Method and system for identifying coil looseness deformation by turn-to-turn vibration difference under frequency sweep excitation

Abstract

The application discloses a method and system for identifying loose deformation of turns under inter-turn vibration difference under sweep excitation, and belongs to the technical field of electrical equipment state detection. The existing coil loose deformation detection method is difficult to position a specific turn. The application applies a sweep voltage to a coil to be detected, collects vibration acceleration signals of each turn at each frequency point, extracts vibration acceleration amplitudes, and generates a vibration amplitude-frequency sequence. The correlation coefficients between each sequence are calculated to obtain the average similarity of each turn. The average similarity between turns of all turns forms a one-dimensional data set, and the K-means clustering is used to take the cluster center as a reference value. The ratio of the average similarity between turns of each turn to the reference value is used as a difference identification coefficient. The turns with a difference identification coefficient less than a preset judgment threshold are determined as loose deformation coils. The application can accurately position each turn without historical data.

Description

Technical Field

[0001] This invention relates to the field of electrical equipment condition monitoring technology. More specifically, this invention relates to a method and system for identifying coil loosening and deformation by detecting inter-turn vibration differences under swept-frequency excitation. Background Technology

[0002] During long-term operation, the windings of coil-type electrical equipment such as transformers and reactors are affected by factors such as the enormous electromagnetic force generated by short-circuit inrush currents, long-term mechanical vibration, and thermal effects, which may cause the windings to loosen or deform. Loosening and deformation of the windings alter the mechanical structural characteristics of the coils, reduce the equipment's short-circuit withstand capability, and in severe cases, lead to equipment failure and threaten the safe operation of the power system. Therefore, effectively detecting the loosening and deformation status of the windings is of great significance for ensuring the safe operation of equipment.

[0003] In existing technologies, the main methods for detecting the condition of transformer windings include the short-circuit impedance method and the frequency response method. The short-circuit impedance method determines whether the winding has deformed by measuring the change in the transformer's short-circuit impedance value. However, this method has low sensitivity. When the winding is slightly loose or slightly deformed, the change in short-circuit impedance is very small and difficult to detect effectively. The frequency response method applies a sweep frequency voltage signal to the winding and measures the winding's transfer function curve. By comparing the current transfer function curve with the historical curve, the winding condition is determined. The frequency response method improves the detection sensitivity to some extent, but its frequency response waveform is complex, the judgment relies heavily on experience, and it still depends on the comparison of historical data.

[0004] To address the shortcomings of the aforementioned methods, Chinese patent CN1776441A proposes a method for detecting the state of transformer windings using a sweep frequency power supply excitation device. This method involves using a variable frequency constant current source to output a sweep frequency excitation signal to the transformer windings via an excitation transformer. A vibration sensor mounted on the transformer housing measures the vibration response signals of the windings at different excitation frequencies. Spectral analysis is used to obtain the resonant frequency curve of the transformer windings. By comparing the current resonant frequency curve with historical curves from the same equipment or with the resonant frequency curves of other phase windings from the same equipment, it is determined whether the windings are loose or deformed. Treating the windings as an elastic system and utilizing the vibration response to detect changes in the mechanical dynamics of the windings improves the sensitivity of loosening detection to some extent, but it still relies on the comparison of historical or inter-phase reference data.

[0005] Chinese patent CN105699838A, a method and apparatus for detecting the condition of transformer windings, further proposes a detection scheme based on a vibration frequency response matrix. This involves obtaining the current vibration frequency response curves at each measuring point through a sweep frequency excitation test on the transformer windings, constructing a current vibration frequency response matrix, and normalizing this matrix. Simultaneously, a normalized historical vibration frequency response matrix is ​​obtained based on historical transformer winding vibration frequency response curves, and this matrix is ​​decomposed to obtain a historical frequency response basis matrix. Statistics are calculated based on the current vibration frequency response matrix and the historical frequency response basis matrix. The average value of these statistics is compared with the upper limit value to determine whether an abnormality has occurred in the transformer windings. This method converts the vibration frequency response curves into matrix form for comprehensive processing, improving detection accuracy. However, it still relies on the basis matrix established from historical data as a judgment benchmark.

[0006] Chinese patent CN115905836B, "A Method for Extracting Features of Transformer Winding Looseness Faults Based on Chaos Theory and Ephemeral Optimization K-means Algorithm," proposes a feature extraction scheme for winding looseness faults from the perspective of analyzing the chaotic characteristics of vibration signals. It calculates the maximum Lyapunov exponent of the transformer vibration signal to determine its chaotic characteristics, and extracts the correlation dimension and Kolmogorov entropy as chaotic features. Simultaneously, it employs a K-means clustering algorithm optimized by the Ephemeral Optimization algorithm to cluster the phase space trajectory of the vibration signal, calculating the sum of cluster central moments and the radius vector offset as geometric features. Combining the chaotic features with the geometric features to construct a feature vector is used to identify transformer winding looseness faults. This expands the fault feature extraction approach from the perspective of chaotic time series analysis, but it still requires extracting features as reference values ​​under normal transformer conditions.

[0007] While the aforementioned existing technologies have achieved winding condition detection to some extent, they all share the following common drawbacks: All methods rely on historical data or reference data under healthy conditions as a comparison benchmark. However, in actual engineering projects, many devices under test lack complete factory or historical testing data for reference. Differences between different operating conditions and individual devices also affect the reliability of the comparison results. These dependencies severely limit the applicability of the methods. Furthermore, the aforementioned methods primarily assess the condition of the entire transformer winding or a specific phase winding, failing to accurately pinpoint the specific coil turn where loosening or deformation has occurred. These shortcomings limit the application scope and diagnostic accuracy of existing detection methods. Summary of the Invention

[0008] One object of the present invention is to solve at least the above-mentioned problems and to provide at least the advantages that will be described later.

[0009] To achieve these objectives and other advantages according to the present invention, a method for identifying coil loosening and deformation under frequency sweep excitation and inter-turn vibration difference is provided, comprising: S1. Apply the sweep frequency voltage signal to both ends of the coil under test, and collect the vibration acceleration signal of multiple preset measurement points of the coil under test at each frequency point through a non-contact vibration measurement device. Each turn of the coil has a preset measurement point, and each preset measurement point is located on the same circumferential angle generatrix at the corresponding axial position of each turn of the coil under test. S2. Extract the vibration acceleration amplitude of each preset measurement point at each frequency point from the vibration acceleration signal, and arrange the vibration acceleration amplitude of the preset measurement points corresponding to the same coil at all frequency points in ascending order of frequency to generate the vibration amplitude-frequency sequence of this coil. S3. Calculate the correlation coefficient between the vibration amplitude-frequency sequence of any two turns of coil to obtain the correlation coefficient matrix. For each turn of coil, according to the correlation coefficient matrix, remove the correlation coefficient between this turn of coil and itself, and calculate the average value of the correlation coefficient between this turn of coil and all other turns of coil as the average inter-turn similarity of this turn of coil. S4. The average inter-turn similarity of all coils is taken as a one-dimensional dataset. The K-means clustering algorithm is used to divide the one-dimensional dataset into two clusters. The cluster with the larger cluster center value is selected as the normal cluster, and the cluster center value of the normal cluster is used as the benchmark value. S5. Calculate the ratio of the average inter-turn similarity of each coil to the benchmark value, and use it as the difference identification coefficient. Coils with a difference identification coefficient less than the preset judgment threshold are judged as loose and deformed coils.

[0010] Preferably, the sweep frequency range of the sweep voltage signal is 20Hz~2000Hz, and the sweep step size is no more than 5Hz.

[0011] Preferably, the preset judgment threshold ranges from 0.75 to 0.90.

[0012] Preferably, the correlation coefficient between the vibration amplitude and frequency sequences of any two turns of the coil is calculated using the Pearson correlation coefficient algorithm.

[0013] Preferably, the non-contact vibration measurement device is a laser vibration meter.

[0014] Preferably, the duration of the sweep voltage signal at each frequency point is 0.5s to 2s, and the adjacent frequency points are continuously swept and switched at a rate not exceeding 10Hz / s.

[0015] Preferably, in step S3, before calculating the average inter-turn similarity of each coil turn, the correlation coefficient matrix undergoes spatial neighborhood decorrelation preprocessing, including the following steps: For each coil i, the turn spacing is defined by the absolute value of the difference between the turn numbers of two coils along the axial direction, according to their axial physical position. All other coils whose turn spacing from coil i is greater than or equal to the preset spatial step size parameter L are selected to form the reference sample set G. i Where L takes the value of 2, 3 or 4, and N is a positive integer greater than 5; If the reference sample set G i If the number of coils in the sample is greater than or equal to the preset minimum statistical number T, then only this coil i and the reference sample set G are considered. i The arithmetic mean of the correlation coefficients of each coil is calculated and used as the average inter-turn similarity for that turn. If the reference sample set G... i If the number of coils in the loop is less than the preset minimum statistical number T, then the arithmetic mean of the correlation coefficients between coil i and all other coils is taken as the average inter-turn similarity of this coil, where T can be 3, 4 or 5.

