Methods and devices for diagnosing converter transformer faults

Abstract

This invention discloses a method and apparatus for fault diagnosis of converter transformers based on sensor location optimization and physical graph neural networks. The method includes: synchronously acquiring multi-channel vibration signals and corresponding voltage and current signals of the converter transformer; obtaining the multi-channel vibration signals through a sensor network arranged at an optimal position on the surface of the converter transformer tank; obtaining the optimal position through an intelligent optimization algorithm; performing frequency domain transformation on the multi-channel vibration signals and corresponding voltage and current signals to obtain the feature matrix of the converter transformer; inputting the feature matrix into the first and second channels of a trained dual-channel physical graph neural network to obtain a predicted vibration amplitude matrix and an adjacency matrix; determining the fault type and locating the physical region of the fault through the adjacency matrix, accurately decoupling and locating the physical region of mechanical loosening faults in the core and windings of the converter transformer, and improving the robustness of the system under complex and variable operating conditions.

Description

Technical Field

[0001] This invention relates to the field of power equipment fault diagnosis technology, and in particular to a method and device for diagnosing converter transformer faults based on sensor location optimization and physical graph neural network. Background Technology

[0002] As a core component of high-voltage direct current (HVDC) transmission systems, converter transformers operate in a complex environment, enduring long-term impacts from superimposed AC and DC voltages and high-order harmonic currents. Mechanical abnormalities (such as insufficient winding clamping force or core loosening) are among the main failure modes of converter transformers. Non-intrusive condition monitoring using vibration signals is currently an important method for assessing the mechanical health of converter transformers.

[0003] However, existing technologies for monitoring and assessing the vibration of converter transformers still have significant limitations in terms of measurement point layout and diagnostic algorithms.

[0004] (1) Regarding the arrangement of sensor measurement points, existing technologies rely on manual experience (such as the four corners of the fuel tank) or uniform distribution. Due to the complexity of vibration modes, sensors are easily placed in the "vibration dead zone" (missed detection) or the "information redundancy zone" (data duplication), resulting in wasted costs and monitoring blind spots. Even if standard intelligent optimization algorithms are used, they are prone to getting stuck in local optima and cannot find the globally optimal measurement point that balances high energy and independence.

[0005] (2) In terms of fault diagnosis, there are three defects: First, it is impossible to eliminate the interference of operating conditions, which easily leads to a high false alarm rate; second, it is difficult to distinguish between core and winding faults, and the fault sources are mixed; third, it can only output fault labels, lacks physical positioning capability, and cannot guide accurate maintenance. Summary of the Invention

[0006] This invention provides a method and apparatus for fault diagnosis of converter transformers based on sensor location optimization and physical graph neural networks. It addresses the shortcomings of existing technologies, such as sensor placement relying on experience leading to monitoring blind spots and redundancy, high false alarm rates caused by vibration amplitude fluctuations with load under varying operating conditions, difficulty in distinguishing fault types, and difficulty in fault location. It achieves accurate decoupling and physical location of mechanical loosening faults in the core and windings of converter transformers.

[0007] This invention provides a converter transformer fault diagnosis method based on sensor location optimization and physical graph neural network, comprising: The system synchronously acquires multi-channel vibration signals from the converter transformer, along with corresponding voltage and current signals. These multi-channel vibration signals are obtained through a sensor network strategically positioned on the surface of the converter transformer's tank. The optimal positions are determined using an intelligent optimization algorithm. The multi-channel vibration signal and the corresponding voltage and current signals are subjected to frequency domain transformation, and feature separation is performed by frequency masking to obtain the first feature matrix and the second feature matrix of the converter transformer. The first feature matrix and the second feature matrix are respectively input into the first and second channels of the trained dual-channel physical graph neural network to obtain the first predicted vibration amplitude matrix, the first adjacency matrix, the second predicted vibration amplitude matrix, and the second adjacency matrix; the first adjacency matrix and the second adjacency matrix are used to characterize the mechanical transmission relationship between the nodes of the sensor network; Based on the first predicted vibration amplitude matrix, the first adjacency matrix, the second predicted vibration amplitude matrix, and the second adjacency matrix, the fault type and the physical area of ​​the fault are determined.

