A transformer core looseness fault diagnosis method and device

Abstract

This invention discloses a method and apparatus for diagnosing transformer core loosening faults, relating to the field of transformer technology. The method includes the following steps: acquiring standard edge magnetic flux signals; collecting real-time edge magnetic flux signals and comparing them with standard edge magnetic flux signals to calculate magnetic flux distortion components; extracting features from the magnetic flux distortion components using a singular value decomposition method based on truncation regularization to obtain magnetic flux distortion feature values; constructing a core magnetic flux topology loop; performing continuous trajectory interpolation on the core magnetic flux topology loop using a three-wire interpolation method to obtain a magnetic flux distortion topology loop, and performing abrupt change analysis to obtain loosening anomaly points; inputting the magnetic flux distortion feature values ​​of the loosening anomaly points into a discrete grey prediction model based on residual correction to output predicted magnetic flux distortion feature values; performing neighborhood analysis on the loosening anomaly points to obtain the length of the loosening interval; and combining the loosening interval length and the predicted magnetic flux distortion values ​​to calculate the transformer loosening risk value and classify the risk level.

Description

Technical Field

[0001] This invention relates to the field of transformer technology, specifically to a method and device for diagnosing transformer core loosening faults. Background Technology

[0002] As a core component of the power grid, the long-term stable operation of transformers is crucial for power supply reliability. The transformer core, as the magnetic circuit and structural core, is susceptible to problems such as increased vibration, noise, and localized overheating due to common core loosening faults. In severe cases, this can lead to multiple core grounding points or even insulation damage, threatening power grid safety. Therefore, early and accurate diagnosis of transformer core loosening is of great significance. In recent years, non-invasive detection technology based on magnetic field measurement has become a research hotspot for core condition monitoring due to its advantages of flexible deployment and high safety. This technology involves deploying a magnetic sensor array on the outer wall of the transformer tank to collect edge magnetic flux signals generated by the magnetostrictive effect and magnetic flux distribution of the core, and then extracting features related to the mechanical state from these signals.

[0003] Traditional methods involve performing singular value decomposition on the acquired magnetic flux distortion signal sequence. The resulting singular values ​​represent the energy intensity of different mode components in the signal. The top k largest singular values ​​(i.e. dominant modes) are extracted to reconstruct the signal, thereby achieving feature extraction and data dimensionality reduction. This results in a feature vector characterizing the main energy of the signal, which can be used for subsequent state assessment and diagnosis.

[0004] However, in the traditional singular value decomposition technique, in order to achieve dimensionality reduction and noise reduction, a cutoff point k must be set, discarding a large number of remaining small singular values. This direct truncation process treats all signal components (including noise and potential weak fault information) other than the k dominant singular values ​​as useless noise. As a result, the extracted fault feature vector loses the signal details related to loosening in the small singular values, which restricts the completeness and sensitivity of feature extraction and makes it difficult to accurately capture the initial fault. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and apparatus for diagnosing transformer core loosening faults, thereby resolving the problems existing in the background technology.

[0006] To achieve the above objectives, the present invention provides a method and apparatus for diagnosing transformer core loosening faults, comprising the following steps: Step S1: Deploy several monitoring nodes on the transformer and collect magnetic flux data of the monitoring nodes under normal operating conditions to obtain standard edge magnetic flux signals; Step S2: Collect real-time edge magnetic flux signals and compare them with standard edge magnetic flux signals to calculate the magnetic flux distortion components; extract features from the magnetic flux distortion components using singular value decomposition based on truncation regularization to obtain magnetic flux distortion feature values. Step S3: Based on the topological sequence of the transformer, construct the core flux topology loop; based on the flux distortion characteristic value of the monitoring node, use the three-line interpolation method to perform continuous trajectory interpolation on the core flux topology loop to obtain the flux distortion topology loop. Step S4: Perform abrupt change analysis on the flux distortion topology loop to obtain loose anomaly points; input the flux distortion feature values ​​of the loose anomaly points into the discrete grey prediction model based on residual correction, and output the predicted value of flux distortion features. Step S5: Perform neighborhood analysis on the loosening anomaly points to obtain the loosening interval length; combine the loosening interval length and the magnetic flux distortion prediction value to calculate the transformer loosening risk value and classify the risk level, thereby realizing the identification of transformer loosening.

[0007] Preferably, the acquisition of magnetic flux data from the monitoring node under normal operating conditions to obtain a standard edge magnetic flux signal includes the following specific steps: A sensor is installed at each monitoring node to convert the voltage signal collected by the sensor into a magnetic induction intensity signal.

[0008] in, Let represent the magnetic flux density output by the i-th sensor during the t-th power frequency cycle. This represents the sensitivity coefficient of the i-th sensor. This represents the voltage signal output by the i-th sensor during the t-th power frequency cycle; The reference signal is obtained using power frequency synchronous phase-locked loop technology. Thus, the edge fluctuation signal after eliminating the main magnetic flux is obtained:

[0009] in, This represents the edge fluctuation signal of the i-th sensor after the main magnetic flux in the t-th power frequency cycle. This represents the amplitude scaling factor between the i-th sensor signal and the reference signal. This represents the phase difference between the i-th sensor signal and the reference signal; Cut off The continuous power frequency periodic waveform is subjected to an arithmetic mean to suppress random noise, ultimately yielding the standard edge flux signal under normal operating conditions:

[0010] in, This represents the standard edge magnetic flux signal of the i-th sensor. This represents the total number of power frequency cycles, and t represents the t-th power frequency cycle.

[0011] Preferably, the step of collecting real-time edge magnetic flux signals and comparing them with standard edge magnetic flux signals to calculate the magnetic flux distortion component specifically involves:

[0012] in, This represents the magnetic flux distortion component of the i-th sensor. This represents the real-time edge magnetic flux signal of the i-th sensor. This represents the standard edge magnetic flux signal of the i-th sensor.

[0013] Preferably, the step of extracting features from the magnetic flux distortion components using a singular value decomposition method based on truncation regularization to obtain magnetic flux distortion feature values ​​includes the following steps: The magnetic flux distortion component of the i-th sensor within one power frequency cycle is discretized into a magnetic flux distortion sequence [d1,d2,d3,...,dN] based on the sensor's sampling frequency. [d1,d2,d3,...,dN] represents the first to Nth sampling points of the magnetic flux distortion component, and N represents the total number of sampling points within the power frequency cycle. A Hankel matrix H is constructed based on the magnetic flux distortion sequence. Perform singular value decomposition on the Hankel matrix H:

[0014] in, It is a singular value diagonal matrix, which is an m×p rectangular diagonal matrix, and the elements on its main diagonal are singular values. Let represent the left singular vector matrix, which is an m×m matrix. The transpose of the right singular vector matrix is ​​a p×p matrix, where m is the number of rows and p is the number of columns. Extract the first k singular values ​​to form the dominant distortion feature vector. , =[ , ,..., ], This represents the k-th singular value; By introducing a regularization matrix, effective signal compensation components are extracted. ; Compensation components of the extracted effective signal Superimposed on the initial solution The final solution vector after feature enhancement is obtained. :

[0015] in, This represents the final solution vector after feature enhancement. This represents the initial solution reconstructed from the first k singular values. Indicates the effective signal compensation component; Finally, the final solution vector after feature enhancement is calculated. The L2 norm is used as the characteristic value of magnetic flux distortion of the i-th sensor in the current power frequency cycle. :

[0016] in, Let be the magnetic flux distortion characteristic value of the i-th sensor.