[0016] Preferably, between steps S4 and S5, an absolute validity check and adaptive correction of the benchmark value are performed, including the following steps: The reference value obtained in step S4 is compared with the preset absolute reference lower limit threshold γ. If the reference value is greater than or equal to γ, the reference value is kept unchanged and the process proceeds to step S5. If the reference value is less than γ, the coil under test is determined to be in an abnormal state. The reference value is corrected to γ, and the process proceeds to step S5 with the corrected reference value. An overall status warning signal is output to mark that the current coil under test has a global health degradation. The value of γ ranges from 0.70 to 0.85, and γ is less than or equal to the preset judgment threshold. The overall status warning signal is used as the confidence level marker for the determination result in step S5. When no overall status warning signal is output, the determination result in step S5 is the final deterministic result. When an overall status warning signal is output, the determination result in step S5 is the suspected result.

[0017] This invention provides a system for identifying coil loosening and deformation under frequency sweep excitation and inter-turn vibration difference, comprising: A sweep frequency signal generator is used to apply a sweep frequency voltage signal to both ends of the coil under test; A non-contact vibration measurement device is used to collect vibration acceleration signals at each frequency point from multiple preset measurement points of the coil under test. Each coil has a preset measurement point, and each preset measurement point is located on the same circumferential angle generatrix at the corresponding axial position of each coil of the coil under test. The processor is communicatively connected to the sweep frequency signal generator and the non-contact vibration measurement device. It controls the sweep frequency signal generator to output a sweep frequency voltage signal and receives the vibration acceleration signal collected by the non-contact vibration measurement device. It extracts the vibration acceleration amplitude of each preset measurement point at each frequency point from the vibration acceleration signal, and arranges the vibration acceleration amplitude of the preset measurement points corresponding to the same coil at all frequency points in ascending order of frequency to generate the vibration amplitude-frequency sequence of the coil. The correlation coefficient matrix is ​​used to calculate the correlation coefficient between the vibration amplitude-frequency sequence of any two coil turns. For each coil turn, the correlation coefficient between the coil turn and itself is removed according to the correlation coefficient matrix. The average correlation coefficient between the coil turn and all other coil turns is then calculated as the average inter-turn similarity of the coil turn. The average inter-turn similarity of all coils is used as a one-dimensional dataset. The K-means clustering algorithm is used to divide the one-dimensional dataset into two clusters. The cluster with the larger cluster center value is selected as the normal cluster, and the cluster center value of the normal cluster is used as the benchmark value. And the ratio of the average inter-turn similarity of each coil to the benchmark value is used as the difference identification coefficient, and coils with a difference identification coefficient less than a preset judgment threshold are judged as loose and deformed coils.

[0018] Preferably, before calculating the average inter-turn similarity of each coil turn, the processor also performs spatial neighborhood decorrelation preprocessing on the correlation coefficient matrix, including: For each coil i, the turn spacing is defined by the absolute value of the difference between the turn numbers of two coils along the axial direction, according to their axial physical position. All other coils whose turn spacing from coil i is greater than or equal to the preset spatial step size parameter L are selected to form the reference sample set G. i Where L takes the value of 2, 3 or 4, and N is a positive integer greater than 5; If the reference sample set G i If the number of coils in the sample is greater than or equal to the preset minimum statistical number T, then only this coil i and the reference sample set G are considered. i The arithmetic mean of the correlation coefficients of each coil is calculated and used as the average inter-turn similarity for that turn. If the reference sample set G... i If the number of coils in the loop is less than the preset minimum statistical number T, then the arithmetic mean of the correlation coefficients between coil i and all other coils is taken as the average inter-turn similarity of this coil, where T can be 3, 4 or 5.

[0019] The present invention has at least the following beneficial effects: First, by constructing the average similarity between turns as a self-reference index and using K-means clustering to extract benchmark values ​​from the data of the coil under test itself, this invention eliminates the dependence on historical data from the factory or health status reference curves. It can improve the positioning accuracy of loosening deformation from the traditional winding phase level or overall level to the specific coil turn level, effectively solving the technical problem that a large number of devices under test lack historical records and are difficult to assess in actual engineering.

[0020] Secondly, before calculating the average similarity between turns, the present invention employs a spatial neighborhood decorrelation preprocessing step. By using a preset spatial step size parameter, coil data that are physically adjacent to the target turn are eliminated. This effectively weakens the false high correlation between adjacent turns caused by mechanical vibration coupling in a healthy state, making the calculated average similarity between turns more accurately reflect the differences in vibration characteristics caused by changes in the stiffness of each turn's own structure. This significantly improves the accuracy and reliability of loosening deformation detection.

[0021] Third, in response to the potential failure of the pure relative comparison method under global degradation conditions, this invention uses an absolute benchmark lower limit threshold to verify and adaptively correct the benchmark value obtained from clustering. This ensures that when the vibration consistency of all turns declines synchronously due to overall aging or uniform attenuation of preload, the judgment benchmark will not collapse accordingly, thereby avoiding the omission of defects caused by false amplification of ratios, and adding an overall state warning confidence mark to the diagnostic results.

[0022] Fourth, by defining the difference identification coefficient, which is the ratio of the average similarity of each turn to the normal cluster reference value, the present invention quantifies the degree of loosening and deformation of the coil into a pure relative deviation index, effectively suppressing the influence of common-mode errors in field measurement such as excitation voltage amplitude fluctuations and sensor absolute sensitivity differences on the judgment results, so that the judgment threshold has good working condition adaptability and engineering operability.

[0023] Fifth, this invention addresses situations where the axial end boundary conditions of the coil are special and the available reference samples are insufficient. It sets a minimum statistical quantity threshold as the switching condition for the judgment logic. When the target turn is located at the end, resulting in too few effective reference samples, the algorithm automatically switches back to the full sample averaging strategy. While ensuring statistical stability, it retains the identifiable information of loose and deformed end turns to the greatest extent, making the whole method widely applicable to coils with different number of turns and different spatial positions.

[0024] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0025] Figure 1This is a flowchart of a method for identifying coil loosening and deformation under frequency sweep excitation for inter-turn vibration difference identification, one of the technical solutions of the present invention. Detailed Implementation

[0026] The present invention will now be described in further detail with reference to examples, so that those skilled in the art can implement it based on the description.

[0027] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation plan are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified.

[0028] like Figure 1 As shown, the present invention provides a method for identifying coil loosening and deformation by inter-turn vibration difference under swept frequency excitation, comprising the following steps: S1. Apply the sweep frequency voltage signal to both ends of the coil under test, and collect the vibration acceleration signal of multiple preset measurement points of the coil under test at each frequency point through a non-contact vibration measurement device. Each turn of the coil has a preset measurement point, and each preset measurement point is located on the same circumferential angle generatrix at the corresponding axial position of each turn of the coil under test. S2. Extract the vibration acceleration amplitude of each preset measurement point at each frequency point from the vibration acceleration signal, and arrange the vibration acceleration amplitude of the preset measurement points corresponding to the same coil at all frequency points in ascending order of frequency to generate the vibration amplitude-frequency sequence of this coil. S3. Calculate the correlation coefficient between the vibration amplitude-frequency sequence of any two turns of coil to obtain the correlation coefficient matrix. For each turn of coil, according to the correlation coefficient matrix, remove the correlation coefficient between this turn of coil and itself, and calculate the average value of the correlation coefficient between this turn of coil and all other turns of coil as the average inter-turn similarity of this turn of coil. S4. The average inter-turn similarity of all coils is taken as a one-dimensional dataset. The K-means clustering algorithm is used to divide the one-dimensional dataset into two clusters. The cluster with the larger cluster center value is selected as the normal cluster, and the cluster center value of the normal cluster is used as the benchmark value. S5. Calculate the ratio of the average inter-turn similarity of each coil to the benchmark value, and use it as the difference identification coefficient. Coils with a difference identification coefficient less than the preset judgment threshold are judged as loose and deformed coils.