[0008] According to the present invention, a converter transformer fault diagnosis method based on sensor location optimization and physical graph neural network is provided, wherein the intelligent optimization algorithm includes: a multi-strategy improved sparrow search algorithm; Correspondingly, the optimal position is obtained through an intelligent optimization algorithm, specifically including: Obtain the modal kinetic energy distribution on the surface of the converter transformer tank; Based on the modal kinetic energy distribution, a weighted roulette wheel betting strategy is used to generate the initial population of the sensor network deployment scheme; Based on the composite fitness value of each placement scheme in the current population, the placement scheme is updated to obtain a new generation of population; for the globally optimal placement scheme of the new generation of population, a local lens strategy based on modal contribution is used for local update; the iteration continues until the maximum number of iterations is reached or the composite fitness value no longer decreases, and the placement scheme with the smallest fitness value is output, which is the globally optimal location set of the sensor network.

[0009] According to the present invention, a converter transformer fault diagnosis method based on sensor location optimization and physical graph neural network is provided. The method for locally updating the globally optimal placement scheme of the new generation population using a local lensing strategy based on modal contribution specifically includes: Based on the modal contribution, identify the weakest sensor network node; Local lens variations are applied to the nodes of the short-board sensor network to generate candidate placement schemes; If the composite fitness value of the candidate placement scheme is better than that of the original globally optimal placement scheme, then the globally optimal placement scheme is updated.

[0010] According to the present invention, a converter transformer fault diagnosis method based on sensor location optimization and physical graph neural network is provided, wherein the composite fitness function Fit(X) is defined as: ; Where X represents the deployment scheme of a sensor network containing m nodes; modal matrix elements The i-th node represents the mode amplitude in the j-th mode; w1 and w2 represent the weighting coefficients. From the mode matrix The extracted first Modal row vectors corresponding to the positions of each measurement point; From the mode matrix The extracted first Modal row vectors corresponding to the positions of each measurement point; ) represents the maximum modal confidence criterion among all node pairs in the node placement scheme X.

[0011] According to the present invention, a converter transformer fault diagnosis method based on sensor location optimization and physical graph neural network is provided, wherein the modal contribution degree Characterized by the following formula: ; in, For the first The total kinetic energy of a node in all modes; This represents the maximum kinetic energy value among all nodes in the sensor network. This represents the maximum similarity between the i-th node and all nodes except itself. This represents the weighting coefficient.

[0012] According to the present invention, a converter transformer fault diagnosis method based on sensor position optimization and physical graph neural network is provided. The dual-channel physical graph neural network model includes: a physical-guided self-attention layer and a graph convolutional regression prediction layer. Correspondingly, the first feature matrix and the second feature matrix are respectively input into the first and second channels of the trained dual-channel physical graph neural network to obtain the first predicted vibration amplitude matrix, the first adjacency matrix, the second predicted vibration amplitude matrix, and the second adjacency matrix, specifically including: The first feature matrix and the second feature matrix are respectively input into the self-attention layer of the first channel and the second channel, and the dynamic mechanical transmission relationship between the sensor network nodes of the first channel and the second channel is calculated to obtain the first adjacency matrix and the second adjacency matrix. The first feature matrix and the first adjacency matrix are input into the physical graph convolutional regression prediction layer of the first channel to obtain the first predicted vibration amplitude matrix of the sensor network; the second feature matrix and the second adjacency matrix are input into the physical graph convolutional regression prediction layer of the second channel to obtain the second predicted vibration amplitude matrix of the sensor network.

