source data fusion, such as... Figure 1 As shown, it includes the following steps:

[0032] Step S1: Event Anchor-Driven Multi-Source Time Alignment and Uncertainty Assessment

[0033] This step is performed under conditions where multi-source synchronous acquisition and timestamp recording capabilities are available. The goal is to achieve millisecond-level or even sub-sampling-level consistent alignment of channels such as current, harmonic phase, vibration, and AE (acoustic emission) under a unified time reference, and to quantitatively evaluate the alignment quality. The output of this step serves as a pre-constraint and input for subsequent robust feature extraction and reliability discount fusion, ensuring that the real-time monitoring link has traceable temporal consistency.

[0034] Within the sliding window, automatic anchor point event detection is performed on each channel: the current channel uses first-order differential and adaptive thresholding to identify abrupt change points; the harmonic phase channel performs short-time discrete Fourier transform on the power frequency and several harmonics, locating phase inflection points based on the sign reversal of the first derivative of the phase curve; the AE channel obtains the envelope through Hilbert transform and uses median absolute deviation (MAD) to drive an adaptive threshold to extract sudden peak values; the vibration channel uses bandpass filtering and envelope detection to locate the arrival time of the impact. The above-mentioned cross-channel observable transients are uniformly defined as candidate "event anchor points".

[0035] In the event-level coarse alignment stage, the system selects the current or harmonic phase as the reference channel, pairs the nearest neighbor anchor points for other channels within the ±Δ search window of the predicted delay, and uses RANSAC to eliminate mismatches. Subsequently, an observation model is established with "channel propagation delay" as the latent variable. Online EM is used to estimate the posterior expectation of the fixed component delay, and particle filtering is used to track the slowly drifting time-varying components, thereby obtaining the delay estimates, variances, and smoothing trajectories of each channel relative to the reference channel, forming a stable coarse alignment result.

[0036] In the refinement alignment stage, the system first performs dynamic time warping (DTW) on the coarsely aligned segments within the Sakoe-Chiba constraint band to limit the time scaling ratio to avoid overfitting and obtain discrete alignment paths. Then, constrained Kalman filtering is used to perform continuous domain smoothing correction on the "residual time bias and its first-order drift". The measurement noise covariance is adaptively set according to the channel signal-to-noise ratio and event sparsity, and the process noise covariance is adjusted online according to the alignment residual statistics. In this way, the discrete registration result is transformed into a continuous and smooth small time correction.

[0037] For short-duration, impact-type events, the system calculates wavelet coherence or cross-power spectrum coherence to estimate the minimum phase difference between channels and converts it into a subsampling-level time shift based on the dominant frequency. Fractional delay interpolation (preferably Farrow structure or third- to fifth-order Lagrange interpolation) is used to perform fine correction within the sampling grid, ensuring that the sharp pulse and high-frequency details are consistent in position across multiple channels.

[0038] This embodiment proposes a channel-level alignment confidence calculation method with "Peak Sharpness Ratio (PSR)" and "Group Delay Consistency (GDC)" as its core, which takes into account both feasibility and auditability.

[0039] To improve the physical reliability of anchor point alignment and suppress false alignment, this embodiment introduces two domain features—Phase-Locked Consistency (PLC) and Geometric Consistency with Temperature Compensation (GCS)—without changing the external fields and value ranges. Using the power frequency voltage (or current) phase as a reference, the phase angle set is calculated for registered AE / UHF pulses within the same event, and the phase clustering degree is defined by the following formula: ; n represents the number of transient pulses participating in the statistics of the same "event"; This represents the power frequency reference phase corresponding to the k-th transient pulse; For phase The corresponding unit complex vector is used to calculate the phase average in the complex plane.

[0040] And according to the quantiles of the healthy period sample Piecewise linear mapping is obtained The more stable the phase cluster (conforming to typical PRPD patterns), the higher the score. Regarding the geometric consistency of temperature compensation, the system performs least-squares positioning on the cross-channel arrival time difference based on the station-calibrated medium sound velocity-temperature function and sensor geometry, calculating the normalized residual. And mapped as: ; For use in Normalized scaling constant / upper limit.