[0029] In the above technical solution, the method of the present invention is applicable to coil-type electrical equipment, such as transformer windings, reactor windings, etc. The coil to be tested is composed of several turns of coil arranged sequentially along the axial direction. The total number of turns of the coil is denoted as N, where N is a positive integer greater than or equal to 3. Each turn of coil is a ring conductor wrapped around an iron core or frame, and each turn of coil has a one-to-one corresponding spatial position along the axial direction. A preset measuring point is set on each turn of the coil to be tested. Each preset measuring point is set on the same circumferential angle generatrix at the corresponding position of each turn of the coil to be tested in the axial direction. The specific setting method is as follows: with the central axis of the coil as the reference, the same angular position is selected on the outer circumference of each turn of the coil, for example, the 12 o'clock position directly above the top of the coil, as the measuring point position, to ensure that the angle of each measuring point in the circumferential direction is completely consistent, and there is only a positional difference in the axial height. The specific implementation method of each preset measuring point is to paste reflective marks or set reflective targets on the surface of each turn of the coil to cooperate with the subsequent non-contact vibration measurement. The non-contact vibration measurement device uses a laser vibrometer. The laser beam of the laser vibrometer is aimed at the reflective marking surface at each preset measurement point. The laser vibrometer emits a laser beam and receives the reflected light signal. It uses the Doppler effect principle to measure the vibration acceleration at each preset measurement point. The sweep frequency signal generator is a programmable sweep frequency signal generator. Its output is electrically connected to the two ends of the coil under test through a test cable. S1. Apply a sweep voltage signal to both ends of the coil under test. Collect vibration acceleration signals at each frequency point from multiple preset measurement points of the coil under test using a non-contact vibration measurement device. Specifically, the sweep signal generator outputs a sweep voltage signal according to the preset start frequency, end frequency, and frequency step size. Before the test begins, the operator sets the sweep parameters through the processor's control interface. The processor sends control commands to the sweep signal generator. After receiving the commands, the sweep signal generator increases the output frequency point by point from the start frequency with a fixed frequency step size until the end frequency. At each frequency point, the sweep signal generator maintains the output voltage at this frequency stably for a period of time to ensure that the mechanical vibration response of the coil under test is fully established. The sweep signal generator maintains a constant output voltage amplitude throughout the sweep process. Since the correlation coefficient calculated in the subsequent steps of this method reflects the relative consistency between the vibration amplitude and frequency sequence of each coil at all frequency points, rather than depending on the absolute amplitude at a certain frequency point, as long as each coil measurement point is subjected to the same excitation voltage amplitude at the same frequency point during the sweep process, the relative comparison relationship between each coil will not be affected. At each frequency point, the laser vibrometer synchronously measures each preset measuring point. The controller of the laser vibrometer receives the synchronous trigger signal from the processor and collects the vibration acceleration signal at each preset measuring point during the stable period of each frequency point. The vibration acceleration signal is output to the processor in the form of an analog voltage signal or a digital signal. The amplitude of this signal reflects the vibration intensity of the measuring point under excitation at that frequency point. S2. Extract the vibration acceleration amplitude of each preset measuring point at each frequency point from the vibration acceleration signal. Arrange the vibration acceleration amplitudes of the preset measuring points corresponding to the same coil at all frequency points in ascending order of frequency to generate the vibration amplitude-frequency sequence of this coil. Specifically, after the processor receives the vibration acceleration signals of each preset measuring point at each frequency point output by the laser vibrometer, it performs signal processing. For the vibration acceleration signal at each frequency point, the processor uses the peak detection method or the effective value detection method to extract the vibration acceleration amplitude corresponding to that frequency point. Specifically, the processor filters the acquired time-domain vibration acceleration waveform to eliminate high-frequency noise interference, and then extracts the peak value of the waveform or calculates the effective value as the vibration acceleration amplitude at this frequency point. For the i-th turn of the coil, i = 1, 2, ..., N, the vibration acceleration amplitudes at all preset measurement points at all frequencies form a set of values, denoted as f1, f2, ..., f M Where M is the total number of frequency points, the processor arranges these values ​​in ascending order of frequency to construct the vibration amplitude-frequency sequence of this coil. ,in This represents the vibration acceleration amplitude of the i-th coil at the k-th frequency point. Following the above method, the processor sequentially constructs sequences for all N-turn coils, obtaining a total of N vibration amplitude-frequency sequences, each corresponding to an N-turn coil. S3. Calculate the correlation coefficient between the vibration amplitude-frequency sequences of any two coil turns to obtain the correlation coefficient matrix. For each coil turn, based on the correlation coefficient matrix, remove the correlation coefficient between this coil turn and itself, and calculate the average of the correlation coefficients between this coil turn and all other coil turns as the average inter-coil similarity of this coil turn. Specifically, for any two coil turns p and q, where p, q = 1, 2, ..., N, p ≠ q, the processor obtains their corresponding vibration amplitude-frequency sequence A. p and A q Calculate the correlation coefficient r between the two. pq ; Taking the Pearson correlation coefficient as an example, its calculation formula is as follows: in, and Let r be the arithmetic mean of the vibration acceleration amplitudes of the p-th and q-th turns of the coil, respectively, over the entire swept frequency range, and let r be the correlation coefficient. pq The value of is in the range of [-1, 1]. The closer its absolute value is to 1, the stronger the correlation between the two sequences, that is, the more similar the vibration amplitude-frequency characteristics of the two coils are; the closer its value is to 0, the weaker the correlation between the two sequences, that is, the greater the difference in vibration amplitude-frequency characteristics between the two coils. The processor performs pairwise calculations on all N-turn coils, obtaining a total of N×N correlation coefficients, which form a correlation coefficient matrix R. In the correlation coefficient matrix R, the element in the i-th row and j-th column is R. ij R represents the correlation coefficient between the vibration amplitude-frequency sequences of the i-th and j-th turns of the coil. When i = j, R ii Let be the correlation coefficient between the i-th turn sequence and itself, with a value of 1; For each coil i, the processor removes the diagonal element R from the i-th row of the correlation coefficient matrix R. ii That is, the correlation coefficient between this turn and itself, and then take out the remaining N-1 correlation coefficients R. ij Let j = 1, 2, ..., N, j ≠ i, which represents the correlation coefficient between this turn of coil and all other turns of coil. The processor calculates the arithmetic mean S of these N-1 correlation coefficients. i : Arithmetic mean S i This refers to the average inter-turn similarity of the i-th turn of the coil, where S is the average inter-turn similarity. i This reflects the average similarity between the vibration amplitude-frequency sequence of the i-th turn and the vibration amplitude-frequency sequences of all other turns in the coil. When the coil under test is in good overall condition and the mechanical structural characteristics of each turn are consistent, the vibration amplitude-frequency sequences of each turn have high similarity to each other, so the average similarity between turns is at a high level. When a turn becomes loose or deformed, the mechanical structural characteristics of this turn, such as stiffness and damping, change, and its vibration amplitude-frequency sequence differs significantly from other turns that have not become loose or deformed. This leads to an overall decrease in the correlation coefficient between this turn and other turns, resulting in a significantly lower average similarity between this turn and other turns. During this process, for the turns located at both ends of the coil axis, there are differences in boundary conditions due to the lack of physical constraints between adjacent turns on one side. The differences in boundary conditions are inherent to the gradual change characteristics of the structural layout. The overall shape of its vibration amplitude-frequency sequence is still consistent with the vast majority of normal turns in the coil, which is fundamentally different from the loosening deformation caused by local stiffness abrupt change. In the subsequent cluster analysis, the correlation coefficient deviation between the end turns and the middle turns is included in the statistical distribution range of the normal group and will not be misjudged as loosening deformation turns due to the differences in boundary conditions. S4. The average inter-turn similarity of all coils is taken as a one-dimensional dataset. The K-means clustering algorithm is used to divide the one-dimensional dataset into two clusters. The cluster with the larger cluster center value is selected as the normal cluster, and the cluster center value of the normal cluster is used as the benchmark value. Specifically, the processor calculates the average inter-turn similarity S of all N-turn coils. i The numbers i = 1, 2, ..., N constitute a one-dimensional dataset. The data in the one-dimensional dataset are all dimensionless numerical values, reflecting the average similarity of each coil relative to the overall coil. The processor uses the K-means clustering algorithm to perform cluster analysis on the one-dimensional dataset D, with the number of clusters K set to 2. The specific execution process of the K-means clustering algorithm is as follows: (1) Initialization: Randomly select two samples from dataset D as initial cluster centers, denoted as μ1 respectively. （0） and μ2 （0） ; (2) Allocation steps: For each sample S in dataset D i Calculate the distance from the sample to the two cluster centers μ1 and μ2. Since it is one-dimensional data, the distance is the absolute value of the difference. i They are assigned to clusters represented by the nearest cluster centers; (3) Update step: Calculate the arithmetic mean of all samples in each of the two clusters, and use it as the new cluster center μ1. （1） and μ2 （1） ; (4) Repeat steps (2) and (3) until the cluster centers no longer change, that is, the absolute value of the difference between the cluster centers of two adjacent iterations is less than the preset convergence accuracy, or the preset maximum number of iterations is reached; After clustering, two clusters C1 and C2 are obtained, with their corresponding final cluster center values ​​μ1 and μ2, respectively. The processor compares the values ​​of μ1 and μ2 and selects the cluster with the larger cluster center value as the normal cluster. The cluster center value of this normal cluster is defined as the baseline value S. ref The principle is that in the coil under test, most coils are usually in a normal state, and their average similarity between turns is high and they are close to each other, so they will naturally cluster in the range of higher values; while a few coils that are loose or deformed have significantly lower average similarity between turns, forming low-value outliers. K-means clustering K=2, after forcibly dividing the dataset into two clusters, the high-value cluster corresponds to the normal coil group and the low-value cluster corresponds to the abnormal coil group. Since the number of normal coils is usually much greater than that of abnormal coils, and the similarity value of normal coils is generally higher than that of abnormal coils, the cluster with the larger cluster center value is the cluster composed of normal coils. S5. Calculate the ratio of the average inter-turn similarity of each coil to the benchmark value, and use it as the difference identification coefficient. Coils with a difference identification coefficient less than the preset judgment threshold are judged as loose and deformed coils. Specifically, the processor calculates the average inter-turn similarity S for each coil i. i Compared with the benchmark value S ref The ratio λ i : λ iThis refers to the difference identification coefficient of the i-th turn coil. The difference identification coefficient is a dimensionless relative value, and its physical meaning is: the degree of deviation of the average inter-turn similarity of the i-th turn coil from the baseline level of the normal turn group. When the i-th turn coil is in a healthy state, its average inter-turn similarity S i Compared with the benchmark value S ref Approaching, λ i Approximately equal to 1; when the i-th turn of the coil becomes loose or deformed, its average inter-turn similarity S i Significantly reduced, leading to λ i Clearly less than 1; The operator or the processor has a preset judgment threshold λ. th Before the test begins, the operator sets the judgment threshold through the processor's input interface based on the type and number of turns of the coil under test and their engineering experience. The processor then assigns the difference identification coefficient λ to each coil. i Compared with the preset judgment threshold λ th Comparison: If λ i <λ th If λ i ≥λ th If so, the coil is determined to be in a normal state; After all N turns of the coil are judged, the processor outputs the judgment result, which includes the number of each turn of the coil and its corresponding status label, indicating normal or loose / deformed. The operator can view the judgment result through the display terminal to locate the specific coil turn that is loose or deformed. The method of this invention realizes the detection and location of coil loosening and deformation through inter-turn self-reference. The entire detection process does not require historical detection data of the coil under test, nor does it require the prior acquisition of reference samples in a healthy state. The judgment can be completed solely based on the difference in vibration response between the turns of the coil under test. It is suitable for on-site detection scenarios where historical detection data is lacking.