[0013] According to the present invention, a method for diagnosing converter transformer faults based on sensor location optimization and physical graph neural networks, the method for calculating the dynamic mechanical transmission relationship between sensor network nodes further includes: ; ; in, Representative sensor node To the node Mechanical transmission weight; The attention coefficient represents the sensor's... For sensors The initial level of attention; Represents the feature vectors of the i-th and j-th nodes; Represents the attention weight vector; The weight matrix is ​​a learnable matrix; This represents a vector concatenation operation; r is the loop variable; m represents the number of nodes in the sensor network.

[0014] According to the present invention, a converter transformer fault diagnosis method based on sensor location optimization and physical graph neural network, wherein determining the fault type and locating the physical area of ​​the fault based on the first predicted vibration amplitude matrix, the first adjacency matrix, the second predicted vibration amplitude matrix, and the second adjacency matrix specifically includes: Construct a structural consistency loss function, which includes: prediction error and structural difference; Based on the first predicted vibration amplitude matrix and the first adjacency matrix, the first loss value is calculated using the structural consistency loss function; Based on the second predicted vibration amplitude matrix and the second adjacency matrix, the second loss value is calculated using the structural consistency loss function; The first loss value is compared with a set first threshold to determine the core state; The second loss value is compared with a set second threshold to determine the winding state.

[0015] According to the present invention, a converter transformer fault diagnosis method based on sensor location optimization and physical graph neural network is provided, wherein locating the physical area where the fault occurs specifically includes: Once the core state and / or winding state are confirmed to be faulty, their respective difference matrices are calculated. The difference matrices are used to characterize the difference between the adjacency matrix and the reference adjacency matrix, which is determined after the model training is completed. Find the element with the largest value in the difference matrix, and determine the area between the sensor nodes represented by the element where the fault point is located.

[0016] This invention also provides a converter transformer fault diagnosis method and apparatus based on sensor location optimization and physical graph neural network, comprising: The acquisition module is used to synchronously acquire multi-channel vibration signals of the converter transformer, as well as the corresponding voltage and current signals; the multi-channel vibration signals are obtained through a sensor network arranged at an optimal position on the surface of the converter transformer tank; the optimal position is obtained through an intelligent optimization algorithm. The feature separation module is used to perform frequency domain transformation on the multi-channel vibration signal and the corresponding voltage and current signals, and to perform feature separation through frequency masking to obtain the first feature matrix and the second feature matrix of the converter transformer. The prediction module is used to input the first feature matrix and the second feature matrix into the first channel and the second channel of the trained dual-channel physical graph neural network, respectively, to obtain a first predicted vibration amplitude matrix, a first adjacency matrix, a second predicted vibration amplitude matrix, and a second adjacency matrix; the first adjacency matrix and the second adjacency matrix are used to characterize the mechanical transmission relationship between the nodes of the sensor network; The fault diagnosis module is used to determine the fault type and locate the physical area of ​​the fault based on the first predicted vibration amplitude matrix, the first adjacency matrix, the second predicted vibration amplitude matrix, and the second adjacency matrix.

[0017] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the converter transformer fault diagnosis method based on sensor position optimization and physical graph neural network as described above.

[0018] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the converter transformer fault diagnosis method based on sensor position optimization and physical graph neural network as described above.

[0019] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the converter transformer fault diagnosis method based on sensor position optimization and physical graph neural network as described above.

[0020] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention optimizes the sensor measurement point location through intelligent optimization algorithm to ensure that the selected measurement point can capture the maximum modal vibration energy and ensure the independence of the signal of each measurement point, thereby realizing high signal-to-noise ratio perception of the vibration state of the entire surface of the converter transformer and greatly reducing the hardware monitoring cost; through frequency masking mechanism and dual-channel parallel diagnostic architecture, the mixed vibration signal is accurately separated in the frequency domain, so that the system can clearly point out whether the abnormal vibration is caused by the abnormal magnetic circuit of the iron core or the insufficient winding clamping force, thus solving the problem of unclear fault source attribute identification; (2) This invention constructs a structural consistency loss function through a graph neural network, and evaluates the equipment status by inverting the mechanical transfer matrix (adjacency matrix) between sensor nodes in real time, thereby reducing false alarms. This diagnostic logic based on physical mechanisms significantly improves the robustness of the system under complex and variable working conditions.

