A transformer winding axial displacement and radial deformation fault diagnosis method
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ANHUI UNIV
- Filing Date
- 2026-03-27
- Publication Date
- 2026-06-26
Smart Images
Figure CN122283535A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of transformer fault diagnosis technology, specifically a method for diagnosing transformer winding axial displacement and radial deformation faults. Background Technology
[0002] Power transformers are core equipment in power grids for power conversion and transmission, and their operating status directly affects power supply reliability and energy utilization efficiency. Under the "dual carbon" target (carbon reduction and emission reduction), ensuring energy-saving and environmentally friendly operation of transformers is particularly important. Abnormal operating conditions such as axial displacement of windings, radial deformation, or inter-turn short circuits in transformers not only increase power transmission losses but may also cause grid load imbalances due to outages, resulting in additional energy waste and carbon emissions.
[0003] Existing methods for oscillating wave signal processing and fault identification have several shortcomings: First, during signal preprocessing, local benchmarks often rely on global statistics (such as global mean and global extreme values) across the entire sequence, leading to a lag in the extraction of local fluctuation features. These methods are susceptible to interference from global outliers and cannot accurately capture subtle fault details. Second, during the generation of two-dimensional feature images, no adaptive focusing optimization is performed on the signal. Strong fluctuation signals are prone to producing image artifacts, and subtle fault features are submerged, affecting the accuracy of subsequent fault identification. Third, fault identification often targets a single fault type, making it difficult to effectively distinguish between axial displacement and radial deformation faults with similar features. Furthermore, it can only qualitatively determine the presence or absence of a fault and cannot achieve quantitative classification of the fault degree, making it difficult to meet the precise maintenance requirements in practical engineering. Fourth, the feature extraction methods are simplistic, the classification models have high computational overhead and poor adaptability, failing to meet the needs of real-time diagnosis.
[0004] To address the shortcomings of the existing technologies, this invention proposes an oscillating wave signal processing and fault identification method based on local neighborhood adaptive focusing. By optimizing the signal preprocessing process and innovating feature extraction and classification methods, this invention solves the technical problems of inaccurate local feature extraction, low fault identification accuracy, and inability to perform hierarchical evaluation in the existing technologies, thereby improving the accuracy and practicality of power transformer winding fault diagnosis. Summary of the Invention
[0005] (a) Technical problems to be solved
[0006] To address the technical challenges in existing transformer winding fault diagnosis, such as large deviations between simulated and actual fault conditions, insufficient accuracy in signal preprocessing and feature extraction, weak ability to distinguish similar faults, and inability to quantify and classify fault severity, this invention discloses a method for identifying and assessing the severity of power transformer winding faults by integrating a controllable fault simulation platform with local neighborhood adaptive feature mapping. The complete technical solution comprises two parts: a fault simulation experimental device structure and a signal processing-image conversion-fault identification and classification method. By standardizing the simulation of two typical fault types—axial displacement and radial deformation—and combining adaptive signal preprocessing and multi-dimensional image feature extraction, it achieves accurate fault type identification and four-level severity assessment: normal, minor, moderate, and severe. The specific technical solution is as follows:
[0007] (II) Technical Solution
[0008] I. Structure of the Fault Simulation Experimental Device
[0009] This invention provides a supporting platform for simulating controllable faults in transformer windings. All mechanical actuators, temperature control components, and electrical testing components are integrated into a single support frame. The structure and connection relationships of each component are as follows: The transformer winding is placed above the winding support disc. The winding support disc is equipped with a winding moving motor and a winding vertical moving device, which can drive the winding to move along the axial direction to accurately locate the fault simulation section and complete the simulation of displacement faults at different axial positions of the winding.
[0010] The fan heating device and the fan motor are rigidly fixed to the overall bracket. The output end of the fan heating device is connected to a multi-degree-of-freedom hot air duct. By adjusting the attitude of the hot air duct, the hot air is precisely guided to the winding fault simulation area to achieve local temperature control at the fault point. This makes the temperature field distribution of the simulated working condition match the thermal working condition of the transformer winding in actual operation, thus improving the authenticity of the sample data.