[0030] In another technical solution, the sweep frequency range of the swept voltage signal is 20Hz to 2000Hz, and the sweep step size is no greater than 5Hz. Specifically, the operator sets the sweep parameters through the processor's control interface, setting the starting frequency to 20Hz, the ending frequency to 2000Hz, and the frequency step size to a specific value no greater than 5Hz, such as 5Hz, 2Hz, or 1Hz. After receiving the control command, the swept signal generator starts from 20Hz and gradually increases the output frequency according to the set step size. During the sweep process, the swept signal generator maintains a constant amplitude of the output voltage. The processing of the sweep termination frequency is as follows: When the frequency is increased to the last frequency point not greater than the termination frequency of 2000Hz according to the set step size, if this frequency point exceeds 2000Hz after adding a step size, then 2000Hz is used as the last frequency point to complete the sweep. That is, the highest frequency of the actual sweep is equal to the termination frequency of 2000Hz. For example, when the starting frequency is 20Hz and the step size is 5Hz, the frequency sequence is 20, 25, 30, ..., 1995, 2000. The step size of 5Hz is divisible by 1980Hz, and the termination frequency falls exactly at 2000Hz. If the step size is 3Hz, 1980Hz is not divisible by 3Hz, and the last value of the frequency sequence not greater than 2000Hz is 1997Hz. At this time, the processor controls the sweep signal generator to output and collect signals with 2000Hz as the last frequency point after 1997Hz, thereby ensuring that the sweep range accurately covers the termination frequency of 2000Hz. In the frequency band below 20Hz, the mechanical vibration noise in the environment, such as the 100Hz and its harmonic components generated by the magnetostriction of the iron core, and the low-frequency vibration of surrounding power equipment, causes strong interference and a low signal-to-noise ratio. This results in a large amount of non-excitation response components in the vibration acceleration signals collected at each turn measurement point, affecting the accuracy of the subsequent amplitude-frequency sequence. In the frequency band below 20Hz, the amplitude of the mechanical vibration response of the coil itself is usually low, making signal acquisition difficult and contributing little to the calculation of the subsequent correlation coefficient. Therefore, setting the lower limit of the sweep frequency to 20Hz can effectively avoid low-frequency environmental noise interference and ensure the quality of the collected signal. The windings of coil-type electrical equipment are elastic structures composed of multiple turns of coils arranged axially by insulating supports. Their vibration characteristics vary complexly with the excitation frequency. Below 2000Hz, the coil as a whole exhibits relatively clear overall bending vibration modes and axial compression vibration modes. The vibration characteristics of these low-order modes are closely related to the overall mechanical integrity of the winding. When the winding becomes loose, the change in axial preload directly alters the natural frequencies and damping characteristics of these low-order modes, which is clearly reflected in the amplitude-frequency sequence of each turn. When the excitation frequency exceeds 2000Hz, local higher-order vibration modes gradually appear in the coil's vibration response. These are relative vibrations between adjacent turns due to local electromagnetic forces or mechanical coupling. The vibration characteristics of these local higher-order modes mainly depend on the local stiffness, local mass distribution, and local properties of the inter-turn insulation layer of each turn. There is no clear correlation between the contact state and the degree of loosening deformation of the winding as a whole. Due to the differences in support conditions and coupling state of adjacent turns, the amplitude and frequency distribution of local high-order modes at different turns are significantly non-uniform. This means that the difference in vibration response between adjacent turns in this frequency band mainly comes from random differences in local structural details, rather than the overall or local loosening deformation to be detected. If high-frequency data above 2000Hz are included in the correlation coefficient calculation, these high-frequency local differences that are not related to the fault will interfere with the overall similarity between turns in the low-frequency band, which should be reflected, and reduce the sensitivity and selectivity of the average similarity between turns to loosening deformation. Above 2000Hz, the high-frequency components of the vibration signal attenuate faster during propagation, and the signal-to-noise ratio decreases. Therefore, setting the upper limit of the sweep frequency to 2000Hz can effectively eliminate the interference of high-order local modes on the calculation of inter-turn similarity, and make the vibration amplitude-frequency sequence concentrated in the low-order mode frequency band that reflects the overall mechanical state of the winding. The sweep step size determines the frequency resolution of the vibration amplitude-frequency sequence. An excessively large step size may cause the resonance peak located between adjacent steps to be missed, reducing the accuracy of the vibration amplitude-frequency sequence in characterizing the winding structure. By controlling the step size within 5Hz, it is possible to ensure that at least a sufficient number of sampling points are obtained at each resonance peak, thereby accurately depicting the complete curve shape of the vibration amplitude changing with frequency, and providing a reliable data basis for subsequent correlation coefficient calculation. The sweep frequency signal generator outputs sweep frequency voltage signals point by point according to the above parameters, and maintains a stable output at each frequency point, so as to allow the non-contact vibration measurement device to collect vibration acceleration signals at each preset measurement point.