[0017] Preferably, the step of extracting effective signal compensation components by introducing a regularization matrix includes the following steps: The first k singular values ​​and their corresponding right singular vector matrices Reconstruct an initial solution By introducing a regularization matrix L, the effective signal compensation component Δx is extracted:

[0018] in, Indicates the effective signal compensation component. Let L represent the right singular vector matrix corresponding to the truncated singular values, and let L represent the regularization matrix. Representation matrix The generalized inverse, Representing the fundamental solution The regularization bias, This represents the initial solution reconstructed from the first k singular values.

[0019] Preferably, the construction of the core flux topology loop based on the transformer topology sequence includes the following steps: Based on the structural relationship of the internal magnetic circuit of the transformer, the core magnetic flux topology loop is constructed: the core magnetic flux topology loop is a closed spatial trajectory that closely follows the geometric center line of the core laminations and strictly follows the main magnetic flux path. In specific construction, firstly, on the three-dimensional model of the transformer core, the path starts and ends at the center of the bottom of the core column, and flows along the center line of the core column through the upper and lower yokes to form a closed loop, thus obtaining the core magnetic flux topology loop. The physical total arc length of the core magnetic flux topology loop is calculated, and the projection position of each monitoring node on this loop is converted into normalized topological coordinates.

[0020] Preferably, the method of using three-line interpolation to perform continuous trajectory interpolation on the magnetic flux topology loop of the iron core to obtain the magnetic flux distortion topology loop includes the following specific steps: Normalized topological coordinates of the i-th sensor It is calculated using the following formula:

[0021] in, These are the normalized topological coordinates of the i-th sensor. It is the arc length from the starting point of the iron core flux topology loop to the projection point of the i-th sensor on the iron core flux topology loop. It is the total arc length of the magnetic flux topology loop in the iron core; Assign normalized topological coordinates to each sensor's flux distortion component. Then, in the t-th power frequency cycle, the magnetic flux distortion topology set Data(t) = {( , ),( , ),...,( , ),...,( , )}; Where Data(t) represents the magnetic flux distortion topology set under the t-th power frequency cycle. Represents the normalized topological coordinates of the i-th sensor. This represents the magnetic flux distortion characteristic value of the i-th sensor during the t-th power frequency cycle; The topological coordinate interval is divided into n-1 sub-intervals according to the normalized topological coordinates of n sensors, and these sub-intervals are strictly incremented. In each sub-interval [ , The cubic polynomial interpolation function on is:

[0022] in, Let represent the cubic polynomial interpolation function over the i-th subinterval. Denotes the coefficient of the constant term of the cubic polynomial over the i-th subinterval. Let the coefficient of the linear term of the cubic polynomial over the i-th subinterval be denoted as . Denotes the coefficient of the quadratic term of the cubic polynomial over the i-th subinterval. Denotes the coefficient of the cubic term of the cubic polynomial over the i-th subinterval; Solve using the chasing method. , , , Thus, a flux distortion topological loop is obtained.

[0023] Preferably, the step of inputting the magnetic flux distortion feature values ​​of the loosening anomaly points into a discrete grey prediction model based on residual correction, and outputting the predicted magnetic flux distortion feature values, includes the following specific steps: For each loosening anomaly point, extract the nearest... The magnetic flux distortion characteristic values ​​of each consecutive power frequency cycle constitute the original sequence. , =( , ,..., ), where CH(t) is the flux distortion characteristic value in the t-th power frequency cycle, and t is the index of the current power frequency cycle; For the original sequence Perform first-order accumulation to generate a first-order sequence. :

[0024] in, Let represent the e-th element in a first-order sequence, where e is the index of an element in the first-order sequence. Represents the original sequence The past element, where past is the index of an element in the original sequence; Finally, a first-order sequence is obtained. =( , ,..., Based on the discrete form of the first-order sequence, a discrete grey prediction equation is established:

[0025] in, This represents the (e+1)th element in a first-order sequence. The development coefficient reflects the overall development trend of the first-order sequence. This is the gray action quantity; To solve for the parameters using the least squares method and Then, by substituting the discrete grey prediction equation, the first-order sequence cumulative value of the (t+1)th power frequency cycle is predicted:

[0026] in, This represents the predicted value of the first-order sequence accumulation value for the (t+1)th power frequency cycle. This represents the first-order sequence accumulation value of the t-th power frequency cycle; By performing a first-order cumulative subtraction inverse operation, the predicted first-order sequence cumulative value is restored to the initial predicted value of magnetic flux distortion characteristics under the (t+1)th power frequency cycle:

[0027] in, This represents the initial predicted value of the magnetic flux distortion characteristics under the (t+1)th power frequency cycle. This represents the predicted value of the first-order sequence accumulation value for the (t+1)th power frequency cycle. This represents the first-order sequence accumulation value of the t-th power frequency cycle; Calculate the residual of the power frequency period e : = - , where e=t-n+1,t-n+2,…,t, The characteristic value of magnetic flux distortion under the power frequency period e. The initial predicted value of magnetic flux distortion characteristics under power frequency period e; Using the power frequency cycle index e as the independent variable, As the dependent variable, construct Polynomial model of degree:

[0028] in, For residuals, , , ,..., These are the polynomial regression coefficients, determined by least squares fitting. After obtaining the well-fitted polynomial model, the index t+1 of the next power frequency cycle is substituted into the polynomial model to obtain the predicted residual value for the (t+1)th power frequency cycle. ; Based on the residual prediction value and the initial prediction value of the magnetic flux distortion feature, the final prediction value of the magnetic flux distortion feature is obtained:

[0029] in, This represents the predicted value of magnetic flux distortion characteristics under the (t+1)th power frequency cycle. This represents the predicted residual value at the (t+1)th power frequency cycle.

[0030] Preferably, the transformer loosening risk value is calculated by combining the loosening interval length and the predicted magnetic flux distortion value, specifically as follows: Neighborhood analysis is performed on the loosening anomalies to obtain the length of the loosening interval, and the maximum distortion risk within the loosening interval is calculated. ; Based on the predicted and actual flux distortion characteristic values, the distortion risk value is calculated:

[0031] in, Distortion risk value, This represents the predicted value of magnetic flux distortion characteristics under the (t+1)th power frequency cycle. This represents the characteristic value of magnetic flux distortion during the (t+1)th power frequency cycle. Indicates the monitoring coefficient; After normalizing the loosening interval length, the transformer loosening risk value is calculated by combining the loosening interval length and the maximum distortion risk.

[0032] in, Indicates the overall risk value of the region. Indicates the first The normalized loosening interval length Indicates the weight of the loosening interval length. Indicates the distortion risk weight. For loose interval index, Indicates the first The maximum distortion risk within each loosening range. () indicates taking all loosening intervals. The maximum value.

[0033] A transformer core loosening fault diagnosis device includes a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps of the above method.

[0034] This invention provides a method for diagnosing transformer core loosening faults, involving machine learning and deep learning technologies, which has the following beneficial effects: (1) Singular value decomposition (SVD) is used to process the flux distortion component, achieving dimensionality reduction, noise reduction, and mode separation of fault features. The flux distortion signal is a complex time-series data mixture of background noise, normal operation fluctuations, and fault features. Direct analysis of this signal makes it difficult to capture stable fault modes directly related to mechanical loosening. Singular value decomposition can decompose the observed data matrix into orthogonal components that characterize mode features, mode energy, and time-series properties, concentrating the energy of the original signal on the first few larger singular values ​​and their corresponding eigenvectors. This allows the method to automatically extract the most essential and stable low-dimensional mode features representing core mechanical loosening from the high-dimensional, noisy original signal, thus providing a pure and physically meaningful input for subsequent quantitative assessment and prediction, laying the foundation for high-precision diagnosis.