[0011] The forward movable crossbar and the backward movable crossbar are respectively located on both sides of the winding axis. The forward rotatable turntable is assembled on the forward movable crossbar, and the backward rotatable turntable is assembled on the backward movable crossbar. The forward movable crossbar is driven by the first crossbar drive motor to achieve forward and backward translation, and the backward movable crossbar is driven by the second crossbar drive motor to achieve forward and backward translation. The forward rotatable turntable and the backward rotatable turntable can freely rotate to switch wedge blocks with different slope angles. By synchronously inserting the forward and backward wedge blocks into the winding gap, axial displacement faults with different deformation angles and different intrusion depths can be simulated.
[0012] The forward and backward vertical moving screws are symmetrically arranged on both sides of the winding radially. The first and second longitudinal movable vertical rods are assembled on the forward vertical moving screw, and the third and fourth longitudinal movable vertical rods are assembled on the backward vertical moving screw. The forward and backward screw drive wheels are synchronously driven by the forward and backward screw linkage belts. The forward screw drive wheel is equipped with a handle. The first radial extrusion bracket is fixed to the first and third longitudinal movable vertical rods, and the second radial extrusion bracket is fixed to the second and fourth longitudinal movable vertical rods. The first and second radial extrusion plates are respectively fixed to the ends of the two sets of radial extrusion brackets. By rotating the handle of the forward screw drive wheel, the two sets of radial extrusion plates can be synchronously driven to perform radial feed extrusion towards the center of the winding, simulating radial deformation faults with different extrusion amounts and different deformation degrees.
[0013] The high-voltage oscillating wave system enclosure integrates a high-voltage oscillating wave power supply and a signal acquisition system. The high-voltage oscillating wave power supply system includes a high-frequency high-voltage switch and a high-frequency high-voltage DC power supply. The high-frequency high-voltage DC power supply, the high-frequency high-voltage switch, the signal acquisition system, and the transformer winding are connected in series to form a closed-loop test circuit. The high-frequency high-voltage switch is used to control the on and off of the high-frequency high-voltage DC power supply to realize the periodic charging and discharging of the transformer winding and to excite a stable oscillating wave signal. The signal acquisition system is used to synchronously acquire the oscillating wave response signals at both ends of the winding.
[0014] To achieve the above objectives, the present invention provides the following technical solution: a method for diagnosing axial displacement and radial deformation faults in transformer windings, characterized in that...
[0015] The transformer winding is placed on the winding support disk and can be moved axially by the winding moving motor and the winding vertical moving device to simulate axial displacement faults at different positions of the winding.
[0016] The fan heating device and fan motor are both fixed on the overall support. The hot air generated can be used to heat the fault simulation location locally through the multi-degree-of-freedom hot air duct to better match the temperature of the transformer winding under actual working conditions.
[0017] II. Complete Technical Implementation Plan
[0018] The method of this invention consists of five main steps: construction of historical fault samples, acquisition and preprocessing of raw signals, adaptive angle mapping and generation of two-dimensional feature images, extraction of multi-dimensional fault-sensitive features, and identification and severity classification of fault types. The specific implementation steps are as follows:
[0019] (I) Construction of historical fault data samples
[0020] Through the aforementioned fault simulation experimental platform, standardized historical fault data covering different fault types, different fault degrees, and different temperature control conditions are obtained, and a historical fault dataset covering all working conditions is constructed as the benchmark for subsequent feature threshold calibration and model training.