[0031] In another technical solution, the preset judgment threshold ranges from 0.75 to 0.90. Specifically, before the test begins, the operator sets the specific value of the judgment threshold within the range of 0.75 to 0.90 through the processor's input interface, based on the type, number of turns, structural dimensions, and on-site testing requirements of the coil under test. For example, for equipment with high safety requirements, the judgment threshold can be set to a higher value, such as 0.88 or 0.90, to improve detection sensitivity and detect as many loose or deformed coils as possible. For general testing or situations with high ambient noise, the judgment threshold can be set to a lower value, such as 0.75 or 0.78, to reduce the risk of false alarms caused by environmental interference. Difference identification coefficient λ i S is the average inter-turn similarity of the i-th turn of the coil. i Compared with the benchmark value S ref The ratio, whose physical meaning is the relative deviation of the average inter-turn similarity of the i-th turn coil from the baseline level of the normal turn group; under healthy conditions, the mechanical structural characteristics of each turn coil are basically the same, and the vibration amplitude-frequency sequences of any two turns maintain a high degree of similarity. The similarity comes from the fact that each turn coil has the same geometric configuration, the same material properties, and similar boundary conditions in the axial arrangement. Under this condition, the average value S of the correlation coefficient between a certain turn coil and all other turn coils is... i The reference value S of the normal turn group ref The differences between them are mainly caused by random factors such as coil manufacturing tolerances, material inhomogeneity, and measurement noise. The cumulative impact of these random factors on the correlation coefficient is limited, making S... i With S ref The ratio λ i In general engineering statistics, the values ​​are mainly distributed in the range above 0.90. When a coil becomes loose, the axial preload of that coil decreases, and its mechanical stiffness decreases accordingly. This causes the natural vibration frequency of that coil to shift, and the damping characteristics also change simultaneously. The resonance peak position in the vibration amplitude-frequency sequence of that coil, which was originally consistent with the whole sequence, will experience frequency shift or amplitude attenuation, thus increasing the difference between the vibration amplitude-frequency sequence of that coil and the sequence of healthy coils. When the loosening of this coil reaches a level that requires attention in engineering, the change in its mechanical structural characteristics is sufficient to significantly reduce the correlation coefficient between it and healthy coils, thereby reducing the average similarity S between the coils. i Relative to the reference value S ref The decrease in λ is generally over 25% in general engineering statistics. i It is mainly distributed in the range below 0.75; When a coil deforms, such as through localized axial bending or radial expansion, the geometry of that turn changes, and its distributed parameters, such as the equivalent inductance and capacitance to ground, change accordingly. This affects its vibration response characteristics under high-frequency excitation, and the correlation coefficient between the vibration amplitude-frequency sequence of the deformed turn and the sequence of the healthy turn also decreases significantly. i The distribution characteristics are similar to those of loose turns, mainly falling within the range below 0.75; Based on the above analysis, the λ of a healthy turn i The λ value is mainly distributed in the range above 0.90, while the λ value of loose or deformed coils is... i It is mainly distributed in the range below 0.75, and the two types of turns are in λ. i The indicators show a discernible distribution interval gap. Therefore, by setting the judgment threshold between 0.75 and 0.90, this distribution interval gap can be used to effectively distinguish between healthy and faulty turns. The specific selection direction for the judgment threshold within the range of 0.75 to 0.90 is as follows: when the judgment threshold setting is low, i.e., close to 0.75, only λ... i Only coils with a significant drop in speed are considered faulty. This reduces the probability of misclassifying healthy coils as faulty, resulting in a low false positive rate. However, it may lead to some coils with minor loosening or deformation being affected by λ. i If the value exceeds the threshold, it will be missed, increasing the false negative rate; when the threshold value is set high, i.e., close to 0.90, λ i A slight decrease in the coil value is considered a fault, which improves the sensitivity of detecting minor loosening and deformation. The false negative rate is low, but normal data fluctuations or measurement noise may cause some healthy coils to be misjudged, increasing the false positive rate. Operators can flexibly select an appropriate judgment threshold in the range of 0.75 to 0.90 according to the actual needs of the specific testing scenario.

[0032] In another technical solution, the Pearson correlation coefficient algorithm is used to calculate the correlation coefficient between the vibration amplitude and frequency sequences of any two turns of the coil; Specifically, the processor calculates the correlation coefficient between the vibration amplitude-frequency sequences of the p-th turn and the q-th turn of the coil according to the definition of the Pearson correlation coefficient and the following formula: Where M is the total number of frequency points throughout the entire frequency sweep process; a p (f k ) and a q (f k ) represent the p-th and q-th turns of the coil at the k-th frequency point f. k The amplitude of vibration acceleration under the following conditions; and , respectively, are the arithmetic mean of the vibration acceleration amplitudes of the p-th and q-th turns of the coil over the entire swept frequency range; correlation coefficient r pq The value range of r is [-1, 1]. pq When r is positive and close to 1, it indicates a strong positive correlation between the vibration amplitude and frequency sequences of the two coils, meaning that the vibration acceleration amplitudes of the two coils have the same trend at each frequency point, reflecting similar mechanical structural characteristics of the two coils; when r pq When the value is close to 0, it indicates that there is no significant linear correlation between the vibration amplitude-frequency sequence of the two coils, reflecting a difference in the mechanical structural characteristics of the two coils; when r pq When the value is negative, it indicates that the vibration amplitude-frequency sequence of the two turns of the coil is negatively correlated. That is, when the amplitude of one turn increases at a certain frequency point, the amplitude of the other turn decreases. This situation occasionally occurs in coils with symmetrical structures. However, in the practical application of this method, since the physical structure and boundary conditions of each turn of the coil are basically the same, the correlation coefficient between healthy turns is usually positive. The Pearson correlation coefficient algorithm is insensitive to the dimensions of the data, effectively measuring the degree of linear correlation between two sequences. It is simple to calculate, has clear physical meaning, and is suitable for comparing the similarity of continuous numerical sequences such as coil vibration amplitude-frequency sequences. The Pearson correlation coefficient is insensitive to the overall amplitude shift in the sequence. That is, if the vibration amplitudes of two coils differ by a multiple but their trends are consistent, the correlation coefficient is still high. This characteristic makes this method independent of the accurate calibration of absolute amplitude, reducing the impact of systematic errors caused by differences in sensor sensitivity or changes in measurement distance on the judgment results during field measurement. The processor combines all N coils one by one in pairs and calculates their correlation coefficients according to the above Pearson correlation coefficient formula to construct the correlation coefficient matrix R.

[0033] In another technical solution, the non-contact vibration measurement device is a laser vibrometer. The laser vibrometer consists of a laser head, a controller, and a signal processing unit. The laser head emits a stable continuous-wave laser beam, which illuminates a pre-set reflective mark or reflective target surface at each preset measurement point of the coil under test. The laser beam is reflected from the reflective mark surface, and the reflected light is received by a photodetector inside the laser head. When the reflective mark is displaced due to the vibration of the coil under test, the frequency of the reflected light undergoes a Doppler shift relative to the emitted light. The amount of the Doppler shift is proportional to the vibration velocity of the reflective mark. The controller of the laser vibrometer demodulates the frequency difference between the emitted and reflected light to obtain the real-time vibration velocity signal at the corresponding measurement point, and then performs differential calculations to obtain the vibration acceleration signal. The processor communicates with the controller of the laser vibrometer via a synchronous trigger line. When the frequency sweep signal generator switches to each new frequency point, the processor sends a synchronous acquisition trigger signal to the laser vibrometer. After receiving the trigger signal, the laser vibrometer samples the vibration acceleration signal of each preset measurement point at that frequency point and transmits the sampled data to the processor. The laser vibrometer does not need to have physical contact with the coil under test, and the vibration characteristics of the coil will not be changed due to the added mass of the sensor, ensuring the authenticity and accuracy of the measurement results. The laser vibrometer has the characteristics of high bandwidth, high resolution, and high sensitivity, and can accurately measure small vibration signals. It is suitable for detection scenarios where the vibration response amplitude of coil-type equipment is relatively small under frequency sweep excitation. The laser beam of the laser vibrometer can be flexibly aimed at measurement points in different spatial locations. Each preset measurement point can be measured sequentially by switching the beam or scanning, without the need to pre-install sensors on each turn of the coil, simplifying the preparation work for on-site testing. The reflective marks at each preset measuring point are made of a special reflective film with high reflectivity and are pasted on the outer surface of each coil. To ensure the stability of the measurement signal, the pasting direction of each reflective mark should be such that the normal direction of its surface is basically consistent with the incident direction of the laser beam to obtain the strongest reflected light signal. For coils with iron cores, the reflective marks are pasted on the surface of the coil exposed on the outside. For multi-layer winding structures, each preset measuring point is selected at the axial corresponding position of each coil in the outermost layer to ensure that the laser beam can be incident on each measuring point without obstruction. The vibration acceleration signal collected by the laser vibrometer is transmitted to the processor in the form of a digital signal after analog-to-digital conversion.

[0034] In another technical solution, the duration of the sweep voltage signal at each frequency point is 0.5s to 2s, and adjacent frequency points are continuously swept and switched at a rate not exceeding 10Hz / s.