[0035] (2) The introduction of regularized truncation technique based on singular value decomposition resolves the contradiction between information compression and detail loss in feature extraction. Directly discarding smaller singular values ​​can reduce dimensionality and noise, but it may also lose key information related to early and weak faults contained in these components. By introducing a regularization matrix, effective signals are purposefully extracted and compensated from the truncated residual components, ensuring that the extracted fault features contain both the main energy of the dominant mode and recover subtle fault traces that may be misjudged as noise. This significantly improves the completeness and accuracy of feature extraction, especially enhancing the sensitivity to early and weak faults.

[0036] (3) The significance of using the grey prediction model lies in solving the problem of effectively predicting the trend of core loosening under limited data conditions. Transformer fault development is a dynamic process, but effective data samples with obvious trends that can be used for modeling are usually scarce. The core advantage of the grey prediction model is precisely its small sample and information-poor modeling. It strengthens the data regularity by generating sequences through first-order cumulative generation and establishes differential equations to describe its inherent development trend. This allows the method to quantitatively extrapolate the development trend of core flux distortion characteristic values ​​using only monitoring data from the most recent few periods, realizing a leap from static current state diagnosis to dynamic future trend prediction, and providing a time window for early warning.

[0037] (4) The introduction of polynomial regression residual correction technology into the grey prediction model significantly improves the prediction accuracy and model adaptability for nonlinear and fluctuating flux distortion sequences. Traditional grey prediction models perform well in predicting sequences with exponential growth trends, but the flux distortion characteristics during transformer operation often exhibit strong nonlinear and non-stationary characteristics due to load changes, transient events, etc. Directly using traditional models can lead to large prediction deviations. This invention effectively solves the inherent defect of insufficient fitting of nonlinear data by performing polynomial regression fitting on the initial prediction residual sequence of the grey prediction model and using the fitted values ​​to adaptively compensate for the initial prediction results. Attached Figure Description

[0038] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart of the steps in the method for diagnosing transformer core loosening faults proposed in this invention; Figure 2This is a step hierarchy diagram of obtaining magnetic flux distortion characteristic values ​​in a transformer core loosening fault diagnosis method proposed in this invention; Figure 3 This is a step hierarchy diagram of obtaining the predicted value of magnetic flux distortion characteristics in a transformer core loosening fault diagnosis method proposed in this invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Please see Figures 1-3 The present invention provides a technical solution: a method for diagnosing transformer core loosening faults.

[0042] Step S1: Deploy several monitoring nodes on the transformer and collect magnetic flux data of the monitoring nodes under normal operating conditions to obtain standard edge magnetic flux signals.

[0043] A magnetic field sensing network is deployed at key structural locations within the transformer core. This network consists of N highly sensitive magnetic elements, such as anisotropic magnetoresistive sensors or giant magnetoresistive sensors. These N sensors are non-invasively fixed to the outer wall of the transformer tank, with their installation positions strictly corresponding to the internal geometry of the core. Key areas covered include the sidewall surfaces of the core columns, the ends of the upper and lower yokes, the corresponding areas of the core lamination joints, and the projected area where the core clamps connect to the core body. After installation, the spatial coordinates of each sensor and its corresponding structural location within the core must be accurately recorded to establish a precise measurement point-structure mapping. Each sensor must be calibrated before use to determine its sensitivity coefficient. (Unit: V / T) This represents the sensitivity coefficient of the i-th sensor, used to convert the subsequently acquired voltage signal into a magnetic induction intensity signal.

[0044] Subsequently, after the transformer is put into normal operation, data from all sensors are recorded synchronously under different operating condition combinations. The required operating condition variables include: different stable load states where the load current increases in steps from 20% to 100% of the rated value; voltage fluctuations within ±5% of the rated value; the operating status of the cooling system under different switching combinations; and the natural variation of ambient temperature during typical day-night and seasonal cycles. Under each target operating condition, after the transformer operating parameters (load current, voltage, cooling system status, ambient temperature) have stabilized, the raw voltage signals output by all sensors are synchronously and continuously collected at a sampling rate of no less than 10 kS / s (thousands of samples per second), with a collection duration sufficient to cover hundreds of power frequency cycles. Simultaneously, the load current, voltage, cooling system status, ambient temperature, and other parameters under this stable operating condition are accurately recorded as operating condition labels for this data set, and each operating condition label is assigned an index number j (j=1,2,...,J), where J is the number of operating condition labels.

[0045] It should be noted that the operating condition labels should be selected from representative values ​​of various operating condition variables. For example, the load current can be specifically set as 100A (approximately 20% of the rated value), 250A (approximately 50% of the rated value), 400A (approximately 80% of the rated value), and 500A (100% of the rated value); the voltage level can be set as 209kV (approximately 95% of the rated value), 220kV (100% of the rated value), and 231kV (approximately 105% of the rated value); the cooling system status can be defined as the specific number of cooler groups in operation, such as 1 group (minimum operating mode) and 4 groups (all operating modes); the ambient temperature can be set as a specific temperature value, such as -10°C (low temperature in winter), 25°C (normal temperature), and 40°C (high temperature in summer). Different operating condition variables are combined to form specific operating condition labels. For example, a specific operating condition label such as {Load Current: 500A, Voltage: 220kV, Number of Cooling Groups: 4, Ambient Temperature: 40°C} can be formed, along with other operating condition labels such as {Load Current: 250A, Voltage: 220kV, Number of Cooling Groups: 1, Ambient Temperature: 25°C}. Each operating condition label j corresponds to a set of parameters determined by specific values. Therefore, each operating condition label j corresponds to a set of parameter vectors. For example, if the operating condition label is {Load Current: 250A, Voltage: 220kV, Number of Cooling Groups: 1, Ambient Temperature: 25°C}, then the parameter vector corresponding to this operating condition label is [250, 220, 1, 25]. The parameter vectors under all operating condition labels are then Z-score standardized.

[0046] The acquired raw voltage signals are first converted into magnetic flux density data based on the sensitivity coefficients of each sensor:

[0047] in, Let represent the magnetic flux density output by the i-th sensor during the t-th power frequency cycle. This represents the sensitivity coefficient of the i-th sensor. This represents the voltage signal output by the i-th sensor during the t-th power frequency cycle.

[0048] Subsequently, the magnetic field strength data needs to undergo a preprocessing process: using power frequency synchronous phase-locked loop technology, a reference signal that is strictly synchronized with the fundamental frequency and phase of the power grid is generated. The amplitude ratio is calculated through adaptive filtering (such as the LMS algorithm). and phase difference The edge fluctuation signal after eliminating the main magnetic flux is obtained:

[0049] in, This represents the edge fluctuation signal of the i-th sensor after the main magnetic flux in the t-th power frequency cycle. This represents the amplitude scaling factor between the i-th sensor signal and the reference signal. This represents the phase difference between the i-th sensor signal and the reference signal.

[0050] It should be noted that the edge fluctuation signal obtained after eliminating the main magnetic flux is... The technical prerequisite for this is generating a reference signal that is strictly synchronized with the fundamental frequency of the power grid. The reference signal originates from the secondary signal of the voltage transformer installed on the transformer bushing or busbar. A pure sine wave signal, perfectly locked to the grid voltage frequency and phase, is generated through a high-precision digital phase-locked loop (this closed-loop control system dynamically adjusts the output until complete synchronization by comparing the phase difference between the input voltage and the internal oscillation signal). The transformer core vibration mainly stems from the magnetostrictive effect, but the strong main power frequency magnetic flux generated by the winding current masks the weak flux edge fluctuations caused by mechanical defects such as loose core joints and attenuated clamp pressure. This is achieved through calculation formulas. Cancelling the main magnetic flux can significantly improve the signal-to-noise ratio of the signal, making the magnetic flux distortion components that were originally masked and are directly related to the mechanical state (such as magnetic flux jumps at the joints and magnetic flux redistribution in the clamping area) stand out.