[0021] (2) Oscillatory wave signal acquisition and adaptive preprocessing steps
[0022] Step 1: Obtain the original oscillatory wave sampling signal:
[0023] Collect the original discrete sampling signal of the transformer oscillatory wave detection to obtain a sampling point sequence , where is the i-th discrete sampling point and N is the total number of sampling points in the sampling sequence;
[0024] Step 2: Calculate the global fluctuation intensity D:
[0025] Based on the sampling point sequence , the global fluctuation intensity D of the entire signal is calculated using the mean absolute difference method of adjacent sampling points. The calculation formula is: , where is the i-th discrete sampling point and N is the total number of sampling points in the sampling sequence;
[0026] Step 3: Construct a sub-boundary weighted local neighborhood window For each sampling point , according to its position in the sampling sequence, construct a sub-boundary local neighborhood window and assign Gaussian distance weights to each sampling point within the window , specifically:
[0027] When i ≤ W (the first segment boundary point), the window ;
[0028] When W < i ≤ N - W (the middle segment point), the window ;
[0029] When i > N - W (the last segment boundary point), the window ;
[0030] Among them, W is the preset local neighborhood half-window length (a positive integer); is the Gaussian weight coefficient of the j-th sampling point within the window <00Based on the local neighborhood window constructed in step 3 and Gaussian weights Combined with the exponential smoothing mechanism, the calculation is performed for each sampling point. Corresponding local neighborhood magnitude benchmark This is used to characterize the local amplitude level around the current point, and does not depend on the global statistics of the entire sequence. The calculation formula is:
[0033] ;
[0034] in, >0 represents a very small positive number, used to prevent the denominator from being 0 and to avoid calculation oddities; >0 represents a local amplitude smoothing coefficient, used to weaken isolated spike outlier pairs within the window. To mitigate interference and improve the robustness of local benchmarks;
[0035] Step 5: Calculate the local fluctuation homogeneity correction factor :
[0036] Based on local neighborhood window Add a local fluctuation uniformity correction factor This is used to correct the impact of local fluctuation distribution differences on adaptive adjustment. It is calculated only using sampling points within a window, without global statistics involved. The calculation formula is: ;
[0037] in, , for window The simple averaging of the amplitudes of the inner sampling points is only used for the normalization of the correction factor; The value range is [0,1]. The more uniform the fluctuation within the window, the better. The closer to 1;
[0038] Step 6: Calculate the point-by-point reverse adaptive focusing coefficient ;
[0039] Based on the global fluctuation intensity D calculated in step 2 and the local neighborhood amplitude benchmark in step 4. Step 5 Local fluctuation homogeneity correction factor Constructing a multi-layered coupled point-by-point adaptive focusing coefficient This achieves precise inverse adjustment of local fluctuations and adaptive coefficients, calculated using the following formula:
[0040] ;
[0041] in, , These are the preset lower and upper limits (fixed hyperparameters) of the adaptive focusing coefficient; this coefficient satisfies the following condition: the smaller the local neighborhood fluctuation, the better. The larger the value, the more it enhances the details of weak signals; the greater the fluctuation in the local neighborhood, The smaller the value, the better it suppresses distortion of strong fluctuation signals;
[0042] Step 7: Complete the angle mapping of the oscillating wave signal;
[0043] The point-by-point adaptive focusing coefficient calculated in step 6 Combined with normalized sampling points Complete each sampling point to polar coordinate angle The mapping provides adaptively optimized input for subsequent local neighborhood adaptive focusing mapping feature extraction, where normalized sampling points... and angle The calculation formulas are as follows: ; .
[0044] Step 8: Generate adaptively optimized 2D feature image
[0045] Based on angle sequence Two-dimensional feature images of GASF and GADF are constructed respectively to complete the conversion from one-dimensional signal to two-dimensional image: The adaptively preprocessed two-dimensional feature image can highlight weak fault textures, suppress strong fluctuation artifacts, and improve the distinguishability of fault features.
[0046] (III) Extraction of multi-dimensional fault-sensitive features
[0047] Step 1: Simultaneously extract GLCM texture features, grayscale statistical features, and geometric morphological features from the 2D feature image. All features are calculated directly from the image without any additional global priors. The definitions, physical meanings, and fault responses of each feature are as follows:
[0048] Gray-level co-occurrence matrix (GLCM) texture features
[0049] Pick Four-directional mean, where d is the pixel step size, which refers to the spatial distance between two pixels participating in the co-occurrence statistics. A value of 1 indicates that only directly adjacent pixels are counted.