[0035] Specifically, the sweep frequency signal generator outputs sweep voltage signals point by point according to a sweep frequency range of 20Hz to 2000Hz and a sweep step size of no more than 5Hz. At each frequency point, the sweep frequency signal generator maintains a constant frequency and amplitude of the output voltage, with the duration set to a specific value within the range of 0.5s to 2s, such as 1s. The operator sets this duration parameter through the processor's control interface, and the sweep frequency signal generator executes the output according to this setting. After completing the output at a certain frequency point, the sweep frequency signal generator switches to the next frequency point. The frequency switching is not a step-like abrupt change, but a continuous and smooth process. Specifically, starting from the current frequency value, the sweep frequency signal generator continuously increases the output frequency at a rate of no more than 10Hz / s until the next preset frequency point is reached. After reaching the next frequency point, the sweep frequency signal generator again maintains a constant output at this frequency point, with the duration consistent with the aforementioned setting. The limitation of the above duration and switching rate is to ensure that the acquired vibration acceleration signal can truly reflect the vibration response characteristics of the coil under test under steady-state excitation. The coil under test is a mechanically elastic structure consisting of several turns of coil arranged axially by insulating supports. When a sweep frequency voltage signal is applied to both ends of the coil, the coil generates mechanical vibration under the action of alternating electromagnetic force. The vibration response requires a certain settling time from the start of excitation to amplitude stabilization. After the vibration system is subjected to simple harmonic excitation, its response is composed of the superposition of transient response components and steady-state response components. The transient response component decays exponentially with time. When it decays to a negligible level, the system enters the steady-state vibration stage. At this time, the vibration acceleration amplitude is determined only by the excitation frequency and the inherent characteristics of the system and no longer changes with time. Therefore, at each frequency point, it is necessary to wait for the transient response to decay sufficiently before collecting the vibration signal. Only then can the obtained amplitude data truly reflect the steady-state vibration characteristics of the coil at that frequency. If the duration at a single frequency point is too short, the mechanical vibration response of the coil is recorded before it reaches steady state. The extracted vibration acceleration amplitude will contain non-steady-state transient components, causing the constructed vibration amplitude-frequency sequence to be distorted, which in turn affects the accuracy of subsequent inter-turn correlation coefficient calculations. The mechanical vibration system of coil-type electrical equipment, such as transformer windings and reactor windings, consists of copper wires, insulating materials, and supporting structures. The decay time constant of its transient response components is usually between 0.1s and 0.5s, and the duration is not less than 0.5s, which is sufficient to ensure that the transient response components of most coil structures decay sufficiently and the system enters the steady-state vibration stage, thereby acquiring accurate steady-state vibration acceleration amplitude. The duration does not exceed 2s, thus avoiding the engineering problem of excessive overall frequency sweep test time due to excessive dwell time at a single point. Considering the large number of frequency points in the frequency sweep range, there are 397 frequency points in the range of 20Hz to 2000Hz with a step size of 5Hz. If the dwell time at each point is too long, it will seriously affect the detection efficiency and may even cause coil temperature rise due to prolonged excitation, affecting the repeatability of measurement results. A time window of 0.5s to 2s is a reasonable objective time window, which ensures the full establishment of vibration response while taking into account test efficiency. When the frequency sweep signal generator switches from the current frequency to the next frequency, if the switching rate is too fast, it is equivalent to applying a frequency change excitation between adjacent frequency points. For mechanical vibration systems, frequency change is similar to amplitude change, which will apply a transient impact force to the system and excite additional mechanical transient disturbance components. The disturbance also needs a certain amount of time to decay. If the signal is acquired before the transient disturbance has decayed, the obtained amplitude data will also deviate from the steady-state value. The upper limit of the switching rate is limited to no more than 10Hz / s to ensure smooth and continuous frequency transition and avoid introducing additional mechanical transient impact due to frequency change, so that the system always maintains a quasi-steady-state working state throughout the frequency sweep process. When the sweep step size is set to 5Hz, the time required to transition from one frequency point to the next frequency point at a switching rate of 10Hz / s is 0.5s, which exactly matches the lower limit of the duration of 0.5s. This ensures that the transient disturbances introduced during the frequency switching process have sufficient decay time before entering the stable holding phase of the next frequency point. When the sweep step size is set to a smaller value, such as 2Hz or 1Hz, the switching time is shortened accordingly, the frequency change of the system during the transition is smoother, the transient disturbances are smaller, and the detection effect is better. The frequency sweep signal generator executes frequency sweep output according to the above-mentioned duration and switching rate settings. During the steady-state holding period at each frequency point, that is, in the latter part of the duration, after the transient response has fully decayed, the processor controls the laser vibrometer to collect the vibration acceleration signal of each preset measurement point through the synchronous trigger signal. The vibration acceleration signal collected by the laser vibrometer no longer contains transient components. The amplitude data of each turn measurement point at the same frequency point truly reflects the vibration response amplitude of each turn coil under steady-state excitation at that frequency.

[0036] In another technical solution, in step S3, the processor calculates the correlation coefficient between the vibration amplitude and frequency sequences of any two turns of the coil, obtaining the correlation coefficient matrix R, where the element in the i-th row and j-th column is R. ij Let represent the correlation coefficient between the vibration amplitude-frequency sequences of the i-th turn and the j-th turn. For each i-th turn, the average correlation coefficient with all other i-th turns is calculated as the average inter-turn similarity. However, there is an inherent mechanical vibration coupling effect between adjacent turns in coil-type electrical equipment. Vibration coupling occurs between adjacent turns through the interaction of insulating pads, supports, and electromagnetic forces, resulting in high similarity in the vibration amplitude-frequency sequences of adjacent turns under healthy conditions. This high similarity does not stem from the healthy consistency of the structural state of each turn, but rather from their spatial proximity. If all other i-th turns are directly used as... The average similarity between turns is calculated using reference samples. For a given coil, the correlation coefficients of several coils spatially adjacent to it, especially adjacent turns, are high in a healthy state due to the mechanical coupling effect. However, the correlation coefficients of coils spatially far away from it may be low due to the attenuation of the vibration transmission path. This difference in correlation caused by spatial distance will interfere with the sensitivity of the average similarity between turns to changes in the coil's own structural state. When a coil becomes loose and deformed, its correlation coefficient with neighboring turns may still remain at a high level due to the coupling effect, thereby weakening the decrease in the average similarity between turns and increasing the risk of missed detection. In a preferred embodiment, step S3, before calculating the average inter-turn similarity of each coil turn, performs spatial neighborhood decorrelation preprocessing on the correlation coefficient matrix, including the following steps: For each coil i, the turn spacing is defined by the absolute value of the difference between the turn numbers of two coils along the axial direction, according to their axial physical position. All other coils whose turn spacing from coil i is greater than or equal to the preset spatial step size parameter L are selected to form the reference sample set G. i Where L takes the value of 2, 3 or 4, and N is a positive integer greater than 5; If the reference sample set G i If the number of coils in the sample is greater than or equal to the preset minimum statistical number T, then only this coil i and the reference sample set G are considered. i The arithmetic mean of the correlation coefficients of each coil is calculated and used as the average inter-turn similarity for that turn. If the reference sample set G... i If the number of coils in the loop is less than the preset minimum statistical number T, then the arithmetic mean of the correlation coefficients between coil i and all other coils is taken as the average inter-turn similarity of this coil, where T can be 3, 4 or 5.

[0037] Specifically, the preset spatial step size parameter L is a positive integer greater than or equal to 2 and less than the total number of turns N of the coil, and the value of L is in the range of 2, 3 or 4. The spatial step size parameter L represents the minimum turn spacing of adjacent coils that are too close to the target coil in the axial direction when performing correlation coefficient statistics. For each turn of coil i, construct a reference sample set G. i The processor defines the turn spacing based on the absolute value of the difference between the turn numbers of two turns along the axial direction, according to the axial physical position of each turn coil. Select all other coils whose turn spacing from coil i is greater than or equal to L to form the reference sample set G. i : Reference sample set G i Each turn of the coil in the reference sample set G maintains an axial spacing of at least L turns from the target turn of the coil i. Taking L=2 as an example, for the i-th turn of the coil, its reference sample set G i The sample includes all turns whose absolute difference from the turn number of turn i is not less than 2, excluding turns i-1 and i+1. This exclusion effectively avoids direct interference from mechanical vibration coupling effects between adjacent turns on the correlation coefficient calculation. A minimum statistical quantity T is preset, with a value ranging from 3, 4, or 5. The minimum statistical quantity T ensures a sufficient sample size for statistical averaging, improving the reliability of the statistical results. For turns near the axial end of the coil, such as turns 1, 2, N-1, and N, since there are not enough turns on one or both sides, G is selected according to the above reference sample set construction rules. i The number of coils in the circuit may be less than T. In this case, if G is forcibly used... i Calculating the arithmetic mean will result in unstable statistical results due to the small sample size. Wherein, if the reference sample set G i If the number of coils in the sample is greater than or equal to T, the processor will only base its judgment on this coil i and the reference sample set G. i The arithmetic mean of the correlation coefficients of each coil is calculated and used as the average inter-turn similarity of this turn: in For the reference sample set G i The number of coils in the sample was reduced by eliminating adjacent coils that had coupling effects due to spatial proximity. Only coils that were axially far from the target coil were retained as reference samples. This made the calculated average similarity between coils more accurately reflect the difference in vibration characteristics caused by changes in the coil's own structural state, rather than by spatial proximity coupling. If the reference sample set G iIf the number of coils in the loop is less than T, the processor abandons spatial neighborhood decorrelation preprocessing and directly uses the arithmetic mean of the correlation coefficients between coil i and all other coils as the average inter-turn similarity of this coil: When the target coil is near the axial end of the coil, there are no adjacent coils on one side, and the number of reference coils is already small. If further screening is performed using a spatial step size L, the number of remaining reference samples may not be sufficient to obtain a statistically stable average. In this case, abandoning preprocessing and using all other coils as reference samples may introduce some of the influence of adjacent coil coupling effects, but it ensures the stability and reliability of the statistics. Since the coil end coil has no adjacent coils on one side, its boundary conditions are different from those of the middle coils. The influence of adjacent coil coupling effects on it is also less than that on the middle coils. Therefore, using the full sample average is a reasonable compromise when the sample size is insufficient. Through the aforementioned spatial neighborhood decorrelation preprocessing, the processor can effectively eliminate the interference of mechanical coupling effect between adjacent turns on the correlation coefficient calculation while ensuring statistical stability. This makes the average similarity between turns more accurately reflect the changes in vibration characteristics of each coil caused by its own loosening or deformation. The average similarity between turns obtained after preprocessing is input into the cluster analysis in the subsequent step S4, which improves the accuracy of loosening and deformation determination.