[0051] After preprocessing, digital signal processing algorithms (such as Butterworth filters or Chebyshev filters) are used to process the signal. After filtering, the edge fluctuation signal of each sensor is extracted. A continuous power frequency periodic waveform, t=1,2,..., The waveform is then time-aligned and arithmetic-averaged to suppress random noise, ultimately extracting representative characteristic waveforms under this stable operating condition.

[0052] in, This represents the standard edge magnetic flux signal of the i-th sensor. This represents the total number of power frequency cycles, where t is the t-th power frequency cycle. For example, for a 50Hz power grid, the power frequency cycle is equal to 0.02s.

[0053] Finally, the standard edge magnetic flux signal of the j-th working condition tag under different monitoring nodes is obtained. The standard edge magnetic flux signal is associated with the location of the monitoring point to form a complete basic data record.

[0054] It should be noted that under normal conditions, edge magnetic flux signals inevitably exist in the transformer core. The physical origins of this are mainly twofold: First, it stems from the inherent magnetostrictive effect of the core material. Under the influence of an alternating magnetic field, silicon steel sheets undergo periodic micro-deformation, triggering inherent magnetostrictive vibrations and associated magnetic field fluctuations. These fluctuations, determined by the material's physical properties, constitute the background signal during normal operation. Second, it originates from the structural characteristics of the core itself. Even under ideal manufacturing and fastening conditions, discontinuities in the core structure, such as seams and clamps, can cause minute but repeatable distortions in the distribution of magnetic field lines, thus forming specific edge magnetic flux fluctuation patterns. Therefore, when constructing the standard edge magnetic flux signal in step S1, what is collected and stored is precisely this inherent, stable edge magnetic flux signal under healthy conditions. Its core significance lies in establishing a quantifiable benchmark for subsequent diagnosis.

[0055] Step S2: Collect real-time edge magnetic flux signals and compare them with standard edge magnetic flux signals to calculate the magnetic flux distortion components; extract features from the magnetic flux distortion components using singular value decomposition based on truncation regularization to obtain magnetic flux distortion feature values.

[0056] After acquiring the magnetic flux density signals of each sensor under the current operating state, a preprocessing procedure identical to that in step S1 is first performed, including power frequency synchronous phase-locking to eliminate the fundamental component and bandpass filtering, thereby obtaining the real-time edge magnetic flux signals representing each measuring point under the current state. .

[0057] It should be noted that the standard edge flux signal is baseline data collected and stored under stable operating conditions when the transformer is initially put into operation or after maintenance confirms its mechanical condition is intact. After multi-cycle averaging, it represents the inherent flux fluctuations under conditions without loosening faults. The real-time edge flux signal, on the other hand, is the actual flux data continuously collected by the online monitoring system during the transformer's daily operation at the current power frequency cycle. This actual flux data not only includes the inherent flux fluctuations under stable operating conditions but may also include abnormal flux distortion components caused by core loosening faults. For example, when the core lamination joints are loose or the clamping pressure is insufficient, it can lead to changes in local magnetic reluctance, resulting in a redistribution of the flux path or abrupt flux changes.

[0058] Subsequently, the real-time edge magnetic flux signal is compared with the standard edge magnetic flux signal of the corresponding operating condition label to calculate the magnetic flux distortion component:

[0059] in, This represents the magnetic flux distortion component of the i-th sensor. This represents the real-time edge magnetic flux signal of the i-th sensor. This represents the standard edge magnetic flux signal of the i-th sensor.

[0060] It should be noted that the flux distortion component represents the difference in flux between the current mechanical state of the transformer core and the standard edge flux signal; it is the difference between the real-time edge flux signal and the standard edge flux signal under the same operating conditions. Ideally, when the mechanical state of the core remains unchanged, this difference should be close to zero. Therefore, when this difference (i.e., the flux distortion component) shows a significantly non-zero value, the flux distortion component directly reflects the change in flux caused by a core loosening fault.

[0061] It should be noted that after obtaining the magnetic flux density signal of the transformer under the current operating state, it is first necessary to identify its current operating condition and match it with the standard edge flux signal with the same operating condition label established in step S1. Specifically, after obtaining the real-time edge flux signal, the current operating parameters of the transformer (such as load current, voltage level, cooling status and ambient temperature) are identified as the real-time parameter vector. After the real-time parameter vector is Z-score standardized, the Euclidean distance between the real-time parameter vector and the parameter vector of each operating condition label in the standard signal library is calculated. When the calculated minimum Euclidean distance is ≤ the preset Euclidean distance threshold (which can be obtained by collecting the values ​​of all minimum Euclidean distances in historical data and taking their 98th percentile as the preset Euclidean distance threshold), it is determined to be an exact match. Then, the standard edge flux signal under the operating condition label corresponding to the minimum Euclidean distance is called as the currently calculated standard edge flux signal. When all Euclidean distances are greater than the preset Euclidean distance threshold, it is determined that a precise match cannot be made. In this case, Num (e.g., Num=3) working condition labels with the smallest Euclidean distance to the real-time parameter vector in the standard signal library are selected as adjacent working conditions. The Euclidean distance between each adjacent working condition and the real-time parameter vector is used as a weight, and the weights are normalized (each weight is divided by the sum of the weights of all adjacent working conditions). Finally, the standard edge magnetic flux signals of the Num adjacent working conditions are weighted and summed with the corresponding normalized weights. The weighted sum is used as the standard edge magnetic flux signal currently calculated.

[0062] To extract stable characteristic quantities representing the overall energy of the magnetic flux distortion from the magnetic flux distortion components, singular value decomposition is required. The magnetic flux distortion component of the i-th sensor within one power frequency cycle is discretized into a magnetic flux distortion sequence [d1,d2,d3,...,dN] based on the sensor's sampling frequency, where [d1,d2,d3,...,dN] represents... The first to Nth sampling points are used, where N represents the total number of sampling points within the power frequency cycle. A Hankel matrix H is constructed based on the flux distortion sequence to explore its inherent spatiotemporal correlation. The number of rows m and columns p of matrix H are determined to satisfy the condition: m + p - 1 = N, making the matrix structure more conducive to subsequent decomposition. Typically, m = ... N / 2 p= N / 2 +1, where The sign for rounding up approximates a square matrix. Each element in matrix H is determined by the following rule: H[m1,p1]=d[m1+p1-1], where m1 is the row index (from 1 to m) and p1 is the column index (from 1 to p).

[0063] It should be noted that constructing the Hankel matrix H essentially embeds the temporal shift information of the magnetic flux distortion component into the matrix's row and column structure. Each row of the matrix can be viewed as a delayed sample of the magnetic flux distortion component. This reconstruction effectively reveals the potential linear dynamic characteristics in the time series, laying a solid foundation for subsequent matrix analysis methods such as singular value decomposition to extract the main mode components of the magnetic flux distortion component.

[0064] To separate modal features of different energy levels from the constructed Hankel matrix H, singular value decomposition is performed on the Hankel matrix H:

[0065] in, It is a singular value diagonal matrix, which is an m×p rectangular diagonal matrix, and the elements on its main diagonal are singular values. Let represent the left singular vector matrix, which is an m×m matrix. The transpose of the right singular vector matrix is ​​a p×p matrix.