[0050] Extract 4 core texture parameters:
[0051] 1. Contrast Ratio: : Characterizes the intensity of grayscale jumps; the larger the value, the more obvious the texture contrast.
[0052] 2. Correlation: : Represents the continuity of texture space; the closer the value is to 1, the more regular the texture.
[0053] 3. Energy: : Characterizes texture uniformity; the larger the value, the more regular the texture.
[0054] 4. Entropy: : Represents texture complexity; the larger the value, the more chaotic the texture.
[0055] Where: d is the pixel step size, which refers to the spatial distance between two pixels participating in the co-occurrence statistics. A value of 1 indicates that only directly adjacent pixels are counted; i and j are the quantized gray level indices (not pixel coordinates), with a value range of [0, L−1] (L is the total number of quantized gray levels in the image), representing different gray levels; Let be the normalized co-occurrence probability of gray levels i and j, that is, the proportion of the occurrence frequency of gray levels i and j in adjacent pixels out of the total statistical frequency, satisfying the following condition: ; (Appearing only in correlation) are the weighted mean and standard deviation of gray level i, respectively, given by The weighted calculation yielded the result.
[0056] Gray-scale statistical characteristics
[0057] 1. Gray-scale mean: : Characterizes the average gray level of the entire image, reflecting the overall energy strength of the signal.
[0058] 2. Gray-scale variance: : Characterizes the degree of dispersion of grayscale relative to the mean, and characterizes the intensity of signal fluctuation and distortion.
[0059] 3. Grayscale peak value: : Represents the maximum gray level of an image, corresponding to the extreme value and impulse component in the signal.
[0060] 4. Gray-level histogram entropy: Characterizes the degree of disorder and unevenness in grayscale distribution, and distinguishes between regular and chaotic fault modes.
[0061] Where: M and N are the total number of rows and columns of the image, and the image space size is M×N; I(x,y) is the gray value at pixel coordinates (x,y); L is the total gray level of the image; p(l) is the normalized probability of gray level l.
[0062] Step 2: Feature Cascading and Standardization
[0063] 1. Feature Concatenation: The GLCM texture features and grayscale statistical features of the GASF and GADF images are concatenated sequentially to form a single high-dimensional original feature vector. The subscript G corresponds to the GASF image, and D corresponds to the GADF image; each segment represents the same type of feature of the corresponding image.
[0064] 2. Min - Max Normalization: Eliminate the differences in different feature dimensions and numerical ranges, and ensure the balanced contribution of each dimension to classification and grading: , is the lower limit of the k - th dimension feature, is the upper limit; Map the features to the [0, 1] interval through linear transformation.
[0065] 3. Standardize the feature set: Normalize each dimension of features and combine them into the final input features: .
[0066] (IV) Fault type identification and degree grading steps
[0067] Step 1: Fault type identification based on feature weighted fusion and lightweight CNN
[0068] Introduce the fault type sensitive weight , strengthen the highly relevant features and weaken the redundant features: , y1 and y2 are the labels of axial displacement and radial deformation respectively; Input the weighted features into the lightweight CNN with a local attention module, and output the fault type determination result.
[0069] Step 2: Four - level quantitative grading of fault degree
[0070] Construct the fault degree quantitative index K: , is the normal feature reference value, is the maximum value of the feature in the severe fault state; Set the grading threshold , decision rule: K < K1: normal state; K1 ≤ K < K2: minor fault; K2 ≤ K < K3: medium fault; K ≥ K3: severe fault.
[0071] Step 3: Output the diagnosis result, integrate the fault type (axial displacement / radial deformation) and the fault degree (normal / minor / medium / severe), generate a standardized diagnosis report and give corresponding maintenance suggestions.
