A high-precision fault diagnosis method and system for wind farm cable insulation state

Abstract

The application relates to the technical field of wind farm cable insulation fault diagnosis, and discloses a high-precision fault diagnosis method and system for the insulation state of a wind farm cable, which comprises the following steps: acquiring radial multi-measurement-point electric-thermal response data, using a cross-correlation algorithm to calculate a time delay difference to generate a radial time delay gradient vector; acquiring axial multi-measurement-point electrical measurement values to generate an axial spatial distribution measurement vector; constructing a three-dimensional grid discretization space to generate a three-dimensional parameter tensor; based on a multi-physical field coupling simulator, using the radial time delay gradient vector and the axial spatial distribution measurement vector to inversely calculate radial and axial parameter distributions respectively; using a sparse regularization tensor completion algorithm to fuse bidirectional constraints to complete a complete three-dimensional parameter tensor; identifying abnormal grid units and extracting a defect area to generate a three-dimensional defect geometric body; and calculating characteristic parameters and outputting a hazard degree evaluation. The application solves the technical problem that the prior art cannot accurately identify the three-dimensional spatial position and geometric morphology of internal defects of insulation.

Description

Technical Field

[0001] This invention relates to the field of wind farm cable insulation fault diagnosis technology, and more specifically, to a high-precision fault diagnosis method and system for wind farm cable insulation condition. Background Technology

[0002] During long-term operation, wind farm cable insulation systems are subject to the coupling effects of electric fields, thermal fields, and environmental stresses, leading to various types of insulation defects such as voids, water trees, and cracks. These defects exhibit complex three-dimensional distribution characteristics, including radial depth (which layer of the multi-layer insulation structure has deteriorated), axial location (which segment along the cable length has the defect), and defect geometry. Accurately identifying the three-dimensional spatial distribution of internal insulation defects is crucial for assessing the cable's safety status, predicting its remaining life, and developing precise maintenance strategies.

[0003] Existing insulation diagnostic technologies mainly include methods such as dielectric loss measurement, partial discharge detection, and capacitance measurement. These methods obtain average parameters reflecting the overall insulation performance by performing electrical measurements on the cable surface or exterior. However, these methods have the following technical limitations: First, they only obtain average parameters based on overall surface measurements, which cannot distinguish the degradation depth of the inner and outer insulation layers, nor can they identify the location of defects in the radial direction; second, they only perform analysis in a single direction (radial or axial), which cannot obtain complete three-dimensional spatial information and makes it difficult to reconstruct the true geometric morphology of defects; third, they lack in-depth analysis of multi-physics coupling mechanisms and fail to fully utilize the depth information contained in the electrothermal coupling response.

[0004] The aforementioned technical deficiencies prevent existing methods from accurately identifying key characteristics such as the spatial location, geometry, size, and connectivity of internal insulation defects, making it impossible to assess the true severity and penetration risk of defects and to guide precise maintenance and remaining life prediction. Summary of the Invention

[0005] This invention provides a high-precision fault diagnosis method and system for the insulation condition of wind farm cables, solving the technical problems in related technologies that cannot accurately identify the three-dimensional spatial location, geometric shape, size and connectivity characteristics of internal insulation defects, and cannot assess the true degree of harm and penetration risk of defects.

[0006] This invention discloses a high-precision fault diagnosis method for the insulation condition of wind farm cables, comprising the following steps:

[0007] The temperature and capacitance time series of multiple measuring points along the radial direction of the cable surface are obtained. The time delay difference between the electrical and thermal responses at each radial position is calculated using a cross-correlation algorithm to generate a radial time delay distribution curve. The difference in time delay values ​​between adjacent measuring points is calculated to generate a radial time delay gradient vector.

[0008] The external electrical measurement values ​​of multiple measuring points along the cable axis are acquired. The external electrical measurement values ​​include local capacitance, local dielectric loss and partial discharge amplitude. An axial spatial distribution measurement vector is generated.

[0009] The cable insulation space is divided into multiple layers radially and multiple segments axially to construct a three-dimensional mesh discretization space. Internal parameter variables are set for each mesh unit to generate a three-dimensional parameter tensor.

[0010] Based on the multiphysics coupling simulator, the radial time delay gradient vector is used as a constraint condition to invert the radial parameter distribution of each axial segment and generate a radial constraint tensor.

[0011] Based on the multiphysics coupling simulator, the axial spatial distribution measurement vector is used as a constraint condition to invert the axial parameter distribution of each radial depth layer and generate the axial constraint tensor.

[0012] Using a sparse regularized tensor completion algorithm, the slice data of the radial constraint tensor and the axial constraint tensor are used as constraints. Combined with the prior knowledge of defect sparsity and spatial continuity constraints, the complete three-dimensional parameter tensor is completed iteratively by the alternating direction multiplier method.

[0013] Abnormal mesh cells with parameters deviating from the normal range are identified from the completed 3D parameter tensor. Abnormal regions with spatial connectivity are extracted using a 3D connectivity analysis algorithm to generate a 3D defect geometry.

[0014] Calculate the characteristic parameters of the three-dimensional defect geometry, and output the spatial reconstruction result and hazard assessment of the three-dimensional insulation defect.

[0015] Furthermore, the calculation of the time delay difference between the electrical and thermal responses at each radial position using the cross-correlation algorithm includes:

[0016] Cross-correlation calculations are performed on the temperature time series and capacitance time series of each radial measuring point, and the time offset corresponding to the peak value of the cross-correlation function is taken as the time delay value of that measuring point.

[0017] The cross-correlation operation is performed in a normalized form, which divides the cross-correlation function by the root mean square product of the two signals to eliminate the influence of signal amplitude differences on time delay estimation.

[0018] Furthermore, the internal parameter variables include dielectric constant, conductivity, and porosity.

[0019] Furthermore, the multiphysics coupled simulator is constructed based on the heat transfer and diffusion equation and the electric field propagation equation, wherein:

[0020] The heat transfer and diffusion equation describes the relationship between material density, specific heat capacity, thermal conductivity and the spatiotemporal distribution of temperature.

[0021] The electric field propagation equation describes the relationship between the dielectric constant, potential, and free charge density.

[0022] Furthermore, the inversion of the radial parameter distribution of each axial segment includes:

[0023] Using the radial time delay gradient vector as a constraint, the theoretical time delay gradient corresponding to different radial parameter distributions is calculated in the forward direction by the simulator.

[0024] The least squares optimization method is used to find the radial parameter distribution that minimizes the error between the theoretical and measured time delay gradients.

[0025] In least squares optimization, a Tikhonov regularization term is introduced to constrain the parameter changes of adjacent radial layers to be smooth.

[0026] Furthermore, the inversion of the axial parameter distribution of each radial depth layer includes:

[0027] Using the axial spatial distribution measurement vector as a constraint, the theoretical measurement values ​​corresponding to different axial parameter distributions are calculated in the forward direction by the simulator.