[0066] The inflection point method is used to plot the singular value descent curve, identify the inflection points in the curve (which can be identified using the second-order difference method), and take the x-coordinate value corresponding to the first identified inflection point as the k-value, thereby extracting the singular value diagonal matrix. The top k singular values ​​arranged in descending order on the main diagonal constitute the dominant distortion feature vector. , =[ , ,..., ], This represents the k-th singular value, and the vector represents the most significant and stable mode energy in the flux distortion component.

[0067] To avoid the potential loss of useful information if the remaining mk small singular values ​​are directly truncated, a regularization compensation technique is introduced. First, the first k singular values ​​and their corresponding right singular vector matrices are used... Reconstruct an initial solution Then, a regularization matrix L is introduced to extract the effective signal compensation component Δx from the truncated residual matrix:

[0068] in, Indicates the effective signal compensation component. Let L represent the right singular vector matrix corresponding to the truncated singular values, and let L represent the regularization matrix. Representation matrix The generalized inverse, Representing the fundamental solution The regularization bias, This represents the initial solution reconstructed from the first k singular values.

[0069] It should be noted that the regularization matrix L is essentially a linear constraint operator. Its function is to impose specific prior assumptions (such as smoothness constraints) on the distribution characteristics of the truncated residual signal, thereby selectively extracting potentially relevant information related to real physical faults from the discarded small singular value components, rather than simply discarding the entire residual part as noise. Matrix L, through a specific structure (commonly in the form of a first-order or second-order difference matrix), forces the compensation amount... Meeting certain smoothness or gradual change conditions, the magnetic flux distortion caused by mechanical faults such as loose iron core often exhibits continuous and gradual characteristics in its spatiotemporal distribution.

[0070] Compensation components of the extracted effective signal Superimposed on the initial solution The final solution vector after feature enhancement is obtained as follows:

[0071] in, This represents the final solution vector after feature enhancement. This represents the initial solution reconstructed from the first k singular values. This indicates the effective signal compensation component.

[0072] Finally, the final solution vector after feature enhancement is calculated. The L2 norm is used as the characteristic value of magnetic flux distortion of the i-th sensor in the current power frequency cycle. :

[0073] in, Let be the magnetic flux distortion characteristic value of the i-th sensor.

[0074] Step S3: Based on the topological sequence of the transformer, construct the core flux topology loop; based on the flux distortion characteristic value of the monitoring node, use the three-line interpolation method to perform continuous trajectory interpolation on the core flux topology loop to obtain the flux distortion topology loop.

[0075] To achieve continuous mapping of sensor data in the core space, a magnetic flux topology loop for the transformer core is defined. On the 3D model or drawing of the transformer core, a closed path is defined, connecting the beginning and end of the loop. This path should closely follow the geometric centerline of the core laminations and traverse the main magnetic flux loop. For example, for a three-phase transformer, a typical path is: starting from the center point of the bottom of phase A core column (topological coordinate 0) and ending at the center point (topological coordinate 1) → ascending along the centerline of phase A core column to the upper yoke → passing through phases B and C along the centerline of the upper yoke → descending along the centerline of phase C core column to the lower yoke → returning to the starting point along the centerline of the lower yoke through phases B and A, forming a complete magnetic circuit loop.

[0076] The total physical length of the magnetic flux topology loop in the iron core can be obtained using geometric methods (either in CAD software or through mathematical integration). Normalized topological coordinates of the i-th sensor It is calculated using the following formula:

[0077] in, These are the normalized topological coordinates of the i-th sensor. It is the arc length from the starting point of the iron core flux topology loop to the projection point of the i-th sensor on the iron core flux topology loop. It is the total arc length of the iron core flux topology circuit.

[0078] Next, the discrete sensor data is correlated with the physical location of the iron core to prepare for spatial interpolation. Normalized topological coordinates are assigned to the flux distortion component of each sensor. Then, under any power frequency cycle, the magnetic flux distortion topology set Data(t) consisting of the positions of all sensors and the magnetic flux distortion components at that moment is obtained as follows: , ),( , ),...,( , ),...,( , )}, where Data(t) represents the set of magnetic flux distortion topology under the t-th power frequency cycle, Represents the normalized topological coordinates of the i-th sensor. This represents the magnetic flux distortion characteristic value of the i-th sensor during the t-th power frequency cycle.

[0079] In order to convert discrete sensor data points ( , The magnetic flux distortion characteristic value is transformed into a continuous smooth trajectory defined on the entire iron core loop [0,1]. This invention uses the cubic spline interpolation method. This ensures that the fitted curve not only passes through every data point, but also has continuous first and second derivatives, thereby accurately reflecting the smooth change trend of the magnetic flux distortion characteristic value along the iron core space.

[0080] The topological coordinate interval [0,1] is normalized according to the topological coordinates of n sensors. Divide the interval into n-1 sub-intervals, and require... Strictly increasing, 0 = < <...< <...< =1, where and These represent the start and end points of the iron core flux topology circuit, respectively.

[0081] In each sub-interval [ , Define a cubic polynomial on [the surface]. As an interpolation function:

[0082] in, Let represent the cubic polynomial interpolation function over the i-th subinterval. Denotes the coefficient of the constant term of the cubic polynomial over the i-th subinterval. Let the coefficient of the linear term of the cubic polynomial over the i-th subinterval be denoted as . Denotes the coefficient of the quadratic term of the cubic polynomial over the i-th subinterval. Let represent the coefficient of the cubic term of the cubic polynomial over the i-th subinterval.

[0083] According to the definition of cubic splines, the equation must be determined under the following conditions: the curve must pass through every data point and at the left endpoint of the interval. Location: At the right endpoint of the interval (also the left endpoint of the next interval): ; in each internal topologically normalized coordinate ( At the position =2,…,n-1), the previous interval polynomial first derivative Equal to the next interval polynomial first derivative : = and +2 +3 = , = - , This represents the length of the i-th subinterval. This represents the length of the (i-1)th subinterval; within each data point... ( At the position =2,…,n-1), the previous interval polynomial The second derivative Equal to the next interval polynomial The second derivative : = And 2 +6 =2 .

[0084] To ensure a unique solution to the system of equations, boundary conditions must be applied at the endpoints of the interval. Since the iron core magnetic circuit is a closed loop with its ends connected, the starting point of the path ( =0) and the endpoint ( =1) Physically, they are the same point. Therefore, using natural boundary conditions is a reasonable choice, i.e., assuming that the curvature of the trajectory at the start and end points is zero (the second derivative is zero): at the start... Second derivative And at the finish line Second derivative =2 +6 =0.

[0085] At internal nodes, the second derivative of the preceding interval must be equal to that of the following interval, i.e. = Combining the condition for continuity of the first derivative, we can derive a tridiagonal linear equation for solving the continuity of the second derivative, let... = - , Let represent the length of the i-th subinterval. Then, the system of equations holds for every internal node i = 2, 3, ..., n-1. The tridiagonal linear equations are as follows:

[0086] Using the chasing method to solve the tridiagonal linear equations, we obtain all and based on Calculations yielded and .

[0087] After obtaining all intervals [ , polynomial coefficients on ] ( = ), , , Then, the complete core flux interpolation function was obtained. This yields the flux distortion topology loop (i.e., each point in the iron core flux distortion topology loop corresponds to a flux distortion eigenvalue). For any point on the flux distortion topology loop, find... The interval in which it is located satisfies ≤ < +1, then the interpolated flux distortion characteristic value of point is = .