[0072] (III) Beneficial effects
[0073] The present invention provides a method for diagnosing the axial displacement and radial deformation faults of a transformer winding. It has the following beneficial effects:
[0074] 1. Highly restore the fault simulation conditions: Adopt an integrated platform with a switchable slope wedge block, radially adjustable extrusion sheet, and local hot - air temperature control, which can accurately simulate the axial displacement and radial deformation faults at different positions, different degrees, and different temperature conditions. The sample data fits the engineering reality, solving the problems of single traditional simulation conditions and lack of temperature boundaries.
[0075] 2. Signal preprocessing without global statistical bias: It adopts a boundary Gaussian weighted local window, anomaly suppression amplitude benchmark, and fluctuation uniformity correction factor. The entire process relies only on local neighborhood calculations and does not introduce easily disturbed statistics such as global mean and global extrema, which significantly improves the accuracy of weak fault feature extraction.
[0076] 3. Adaptive mapping bidirectional optimization: It achieves inverse adjustment of local fluctuations and mapping gain, enhances details of weak signals, suppresses distortion caused by severe fluctuations, and produces clear image textures with fewer artifacts, providing high-quality input for subsequent feature extraction.
[0077] 4. Strong fault differentiation and classification capabilities: Constructing a GLCM texture and grayscale statistical feature system to specifically distinguish between linear regular axial displacement and nonlinear chaotic radial deformation; achieving accurate classification at four levels through quantitative index K, breaking through the limitation of existing technologies that only make qualitative judgments.
[0078] 5. Excellent engineering practicality: The lightweight CNN model has low computational overhead and fast response. The device and algorithm are integrated in a closed loop, which can be used for laboratory sample construction and can also be adapted to rapid on-site detection, meeting the dual needs of online diagnosis and offline analysis of transformer winding faults. Attached Figure Description
[0079] Figure 1 This is a front view of the transformer winding axial displacement and radial deformation fault simulation device in an embodiment of the present invention;
[0080] Figure 2 This is a rear front view of the transformer winding axial displacement and radial deformation fault simulation device in an embodiment of the present invention;
[0081] Figure 3 This is a top view of the transformer winding axial displacement and radial deformation fault simulation device in an embodiment of the present invention;
[0082] Figure 4 This is a right-side view of the transformer winding axial displacement and radial deformation fault simulation device in an embodiment of the present invention.
[0083] Figure 5 This is a left-side view of the transformer winding axial displacement and radial deformation fault simulation device in an embodiment of the present invention. Figure 6 This is a flowchart of the transformer winding axial displacement and radial deformation fault diagnosis method in an embodiment of the present invention.
[0084] In the diagram, 1. Transformer winding; 2. High-voltage oscillating wave power supply and signal acquisition system; 3. Multi-degree-of-freedom hot air duct; 4. Forward rotatable turntable; 5. Forward movable horizontal bar; 6. First longitudinal movable vertical bar; 7. Second longitudinal movable vertical bar; 8. First horizontal bar drive motor; 9. Second horizontal bar drive motor; 10. Forward vertical bar moving screw; 11. Winding moving motor; 12. Fan heating device; 13. Fan motor; 14. Reverse rotatable turntable; 15. Reverse movable horizontal bar; 16. Winding vertical moving device; 17. Third longitudinal movable vertical bar; 18. Fourth longitudinal movable vertical bar; 19. Reverse vertical bar moving screw; 20. Reverse first slope angle wedge block; 21. Reverse second slope angle wedge block. 22. Rearward third-angle wedge block; 23. Rearward fourth-angle wedge block; 24. First radial extrusion bracket; 25. Second radial extrusion bracket; 26. First radial extrusion plate; 27. Second radial extrusion plate; 28. Winding support disc; 29. Forward first-angle wedge block; 30. Forward second-angle wedge block; 31. Forward third-angle wedge block; 32. Forward fourth-angle wedge block; 33. Emergency stop button; 34. Control system display screen; 35. High-voltage oscillating wave system housing; 36. Forward rotatable turntable drive belt; 37. Rearward rotatable turntable drive belt; 38. Rearward lead screw drive wheel; 39. Forward lead screw drive wheel handle; 40. Forward lead screw drive wheel; 41. Forward and backward lead screw linkage belt Detailed Implementation
[0085] 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.