[0041] In calculating channel alignment confidence, the system maintains the monotonic relationship of "the larger the confidence, the more reliable" and the interval remains unchanged ([0,1]). First, existing consistency metrics (such as scores based on PSR and Group Delay Consistency (GDC)) are synthesized into... Then, the original confidence level is obtained by weighting and fusing it with the two new types of consistency quantities added in this embodiment: Where α1, β1, and γ1 are preset weighting coefficients, for example, α1=0.6, β1=0.2, and γ1=0.2.

[0042] To ensure compatibility with existing systems in terms of statistical distribution, a distribution-conformal quantile mapping is used. Reflecting back to the empirical distribution of the alignment confidence from the previous version: ;in The old version of the historical healthy period Sample distribution This is an example. The online estimated distribution. Channel alignment confidence.

[0043] Within each analysis window, cross-correlation sequences are obtained using generalized cross-correlation (such as GCC-PHAT). Its main peak amplitude is The amplitude of the second largest side peak is (After removing the main lobe containing the main peak, take the maximum among the remaining local maxima), to avoid zero denominator, let The sharpness ratio of the main peak is defined as follows: ;

[0044] PSR reflects the separability of delay peaks; a higher value indicates more reliable alignment. To map it to a confidence score of 0-1, a three-segment linear rule is used: given two threshold levels, "poor" and "excellent". (By default, the P80 and P95 values ​​of the healthy period distribution are used as the calibration), let Here, clip(⋅) means clipping the result to the interval [0,1]. Thus, when... A perfect score is 1 when... The time is recorded as 0 to ensure the monotonicity and auditability of the principle that "the more separable the peak shape, the higher the score".

[0045] Group delay consistency is achieved through group delay sequence Depiction, ,in This is cross-spectral phase (preceding phase expansion). Discrete frequency points are selected within the service-relevant frequency band B. The in-band dispersion of group delay is measured by interquartile range and scaled according to a reference delay. Return to One: .

[0046] The representative value of the standard deviation of latency during the healthy period or the maximum acceptable fluctuation for the business can be taken. Lower dispersion indicates better cross-frequency consistency. A dual three-segment linear mapping is used: given a threshold... (For example, using healthy P80 and P95 as calibration), let thereby Note 1, Marking it as 0, the frequency decreases linearly within the interval, ensuring that "the more consistent the group delay, the higher the score." To suppress the influence of isolated frequency points, adjustments can be made at the implementation layer. First, perform 3–5 point median smoothing; this process does not change the above definition.

[0047] Take the weighted average of the two scores and crop it to [0,1]: ;For example =0.5, =0.5.

[0048] In engineering implementation, the window length and overlap rate are configured by balancing real-time performance and accuracy, with the end-to-end processing latency preferably not exceeding one analysis window period. The final output includes: aligned data segments under a unified time base, relative delays and uncertainties of each channel, alignment residual statistics and quality indicators, channel-level alignment confidence, and traceable time-aligned metadata, providing a clear and reproducible temporal basis for robust feature extraction and fusion decisions in subsequent steps.

[0049] Step S2: Confidence-driven robust feature extraction and context threshold tuning

[0050] This step is performed under the unified time reference and alignment confidence conditions output in step S1. The goal is to extract fault-sensitive and noise-robust features from multi-channel data such as AE (acoustic emission / ultrasound), vibration, and temperature / current, and to establish a dynamic threshold model in combination with load and environmental context. Finally, a three-element output of "robust feature set - threshold - deviation" can be directly used for subsequent fusion judgment. The system first inherits the time alignment results of step S1, processes each channel synchronously under the same analysis window and step size, and uses the channel-level alignment confidence as a quality weight to participate in the adaptive tuning and validity labeling of subsequent statistics, thereby ensuring the time consistency and comparability of different channels.