[0028] The regularized gradient descent algorithm is used for iterative optimization. In each iteration, the parameter values ​​are updated based on the gradient of the loss function and the gradient of the regularization term between the measured and simulated values ​​until convergence.

[0029] Furthermore, the optimization objective of the sparse regularized tensor completion algorithm includes:

[0030] The observation data fitting term ensures that the completed tensor matches the measured data at the observation location.

[0031] The nuclear norm constraint term utilizes the sparsity prior with a small defect ratio to achieve low-rank constraint.

[0032] The total variation regularization term utilizes the spatial continuity constraint of smooth changes in parameters between adjacent grids.

[0033] Furthermore, the extraction of spatially connected anomalous regions using a three-dimensional connectivity analysis algorithm includes:

[0034] For each grid cell, if the deviation of its parameter value from the average value of normal insulation parameters exceeds the product of a set threshold and the standard deviation, the grid cell is marked as an abnormal cell.

[0035] Using the 26-neighborhood connectivity criterion, adjacent anomalous mesh cells that are face-adjacent, edge-adjacent, and corner-adjacent are aggregated into independent three-dimensional defect regions.

[0036] Furthermore, the characteristic parameters include:

[0037] Volume is the number of mesh cells contained in the defect region multiplied by the volume of a single mesh cell;

[0038] Maximum radial depth is the radial position of the grid cell furthest from the outer surface of the cable in the defect region;

[0039] Axial span is the length of the defect area covered in the axial direction.

[0040] Morphological characteristic parameters include sphericity, which is the ratio of the defect volume to the volume of its smallest circumscribed sphere, and aspect ratio, which is the ratio of the axial span to the radial thickness.

[0041] The hazard assessment is based on the ratio of the maximum radial depth of the defect to the total insulation thickness, which evaluates the penetration risk level of the defect.

[0042] This invention discloses a high-precision fault diagnosis system for the insulation condition of wind farm cables, used to perform the above-mentioned method, including:

[0043] The radial data acquisition module is used to acquire the temperature time series and capacitance time series of multiple measuring points along the radial direction of the cable surface, calculate the time delay difference using a cross-correlation algorithm, and generate a radial time delay gradient vector.

[0044] The axial data acquisition module is used to acquire external electrical measurement values ​​of multiple measuring points along the axial direction of the cable and generate an axial spatial distribution measurement vector.

[0045] The 3D mesh construction module is used to discretize the cable insulation space into a 3D mesh, set internal parameter variables for each mesh cell, and generate a 3D parameter tensor.

[0046] The radial inversion module is used to invert the radial parameter distribution of each axial segment based on a multiphysics coupled simulator and using radial time delay gradient vector constraints to generate a radial constraint tensor.

[0047] The axial inversion module is used to invert the axial parameter distribution of each radial depth layer based on a multiphysics coupled simulator and by utilizing the axial spatial distribution measurement vector constraint to generate an axial constraint tensor.

[0048] The tensor completion module is used to complete the complete three-dimensional parameter tensor using a sparse regularized tensor completion algorithm, with radial and axial constraint tensors as constraints, and iteratively completes the complete three-dimensional parameter tensor through the alternating direction multiplier method.

[0049] The defect identification module is used to identify abnormal mesh cells from the completed 3D parameter tensor, extract spatially connected abnormal regions using a 3D connectivity analysis algorithm, and generate 3D defect geometry.

[0050] The evaluation output module is used to calculate the characteristic parameters of the three-dimensional defect geometry and output the spatial reconstruction results and hazard assessment of the three-dimensional insulation defect.

[0051] This invention introduces a tensor completion method with bidirectional constraints of electrothermal coupling gradient and simulation inversion. It utilizes radial time delay gradient vector and axial spatial distribution measurement vector to provide constraint information in the radial and axial directions, respectively. Combined with a sparse regularized tensor completion algorithm, the two methods mutually verify and complement each other within the tensor completion framework. This reconstructs a complete three-dimensional distribution from sparse observations, solving the technical problem in existing technologies that cannot accurately identify the three-dimensional spatial location, geometric shape, size, and connectivity characteristics of internal insulation defects based solely on overall surface measurement or single-direction analysis. This invention achieves the technical effect of enabling precise three-dimensional spatial positioning and geometric reconstruction of internal insulation defects, accurately assessing the true hazard level and penetration risk of defects, and providing a reliable basis for precise maintenance and remaining life prediction.

[0052] Specifically, the present invention has the following beneficial effects:

[0053] By utilizing the difference in propagation of electrothermal response along the insulation depth direction to calculate the radial time delay gradient vector, the defect location information in the radial direction can be reflected, overcoming the limitation of existing technologies that cannot distinguish the degradation depth of the inner and outer insulation layers.

[0054] By using electrical measurements from multiple measuring points arranged along the axial direction for axial simulation inversion, the parameter distribution of each axial segment can be separated, providing constraint information in the axial direction, thus overcoming the limitation of existing technologies that only perform single-direction analysis.

[0055] By fusing slice information from two orthogonal directions, radial and axial constraints, using the sparse regularized tensor completion algorithm, and leveraging the prior knowledge of defect sparsity and spatial continuity constraints, the complete three-dimensional parameter distribution was reconstructed from sparse observations, thus obtaining complete three-dimensional spatial information of internal defects in the insulation.

[0056] By extracting spatially connected abnormal regions through a three-dimensional connectivity analysis algorithm, generating a three-dimensional defect geometry, and calculating its volume, maximum radial depth, axial span, and other characteristic parameters, the penetration risk level and severity of defects can be accurately assessed, providing a quantitative basis for maintenance decisions. Attached Figure Description

[0057] Figure 1 This is a flowchart of the high-precision fault diagnosis method for the insulation status of wind farm cables according to the present invention. Detailed Implementation

[0058] This embodiment provides a high-precision fault diagnosis method for the insulation status of wind farm cables. It should be understood that the hardware environment for executing this method includes: multiple temperature sensors and capacitance sensors arranged radially along the cable surface, multiple electrical measuring devices arranged axially along the cable (for measuring local capacitance, local dielectric loss and partial discharge amplitude), and a processor with numerical calculation capabilities.

[0059] Step 1: Obtain the electrothermal response data of multiple radial measurement points of the cable, calculate the time delay difference using the cross-correlation algorithm, and generate the radial time delay gradient vector.

[0060] Acquire temperature time series at multiple measuring points along the radial direction on the cable surface and capacitance time series ,in Indicates the radial measuring point number, The time is represented. Before calculating the cross-correlation function, the temperature time series and the capacitance time series are first subjected to Z-score normalization to eliminate the differences in dimensions and amplitudes between the two physical quantities.

[0061] Furthermore, the temperature and capacitance time series data were acquired during the cable load transient process, with a sampling frequency of 100Hz and an acquisition duration of 10 seconds to ensure the complete electrothermal coupling response process was captured. Temperature measurement employed a PT100 platinum resistance temperature sensor with a measurement accuracy of ±0.1℃; capacitance measurement used a high-frequency capacitance sensor with a measurement frequency of 1MHz and a measurement accuracy of ±0.1pF.