[0021] (3) By using the interpretability of graph neural networks, this invention can calculate the difference between the current dynamic adjacency matrix and the health baseline matrix, locate the physical area where mechanical stiffness changes abruptly, and directly point out to maintenance personnel the fault location "between which two sensors", thus narrowing the scope of maintenance from the entire equipment to a local area, significantly improving maintenance efficiency and reducing maintenance costs. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 This is one of the flowcharts of a converter transformer fault diagnosis method based on sensor position optimization and physical graph neural network according to an embodiment of the present invention.

[0024] Figure 2 This is the second flowchart of the converter transformer fault diagnosis method based on sensor position optimization and physical graph neural network according to an embodiment of the present invention.

[0025] Figure 3 This is the third flowchart of the converter transformer fault diagnosis method based on sensor position optimization and physical graph neural network according to an embodiment of the present invention.

[0026] Figure 4 This is the fourth flowchart of the converter transformer fault diagnosis method based on sensor position optimization and physical graph neural network according to an embodiment of the present invention.

[0027] Figure 5This is a schematic diagram of the structure of a converter transformer fault diagnosis device based on sensor position optimization and physical graph neural network according to an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0029] As a core component of high-voltage direct current (HVDC) transmission systems, converter transformers operate in a complex environment, enduring long-term impacts from superimposed AC and DC voltages and high-order harmonic currents. Mechanical abnormalities (such as insufficient winding clamping force or core loosening) are among the main failure modes of converter transformers. Non-intrusive condition monitoring using vibration signals is currently an important method for assessing the mechanical health of converter transformers. However, existing vibration monitoring and assessment technologies for converter transformers still have significant limitations in terms of measurement point placement and diagnostic algorithms.

[0030] Regarding sensor placement, current technologies largely rely on manual experience (such as fixing them at the four corners of the tank) or simple geometric uniform distribution. Due to the complex internal structure of converter transformers, the vibration modes on the tank surface exhibit high non-uniformity, with numerous "vibration dead zones" (nodes) and "information redundancy zones." Experience-based placement easily leads to sensors being installed in areas with weak vibration responses, missing crucial modal information; or multiple sensors being placed in areas with strong modal correlation, resulting in redundant and repetitive data collection and wasted hardware costs. While some studies have attempted to introduce swarm intelligence optimization algorithms, such as the standard Sparrow Search Algorithm (SSA) and particle swarm optimization, for optimal placement, standard algorithms are prone to getting trapped in local optima when dealing with complex multimodal functions, making it difficult to find the globally optimal observation location and thus hindering the balance between maximizing kinetic energy and modal independence.

[0031] In terms of fault diagnosis algorithms, existing technologies mainly employ statistical threshold-based discrimination methods or traditional "black box" deep learning models (such as Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM)). These methods have several problems in practical applications. First, they cannot eliminate operating condition interference. The vibration amplitude of a converter transformer naturally fluctuates with the load (current, voltage). When the load changes drastically, the vibration amplitude increases accordingly, making it easy for traditional methods to misjudge it as a mechanical fault (false alarm). Second, fault sources are often unclear, with core vibration and winding vibration overlapping in the signal. Existing methods lack effective physical feature separation mechanisms, making it difficult to accurately distinguish whether abnormal vibration originates from the core or the winding. Third, they lack physical location capabilities. Traditional neural networks can only output a "fault" label and cannot infer changes in the mechanical transmission path within the equipment, nor can they point out the specific physical fault area to maintenance personnel (such as which two measuring points have a loose connection structure), resulting in a lack of clear direction in maintenance work and low maintenance efficiency.