[0051] On the AE channel, the system performs bandpass and spectral density estimation on the 15–160 kHz frequency band, and performs background subtraction using the robust median spectrum as the background baseline. After calculating the net spectral energy, it obtains core features such as N (the ratio of net energy of the target frequency band to the net energy of the reference subband) and Len (the number of independent pulses exceeding the adaptive threshold within the analysis window). At the same time, it extracts consistency features such as arrival time difference, cross-sensor coherence, and amplitude ratio to characterize spatial source and reliability. The threshold estimation adopts a rolling strategy of MAD or conditional quantile to adapt to changes in noise floor and interference injection. When a decrease in alignment confidence or an increase in noise proportion is detected, the threshold bandwidth is automatically expanded and a reliability label is added to the features of the window for discounting processing in the subsequent fusion stage.

[0052] To ensure consistent and robust statistical standards across channels, all features are dimensionless processed using the rolling median and MAD. The outlier truncation ratio, window length, and overlap rate are adaptively tuned based on alignment confidence, noise percentage, and event sparsity. Normalized features and their corresponding quality weights are encapsulated as time-series records with timestamps, window indices, and sensor identifiers to support local playback and auditing. On the threshold side, a layered strategy of "offline calibration + online calibration" is adopted: In the offline stage, the conditional relationship between each feature dimension and context variables is learned based on representative environment / load samples, outputting the basic threshold function and its confidence band; in the online stage, distribution drift is identified and slightly calibrated at the site level using a rolling window, with the calibration amplitude limited to a safe range to avoid misadjustment caused by short-term noise; for multiple features requiring joint judgment, joint tolerance information for judgment combinations is also provided to support subsequent consistency and FDR control. In scenarios with incomplete data, the system outputs a feature missing mask and enables controlled interpolation to maintain consistent input dimensions; when key features are completely unavailable, the threshold deviation is not calculated, only an unavailable flag is output, and the weights are automatically reduced in subsequent fusion.

[0053] This step uses alignment confidence for small-step threshold tuning to close the "alignment-feature-threshold-deviation" link: Definition ,in When the tuning direction is reversed, a hysteresis half-width H is applied to θ to suppress jitter, and the update period Ts is the duration of the analysis window; when key features are missing or the quality weight is below the threshold, this window executes "hold" and does not update θ, and records the status bit.

[0054] The external output maintains consistency with the interface of subsequent steps: it provides a robust feature set, the context threshold and uncertainty of the current window instantiation, the corresponding deviation, and quality and availability metadata (including alignment confidence projection weights, noise ratio, missing mask / imputation flag, threshold version number and saturation / hysteresis status bits), which are stored and called only locally and are directly used for the gated weighted fusion in step S3 and the statistical judgment and physical consistency verification in step S4.

[0055] Step S3: Gated weighted fusion based on channel health factors

[0056] This step is performed based on the unified time base and alignment confidence of step S1, and the robust features, context thresholds, and deviations output from step S2, retaining only the "gated weighted probability stacking" fusion route. The system first calculates the channel health factor for each channel i. It is obtained by monotonic mapping of channel alignment confidence and noise ratio, and then cropped to the [0,1] interval; when a channel has missing key features or its quality weight is below a threshold, it will be... The value is lowered and the cause code is recorded, but the original feature value is not changed. Then, each target fault type j is scored within the current window's channel. (Obtained from the robust features of step S2 through offline calibration of a small model, with only minor online calibration), and then proceeds to a three-stage fusion: the first is discounting, multiplying the gating weights into the evidence formation. The first step involves performing a one-time decorrelation (within-group whitening or shrinkage estimation) on channel groups with significant correlations to suppress double counting; the second step is synthesis, which calculates normalized weights. Linear-nonlinear synthesis of discounted evidence using differentiable stacking functions Where g(⋅) takes the log-likelihood or piecewise saturated generalized linear unit to ensure numerical stability; the third is calibration, which uses temperature scaling or Platt scaling to obtain the calibrated probability. .