[0062] The time delay difference between the electrical and thermal responses at each radial position was calculated using a cross-correlation algorithm. The cross-correlation function is defined as:

[0063]

[0064] Furthermore, for discretely sampled time series, the cross-correlation function is calculated in discrete form: ,in The time series length is calculated based on a data acquisition duration of 10 seconds and a sampling frequency of 100Hz. One sampling point, The number of sampling points corresponding to the time delay. The sampling interval (for a sampling frequency of 100Hz), s), For summation index.

[0065] For each radial measurement point, the value of the cross-correlation function at its peak value is taken. The value is used as the time delay value at this measuring point to generate a radial time delay distribution curve. ,in Let represent the time delay values ​​of the 1st, 2nd, ..., Kth radial measurement points, respectively. Calculate the difference in time delay values ​​between adjacent measurement points to generate the radial time delay gradient vector. ,in , These represent the time delay differences between adjacent measuring points 1 to 2, 2 to 3, ..., K-1 to K, respectively.

[0066] Furthermore, the search for the peak value of the cross-correlation function is achieved through the following steps: within the time delay range Within the sampling interval Calculate the cross-correlation function value, where Based on the cable insulation thickness and the velocity of electrothermal propagation, it is estimated that for an insulation layer with a thickness of 10mm, The value is set to 100ms. From the calculated sequence of cross-correlation function values, the time delay corresponding to the maximum value is taken as the time delay value for that measurement point. .

[0067] Routine insulation testing was performed on the 10kV main cable (model YJV3×240, insulation thickness 10mm, 5 years of operation) of Unit 3 of a wind farm. The testing section was 20m long. During the transient process of the load suddenly increasing from 30% to 80% of the rated power, 5 measuring points were arranged radially along the cable. Electrothermal response data were collected at a sampling frequency of 100Hz for 10 seconds. After Z-score normalization and cross-correlation calculation, the electrothermal delay values ​​at each radial measuring point were obtained, as shown in Table 1.

[0068] Table 1. Measurement results of radial heating time delay at measuring points;

[0069]

[0070] Based on the data in Table 1, calculate the difference in time delay values ​​between adjacent measurement points and generate a radial time delay gradient vector. ms / layer. The time delay gradient between the 3rd and 4th measurement points (11.7 ms / layer) was observed to be significantly greater than that at other locations, initially indicating that there may be an anomaly in dielectric constant or thermal conductivity at this radial depth.

[0071] Step 2: Obtain the external electrical measurement values ​​of multiple axial measurement points of the cable and generate an axial spatial distribution measurement vector.

[0072] Acquire external electrical measurements at multiple measuring points along the cable's axial direction, including local capacitance. Local dielectric loss and partial discharge amplitude ,in This indicates the axial measurement point number. The measured values ​​from each measurement point are arranged according to their axial position to generate an axial spatial distribution measurement vector.

[0073]

[0074] in These represent the local capacitance, local dielectric loss, and partial discharge amplitude at the 1st, 2nd, ..., Nth axial measuring points, respectively.

[0075] Furthermore, the local capacitance is measured using the capacitance bridge method at a frequency of 50 Hz and an accuracy of ±0.5%; the local dielectric loss is measured using a high-voltage Schering bridge at a voltage of 10 kV and an accuracy of ±0.0001; and the partial discharge amplitude is measured using the pulse current method with a high-frequency current sensor, a detection frequency range of 100 kHz to 30 MHz, and a detection sensitivity of 5 pC.

[0076] An electrical measuring device is arranged every 5m along the cable axis, for a total of 4 measuring points. The points were located at 5m, 10m, 15m, and 20m within the diagnostic zone. The local capacitance, local dielectric loss, and partial discharge amplitude at each point were measured using the capacitance bridge method, the high-voltage Schering bridge method, and the pulse current method, respectively. The measurement results are shown in Table 2.

[0077] Table 2. External electrical measurement results of axial measuring points;

[0078]

[0079] Table 2 shows that at the second axial measuring point (10m position), the local capacitance increased by 15.5%, the dielectric loss increased by 78.3%, and the partial discharge amplitude reached 85pC, significantly higher than other measuring points, indicating significant insulation degradation in this axial section. Axial spatial distribution measurement vectors were generated.

[0080] .

[0081] Step 3: Discretize the cable insulation space into a three-dimensional mesh to generate a three-dimensional parametric tensor.

[0082] The cable insulation space is divided radially into Layers, divided along the axial direction Segments are used to construct a three-dimensional mesh discretization space. Each mesh element is then used as a unit. Set internal parameter variables, including dielectric constant. Electrical conductivity and porosity Generate a three-dimensional parameter tensor .

[0083] Furthermore, the three-dimensional parameter tensor is initialized using the nominal parameter values ​​of normal insulating materials; the dielectric constant is initialized to the nominal value of cross-linked polyethylene (XLPE), 2.3, and the conductivity is initialized to... S / m, porosity is initialized to 0, and the parameter values ​​of each grid cell are subsequently updated through the inversion process in steps 4 and 5.

[0084] It should be noted that the radial division number of layers The number of axially divided segments can be determined based on the actual number of layers in the cable insulation structure. The value can be determined based on the axial measurement point spacing and the desired spatial resolution. To eliminate the influence of the dimensional differences of the three physical quantities—dielectric constant, conductivity, and porosity—on subsequent tensor calculations, the three parameters in the three-dimensional parametric tensor are respectively subjected to mean normalization based on the range.

[0085] Step 4: Based on the multiphysics coupling simulator, the radial parameter distribution of each axial segment is inverted using the radial time delay gradient vector constraint to generate the radial constraint tensor.

[0086] A multiphysics coupled simulator is constructed based on the heat transfer and diffusion equation and the electric field propagation equation. The heat transfer and diffusion equation is:

[0087]

[0088] in, For material density, For specific heat capacity, Thermal conductivity, It is an internal heat source. The electric field propagation equation is:

[0089]

[0090] in, Where is the dielectric constant. Potential, This represents the free charge density.

[0091] Furthermore, the material density Take the nominal value of cross-linked polyethylene as 930 kg / m 3 Specific heat capacity Take 2300 J / (kg·K), thermal conductivity Taking 0.4 W / (m·K) as the initial value, during the inversion process, the dielectric constant is used as the basis for empirical relationships. Update, in which W / (m·K) is the nominal thermal conductivity of a normal insulating material. This is the nominal dielectric constant of a normal insulating material. Internal heat source. It is generated by dielectric loss and is calculated as follows: ,in Electric field strength. Free charge density. The value is 0 in the normal insulation region, and the value is calculated based on the porosity and electric field strength in the defect region where space charge accumulates.