[0032] Based on this, such as Figure 1 As shown, this invention provides a converter transformer fault diagnosis method based on sensor location optimization and physical graph neural network, comprising the following steps: Step 100: Synchronously acquire multi-channel vibration signals of the converter transformer and corresponding voltage and current signals; the multi-channel vibration signals are obtained through a sensor network arranged at the optimal position on the surface of the converter transformer tank; the optimal position is obtained through an intelligent optimization algorithm.

[0033] Specifically, a sensor network refers to a network consisting of one or more vibration sensors deployed on the surface of the converter transformer tank. Voltage and current signals are also acquired synchronously with the multi-channel vibration signals within the same time window / operating condition. For example, the vibration sensor output signal, voltage transformer (PT) signal, and current transformer (CT) signal can be connected to the same data acquisition card (DAQ). The sampling rate is set as follows: Obtain vector sequence , where m represents the number of vibration sensors in the sensor network. Only when there is only one clock source will the time corresponding to the a-th sampling point be completely consistent.

[0034] The intelligent optimization algorithm can be a genetic algorithm, a particle swarm optimization algorithm, or a sparrow search algorithm. This embodiment of the invention does not limit the specific algorithms. It should be noted that due to the complexity of the modal distribution of the converter transformer tank, directly using unmodified conventional genetic algorithms or particle swarm optimization algorithms often results in slow convergence speed and easy getting trapped in local optima.

[0035] Optionally, the intelligent optimization algorithm in this embodiment of the invention employs the Multi-strategy improved sparrow search algorithm (MS-SSA). Correspondingly, the optimal position is obtained through the intelligent optimization algorithm, and the method includes the following steps: Step 1001: Obtain the modal kinetic energy distribution on the surface of the converter transformer tank.

[0036] Specifically, traditional vibration measurement point placement often relies on manual experience (such as fixing them at the four corners of the tank) or simple uniform distribution. Converter transformers exhibit complex vibration modes, with numerous "vibration dead zones" (nodes) and "information redundancy zones." Experience-based point placement often misses key modes or results in multiple sensors measuring duplicate information, leading to wasted hardware costs and insufficient effective data. The Standard Sparrow Algorithm (SSA) is prone to getting trapped in local optima, failing to find the truly optimal location. From thousands of potential locations on the tank surface, k (e.g., k=8 in this embodiment) "golden measurement points" are calculated to capture maximum vibration energy while ensuring the independence (non-repetition) of signals from each measurement point. This is achieved by constructing a composite objective function that includes a Modal Confidence Criterion (MAC). From thousands of locations on the converter transformer tank surface, the algorithm automatically searches for 8 "golden measurement points."

[0037] Modal characteristics refer to the inherent vibration characteristics of a transformer system. The location of a sensor can be determined by calculating the modal characteristics of the transformer system. According to the principle of mechanical vibration, the mode shape of a system is entirely determined by its mass distribution and stiffness distribution. Therefore, a complete model including the core, windings, and insulating oil must be established to account for the additional mass effect and fluid-structure interaction effect generated by the internal entities on the tank wall through the fluid medium, thereby accurately identifying the vibration-sensitive areas (antinodes) on the tank surface.

[0038] A 3D model of the converter transformer tank was created using finite element simulation software (such as ANSYS or COMSOL) for modal analysis. Simplified 3D geometries of the tank, core, windings, and insulating oil were established based on the actual transformer's geometry, spatial location, and material parameters. The contact surfaces between the outer surfaces of the core, windings, and tank, and the insulating oil were defined as "fluid-structure interaction interfaces." This ensures that the vibrational energy of the internal entities is accurately transmitted to the tank wall through the oil medium, reproducing the real physical vibration environment. Although the model is not under stress at this stage, it already possesses a realistic mass and stiffness distribution.