[0057] The system outputs confidence intervals based on rolling quantiles or Bayesian approximations. To address inconsistencies across channels, the system calculates residual inconsistency and divergence indices after synthesis.

[0058] Step S4: FDR-controlled multi-level judgment and physical consistency verification

[0059] This step is performed based on the outputs of steps S2 and S3. The goal is to provide a statistically controllable and physically verifiable final classification result under the constraints of a unified time base and calibrated fusion probabilities, and to solidify all evidence and quality traces in the local event-level record (ELR), independent of any uplink. Inputs include: calibrated probabilities (including confidence intervals) and inconsistencies / handling status bits for each fault type, robustness features such as AE / vibration and consistency features, and quality metadata such as alignment confidence and channel health factors.

[0060] On the statistical side, the Benjamini–Hochberg (BH) procedure is used to control the false detection rate (FDR) of peer candidates: first, a monotonic mapping is used. The p-value is obtained for those that meet the pre-screening threshold. Candidates enter multiple tests ( (Used to reduce computing power without changing the legitimacy of FDR); Let the number of candidates be M, and sort the p values ​​in ascending order. ,Pick Based on this, a statistical rejection set was obtained. This serves as the trigger set for "statistical significance" and outputs a count of significant evidence. And the rank / threshold comparison information for each rejection item. On the physical side, for Candidate physical consistency checks are performed to suppress spurious triggering: the check gate consists of an upper bound on the time difference of arrival, cross-sensor coherence, and amplitude ratio, if and only if If both conditions are met, the triggering continues; otherwise, the candidate is downgraded to "Note / Requires Review" and the "Physical Inconsistency" reason code is recorded.

[0061] In cases of anomalies and missing evidence, the missing mask and controlled imputation constraints from step S2 are applied: candidates with completely missing key consistency evidence are not statistically rejected; candidates with incomplete but acceptable quality evidence are allowed to participate in BH (Browser-Henderson) trials, but the probability threshold is increased accordingly or the threshold is conservatively amplified during final grading. If the overall quality of this window is insufficient (e.g., ...), If the data quality is below the site threshold, only the "Insufficient data quality / requires review" status will be output without issuing a hard trigger.

[0062] Final classification and local record keeping. The system makes a joint decision based on three types of quantitative evidence—(i) calibrated probability. (ii) FDR rejection result and significant evidence count K; (iii) Physical consistency review conclusion and S3 inconsistency degree—generating a four-level judgment of "normal / attention / abnormal / serious": if there is And the review was approved, at the same time And if the inconsistency is low, it is marked as "abnormal / serious" (subdivided by hazard level); if However, if the review fails or the inconsistency is high, it will be marked as "Note / Requires Review"; if Then mark it as "normal".

[0063] After the ruling is completed, an event-level record (ELR) is generated locally, with fields including at least: Uniform Time Index and Window Number, Fault Type and Classification. It includes confidence intervals, K and rank / threshold comparison, physical review conclusions and reason codes, S3 inconsistency and handling status bits, alignment confidence and channel health factor snapshots, context thresholds and model version numbers, and missing mask / imputation flags; ELR uses incremental writing and supports playback by event number for auditing and post-event review.

[0064] Example 2

[0065] The design of the real-time monitoring system for transformer equipment status based on multi-source data fusion in this invention is based on the method in Embodiment 1, specifically as follows: Figure 2 The following modules are shown:

[0066] The time alignment and evaluation module is used to perform event anchor point detection, coarse and fine two-level alignment, and confidence evaluation for multiple channels such as current, harmonic phase, vibration, and AE under a unified time reference. It outputs aligned data segments with a unified time reference, relative time delays and uncertainties for each channel, alignment residual statistics and quality indicators, channel-level alignment confidence, and traceable alignment metadata, serving as pre-constraints and inputs for subsequent modules. This module corresponds to step S1 in Example 1, including anchor point detection, RANSAC mismatch removal, time delay estimation based on EM / particle filter modeling, fine alignment using DTW+ constrained Kalman smoothing, and consistency scoring and gating strategies based on PSR / GDC quantization.