[0088] It should be noted that cubic spline interpolation is introduced to resolve the mapping discrepancy between the finite sensor measurement points and the continuous iron core space. The physical basis for this is that the transformer iron core, as a continuous medium, exhibits spatial continuity in its magnetic flux distribution and distortions caused by mechanical loosening. Therefore, cubic spline interpolation can reveal the masked distortion gradient information between discrete measurement points.

[0089] Step S4: Perform abrupt change analysis on the flux distortion topology loop to obtain loose anomalies; input the flux distortion feature values ​​of the loose anomalies into the discrete grey prediction model based on residual correction, and output the predicted values ​​of flux distortion features.

[0090] In the t-th power frequency cycle, using the projection points of n sensors on the flux distortion topology loop (i.e., the points corresponding to the normalized topological coordinates) as the calculation benchmark, the average value of the flux distortion eigenvalues ​​of these projection points is calculated. and standard deviation When the flux distortion eigenvalue of point on the flux topology loop of the iron core is > + (like =2.5, this The parameters are based on the principle of significance testing in statistics. Under the assumption of an approximately normal distribution, =2.5 corresponds to a high confidence level (approximately 99%), meaning that the probability of misjudging normal fluctuations as abnormalities is extremely low (approximately 1%), thus ensuring that the marked outliers have high statistical significance. Therefore, the point is marked as a loosening outlier.

[0091] To predict the characteristic values ​​of magnetic flux distortion in the iron core and identify suspected loose areas, a metabolic discrete grey prediction model is employed. This model uses discrete difference equations as its foundation, improving computational accuracy; and it incorporates a metabolic mechanism, enabling the model to dynamically update using the latest data, thus effectively tracking changes in the characteristic values ​​of magnetic flux distortion.

[0092] For each loosening anomaly, extract the most recent data from its historical data. The magnetic flux distortion characteristic values ​​of each consecutive power frequency cycle constitute the original sequence. , =( , ,..., ), where CH(t) is the characteristic value of magnetic flux distortion under the t-th power frequency cycle.

[0093] For the original sequence Perform first-order accumulation to generate a first-order sequence. :

[0094] in, Let represent the e-th element in a first-order sequence, where e is the index of an element in the first-order sequence. Represents the original sequence The past element is the index of the element in the original sequence.

[0095] Finally, a first-order sequence is obtained. =( , ,..., A discrete grey prediction equation is established based on a first-order sequence in discrete form:

[0096] in, This represents the (e+1)th element in a first-order sequence. The development coefficient reflects the overall development trend of the first-order sequence. This represents the gray action quantity.

[0097] To estimate the parameter vector using the least squares method = First, we need to construct the data matrix B and the observation vector Y, in the following form:

[0098] Where B is the data matrix and Y is the observation vector.

[0099] The parameters were finally solved using the least squares method. and Then, by substituting the discrete grey prediction equation, the first-order sequence cumulative value of the (t+1)th power frequency cycle is predicted:

[0100] in, This represents the predicted value of the first-order sequence accumulation value for the (t+1)th power frequency cycle. This represents the first-order sequence accumulation value of the t-th power frequency cycle.

[0101] By performing a first-order cumulative subtraction inverse operation, the predicted first-order sequence cumulative value is restored to the initial predicted value of magnetic flux distortion characteristics under the (t+1)th power frequency cycle:

[0102] in, This represents the initial predicted value of the magnetic flux distortion characteristics under the (t+1)th power frequency cycle. This represents the predicted value of the first-order sequence accumulation value for the (t+1)th power frequency cycle. This represents the first-order sequence accumulation value of the t-th power frequency cycle.

[0103] Calculate the residual of the power frequency period e : = - , where e = t - +1,t- +2,…,t, The characteristic value of magnetic flux distortion under the power frequency period e. The initial predicted values ​​for magnetic flux distortion characteristics under the power frequency period e are given. To address the problem of insufficient fitting of traditional models to nonlinear data, multinomial regression is introduced to fit the residual sequence: using the power frequency period index e as the independent variable, As the dependent variable, construct Polynomial model of degree:

[0104] in, For residuals, , , ,..., The coefficients are polynomial regression coefficients, determined by least squares fitting.

[0105] After obtaining the well-fitted polynomial model, the index t+1 of the next power frequency cycle is substituted into the polynomial model to obtain the predicted residual value for the (t+1)th power frequency cycle. .

[0106] Based on the residual prediction value and the initial prediction value of the magnetic flux distortion feature, the final prediction value of the magnetic flux distortion feature is obtained:

[0107] in, This represents the initial predicted value of the magnetic flux distortion characteristics under the (t+1)th power frequency cycle. This represents the predicted residual value at the (t+1)th power frequency cycle.

[0108] When the (t+1)th power frequency cycle ends, obtain the true magnetic flux distortion characteristic value CH(t+1), and calculate the prediction relative error to verify the model performance:

[0109] in, To predict relative error, This represents the characteristic value of magnetic flux distortion during the (t+1)th power frequency cycle.

[0110] If the prediction error does not meet the accuracy requirements, that is If the error rate is ≥ a preset relative error threshold (e.g., 5%), an early warning will be issued, indicating that a significant change may have occurred in the transformer. In this case, the time window should be expanded. Alternatively, check the data quality, temporarily suspend model updates, and record the anomaly. If the prediction error meets the accuracy requirements ( If the preset relative error threshold is reached, adaptive update will be initiated: the newly acquired true magnetic flux distortion feature values ​​will be updated. Add to the sequence, and remove the oldest historical data CH(t-) +1), ultimately forming a new original sequence. , =( , ,..., ). new sequence Still maintain length This is used to predict the next power frequency cycle, thereby enabling dynamic adaptive updating of the model.

[0111] It should be noted that if the prediction error does not meet the accuracy requirements, it indicates a significant deviation between the model's predicted values ​​and the actual observed values. In this case, a robust and conservative strategy will be adopted: the model's adaptive update process will be paused to prevent potentially abnormal state data from contaminating the existing health state model baseline. Simultaneously, the abnormal state and its corresponding operating data will be recorded to provide a data foundation for subsequent root cause analysis and trend tracing. Operators should use this as an alert to focus on verifying the quality of sensor data, or consider appropriately extending the model's learning time window. This mechanism captures gradual trends in state changes. It ensures the robustness of the diagnostic system in the face of unknown changes and organically combines early warning with preliminary diagnosis.

[0112] Step S5: Perform neighborhood analysis on the loosening anomaly points to obtain the loosening interval length; combine the loosening interval length and the magnetic flux distortion prediction value to calculate the transformer loosening risk value and classify the risk level, thereby realizing the identification of transformer loosening.

[0113] Based on the predicted and actual flux distortion characteristic values, the distortion risk value is calculated:

[0114] in, Distortion risk value, This represents the predicted value of magnetic flux distortion characteristics under the (t+1)th power frequency cycle. This represents the characteristic value of magnetic flux distortion during the (t+1)th power frequency cycle. This represents the monitoring coefficient.

[0115] It should be noted that the monitoring coefficient , is a preset positive integer used to define the length of the observation time window for calculating the risk value. Its specific value can be determined based on the system's required response speed to changes in risk; for example, if a rapid response is required, it can be set to . =3; to smooth fluctuations and observe trends, you can set... =10. This coefficient is usually set during system initialization.

[0116] Finally, the distortion risk value of each loosening anomaly point is calculated. The set of loosening anomaly points is recorded, and the normalized topological coordinates of the set are sorted. < <...< <...< , This represents the normalized topological coordinates of the i-th loose anomaly point. This represents the total number of loosening anomalies.