[0039] Mesh all the above components. Extract the mesh using the Block Lanczos method. First-order intrinsic modes.

[0040] After the solution is completed, only the data of M grid nodes on the outer surface of the fuel tank that can be fitted with sensors are filtered and exported. The total number of grid nodes extracted from the surface of the fuel tank is denoted as set M. Generate the mode matrix.

[0041] In this embodiment Before extraction The modal shapes at natural frequencies (typically covering the 0-2000Hz range) are used to obtain the modal matrix for calculating the modal kinetic energy distribution in embodiments of the present invention. This is The matrix. Indicates the first The number of physical sensor nodes in the first Normalized displacement amplitude (mode amplitude) under first mode.

[0042] After obtaining the mode matrix Then, the modal kinetic energy distribution on the surface of the converter transformer tank can be obtained.

[0043] Step 1002: Based on the modal kinetic energy distribution, an initial population of the sensor network deployment scheme is generated using a weighted roulette wheel betting strategy; Specifically, it can be understood that traditional SSA (Simultaneous Dynamic Algorithm Analysis) involves uniformly and randomly scattering points within the solution space, neglecting the highly non-uniform distribution of vibration modes on the surface of the converter transformer tank. This results in a large number of initial individuals falling into ineffective regions with weak vibration energy. This invention utilizes modal kinetic energy data obtained from finite element simulation to construct a probabilistic guidance mechanism. Through a weighted roulette wheel, it forces the initial population to preferentially distribute in the "antinode" regions where vibration energy is concentrated. This initialization method, combined with prior physical knowledge, enables the algorithm to possess a high-quality solution space distribution in the early stages of iteration. This overcomes the blindness and inefficiency of traditional algorithm initialization, significantly improves search efficiency, and drastically shortens the time required for the algorithm to converge to the optimal solution.

[0044] (1) Constructing probability intervals 1) Calculate nodal potential energy For the solution space Each node in ( ), calculate its combined modal kinetic energy: (1) 2) Calculate the probability of a single point (Length of each interval): First, normalize the energy and calculate the length of the interval. The probability weights of each node being selected: ; (2) in, This represents the total kinetic energy of all 3000 nodes. For the first The proportion of each node on the roulette wheel, all They add up to 1.

[0045] 3) Calculate the cumulative probability (Interval boundary scale): To perform roulette wheel play, we need to calculate the corresponding interval boundary of each node on the (0,1] number line: (3) in, For summation, the counting variable represents the sum from the first node to the second node. One node; Indicates the first The end scale of the interval for each node.

[0046] 4) The interval for the first node is calculated as follows: The interval of the second node is ;...No. The interval of each node is .

[0047] (2) Perform rejection-based roulette sampling Build size is The initial population matrix. For the t-th sparrow ( The positions of its eight sensors are determined using the following loop logic: 1) Generate a random pointer: Generate a random decimal number that follows a uniform distribution. The value range is (0,1], and the cumulative probability sequence is calculated by iterating through the sequence. Find the node number that uniquely satisfies the following inequality. , (4) in, This represents the right boundary of the target node; The left boundary of the target node (when hour, This formula represents a random decimal. It just happened to fall into the category of the first The node is selected because it falls within the probability range of node number 1.

[0048] 2) Individual deduplication Let the current individual The selected set of nodes is If the calculated If the sampling is invalid, a new random number will be generated. Repeat the above sampling steps until... Add it Repeat this process until the individual... It contains 8 distinct nodes.

[0049] 3) Population diversity verification In generating the first individual Then, combine it with the already generated set of individuals. Perform a comparison. If any... Make and If the node combinations are exactly the same (i.e., the sets are equal), then discard the current set. Then, repeat the above generation process. Finally, the initial population locations (spotting scheme) are obtained.