[0067] The feature and threshold module, under the unified time reference and alignment confidence provided by the time alignment and evaluation module, synchronously extracts fault-sensitive and noise-robust features from multiple channels such as AE, vibration, and temperature / current. It then establishes a dynamic threshold model combining offline calibration and online calibration, forming a "robust feature set—threshold—deviation" output. Simultaneously, it outputs quality and availability metadata (alignment confidence projection weights, noise percentage, missing mask / imputation flag, threshold version number, and saturation / hysteresis state bits) for subsequent fusion and judgment. The feature caliber is uniformly processed using the rolling median and MAD as dimensionless values. The threshold is calibrated in small steps within a safe range, and controlled imputation is performed in missing scenarios to maintain input dimension consistency.

[0068] The fusion and calibration module, based on the deviation and quality metadata output by the feature and threshold module, and combined with the alignment confidence given in Module 1, calculates the channel health factor and performs "gated weighted probability stacking" fusion. First, the evidence is discounted for health factors and intra-group decorrelated. Then, linear-nonlinear synthesis is performed according to a differentiable stacking function, and calibrated probabilities are obtained through temperature scaling or Platt scaling. Simultaneously, rolling quantiles or Bayesian approximation confidence intervals and residual inconsistency / divergence indices are output as inputs for statistical and physical verification.

[0069] The judgment and review module performs multi-level judgments with a controllable error detection rate (FDR) under a unified time base and calibrated probability constraints. It conducts physical consistency reviews based on thresholds such as reaching the upper limit of time difference and cross-sensor coherence and amplitude ratio. For windows with incomplete evidence or insufficient overall quality, it provides safety measures such as "requires review / does not issue hard trigger". Finally, it generates "normal / attention / abnormal / serious" graded results and solidifies them as event-level records (ELR). The records include metadata such as grade, confidence interval, significant evidence count and rank / threshold comparison, physical review conclusion and cause code, channel health factor snapshot, threshold and model version number, missing and imputation flags, etc., for auditing and playback.

[0070] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0071] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0072] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0073] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0074] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for real-time monitoring of transformer equipment status based on multi-source data fusion, characterized in that, Includes the following steps: Event anchor point detection is performed on current, harmonic phase, vibration and acoustic emission channels; erroneous anchor points are eliminated by robust matching algorithm, and inter-channel time delay estimation and uncertainty assessment are completed by combining statistical estimation and random sampling methods to obtain a unified time reference; Constrained dynamic temporal warping and state smoothing are applied to coarsely aligned segments to achieve finer alignment; The original confidence level is obtained by weighting and fusing the main peak sharpness ratio, group delay consistency, phase lock consistency, and geometric consistency combined with temperature compensation according to preset weights. Perform a distribution-preserving quantile mapping on the original confidence scores to make the mapping result consistent with the empirical distribution of the old aligned confidence scores within the interval, and output the result as the channel aligned confidence score; Based on a unified time reference and channel alignment confidence, the spectral energy and pulse number of the acoustic emission channel, the impact characteristics of the vibration channel, and the consistency characteristics of cross-sensor coherence and amplitude ratio are extracted. The channel health factor is constructed based on the channel alignment confidence and noise ratio. Gated weighting and fusion are performed on the multi-channel fault evidence, and the probability calibration of the fused output is performed to obtain the fault probability and inconsistency index. Under the constraints of failure probability and inconsistency index, a multi-level judgment process based on error detection rate control is adopted, and physical consistency is verified by combining the upper limit of arrival time difference, cross-sensor coherence and amplitude ratio, to generate graded results.