[0092] Furthermore, the heat transfer and diffusion equations and the electric field propagation equations are solved using the finite element method. The cable insulation space is discretized into a finite element mesh. The heat transfer and diffusion equations are discretized in time using the backward Euler method, and spatial discretization is performed using the Galerkin weighted residual method, resulting in a discretized system of linear equations. ,in For the quality matrix, Here is the stiffness matrix. Given the load vector, the temperature distribution at each time step is obtained by solving the linear equations using the conjugate gradient method. The electric field propagation equation is an elliptic partial differential equation, and the linear equations are obtained by spatial discretization using the Galerkin weighted residual method. The potential distribution can be obtained by direct methods (such as LU decomposition).

[0093] Furthermore, the time step Based on the stability condition, it satisfies... ,in Let be the characteristic size of the mesh cell. For a radial mesh size of 1 mm, take . The total simulation duration was set to 10 seconds, consistent with the actual measurement acquisition time, to ensure that the simulation results correspond to the measured data in the time dimension.

[0094] For each axial section The radial time delay gradient vector As a constraint, the theoretical time delay gradient corresponding to different radial parameter distributions is calculated forward using a multiphysics coupled simulator, and the radial parameter distribution of this section is inverted using least squares optimization. ,in They represent the first The dielectric constants of the 1st, 2nd, ..., Mth radial layers of each axial segment are used to fill radial slices of a three-dimensional parameter tensor, generating a radial constraint tensor. .

[0095] Furthermore, the steps for the multiphysics coupled simulator to calculate the theoretical time delay gradient are as follows: given a set of radial parameter distributions... The dielectric constant of each layer Substitute the electric field propagation equations to solve for the electric field response time of each layer. The thermal conductivity of each layer (based on the empirical relationship between dielectric constant and thermal conductivity) is determined. (Obtain) Substitute into the heat transfer and diffusion equation to solve for the thermal response time of each layer. Calculate the electrothermal delay of each layer. Thus, the theoretical radial time delay gradient is obtained. ,in These represent the simulation delay differences for layers 1 to 2, ..., and M-1 to M, respectively.

[0096] Furthermore, the electric field response time Defined as from the start of load transient to the [number]th [time period]. The time it takes for the layer capacitance response to reach 63.2% of its steady-state value is calculated as follows: ,in The equivalent resistance is calculated based on the conductivity. , For the first The capacitance of the layer The electrode area is taken as the cable surface area. , For the first The thickness of the layer is taken from the radial grid size. The thermal response time Defined as from the start of internal heat source activation to the [number]th The time when the layer temperature response reaches 63.2% of its steady-state value is obtained by solving the heat transfer and diffusion equations to obtain the time series. Find satisfaction The moment ,in This represents the steady-state temperature. The time dimension is determined through... within the s interval, with step size The solution of the time-discretized heat transfer and diffusion equations is reflected in s.

[0097] The input to the aforementioned least-squares optimization is the theoretical time delay gradient calculated by the multiphysics coupled simulator and the radial time delay gradient vector measured in practice. The output is the radial parameter distribution that minimizes the sum of squared residuals between the theoretical and actual time delay gradients. .

[0098] Furthermore, the least squares optimization problem has the following constraints: the dielectric constant constraint is... The lower bound of 1.5 corresponds to severely degraded insulating materials, and the upper bound of 10 corresponds to defect regions containing water or high dielectric constant impurities; the conductivity constraint is... The lower bound corresponds to normal insulation, and the upper bound corresponds to severe aging or water treeing degradation; the porosity constraint is... The upper bound of 0.3 represents the maximum void ratio in the defect region. When the parameter value exceeds the constraint range during the iteration process, it is truncated to the constraint boundary.

[0099] Furthermore, the least squares optimization is solved using the Levenberg-Marquardt algorithm, which combines the advantages of gradient descent and the Gauss-Newton method. The iterative formula is as follows: ,in Let Jacobian matrix be the residual with respect to parameters. This represents the transpose, calculated using the finite difference method. , For the residual vector, The perturbation step size is set to 0.001 times the current value of the parameter. For the first Unit vectors, This is the damping coefficient, which decreases as the residual decreases. This makes the Levenberg-Marquardt algorithm approximate the Gauss-Newton method, increasing as the residual increases. This makes the Levenberg-Marquardt algorithm approach the gradient descent method.

[0100] Furthermore, the damping coefficient The initial value is set to 0.01, when the first... When the next iteration reduces the sum of squared residuals, update When the first When the next iteration increases the sum of squared residuals, update At the same time, cancel the current iteration and recalculate. The range of values ​​for the damping coefficient is limited to... between.

[0101] In this embodiment of the application, in order to improve the stability of radial inversion, a Tikhonov regularization term is introduced in the least squares optimization to constrain the parameter changes of adjacent radial layers to be smooth.

[0102] Furthermore, the Tikhonov regularization term is defined as the sum of squares of the differences in parameters between adjacent radial layers. Its regularization coefficient is determined based on the measurement noise level, and its value range is [value range missing]. When the signal-to-noise ratio of the radial time delay gradient vector is less than 20dB, a larger regularization coefficient is used to enhance the smoothness constraint.

[0103] The cable insulation space is divided into 5 layers radially. Divided into 4 segments along the axial direction ( A three-dimensional mesh was constructed. For the second axial segment where the anomaly is most significant (corresponding to the 10m position in Table 2), the radial time delay gradient vector from Table 1 was used. As a constraint, radial inversion was performed using a multiphysics coupled simulator and the Levenberg-Marquardt algorithm. The initial dielectric constant was set to 2.3, and the simulation converged after 15 iterations. The sum of squared residuals decreased from the initial 137.6 to 0.8. The radial dielectric constant distribution of this axial segment was obtained through inversion, as shown in Table 3.

[0104] Table 3 shows the radial parameter inversion results for the second axial segment;

[0105]

[0106] Table 3 shows that the dielectric constants of the third and fourth layers (radial depth 4-8 mm) in the second axial segment increased to 3.87 and 4.21, respectively, the conductivity increased by 2-3 orders of magnitude, and the porosity reached 0.18-0.23, which is consistent with the typical characteristics of water tree aging defects. The parameters of the first, second, and fifth layers are close to normal values, indicating that the defects are mainly concentrated in the middle of the insulation.

[0107] Step 5: Based on the multiphysics coupling simulator, the axial parameter distribution of each radial depth layer is inverted using the axial measurement vector constraint to generate the axial constraint tensor.

[0108] Using the multiphysics coupling simulator in step 4, for each radial depth layer , axial spatial distribution measurement vector As a constraint, the theoretical measurements corresponding to different axial parameter distributions are calculated forward using a multiphysics coupled simulator, and the axial parameter distribution of this depth layer is iteratively optimized and inverted using a regularized gradient descent algorithm. ,in They represent the first The dielectric constants of the 1st, 2nd, ..., Nth axial segments of each radial layer are used to fill axial slices of a three-dimensional parametric tensor, generating an axially constrained tensor. .