[0038] In another technical solution, between steps S4 and S5, an absolute validity check and adaptive correction of the reference value are performed, including the following steps: The reference value obtained in step S4 is compared with the preset absolute reference lower limit threshold γ. If the reference value is greater than or equal to γ, the reference value is kept unchanged and the process proceeds to step S5. If the reference value is less than γ, the coil under test is determined to be in an abnormal state. The reference value is corrected to γ, and the process proceeds to step S5 with the corrected reference value. An overall status warning signal is output to mark that the current coil under test has a global health degradation. The value of γ ranges from 0.70 to 0.85, and γ is less than or equal to the preset judgment threshold. The overall status warning signal is used as the confidence level marker for the determination result in step S5. When no overall status warning signal is output, the determination result in step S5 is the final deterministic result. When an overall status warning signal is output, the determination result in step S5 is the suspected result.

[0039] Specifically, in step S4, the processor uses the K-means clustering algorithm (K=2) to divide the average inter-turn similarity of all turns into two clusters, selects the cluster with the larger cluster center value as the normal cluster, and uses the cluster center value of the normal cluster as the benchmark value S. refK-means clustering is essentially a purely relative comparison method. Clustering algorithms can only divide the dataset into two groups: relatively high and relatively low, but cannot determine whether the relatively high group has reached an absolute level of health. When the coil under test experiences overall insulation aging, uniform decay of preload across the entire section, or systematic temperature rise deformation due to long-term operation, the vibration amplitude-frequency sequence of all coil turns will be distorted synchronously, causing the average similarity between all turns to decline overall and proportionally. Under this global degradation condition, K-means clustering will still forcibly select large cluster centers as benchmark values ​​within the lower data range, causing the benchmark value to lose its physical meaning as a health reference. Even if there are locally loose or deformed coils, the ratio of their average similarity between turns to the collapsed benchmark value after the global degradation may be falsely amplified, causing the loose or deformed coils that should have been detected to cross the judgment threshold and be missed. To solve the above-mentioned systematic missed detection problem under global degradation, in this embodiment, the processor performs absolute validity verification and adaptive correction on the benchmark value between steps S4 and S5. The specific steps are as follows: A preset absolute baseline lower limit threshold γ is defined, with a value ranging from 0.70 to 0.85, and γ is less than or equal to a preset judgment threshold λ. th , λ th The value range is 0.75 to 0.90. The absolute reference lower limit threshold γ is an absolute health threshold value set based on the physical common sense of coil mechanical vibration. It represents the lowest acceptable level of average similarity between normal coil turns. Setting γ to be less than or equal to the preset judgment threshold ensures that correction can be triggered when the reference value is lower than the judgment threshold, so that the corrected reference value is within a reasonable range. The baseline value S obtained in step S4 ref Compared with γ, if S ref If ≥ γ, the processor determines that the baseline value obtained from clustering is within a reasonable absolute healthy range, maintains the baseline value unchanged, and proceeds to step S5 to perform subsequent difference identification coefficient calculation and determination; if S ref If the value is less than γ, the processor determines that the coil under test is in an abnormal state, i.e., a global health degradation has occurred, and forcibly corrects the reference value to γ, using the corrected reference value S. ref =γ enters step S5 to perform subsequent calculations, and at the same time the processor outputs an overall status warning signal to mark that the current coil under test has a global health degradation; The processor uses the overall state warning signal as the confidence level marker for the determination result of step S5. When no overall state warning signal is output, i.e., S... ref ≥ γ indicates that the baseline value obtained from clustering is within a reasonable absolute healthy range. The result of determining the loose or deformed coil in step S5 is the final deterministic result. On-site maintenance personnel can directly locate the loose or deformed coil based on this result. When the overall status warning signal is output, i.e., S...ref <γ indicates that the coil under test has global health degradation. The result of the loose and deformed coil in step S5 is marked as a suspected state. That is, this turn does have a significant difference relative to the current degraded group. However, since the global reference value has collapsed, the difference identification coefficient may underestimate the degree of deviation between this turn and the true healthy state. Therefore, the judgment result is output as a suspected result for on-site maintenance personnel to make a comprehensive judgment in combination with other detection methods. Through the aforementioned absolute validity verification and adaptive correction, under global degradation conditions, the benchmark value is clamped at the absolute benchmark lower limit threshold γ, avoiding false amplification of ratios and systematic missed judgments caused by benchmark value collapse. Taking the case where the average similarity between all turns collapses to 0.65 after global degradation as an example: without correction, K-means clustering takes the center of the large cluster at approximately 0.65 as the benchmark value. The actual average similarity between a locally loosened turn is 0.55, and the difference identification coefficient is 0.55 / 0.65≈0.846. When the judgment threshold is 0.85, it will cross the threshold line and be misjudged as normal. After correction, the benchmark value is γ, for example, 0.7. 5. The difference identification coefficient is 0.55 / 0.75≈0.733, which is less than 0.85. This turn was correctly identified as loose and deformed, and missed detection was effectively avoided. The introduction of the overall status warning signal and confidence level mark provides on-site maintenance personnel with additional information about the overall health status of the equipment. When the overall status warning signal is output, the maintenance personnel can know that the coil under test has a global health degradation, so they can pay more attention to the coil marked as suspected, and help them to more accurately judge the credibility of the test results and formulate subsequent maintenance strategies. When no overall status warning signal is output, the judgment result can be directly used as the basis for maintenance. The present invention also provides a system for identifying coil loosening and deformation under frequency sweep excitation and inter-turn vibration difference, comprising: A sweep frequency signal generator is used to apply a sweep frequency voltage signal to both ends of the coil under test. The sweep frequency signal generator is a programmable signal generator that can output a sweep frequency voltage signal according to preset sweep frequency parameters, including the sweep frequency range, sweep frequency step size, duration of each frequency point, and switching rate between adjacent frequency points. The output end of the sweep frequency signal generator is electrically connected to both ends of the coil under test through a test cable. A non-contact vibration measurement device is used to collect vibration acceleration signals at each frequency point from multiple preset measurement points of the coil under test. Each turn of the coil has a preset measurement point, and each preset measurement point is located on the same circumferential angle generatrix at the corresponding axial position of each turn of the coil under test. The non-contact vibration measurement device is a laser vibrometer, which emits a laser beam to illuminate the reflective mark surface of each preset measurement point, receives the reflected light signal, and uses the Doppler effect principle to measure the vibration acceleration at each preset measurement point. The processor is communicatively connected to the sweep frequency signal generator and the non-contact vibration measurement device. The processor is used to control the sweep frequency signal generator to output the sweep frequency voltage signal and to receive the vibration acceleration signal collected by the non-contact vibration measurement device. The processor can be a general-purpose computer, industrial control computer, embedded system or digital signal processor or other hardware device with data acquisition, calculation and control functions. The processor is also connected to a display and input device to display the detection results and receive user instructions. The processor performs the following steps: Extract the vibration acceleration amplitude of each preset measurement point at each frequency point from the vibration acceleration signal, and arrange the vibration acceleration amplitude of the preset measurement points corresponding to the same coil at all frequency points in ascending order of frequency to generate the vibration amplitude-frequency sequence of the coil. Calculate the correlation coefficient between the vibration amplitude-frequency sequences of any two turns of the coil to obtain the correlation coefficient matrix; for each turn of the coil, based on the correlation coefficient matrix, remove the correlation coefficient between the turn of the coil and itself, and calculate the average value of the correlation coefficients between the turn of the coil and all other turns of the coil as the average inter-turn similarity of the turn of the coil. The average inter-turn similarity of all coils is used as a one-dimensional dataset. The K-means clustering algorithm is used to divide the one-dimensional dataset into two clusters. The cluster with the larger cluster center value is selected as the normal cluster, and the cluster center value of the normal cluster is used as the benchmark value. Calculate the ratio of the average inter-turn similarity of each coil to the benchmark value, and use it as the difference identification coefficient. Coils with a difference identification coefficient less than the preset judgment threshold are judged as loose and deformed coils. The operator sets the sweep frequency parameters and judgment thresholds through the processor's input interface. The sweep frequency parameters include the start frequency, end frequency, sweep step size, frequency point duration, and switching rate. The processor sends control commands to the sweep frequency signal generator, which outputs a sweep frequency voltage signal according to the set parameters and applies it to both ends of the coil under test. During the sweep frequency process, the processor synchronously controls the non-contact vibration measurement device to collect the vibration acceleration signal at each preset measurement point at each frequency point and receives the data returned by the measurement device. The processor performs amplitude extraction, sequence generation, correlation coefficient calculation, matrix construction, inter-turn average similarity calculation, cluster analysis, benchmark value extraction, difference identification coefficient calculation, and loosening deformation judgment on the received vibration acceleration signal according to the above functional configuration. Finally, the judgment results of each coil are output through the display.