[0117] Calculate the length of the loosened interval: for point=1 to -1: Special attention should be paid to the arc length between the first and last points: Define the continuity threshold (like =0.03), when ≤ Then point and If they belong to the same continuous region, Then point and They do not belong to the same continuous region; similarly, according to... and The size of the value can be used to determine whether the first and last points are continuous.

[0118] Traverse all consecutive points, merge consecutive loosening anomalies into loosening regions, and calculate the interval length of each loosening region (calculated by normalizing topological coordinates; the maximum coordinate value minus the minimum coordinate value equals the interval length. When an interval crosses point 0, meaning the region contains coordinates close to both 1 and 0, indicating the anomaly is located at the beginning and end of a loop path, the arc length equals 1 minus the maximum coordinate value within the region plus the minimum coordinate value within the region). This yields the interval length of the loosening region, and the maximum distortion risk within the loosening interval is calculated. .

[0119] After normalizing the loosening interval length, the transformer loosening risk value is calculated by combining the loosening interval length and the maximum distortion risk.

[0120] in, Indicates the overall risk value of the region. Indicates the first The normalized loosening interval length Indicates the weight of the loosening interval length. Indicates the distortion risk weight. For loose interval index, Indicates the first The maximum distortion risk within each loosening range. () indicates taking all loosening intervals. The maximum value.

[0121] It should be noted that the weight of the loosening interval length and distortion risk weight The specific distortion risk weight needs to be determined comprehensively based on the transformer's structural characteristics, operating history, and maintenance strategy. If the transformer is a critical piece of equipment or has extremely high requirements for its operational reliability, the distortion risk weight can be appropriately increased. (For example, setting it to 1.2 to 1.5) can give higher sensitivity to rapidly deteriorating fault characteristics and achieve more stringent early warning; conversely, if it is necessary to focus on the overall trend of loosening and diffusion, the weight of the interval length can be appropriately increased. Furthermore, more precise values ​​can be determined based on historical fault data samples using decision-making tools such as the Analytic Hierarchy Process (AHP). The final determined weight values ​​should be verified in actual operation and can be dynamically optimized and adjusted based on accumulated diagnostic results.

[0122] Based on the calculated regional comprehensive risk value This can be divided into three risk levels to provide differentiated operation and maintenance guidance: when When the risk level is below the empirical threshold th1, it is considered a low-risk level, indicating that the current signs of easing are weak or stable, and it is recommended to include it in the routine monitoring plan and record its trend changes; when When the risk level is between threshold th1 and a higher threshold th2, it is classified as medium risk, indicating that the loosening is in a clear development stage, requiring a specific inspection plan to be developed and prioritized for handling during the most recent routine maintenance; when If the threshold th2 is exceeded, it is classified as a high-risk level, indicating that the loosening may have posed a substantial threat to the transformer's insulation or mechanical structure. An immediate warning must be issued and a power outage for maintenance must be arranged. The specific values ​​of thresholds th1 and th2 need to be determined based on historical fault statistics or industry expert experience, and can be dynamically calibrated in practical applications according to specific circumstances such as equipment importance and service life.

[0123] Furthermore, based on the above method embodiments, the present invention also provides an apparatus, including a memory, a processor, and a computer program stored in the memory, which is adapted to be loaded and executed by the processor to implement the above-described method for diagnosing transformer core loosening faults.

[0124] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, the phrase "comprising an element defined as..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0125] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for diagnosing transformer core loosening faults, characterized in that: Includes the following steps: Step S1: Deploy several monitoring nodes on the transformer and collect magnetic flux data of the monitoring nodes under normal operating conditions to obtain standard edge magnetic flux signals; Step S2: Collect real-time edge magnetic flux signals and compare them with standard edge magnetic flux signals to calculate the magnetic flux distortion components; extract features from the magnetic flux distortion components using singular value decomposition based on truncation regularization to obtain magnetic flux distortion feature values. Step S3: Based on the topological sequence of the transformer, construct the core flux topology loop; based on the flux distortion characteristic value of the monitoring node, use the three-line interpolation method to perform continuous trajectory interpolation on the core flux topology loop to obtain the flux distortion topology loop. Step S4: Perform abrupt change analysis on the flux distortion topology loop to obtain loose anomaly points; input the flux distortion feature values ​​of the loose anomaly points into the discrete grey prediction model based on residual correction, and output the predicted value of flux distortion features. Step S5: Perform neighborhood analysis on the loosening anomaly points to obtain the loosening interval length; combine the loosening interval length and the magnetic flux distortion prediction value to calculate the transformer loosening risk value and classify the risk level, thereby realizing the identification of transformer loosening.

2. The method for diagnosing transformer core loosening faults according to claim 1, characterized in that: The acquisition of magnetic flux data from the monitoring nodes under normal operating conditions to obtain standard edge magnetic flux signals includes the following specific steps: A sensor is installed at each monitoring node to convert the voltage signal collected by the sensor into a magnetic induction intensity signal. ； in, Let represent the magnetic flux density output by the i-th sensor during the t-th power frequency cycle. This represents the sensitivity coefficient of the i-th sensor. This represents the voltage signal output by the i-th sensor during the t-th power frequency cycle; The reference signal is obtained using power frequency synchronous phase-locked loop technology. Thus, the edge fluctuation signal after eliminating the main magnetic flux is obtained: ； in, This represents the edge fluctuation signal of the i-th sensor after the main magnetic flux in the t-th power frequency cycle. This represents the amplitude scaling factor between the i-th sensor signal and the reference signal. This represents the phase difference between the i-th sensor signal and the reference signal; Cut off The continuous power frequency periodic waveform is subjected to an arithmetic mean to suppress random noise, ultimately yielding the standard edge flux signal under normal operating conditions: ； in, This represents the standard edge magnetic flux signal of the i-th sensor. This represents the total number of power frequency cycles, and t represents the t-th power frequency cycle.

3. The method for diagnosing transformer core loosening faults according to claim 2, characterized in that: The process of collecting real-time edge magnetic flux signals and comparing them with standard edge magnetic flux signals to calculate the magnetic flux distortion component is as follows: ； in, This represents the magnetic flux distortion component of the i-th sensor. This represents the real-time edge magnetic flux signal of the i-th sensor. This represents the standard edge magnetic flux signal of the i-th sensor.

4. The method for diagnosing transformer core loosening faults according to claim 3, characterized in that: The step of extracting features from the magnetic flux distortion components using a singular value decomposition method based on truncation regularization to obtain magnetic flux distortion feature values ​​includes the following steps: The magnetic flux distortion component of the i-th sensor within one power frequency cycle is discretized into a magnetic flux distortion sequence [d1,d2,d3,...,dN] based on the sensor's sampling frequency. [d1,d2,d3,...,dN] represents the first to Nth sampling points of the magnetic flux distortion component, N represents the total number of sampling points within the power frequency cycle, and dN represents the magnetic flux distortion component value corresponding to the Nth sampling point. A Hankel matrix H is constructed based on the magnetic flux distortion sequence. Perform singular value decomposition on the Hankel matrix H: ； in, It is a singular value diagonal matrix, which is an m×p rectangular diagonal matrix, and the elements on its main diagonal are singular values. Let represent the left singular vector matrix, which is an m×m matrix. The transpose of the right singular vector matrix is ​​a p×p matrix, where m is the number of rows and p is the number of columns. Extract the first k singular values ​​to form the dominant distortion feature vector. , =[ , ,..., ], This represents the k-th singular value; By introducing a regularization matrix, effective signal compensation components are extracted. ; Compensation components of the extracted effective signal Superimposed on the initial solution The final solution vector after feature enhancement is obtained. : ； in, This represents the final solution vector after feature enhancement. This represents the initial solution reconstructed from the first k singular values. Indicates the effective signal compensation component; Finally, the final solution vector after feature enhancement is calculated. The L2 norm is used as the characteristic value of magnetic flux distortion of the i-th sensor in the current power frequency cycle. : ； in, Let be the magnetic flux distortion characteristic value of the i-th sensor.