[0109] Furthermore, the steps for the multiphysics coupled simulator to calculate theoretical measurements are as follows: given the axial position... The parameter distribution of the region and its vicinity is used to solve for the potential distribution at that location using the electric field propagation equation based on the dielectric constant distribution, and then the simulated value of the local capacitance is calculated. ,in The charge amount of this section is obtained by integrating the electric displacement vector of the surface of this section. , The applied measurement voltage is 10 kV. The dielectric loss formula is used based on the dielectric constant and conductivity. Calculate the simulated value of dielectric loss, where Angular frequency, Hz is the measurement frequency. Based on the electric field intensity distribution and porosity distribution, when the electric field intensity exceeds the gap breakdown threshold (3kV / mm for air gaps), the simulated value of the partial discharge amplitude is calculated. ,in The gap equivalent capacitance is... F / m is the vacuum permittivity. The equivalent area of ​​the voids is determined by the porosity. and the radial cross-sectional area of ​​the grid cell Estimated as , Take the radial grid size as the equivalent thickness of the void. , kV / mm is the breakdown threshold.

[0110] The input to the aforementioned regularized gradient descent algorithm is the theoretical measurement value calculated by the multiphysics coupled simulator and the actual axial spatial distribution measurement vector. The output is the axial parameter distribution that minimizes the loss function between the theoretical and actual measurements. To eliminate the influence of dimensional differences between different physical quantities on the calculation of the loss function, the local capacitance, local dielectric loss, and partial discharge amplitude are respectively normalized based on the range before the loss function calculation.

[0111] Furthermore, the gradient descent algorithm is also subject to the same physical constraints as in step 4: dielectric constant constraint. Conductivity constraint Porosity constraint When gradient updates cause parameters to exceed the constraints, the projected gradient method is used to project the parameters back into the feasible region.

[0112] The iterative update formula is:

[0113]

[0114] Furthermore, the gradient of the loss function with respect to the parameters Calculated using the finite difference method, for the th parameter in the distribution Each component Its gradient is calculated as ,in For the first Unit vectors, This is the perturbation step size, which is 0.01 times the current value of the parameter. The gradient of the regularization term. It has an analytical form, for the th One portion, .

[0115] in, The loss function between measured and simulated values ​​is specifically defined as:

[0116]

[0117] in, , and Based on parameter distribution The first result obtained by multiphysics coupling simulator Simulated values ​​of local capacitance, dielectric loss, and partial discharge amplitude at each measuring point. , and This corresponds to the actual measured value. , and These are the weighting coefficients for each measurement.

[0118] Furthermore, the weighting coefficients , and The weighting coefficient is determined based on the sensitivity of each measurement to insulation defects, specifically by using the reciprocal of the normalized standard deviation of each measurement as its weighting coefficient. , , ,in , and These are the normalized standard deviations of local capacitance, local dielectric loss, and partial discharge amplitude at all axial measurement points, respectively.

[0119] The regularization term is defined as the sum of squares of the differences between adjacent axial position parameters:

[0120]

[0121] The sum of squares regularization term of the difference between adjacent axial position parameters constrains the smooth change of parameters between adjacent axial positions. For learning rate, is the regularization coefficient.

[0122] Furthermore, the learning rate The range of values ​​is The loss function is adaptively adjusted based on its rate of decline. When the rate of decline of the loss function is less than 1 / 3 in three consecutive iterations... When the learning rate is halved, the regularization coefficient is used. The value range is determined based on the balance between the data fitting term and the smoothing constraint term. Cross-validation is used to select the option that minimizes the validation set error. value.

[0123] Furthermore, the cross-validation employs a 5-fold cross-validation method, dividing the axial measurement points into 5 subsets. Each time, 4 subsets are selected as the training set for inverting the parameter distribution, and the remaining subset is used as the validation set to evaluate the inversion accuracy. For candidate parameters... value (in) Ten candidate values ​​are selected at logarithmic intervals within a given range. Five-fold cross-validation is then performed on each candidate value to calculate the average validation error. The value with the smallest average validation error is selected. The value is used as the optimal regularization coefficient.

[0124] For the radial third layer (depth 4-6mm) with the most severe defects identified in Table 3, the axial measurement vectors in Table 2 were used. As a constraint, axial inversion is performed using a multiphysics coupled simulator and a regularized gradient descent algorithm. After normalizing the local capacitance, dielectric loss, and partial discharge amplitude, the weighting coefficients are determined as follows: , , The learning rate was initially set to 0.05, and the regularization coefficient was determined through cross-validation. After 42 iterations, the algorithm converged, and the loss function decreased from the initial 28.3 to 1.2. The axial dielectric constant distribution of the radial layer was obtained by inversion, as shown in Table 4.

[0125] Table 4. Axial parameter inversion results for the third radial layer;

[0126]

[0127] The results in Table 4 show that the radial third layer exhibits a significant anomaly in the axial second segment (5-10m position), with a dielectric constant of 3.87 consistent with the value at the same position in Table 3, verifying the mutual consistency of the two-way constraint inversion. The parameters in the axial first and fourth segments are close to normal values, while the axial third segment shows a slight anomaly, but to a minor degree, indicating that the main defects are concentrated in the axial 5-10m segment.

[0128] Step 6: Using the sparse regularized tensor completion algorithm, the radial constraint tensor and the axial constraint tensor are fused, and the complete three-dimensional parameter tensor is completed iteratively by the alternating direction multiplier method.

[0129] radial constraint tensor and axial constraint tensor Using sliced ​​data as constraints, and combining prior knowledge of defect sparsity and spatial continuity constraints, a sparse regularized tensor completion algorithm is employed to reconstruct the complete 3D parametric tensor. The optimization objective function of the sparse regularized tensor completion algorithm is:

[0130]

[0131]

[0132] Furthermore, the constraints of the optimization problem remain consistent with those in steps 4 and 5, ensuring that the parameter values ​​of each mesh element in the reconstructed 3D parameter tensor satisfy physical feasibility. The ADMM algorithm updates the original variables each time. Then, the parameter values ​​that exceed the constraint range are truncated and projected to limit them to the constraint range.

[0133] in, For the projection operator of the observation position, Observational data provided for radial and axial constraints. It is the Frobenius norm. It is the nuclear norm (used for low-rank constraints, reflecting the sparsity of defects). This is the total variational regularization term (used for spatial continuity constraints). and is the regularization coefficient.

[0134] Furthermore, the set of observation locations Includes all mesh cell locations obtained inversion in steps 4 and 5, specifically: radial constraint tensor Provides radial slice observations for all axial segments, corresponding location sets. Axial constraint tensor Provides axial slice observations of all radial depth layers, and corresponding location sets. The set of observation locations is the union of the two. Projection operator Defined as: when Time Retention The value is set to 0 if it is not set to 0 otherwise.

[0135] Furthermore, the regularization coefficient The strength of the constraint to control the sparsity of defects is within the range of values. The ratio of radial and axial constraint observation points to the total grid count is determined based on the completeness of the observation data. When the proportion of observation points to radial and axial constraints is less than 30%, the larger ratio is used. The value is used to enhance the low-rank constraint. The regularization coefficient... The strength of the control space continuity constraint is within the range of values. It is determined based on the expected smoothness of parameter space variation.