[0086] Example 1
[0087] This embodiment uses a simulated minor axial displacement fault as an example to demonstrate how the method of this invention can achieve fault identification and severity assessment. It simulates different types and degrees of faults in transformer windings, including:
[0088] Step 1: Device Failure Simulation
[0089] 1.1: The transformer winding 1 is placed on the winding support disk 28 and the first radial extrusion plate 26 and the second radial extrusion plate 27 are driven by rotating the forward screw drive wheel handle 39 to extrude the transformer winding 1 in a radial manner to achieve radial fixation.
[0090] 1.2: Connect the high-voltage oscillating wave power supply and the signal acquisition system 2 in series across the two ends of the transformer winding 1;
[0091] 1.3: Rotate the forward rotatable turntable 4 and the backward rotatable turntable 14 so that the wedge blocks with the same slope angle point towards the transformer winding 1. Start the first crossbar drive motor 8 and the second crossbar drive motor 9 to gradually move the forward and backward wedge blocks to almost touch the transformer winding 1. Adjust the winding vertical movement device 16 to move the winding 1 to the fault simulation point. Continue to start the first crossbar drive motor 8 and the second crossbar drive motor 9 to gradually insert the forward and backward wedge blocks into the transformer winding 1, causing the transformer winding 1 to experience an axial displacement fault. In the axial displacement fault simulation, different degrees of transformer winding axial displacement faults can be simulated by controlling the degree of insertion of the wedge blocks into the transformer winding 1 and changing the wedge blocks with different slope angles on the forward rotatable turntable 4 and the backward rotatable turntable 14.
[0092] 1.4: Radial Deformation Fault Simulation: By rotating the forward lead screw drive wheel handle 39, the first radial extrusion plate 26 and the second radial extrusion plate 27 installed on the forward vertical rod moving lead screw 10 and the backward vertical rod moving lead screw 19 are driven to further extrude the transformer winding 1, so as to realize the radial deformation fault simulation of the transformer winding 1. The degree of fault is changed by rotating the forward lead screw drive wheel handle 39 to control the degree of extrusion of the first radial extrusion plate 26 and the second radial extrusion plate 27.
[0093] 1.5: Adjust the multi-degree-of-freedom hot air duct 3 so that the air outlet is aligned with the fault simulation point, and perform local heating on the fault simulation point of axial displacement or radial deformation of transformer winding 1 to better match the temperature of transformer winding under actual working conditions.
[0094] 1.6: Repeat steps 2) to 5) to simulate axial displacement or radial deformation faults of transformer windings with different fault degrees and types.
[0095] Step 2: Acquire the original oscillation wave sampling signal
[0096] After completing the fault simulation setup, wires were led out from the beginning and end of transformer winding 1, respectively, and connected to the high-voltage oscillation wave power supply and signal acquisition system 2. The high-frequency high-voltage switch then periodically charged and discharged transformer winding 1, obtaining transformer oscillation wave waveform data through the high-voltage oscillation wave power supply and signal acquisition system 2. The sampling frequency was set to 10kHz, and the sampling duration was 1s, resulting in a sampling point sequence. (i.e., N=10000) Let be the amplitude of the i-th sampling point.
[0097] Step 3: Calculate the global fluctuation intensity D
[0098] Based on sampling point sequence According to the formula , the global fluctuation intensity D = 0.86 is calculated, and this value characterizes the macroscopic fluctuation level of the entire oscillatory wave signal, providing a global benchmark for subsequent adaptive regulation.
[0099] Step 4: Construct a weighted local neighborhood window with sub-boundaries
[0100] Set the local neighborhood half-window length W = 4 (determined according to the sampling frequency and fault characteristic scale), and for each sampling point , construct a local neighborhood window with sub-boundaries , and calculate the Gaussian weight of each sampling point within the window :
[0101] When i ≤ 3000 (the first segment boundary point), the window ;
[0102] When 3000 < i ≤ 7000 (the middle segment points), the window ;
[0103] When i > 7000 (the last segment boundary point), the window , and the Gaussian weight is calculated according to the formula For example, for the sampling point with an index difference of 1 from the current point within the window , the weight , and the closer the distance, the greater the weight. The closer the distance, the greater the weight.