5. The method for diagnosing transformer core loosening faults according to claim 4, characterized in that: The method of extracting effective signal compensation components by introducing a regularization matrix includes the following steps: The first k singular values ​​and their corresponding right singular vector matrices Reconstruct an initial solution By introducing a regularization matrix L, the effective signal compensation component Δx is extracted: ； in, Indicates the effective signal compensation component. Let L represent the right singular vector matrix corresponding to the truncated singular values, and let L represent the regularization matrix. Representation matrix The generalized inverse, Representing the fundamental solution The regularization bias, This represents the initial solution reconstructed from the first k singular values.

6. The method for diagnosing transformer core loosening faults according to claim 5, characterized in that: The method for constructing a core flux topology loop based on the transformer topology sequence includes the following steps: Based on the structural relationship of the internal magnetic circuit of the transformer, the core magnetic flux topology loop is constructed: the core magnetic flux topology loop is a closed spatial trajectory that closely follows the geometric center line of the core laminations and strictly follows the main magnetic flux path. In the specific construction, firstly, on the three-dimensional model of the transformer core, the path starts and ends at the center of the bottom of the core column, and flows along the center line of the core column through the upper and lower yokes to form a closed loop, thus obtaining the core magnetic flux topology loop. The physical total arc length of the core magnetic flux topology loop is calculated, and the projection position of each monitoring node on this loop is converted into normalized topological coordinates.

7. The method for diagnosing transformer core loosening faults according to claim 6, characterized in that: The method of using three-line interpolation to perform continuous trajectory interpolation on the magnetic flux topology loop of the iron core to obtain the magnetic flux distortion topology loop includes the following specific steps: Normalized topological coordinates of the i-th sensor It is calculated using the following formula: ； in, These are the normalized topological coordinates of the i-th sensor. It is the arc length from the starting point of the iron core flux topology loop to the projection point of the i-th sensor on the iron core flux topology loop. It is the total arc length of the magnetic flux topology loop in the iron core; Assign normalized topological coordinates to each sensor's flux distortion component. Then, in the t-th power frequency cycle, the magnetic flux distortion topology set Data(t) = {( , ),( , ),...,( , ),...,( , )}; Where Data(t) represents the magnetic flux distortion topology set under the t-th power frequency cycle. This represents the normalized topological coordinates of the i-th sensor. This represents the magnetic flux distortion characteristic value of the i-th sensor during the t-th power frequency cycle; The topological coordinate interval is divided into n-1 sub-intervals according to the normalized topological coordinates of n sensors, and these sub-intervals are strictly increased. In each sub-interval [ , The cubic polynomial interpolation function on is: ； in, Let represent the cubic polynomial interpolation function over the i-th subinterval. Denotes the coefficient of the constant term of the cubic polynomial over the i-th subinterval. Let the coefficient of the linear term of the cubic polynomial over the i-th subinterval be denoted as . Denotes the coefficient of the quadratic term of the cubic polynomial over the i-th subinterval. Denotes the coefficient of the cubic term of the cubic polynomial over the i-th subinterval; Solve using the chasing method. , , , Thus, a flux distortion topological loop is obtained.

8. The method for diagnosing transformer core loosening faults according to claim 7, characterized in that: The process of inputting the magnetic flux distortion feature values ​​of loosening anomalies into a discrete grey prediction model based on residual correction, and outputting predicted magnetic flux distortion feature values, includes the following specific steps: For each loosening anomaly point, extract the nearest... The magnetic flux distortion characteristic values ​​of each consecutive power frequency cycle constitute the original sequence. , =( , ,..., ), where CH(t) is the characteristic value of magnetic flux distortion under the t-th power frequency cycle; For the original sequence Perform first-order accumulation to generate a first-order sequence. : ； in, Let represent the e-th element in a first-order sequence, where e is the index of an element in the first-order sequence. Represents the original sequence The past element, where past is the index of an element in the original sequence; Finally, a first-order sequence is obtained. =( , ,..., Based on the discrete form of the first-order sequence, a discrete grey prediction equation is established: ； in, This represents the (e+1)th element in a first-order sequence. The development coefficient reflects the overall development trend of the first-order sequence. This is the gray action quantity; To solve for the parameters using the least squares method and Then, by substituting the discrete grey prediction equation, the first-order sequence cumulative value of the (t+1)th power frequency cycle is predicted: ； in, This represents the predicted value of the first-order sequence accumulation value for the (t+1)th power frequency cycle. This represents the first-order sequence accumulation value of the t-th power frequency cycle; By performing a first-order cumulative subtraction inverse operation, the predicted first-order sequence cumulative value is restored to the initial predicted value of magnetic flux distortion characteristics under the (t+1)th power frequency cycle: ； in, This represents the initial predicted value of the magnetic flux distortion characteristics under the (t+1)th power frequency cycle. This represents the predicted value of the first-order sequence accumulation value for the (t+1)th power frequency cycle. This represents the first-order sequence accumulation value of the t-th power frequency cycle; Calculate the residual of the power frequency period e : = - , where e = t - +1,t- +2,…,t, The characteristic value of magnetic flux distortion under the power frequency period e. The initial predicted value of magnetic flux distortion characteristics under power frequency period e; Using the power frequency cycle index e as the independent variable, As the dependent variable, construct Polynomial model of degree: ； in, For residuals, , , ,..., These are the polynomial regression coefficients, determined by least squares fitting. After obtaining the well-fitted polynomial model, the index t+1 of the next power frequency cycle is substituted into the polynomial model to obtain the predicted residual value for the (t+1)th power frequency cycle. ; Based on the residual prediction value and the initial prediction value of the magnetic flux distortion feature, the final prediction value of the magnetic flux distortion feature is obtained: ； in, This represents the predicted value of magnetic flux distortion characteristics under the (t+1)th power frequency cycle. This represents the predicted residual value at the (t+1)th power frequency cycle.

9. The method for diagnosing transformer core loosening faults according to claim 8, characterized in that: The transformer loosening risk value is calculated by combining the loosening interval length and the predicted magnetic flux distortion value, specifically as follows: Neighborhood analysis is performed on the loosening anomalies to obtain the length of the loosening interval, and the maximum distortion risk within the loosening interval is calculated. ; Based on the predicted and actual flux distortion characteristic values, the distortion risk value is calculated: ； in, Distortion risk value, This represents the predicted value of magnetic flux distortion characteristics under the (t+1)th power frequency cycle. This represents the characteristic value of magnetic flux distortion during the (t+1)th power frequency cycle. Indicates the monitoring coefficient; After normalizing the loosening interval length, the transformer loosening risk value is calculated by combining the loosening interval length and the maximum distortion risk. ； in, This represents the overall risk value of the region. Indicates the first The normalized loosening interval length Indicates the weight of the loosening interval length. Indicates the distortion risk weight. For loose interval index, Indicates the first The maximum distortion risk within each loosening range () indicates taking all loosening intervals. The maximum value.

10. A transformer core loosening fault diagnosis device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-9.

Images