[0136] The above optimization problem is solved iteratively using the Alternating Direction Multiplier Method (ADMM). The ADMM iterative process includes:

[0137] Sub-step 601: Update the original variable Solve the minimization subproblem by fixing auxiliary variables;

[0138] Furthermore, the update of the original variable The specific steps are as follows: fix the auxiliary variables and dual variables Solve ,in This is the penalty parameter, and its value range is... The method is determined based on the problem size and convergence speed requirements; in this implementation, we take... This subproblem has a closed-form solution: The observed value is taken for the observation location, and the corresponding value of the following term is taken for the non-observation location.

[0139] Sub-step 602: Update auxiliary variables, perform soft thresholding on singular values ​​for the nuclear norm term, and perform soft thresholding on the total variation term;

[0140] Furthermore, the specific steps for updating the auxiliary variable are as follows: for the tensor Tensor expansion yields the matrix ,right Singular value decomposition yields Applying the soft threshold shrinkage operator to singular values A new singular value matrix is ​​obtained. Reconstruction yields For the total variation term, calculate the gradient of the tensor. Apply the soft threshold shrinkage operator to each component of the gradient. get Final auxiliary variable Pick and The weighted average.

[0141] Furthermore, the weights of the weighted average are determined based on the relative importance of the two regularization terms, and the weighting formula is as follows: The weighting coefficient , ,satisfy .

[0142] Furthermore, the gradient of the tensor The forward difference method is used for calculation, for the mesh element. Its radial gradient component is The axial gradient component is The total variation norm is defined as .

[0143] Furthermore, the tensor expansion adopts a modulo-1 expansion method to expand the three-dimensional tensor. Expand into a matrix along the first dimension Specifically, this involves concatenating the radial slices of the tensor along the column direction. Tensor reconstruction is the inverse process of expansion, transforming the matrix... Rearranged into tensors arranging the column vectors of the matrix according to The dimensions are divided into There are 3 slices, and each slice contains 3 parameter channels.

[0144] Sub-step 603: Update the dual variable according to the dual increase formula;

[0145] Furthermore, the dual variable update formula is as follows: .

[0146] Sub-step 604: Determine the convergence condition. If both the original residual and the dual residual are less than the set threshold, terminate the iteration; otherwise, return to sub-step 601.

[0147] Furthermore, the set thresholds include the original residual threshold and the dual residual threshold, both of which are set to... The original residual is defined as The dual residual is defined as ,in As an auxiliary variable, This represents the number of iterations.

[0148] After iteration terminates, output the completed 3D parameter tensor. .

[0149] In this embodiment of the application, in order to accelerate the convergence speed of the ADMM algorithm, an adaptive step size strategy can be adopted to dynamically adjust the update step size of the dual variable.

[0150] The radial constraint tensor obtained in step 4 (five-layer radial parameter distribution for each of the four axial segments) and the axial constraint tensor obtained in step 5 (four-segment axial parameter distribution for each of the five radial layers) are used as observation data. The observation location accounts for a percentage of the total number of grid cells. 100% of the units. Set the regularization coefficient. , Penalty parameters The ADMM algorithm was used for iterative solving. After 68 iterations, the original residual and the dual residual were reduced to [values ​​missing]. and The convergence condition is met. The dielectric constant distribution of the defect region (radial layers 3-4 and axial segment 2) in the completed three-dimensional parametric tensor is shown in Table 5.

[0151] Table 5 shows the dielectric constants of the defect region in the completed 3D parameter tensor.

[0152]

[0153] Table 5 shows that the completed tensor forms a high dielectric constant region at the position (3-4 layers, 2nd segment), which is completely consistent with the inversion results in Tables 3 and 4 at the intersection (3.87 for the 2nd segment of the 3rd layer and 4.21 for the 2nd segment of the 4th layer), verifying the effectiveness of the tensor completion algorithm. Simultaneously, tensor completion also reveals the extension of the defect in the axial 3rd segment (2.89 for the 3rd layer and 3.12 for the 4th layer), information that cannot be obtained from unidirectional constraints.

[0154] Step 7: Identify abnormal mesh cells from the completed 3D parameter tensor, extract defect regions using a 3D connectivity analysis algorithm, and generate 3D defect geometry.

[0155] From the completed 3D parameter tensor In the process, grid cells with parameters that significantly deviate from the normal range are identified. For each grid cell... Determine the normalized dielectric constant, conductivity, and porosity of each parameter. If any parameter value is... satisfy If so, the mesh cell is marked as an anomalous cell, where This represents the normalized mean of the insulation parameters. The normalized standard deviation This is the threshold for anomaly detection.

[0156] Furthermore, the average value of the normal insulation parameters and standard deviation Obtained through the following method: from a three-dimensional parameter tensor The parameter values ​​of all grid cells are extracted from each parameter component (dielectric constant, conductivity, porosity). The histogram distribution of each parameter component is calculated, and the dominant peak interval with the largest proportion in the histogram is identified. This dominant peak interval corresponds to the parameter distribution of a normal insulating material. The mean of all parameter values ​​within the dominant peak interval is calculated as the parameter distribution. Standard deviation as .

[0157] Furthermore, the method for identifying the main peak interval of the histogram is as follows: the normalized range of parameter values... The histogram is divided into 50 equally wide histogram intervals. The number of parameter values ​​falling into each interval is counted. The interval with the most parameter values ​​is identified as the main peak center interval. Then, adjacent intervals are expanded to the left and right until the number of parameter values ​​in an existing interval is less than 50% of the number in the main peak center interval. All expanded consecutive intervals are then merged to form the main peak interval. The main peak interval typically corresponds to parameter values ​​at... The normal insulation area within the range.

[0158] Furthermore, the anomaly detection threshold The value is determined based on the balance between the detection sensitivity and false alarm rate of insulation defects, and the range is as follows: ,when At a 95% confidence level, when This corresponds to a 99.7% confidence level. In wind farm cable insulation diagnostics applications, the preferred method is... To ensure a low false alarm rate.

[0159] Using a 3D connectivity analysis algorithm and employing the 26-neighborhood connectivity criterion (including adjacent meshes with face adjacency, edge adjacency, and corner adjacency), spatially connected anomalous cells are aggregated into independent 3D defect regions, generating a set of 3D defect geometries. ,in These represent the 1st, 2nd, ..., Lth independent three-dimensional defect geometries, respectively. The number of independent defects identified.

[0160] Furthermore, the specific steps of the three-dimensional connectivity analysis algorithm are as follows: initialize all abnormal mesh cells to an unvisited state, perform a depth-first search on each unvisited abnormal mesh cell, check the 26 neighboring meshes of the current mesh cell (including adjacent meshes with 6 face adjacencies, 12 edge adjacencies, and 8 corner adjacencies), if a neighboring mesh is also an abnormal cell and has not been visited, add it to the current defect region and mark it as visited, recursively execute the above search process until the neighbors of all mesh cells in the current defect region have been checked, save the defect region as an independent three-dimensional defect geometry, and continue to repeat the above process for the next unvisited abnormal mesh cell until all abnormal mesh cells have been visited.