[0104] Step 5: Calculate the local neighborhood amplitude reference with outlier suppression
[0105] Set , , and calculate the local neighborhood amplitude reference corresponding to each sampling point according to the formula ; when there are isolated spike points within the window, , the contribution of this spike point is significantly weakened to ensure the robustness of.
[0106] Step 6: Calculate the local fluctuation uniformity correction factor
[0107] Calculate the local fluctuation uniformity correction factor corresponding to each sampling point according to the formula , where ; for sampling points in the normal state, the fluctuations within the window are uniform, ≈0.92; for sampling points with minor faults, the fluctuations within the window are slightly non-uniform, ≈0.65.
[0108] Step 7: Calculate the point-by-point reverse adaptive focusing coefficient
[0109] set up According to the formula Calculate each sampling point Corresponding adaptive focusing coefficient Under normal conditions, local fluctuations are small. ≈9.3, enhancing subtle details; under minor fault conditions, ≈7.1; Under severe fault conditions, local fluctuations are large. ≈1.8, suppressing distortion.
[0110] Step 8: Complete the angle mapping of the oscillating wave signal
[0111] According to the formula , sampling points Normalizing to the interval [−1,1], we get Then follow the formula Calculate the polar coordinate angle of each sampling point. This completes the angle mapping.
[0112] Step 9: Generate an adaptively optimized 2D feature image
[0113] Angle based on all sampling points According to the formula A matrix (10000×10000) is constructed and transformed into a two-dimensional feature image; this image can clearly present the texture details corresponding to weak faults without obvious artifacts.
[0114] Step 10: Extract multi-dimensional fault-sensitive features from the two-dimensional feature image
[0115] Multi-dimensional feature extraction from two-dimensional feature images:
[0116] Texture features: Contrast, correlation, energy, and entropy of GLCM are extracted to obtain four texture parameters;
[0117] Gray-level features: Extract gray-level mean, gray-level variance, gray-level peak value, and gray-level entropy to obtain four gray-level parameters;
[0118] Eight features were extracted and normalized to obtain a standardized fault-sensitive feature set. .
[0119] Step 11: Identification of Two Types of Winding Faults Based on Feature Fusion
[0120] 11.1 Feature-weighted fusion: According to the formula Calculate the weight coefficient for each feature. ,in (Axial displacement) (Radial deformation), geometric features that are highly correlated with the two types of faults have larger weights;
[0121] 11.2 Fault Classification and Recognition: The weighted and fused features are input into the improved lightweight CNN model. The model removes the two redundant convolutional layers of the traditional CNN and adds a local attention module to output the fault recognition results. According to the test, the model achieves an accuracy of 96.2% in recognizing two types of faults: axial displacement and radial deformation.
[0122] Step 12: Grading and Assessment of Winding Fault Severity
[0123] 12.1 Constructing a quantitative index K for the degree of failure: according to the formula Calculate the value of K, where These are the characteristic standard values for this type of transformer under normal conditions. This represents the maximum characteristic value under severe fault conditions;
[0124] 12.2 Four-level severity classification: setting thresholds If K=0.32 for a certain test sample and the fault identification result is axial displacement, it is judged as a minor axial displacement fault; if K=0.68, it is judged as a moderate axial displacement fault.
[0125] Step 13: Output the fault identification and severity assessment results
[0126] Based on the above results, the following fault diagnosis report is output: The simulated fault is a slight axial displacement fault in the winding. It is recommended to monitor the winding operation status regularly and perform an oscillation wave retest every 3 months to ensure that the fault does not worsen.
[0127] Example 2
[0128] The difference between this embodiment and Embodiment 1 is that the fault simulation is changed to a medium radial deformation fault, while the other parameters are the same as in Embodiment 1.