[0050] Step 1003: Based on the composite fitness value of each placement scheme in the current population, update the placement scheme to obtain a new generation population; for the globally optimal placement scheme of the new generation population, use a local lens strategy based on modal contribution to perform local updates; iterate until the maximum number of iterations is reached or the composite fitness value no longer decreases, and output the placement scheme with the smallest fitness value, which is the globally optimal location set of the sensor network.

[0051] Specifically, a "scoring standard" is provided for the sparrow search algorithm. A formula is needed to evaluate the quality of a set of node placement schemes. Optionally, when the number of nodes in the sensor network is 8, the composite fitness function is constructed as follows: (5) in, X represents the deployment scheme of a sensor network containing 8 nodes (the scheme to be evaluated), including the index numbers of the 8 nodes [n1, n2, ..., n8]; modal matrix elements. The i-th node represents the mode amplitude in the j-th mode; w1 and w2 represent the weighting coefficients. From the mode matrix The extracted first Modal row vectors corresponding to the positions of each measurement point; From the mode matrix The extracted first Modal row vectors corresponding to the positions of each measurement point; ) represents the maximum modal confidence criterion among all node pairs in the node placement scheme X.

[0052] The first term is the energy term, and the denominator is... This represents the total vibrational energy of these 8 points across all modes. A larger denominator results in a smaller reciprocal (better fitness). The guided algorithm searches for regions of intense vibration.

[0053] The second item is redundant. It is the first The modal row vectors corresponding to each measurement point (from) Extracted from ( ). Then it is the first The modal row vectors corresponding to each measurement point (from) Extracted from ( ). This represents the maximum similarity among pairwise combinations of these 8 points. This is key to overcoming the shortcomings of traditional methods. If two points are too close, their waveforms are similar, approaching 1, leading to... Increase (deteriorate). This forced algorithm pushes the sensors apart, making even the most similar pair of 8 points as dissimilar as possible, to avoid duplicate measurements by the two sensors. This represents the weighting coefficient; in this embodiment, we take... .

[0054] Optionally, the globally optimal placement scheme of the new generation population is locally updated using a local lensing strategy based on modal contribution, including: Based on the modal contribution, identify the weakest sensor network node; Local lens variations are applied to the nodes of the short-board sensor network to generate candidate placement schemes; If the composite fitness value of the candidate placement scheme is better than that of the original globally optimal placement scheme, then the globally optimal placement scheme is updated.

[0055] Specifically, it can be understood that traditional SSA mainly relies on random walk or follow strategies, lacking precise improvement methods for inferior dimensions, making it difficult to break local extreme value deadlocks. This invention, by introducing a modal contribution evaluation index, can accurately identify several "weak" nodes (e.g., one or more nodes with the lowest contribution) in the current deployment scheme with low energy contribution or high modal redundancy, and then use the lens imaging principle to perform deterministic back-projection mutation on them. This strategy of "preserving excellent backbones and directionally correcting weaknesses" avoids the blindness of traditional random mutation, provides the deterministic large-span jump capability required to break local convergence deadlocks, solves the problem that traditional algorithms are prone to getting trapped in local optima and lack directional optimization capabilities, and ensures that the final selected sensor layout achieves the best balance between maximizing vibration energy and modal information mutual exclusion on a global scale.

[0056] Before the end of each iteration, the current global optimal solution is... (Including 8 sensor node numbers) Perform the following operations: (1) Evaluation of modal contributions of all staff 1. Input the node numbers of the current optimal solution. ; 2. Calculate the overall contribution score for each sensor: ; (6) in For the first The total kinetic energy of a node in all modes (look up the mode matrix) get); It is the largest kinetic energy value among the current 8 nodes; This represents the maximum similarity between the i-th node and the other 7 nodes; This represents the weighting coefficient; in this example, we take... .

[0057] After calculation, eight scores were obtained. The lower the score, the lower the quality of the sensor, indicating that it has "low energy and overlaps with others," classifying it as a poor-quality node.