[0161] From the completed 3D parameter tensor In this process, anomaly detection is performed on the normalized dielectric constant. The normalized dielectric constant of all 20 grid cells is statistically analyzed, and the main peak interval is identified. Calculate the mean of this interval. Standard deviation If an anomaly detection threshold is set, then the detection threshold is... Four grid cells were identified as anomalous cells, and their locations and parameters are shown in Table 6.

[0162] Table 6 identifies the abnormal mesh cells;

[0163]

[0164] Table 6 shows that the deviation of all four abnormal mesh elements exceeded the judgment threshold of 0.18. Among them, mesh elements (3,2) and (4,2) had the largest deviations, reaching 0.28 and 0.36 respectively, corresponding to original dielectric constants of 3.87 and 4.21. Using the 26-neighborhood connectivity criterion for three-dimensional connectivity analysis, the four abnormal mesh elements formed a single connected defect region through face adjacency. The mesh elements (3,2) and (4,2) are adjacent in the radial direction, (3,3) and (4,3) are adjacent in the radial direction, and the two groups are adjacent in the axial direction, forming a continuous three-dimensional defect body.

[0165] Step 8: Calculate the characteristic parameters of the three-dimensional defect geometry, and output the spatial reconstruction results and hazard assessment of the three-dimensional insulation defect.

[0166] For each three-dimensional defect geometry Calculate its characteristic parameters, including:

[0167] volume The defect region is the number of mesh cells multiplied by the volume of a single mesh cell.

[0168] Maximum radial depth : The radial position of the grid cell furthest from the outer surface of the cable in the defect area;

[0169] Axial span : The maximum length of the defect area covered in the axial direction;

[0170] Morphological characteristic parameters include sphericity (the ratio of defect volume to its smallest circumscribed sphere volume) and aspect ratio (the ratio of axial span to radial thickness).

[0171] Furthermore, the volume of a single grid cell is calculated as follows: ,in Radial grid size (based on total insulation thickness) and radial division of layers Sure, ), Axial grid size (based on cable diagnostic section length and number of axial segments) Sure, ), This is the average radius of the radial layer containing the grid cell. For the cable outer radius... and total insulation thickness , No. The average radius of the layer is .

[0172] Furthermore, the radial thickness is defined as the maximum span of the defect region in the radial direction, calculated as the difference between the maximum and minimum radial coordinates in the defect region. The minimum circumscribed sphere is calculated as follows: using the centroid of all grid cells in the defect region as the initial position of the sphere center, the maximum distance from the center point of all grid cells to the sphere center is calculated as the radius. If any grid cell exceeds the minimum circumscribed sphere, the position of the sphere center is adjusted so that all grid cells are included and the radius is minimized.

[0173] Based on the aforementioned characteristic parameters, the spatial reconstruction results of the three-dimensional insulation defects are output, including the three-dimensional spatial coordinates, geometric shape, and parameter anomaly degree of each defect. The penetration risk level of the defect is assessed based on the ratio of the maximum radial depth of the defect to the total insulation thickness, and the hazard assessment result is output.

[0174] Furthermore, the penetration risk level is based on the maximum radial depth ratio. Divided into three levels: when When the risk is low, it means the defect has not penetrated half the thickness of the insulation layer; when The risk level is medium, indicating that the defect has penetrated deep into the insulation layer but has not yet approached the conductor; when... A timeframe indicating high risk means that the defect is close to the conductor and there is a risk of breakdown.

[0175] For the identified defect areas Calculate its characteristic parameters. The volume of a single mesh cell is... ,in , cable outer radius Average radius of the third layer The average radius of the 4th layer The defect region contains four mesh elements, mainly concentrated in layers 3-4 and segments 2-3. The volume of the element in layer 3 is... m 3 The volume of the 4th layer unit is m 3 The calculated characteristic parameters are shown in Table 7.

[0176] Table 7. Characteristic parameters and hazard assessment of defective areas;

[0177]

[0178] Table 7 shows that the maximum radial depth of the defect reached 8 mm, penetrating 80% of the insulation layer thickness, placing it at a high-risk level with a remaining safety margin of only 20%. The defect exhibits an axially elongated strip-like distribution (length-to-diameter ratio 2500), consistent with the typical morphology of water tree aging developing along the electric field direction. The volume reached... m 3This indicates that a large-scale connectivity defect has formed in the deteriorated area. Based on the comprehensive assessment, this defect poses a high risk of penetration and a high risk of breakdown; therefore, it is recommended to immediately arrange for a shutdown, maintenance, and replacement of this cable section.

[0179] In addition to step 8, the following steps are also included:

[0180] Step 9: Based on the three-dimensional defect characteristic parameters, calculate the remaining safety margin of insulation and output maintenance priority ranking suggestions.

[0181] The remaining safety margin of insulation is calculated based on the maximum radial depth, volume, and morphological characteristics of each three-dimensional defect geometry. ,in This represents the total insulation thickness. Based on the remaining safety margin from smallest to largest, the defects are prioritized for repair, and repair recommendations are output.

[0182] According to the embodiments of this implementation, by introducing a tensor completion method with bidirectional constraints of electrothermal coupling gradient and simulation inversion, the limitation of existing technologies that cannot obtain complete three-dimensional information by unidirectional measurement or unidirectional analysis is overcome.

[0183] Specifically, in step 1, the radial time delay gradient vector is calculated by utilizing the propagation difference of the electrothermal response in the direction of insulation depth. Since the dielectric constant and thermal conductivity of the insulating material determine the propagation speed of the electric and thermal fields, when there is a defect in a certain radial depth layer, the electrothermal coupling delay of that layer will change. Therefore, the radial time delay gradient vector can reflect the defect location information in the radial direction.

[0184] In steps 2 to 5, the sensitivity of internal parameters to electrical measurements from multiple measuring points arranged along the axial direction is used to perform axial simulation inversion. Since the measured values ​​at different axial positions are affected by the internal parameters of that position and its neighboring areas, the parameter distribution of each axial segment can be separated through inversion using a multiphysics coupling simulator, providing constraint information in the axial direction.

[0185] In step 6, radial and axial constraints provide slice information in two orthogonal directions. The sparse regularized tensor completion algorithm utilizes the sparsity prior of small defect proportion through nuclear norm constraint and the spatial continuity prior of smooth change of adjacent mesh parameters through total variational regularization constraint. The constraint information in the two orthogonal directions mutually verify and complement each other under the tensor completion framework, reconstructing the complete three-dimensional distribution from sparse observations, thereby realizing the three-dimensional spatial accurate location and geometric reconstruction of internal insulation defects.

[0186] Therefore, this embodiment solves the technical deficiency of the prior art that it cannot accurately identify the spatial location, geometric shape and connectivity characteristics of internal insulation defects, and can assess the true degree of harm and penetration risk of defects, providing a basis for accurate maintenance and remaining life prediction.