[0129] After processing by the method of this invention, a medium radial deformation fault (K=0.62) in the transformer winding was successfully identified, and the maintenance suggestion was: immediately inspect the winding, adjust the winding position, and avoid aggravating the fault.
[0130] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for diagnosing axial displacement and radial deformation faults in transformer windings, characterized in that, include: Step 1: Acquire the original oscillation wave sampling signal; Step 2: Calculate the global fluctuation intensity D; The calculation formula is: ,in Let be the i-th discrete sampling point, and N be the total number of points in the sampling sequence; Step 3: For each sampling point Based on its position in the sampling sequence, a local neighborhood window for the boundary is constructed. And assign Gaussian distance weights to each sampling point within the window. The calculation formula is: Where W is the preset local neighborhood half-window length (positive integer); For window The j-th sampling point relative to the current sampling point Gaussian weighting coefficients; Step 4: Calculate the local neighborhood magnitude benchmark with outlier suppression ; The calculation formula is: ;in, >0 represents a very small positive number, used to prevent the denominator from being 0 and to avoid calculation oddities; >0 represents a local amplitude smoothing coefficient, used to weaken isolated spike outlier pairs within the window. To mitigate interference and improve the robustness of local benchmarks; Step 5: Calculate the local fluctuation homogeneity correction factor ; The calculation formula is: ;in, , for window The simple averaging of the amplitudes of the inner sampling points is only used for the normalization of the correction factor; Step 6: Calculate the point-by-point reverse adaptive focusing coefficient ; The calculation formula is: ;in, , These are the preset lower and upper limits (fixed hyperparameters) of the adaptive focusing coefficient, respectively. Step 7: Complete the angle mapping of the oscillating wave signal; Normalized sampling points and angle The calculation formulas are as follows: ; ;in, and Let be the amplitude of the i-th and j-th discrete sampling points of the original oscillating wave signal.
2. The signal processing method according to claim 1, characterized in that, In step 3, the Gaussian weighting coefficients Its core function is to enable local neighborhood windows. Inside, distance from the current sampling point The closer the sampling point, the greater the weight, which improves the targeting of local neighborhood features. This is different from the local window with uniform weight in the existing technology, and further improves the accuracy of adaptive adjustment.
3. The signal processing method according to claim 1, characterized in that, In step 4, through The item implements outlier suppression when the window... There are isolated spikes in memory ( When the value increases significantly, the exponential term decays rapidly, reducing the impact of the peak on the local neighborhood amplitude benchmark. The contribution addresses the technical problems of existing local benchmarks being susceptible to outlier interference and having poor robustness.
4. The signal processing method according to claim 1, characterized in that, In step 5, the local fluctuation uniformity correction factor This new design addresses the shortcomings of existing technologies that only consider local amplitudes and neglect the uniformity of local fluctuations, enabling adaptive focusing coefficients. Simultaneously responding to local amplitude levels and the uniformity of fluctuation distribution further enhances the rationality and accuracy of adaptive adjustment.
5. The signal processing method according to claim 1, characterized in that, In step 6, the point-to-point adaptive focusing coefficient The design adopts "global fluctuation constraint + local dual-factor collaborative driving". It does not introduce global statistics such as global mean and global extreme value of the whole sequence. It only calculates based on local neighborhood features and global fluctuation benchmark. This solves the technical defects of existing adaptive methods that rely on global statistics, have local adjustment lag, and are susceptible to global anomaly interference, and achieves accurate adaptive focusing of local signal features.
6. The method for oscillating wave signal processing and fault identification based on local neighborhood adaptive focusing according to claim 1, characterized in that, Also includes: The angle sequence is constructed into a two-dimensional feature image, and image fault-sensitive features are extracted. Based on the features, two types of faults, namely axial displacement and radial deformation of transformer windings, are identified. The corresponding fault severity is determined according to four levels: normal, minor, moderate, and severe. Finally, the diagnostic results of fault type and fault severity are output.