2. The method for real-time monitoring of transformer equipment status based on multi-source data fusion according to claim 1, characterized in that: Anchor point detection includes: for the current channel, first-order difference and adaptive threshold are used to identify abrupt change points; for the harmonic phase channel, short-time discrete Fourier transform is performed on the power frequency and several harmonics, and the phase inflection point is located based on the sign reversal of the first derivative of the phase curve; for the acoustic emission channel, the envelope is obtained through Hilbert transform and the burst peak is extracted by adaptive threshold driven by the median absolute deviation; for the vibration channel, bandpass filtering and envelope detection are used to locate the arrival time of the impact.

3. The method for real-time monitoring of transformer equipment status based on multi-source data fusion according to claim 1, characterized in that: The construction and scoring of the main peak sharpness ratio are as follows: Within each analysis window, the generalized cross-correlation phase transformation method is applied to the reference channel and the aligned channel to obtain the cross-correlation sequence, the main peak is identified and its amplitude is taken as the main peak amplitude; after removing the main lobe where the main peak is located, the side peak with the largest amplitude is selected from the remaining local maxima as a comparison; when the side peak amplitude is zero or close to zero, it is replaced with a preset minimum positive value to avoid anomalies; the main peak sharpness ratio is defined by the relative separation degree of the amplitudes of the main peak and the side peak.

4. The method for real-time monitoring of transformer equipment status based on multi-source data fusion according to claim 1, characterized in that: Group delay consistency is specifically obtained as follows: the cross-spectral phase of the reference channel and the aligned channel is expanded to form a group delay sequence; group delay samples are obtained at discrete frequency points within the frequency band of business interest, and the interquartile range is used to characterize the intra-band dispersion. The reference delay scale is taken as the representative value of the group delay fluctuation during the healthy period or the maximum fluctuation acceptable to the business; a three-segment linear mapping inversely proportional to the dispersion is used for scoring: full marks are given when the score is below the "excellent" threshold, zero marks are given when the score is above the "poor" threshold, and the score decreases linearly when the score is between the two thresholds.

5. The method for real-time monitoring of transformer equipment status based on multi-source data fusion according to claim 1, characterized in that: Phase-locked consistency measures the convergence of reference phases of registered acoustic emissions or UHF pulses within the same event, using the phase of power frequency voltage or current as a reference.

6. The method for real-time monitoring of transformer equipment status based on multi-source data fusion according to claim 1, characterized in that: The geometric consistency of temperature compensation is based on the sound velocity-temperature relationship calibrated at the site and the geometric layout of the sensor. The positioning residual of the cross-channel arrival time difference is evaluated and normalized into a monotonic score where the smaller the residual, the higher the score.

7. The method for real-time monitoring of transformer equipment status based on multi-source data fusion according to claim 1, characterized in that: The specific steps of gated weighted fusion and probability calibration include: Based on channel alignment confidence and noise ratio, a channel health factor ranging from 0 to 1 is generated using a monotonic mapping function. The channel health factor is multiplied by the original fault evidence to obtain discounted channel-level evidence. For channel groups with statistical correlation, intragroup whitening or shrinkage estimation operations are performed to eliminate duplicate information. All discounted evidence is weighted and fused using a log-likelihood function. The fused output probability is calibrated using temperature scaling or Platt scaling techniques, and the quantile output confidence interval is calculated based on a rolling time window.

8. A real-time monitoring system for transformer equipment status based on multi-source data fusion, characterized in that, The monitoring system, based on the method described in any one of claims 1-7, includes the following modules: Time alignment and evaluation module: performs multi-channel event anchor detection, coarse / fine two-level time alignment, and outputs channel-level alignment confidence; Features and threshold module: Extracts acoustic emission spectrum energy, vibration and shock features, and cross-sensor consistency features, and dynamically adjusts the threshold based on the load environment context; Fusion and calibration module: Utilizes channel health factors to gating and weight multi-source evidence, and outputs the failure probability through probability stacking and calibration; Judgment and Verification Module: Implements a multi-level judgment process to control the error detection rate, performs physical verification by combining sound speed propagation constraints and sensor consistency, and generates hierarchical event records.

Images