Claims

1. A high-precision fault diagnosis method for the insulation condition of wind farm cables, characterized in that, Includes the following steps: The temperature and capacitance time series of multiple measuring points along the radial direction of the cable surface are obtained. The time delay difference between the electrical and thermal responses at each radial position is calculated using a cross-correlation algorithm to generate a radial time delay distribution curve. The difference in time delay values ​​between adjacent measuring points is calculated to generate a radial time delay gradient vector. The external electrical measurement values ​​of multiple measuring points along the cable axis are acquired. The external electrical measurement values ​​include local capacitance, local dielectric loss and partial discharge amplitude. An axial spatial distribution measurement vector is generated. The cable insulation space is divided into multiple layers radially and multiple segments axially to construct a three-dimensional mesh discretization space. Internal parameter variables are set for each mesh unit to generate a three-dimensional parameter tensor. Based on the multiphysics coupling simulator, the radial time delay gradient vector is used as a constraint condition to invert the radial parameter distribution of each axial segment and generate a radial constraint tensor. Based on the multiphysics coupling simulator, the axial spatial distribution measurement vector is used as a constraint condition to invert the axial parameter distribution of each radial depth layer and generate the axial constraint tensor. Using a sparse regularized tensor completion algorithm, the slice data of the radial constraint tensor and the axial constraint tensor are used as constraints. Combined with the prior knowledge of defect sparsity and spatial continuity constraints, the complete three-dimensional parameter tensor is completed iteratively by the alternating direction multiplier method. Abnormal mesh cells with parameters deviating from the normal range are identified from the completed 3D parameter tensor. Abnormal regions with spatial connectivity are extracted using a 3D connectivity analysis algorithm to generate a 3D defect geometry. Calculate the characteristic parameters of the three-dimensional defect geometry, and output the spatial reconstruction result and hazard assessment of the three-dimensional insulation defect.

2. The method according to claim 1, characterized in that, The calculation of the time delay difference between the electrical and thermal responses at each radial position using a cross-correlation algorithm includes: Cross-correlation calculations are performed on the temperature time series and capacitance time series of each radial measuring point, and the time offset corresponding to the peak value of the cross-correlation function is taken as the time delay value of that measuring point. The cross-correlation operation is performed in a normalized form, which divides the cross-correlation function by the root mean square product of the two signals to eliminate the influence of signal amplitude differences on time delay estimation.

3. The method according to claim 1, characterized in that, The internal parameter variables include dielectric constant, conductivity, and porosity.

4. The method according to claim 1, characterized in that, The multiphysics coupled simulator is built based on the heat transfer diffusion equation and the electric field propagation equation, wherein: The heat transfer and diffusion equation describes the relationship between material density, specific heat capacity, thermal conductivity and the spatiotemporal distribution of temperature. The electric field propagation equation describes the relationship between the dielectric constant, potential, and free charge density.

5. The method according to claim 1, characterized in that, The inversion of the radial parameter distribution of each axial segment includes: Using the radial time delay gradient vector as a constraint, the theoretical time delay gradient corresponding to different radial parameter distributions is calculated in the forward direction by the simulator. The least squares optimization method is used to find the radial parameter distribution that minimizes the error between the theoretical and measured time delay gradients. In least squares optimization, a Tikhonov regularization term is introduced to constrain the parameter changes of adjacent radial layers to be smooth.

6. The method according to claim 1, characterized in that, The inverted axial parameter distribution of each radial depth layer includes: Using the axial spatial distribution measurement vector as a constraint, the theoretical measurement values ​​corresponding to different axial parameter distributions are calculated in the forward direction by the simulator. The regularized gradient descent algorithm is used for iterative optimization. In each iteration, the parameter values ​​are updated based on the gradient of the loss function and the gradient of the regularization term between the measured and simulated values ​​until convergence.

7. The method according to claim 1, characterized in that, The optimization objectives of the sparse regularized tensor completion algorithm include: The observation data fitting term ensures that the completed tensor matches the measured data at the observation location. The nuclear norm constraint term utilizes the sparsity prior with a small defect ratio to achieve low-rank constraint. The total variation regularization term utilizes the spatial continuity constraint of smooth changes in parameters between adjacent grids.

8. The method according to claim 1, characterized in that, The method of extracting spatial connectivity anomalies using a three-dimensional connectivity analysis algorithm includes: For each grid cell, if the deviation of its parameter value from the average value of normal insulation parameters exceeds the product of a set threshold and the standard deviation, the grid cell is marked as an abnormal cell. Using the 26-neighborhood connectivity criterion, adjacent anomalous mesh cells that are face-adjacent, edge-adjacent, and corner-adjacent are aggregated into independent three-dimensional defect regions.

9. The method according to claim 1, characterized in that, The feature parameters include: Volume is the number of mesh cells contained in the defect region multiplied by the volume of a single mesh cell; Maximum radial depth is the radial position of the grid cell furthest from the outer surface of the cable in the defect region; Axial span is the length of the defect area covered in the axial direction. Morphological characteristic parameters include sphericity, which is the ratio of the defect volume to the volume of its smallest circumscribed sphere, and aspect ratio, which is the ratio of the axial span to the radial thickness. The hazard assessment is based on the ratio of the maximum radial depth of the defect to the total insulation thickness, which evaluates the penetration risk level of the defect.

10. A high-precision fault diagnosis system for the insulation condition of wind farm cables, used to execute the method according to any one of claims 1-9, characterized in that, include: The radial data acquisition module is used to acquire the temperature time series and capacitance time series of multiple measuring points along the radial direction of the cable surface, calculate the time delay difference using a cross-correlation algorithm, and generate a radial time delay gradient vector. The axial data acquisition module is used to acquire external electrical measurement values ​​of multiple measuring points along the axial direction of the cable and generate an axial spatial distribution measurement vector. The 3D mesh construction module is used to discretize the cable insulation space into a 3D mesh, set internal parameter variables for each mesh cell, and generate a 3D parameter tensor. The radial inversion module is used to invert the radial parameter distribution of each axial segment based on a multiphysics coupled simulator and using radial time delay gradient vector constraints to generate a radial constraint tensor. The axial inversion module is used to invert the axial parameter distribution of each radial depth layer based on a multiphysics coupled simulator and by utilizing the axial spatial distribution measurement vector constraint to generate an axial constraint tensor. The tensor completion module is used to complete the complete three-dimensional parameter tensor using a sparse regularized tensor completion algorithm, with radial and axial constraint tensors as constraints, and iteratively completes the complete three-dimensional parameter tensor through the alternating direction multiplier method. The defect identification module is used to identify abnormal mesh cells from the completed 3D parameter tensor, extract spatially connected abnormal regions using a 3D connectivity analysis algorithm, and generate 3D defect geometry. The evaluation output module is used to calculate the characteristic parameters of the three-dimensional defect geometry and output the spatial reconstruction results and hazard assessment of the three-dimensional insulation defect.

Images