[0040] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. Other modifications can be easily made by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and examples shown and described herein.

Claims

1. A method for identifying coil loosening and deformation due to inter-turn vibration differences under swept-frequency excitation, characterized in that, include: S1. Apply the sweep frequency voltage signal to both ends of the coil under test, and collect the vibration acceleration signal of multiple preset measurement points of the coil under test at each frequency point through a non-contact vibration measurement device. Each turn of the coil has a preset measurement point, and each preset measurement point is located on the same circumferential angle generatrix at the corresponding axial position of each turn of the coil under test. S2. Extract the vibration acceleration amplitude of each preset measurement point at each frequency point from the vibration acceleration signal, and arrange the vibration acceleration amplitude of the preset measurement points corresponding to the same coil at all frequency points in ascending order of frequency to generate the vibration amplitude-frequency sequence of this coil. S3. Calculate the correlation coefficient between the vibration amplitude-frequency sequence of any two turns of coil to obtain the correlation coefficient matrix. For each turn of coil, according to the correlation coefficient matrix, remove the correlation coefficient between this turn of coil and itself, and calculate the average value of the correlation coefficient between this turn of coil and all other turns of coil as the average inter-turn similarity of this turn of coil. S4. The average inter-turn similarity of all coils is taken as a one-dimensional dataset. The K-means clustering algorithm is used to divide the one-dimensional dataset into two clusters. The cluster with the larger cluster center value is selected as the normal cluster, and the cluster center value of the normal cluster is used as the benchmark value. S5. Calculate the ratio of the average inter-turn similarity of each coil to the benchmark value, and use it as the difference identification coefficient. Coils with a difference identification coefficient less than the preset judgment threshold are judged as loose and deformed coils.

2. The method for identifying coil loosening and deformation under frequency sweep excitation under the sweep frequency excitation method according to claim 1, characterized in that, The sweep frequency range of the sweep voltage signal is 20Hz~2000Hz, and the sweep step size is no more than 5Hz.

3. The method for identifying coil loosening and deformation under frequency sweep excitation for inter-turn vibration difference as described in claim 1, characterized in that, The preset threshold value ranges from 0.75 to 0.

90.

4. The method for identifying coil loosening and deformation under frequency sweep excitation for inter-turn vibration difference as described in claim 1, characterized in that, The correlation coefficient between the vibration amplitude and frequency sequences of any two turns of the coil is calculated using the Pearson correlation coefficient algorithm.

5. The method for identifying coil loosening and deformation under frequency sweep excitation for inter-turn vibration difference as described in claim 1, characterized in that, The non-contact vibration measurement device is a laser vibration meter.

6. The method for identifying coil loosening and deformation under frequency sweep excitation for inter-turn vibration difference as described in claim 1, characterized in that, The duration of the sweep voltage signal at each frequency point is 0.5s to 2s, and adjacent frequency points are continuously swept and switched at a rate not exceeding 10Hz / s.

7. The method for identifying coil loosening and deformation under frequency sweep excitation under sweep frequency excitation as described in claim 1, characterized in that, In step S3, before calculating the average inter-turn similarity of each coil turn, the correlation coefficient matrix undergoes spatial neighborhood decorrelation preprocessing, including the following steps: For each coil i, the turn spacing is defined by the absolute value of the difference between the turn numbers of two coils along the axial direction, according to their axial physical position. All other coils whose turn spacing from coil i is greater than or equal to the preset spatial step size parameter L are selected to form the reference sample set G. i Where L takes the value of 2, 3 or 4, and N is a positive integer greater than 5; If the reference sample set G i If the number of coils in the sample is greater than or equal to the preset minimum statistical number T, then only this coil i and the reference sample set G are considered. i The arithmetic mean of the correlation coefficients of each coil is calculated and used as the average inter-turn similarity for that turn. If the reference sample set G... i If the number of coils in the loop is less than the preset minimum statistical number T, then the arithmetic mean of the correlation coefficients between coil i and all other coils is taken as the average inter-turn similarity of this coil, where T can be 3, 4 or 5.

8. The method for identifying coil loosening and deformation under frequency sweep excitation for inter-turn vibration difference as described in claim 3, characterized in that, Between steps S4 and S5, the absolute validity of the benchmark value is verified and adaptively corrected, including the following steps: The reference value obtained in step S4 is compared with the preset absolute reference lower limit threshold γ. If the reference value is greater than or equal to γ, the reference value is kept unchanged and the process proceeds to step S5. If the reference value is less than γ, the coil under test is determined to be in an abnormal state. The reference value is corrected to γ, and the process proceeds to step S5 with the corrected reference value. An overall status warning signal is output to mark that the current coil under test has a global health degradation. The value of γ ranges from 0.70 to 0.85, and γ is less than or equal to the preset judgment threshold. The overall status warning signal is used as the confidence level marker for the determination result in step S5. When no overall status warning signal is output, the determination result in step S5 is the final deterministic result. When an overall status warning signal is output, the determination result in step S5 is the suspected result.

9. A system for identifying coil loosening and deformation under frequency sweep excitation and inter-turn vibration difference, characterized in that, include: A sweep frequency signal generator is used to apply a sweep frequency voltage signal to both ends of the coil under test; A non-contact vibration measurement device is used to collect vibration acceleration signals at each frequency point from multiple preset measurement points of the coil under test. Each coil has a preset measurement point, and each preset measurement point is located on the same circumferential angle generatrix at the corresponding axial position of each coil of the coil under test. The processor is communicatively connected to the sweep frequency signal generator and the non-contact vibration measurement device. It controls the sweep frequency signal generator to output a sweep frequency voltage signal and receives the vibration acceleration signal collected by the non-contact vibration measurement device. It extracts the vibration acceleration amplitude of each preset measurement point at each frequency point from the vibration acceleration signal, and arranges the vibration acceleration amplitude of the preset measurement points corresponding to the same coil at all frequency points in ascending order of frequency to generate the vibration amplitude-frequency sequence of the coil. The correlation coefficient matrix is ​​used to calculate the correlation coefficient between the vibration amplitude-frequency sequence of any two coil turns. For each coil turn, the correlation coefficient between the coil turn and itself is removed according to the correlation coefficient matrix. The average correlation coefficient between the coil turn and all other coil turns is then calculated as the average inter-turn similarity of the coil turn. The average inter-turn similarity of all coils is used as a one-dimensional dataset. The K-means clustering algorithm is used to divide the one-dimensional dataset into two clusters. The cluster with the larger cluster center value is selected as the normal cluster, and the cluster center value of the normal cluster is used as the benchmark value. And the ratio of the average inter-turn similarity of each coil to the benchmark value is used as the difference identification coefficient, and coils with a difference identification coefficient less than a preset judgment threshold are judged as loose and deformed coils.

10. The system for identifying coil loosening and deformation under frequency sweep excitation excitation according to claim 9, characterized in that, Before calculating the average inter-turn similarity of each coil, the processor also performs spatial neighborhood decorrelation preprocessing on the correlation coefficient matrix, including: For each coil i, the turn spacing is defined by the absolute value of the difference between the turn numbers of two coils along the axial direction, according to their axial physical position. All other coils whose turn spacing from coil i is greater than or equal to the preset spatial step size parameter L are selected to form the reference sample set G. i Where L takes the value of 2, 3 or 4, and N is a positive integer greater than 5; If the reference sample set G i If the number of coils in the sample is greater than or equal to the preset minimum statistical number T, then only this coil i and the reference sample set G are considered. i The arithmetic mean of the correlation coefficients of each coil is calculated and used as the average inter-turn similarity for that turn. If the reference sample set G... i If the number of coils in the loop is less than the preset minimum statistical number T, then the arithmetic mean of the correlation coefficients between coil i and all other coils is taken as the average inter-turn similarity of this coil, where T can be 3, 4 or 5.

Images