[0058] 3. The 8 sensors Sort the results from smallest to largest. Select the lowest-scoring candidate, LZ (in this example, LZ=2), and denote it as... and These two points are the "weak points" in the current deployment plan and need to be replaced.

[0059] (2) Perform the collaborative lens back projection to simultaneously perform the reverse jump on these two "short board" nodes.

[0060] (7) ; (8) ; (9) In this embodiment of the invention, , indicating the numbering boundary of the grid node; This represents the dynamic scaling factor; t is the current iteration number; The initial maximum number of iterations; If the calculated number is less than 1 or greater than 3000, use the modulo operation to map it back to the valid range and check. and Check if it is a duplicate of the remaining 6 good nodes; if so, reset it randomly.

[0061] (3) Greedy update Keep the 6 high-scoring nodes unchanged, and use the new ones. Replace the old, substandard nodes to form a new solution. Calculate the fitness of the new scheme

[0062] like Then accept the mutation and update the global optimal solution.

[0063] like Then, the mutation is rejected, and the original scheme remains unchanged.

[0064] (4) Continue until the maximum number of iterations is reached (e.g., 100). Output the group with the smallest fitness function value. (Includes 8 coordinate points).

[0065] (5) Discretization correction The coordinates calculated by the algorithm may be decimals. By minimizing the Euclidean distance, they are corrected to the number of the nearest grid node.

[0066] The above steps utilize modal kinetic energy data obtained from finite element simulation to construct a probabilistic guidance mechanism. By using a weighted roulette wheel to force the initial population to preferentially distribute in the "antinode" region where vibration energy is concentrated, the blindness and inefficiency of traditional algorithm initialization are overcome, significantly improving search efficiency and greatly shortening the time required for the algorithm to converge to the optimal solution. By introducing a modal contribution evaluation index, several "short-board" nodes with low energy contribution or high modal redundancy in the current deployment scheme can be accurately identified. The deterministic back-projection mutation is then performed on these nodes using the lens imaging principle, solving the problem that traditional algorithms are prone to getting trapped in local optima and lack directional optimization capabilities. This ensures that the final selected sensor layout achieves the best balance between maximizing vibration energy and modal information mutual exclusion in the global scope, realizing high signal-to-noise ratio perception of the vibration state of the entire surface of the converter transformer and significantly reducing hardware monitoring costs.

[0067] Step 101: Perform frequency domain transformation on the multi-channel vibration signal and the corresponding voltage and current signals, and perform feature separation through frequency masking to obtain the first feature matrix and the second feature matrix of the converter transformer; Specifically, this is achieved through the physical mechanism that "core vibration is related to voltage and its energy is concentrated at low frequencies, while winding vibration is related to current and its energy is concentrated at mid-to-high frequencies." In this embodiment of the invention, low frequency refers to... The range, mid-to-high frequency refers to the range of 600Hz to 2000Hz.

[0068] The obtained 8 channels of raw vibration data, 1 channel of raw voltage data, and 1 channel of raw current data are each subjected to FFT (Fast Fourier Transform) to generate a spectrum, resulting in discrete "frequency buckets". For example: 0Hz, 10Hz, 20Hz, ..., 100Hz, ..., 5000Hz.

[0069] (1) Core characteristics: Core vibration originates from magnetostriction. Magnetostriction is related to voltage flux. Its fundamental frequency is twice the power supply frequency (100Hz). Due to nonlinearity, it generates higher harmonics (200Hz, 300Hz...). The energy of core vibration is concentrated at low frequencies. At medium and high frequencies, core vibration decays significantly and can be almost ignored.

[0070] Constructing the core feature vector (Regarding the first) (Sensors) From the FFT results, only retain the amplitude values ​​at 100Hz, 200Hz, 300Hz, 400Hz, and 500Hz. Set the rest to zero.

[0071] Vibration components:

[0072] Voltage component (excitation):

[0073] Get Node Input features:

[0074] (2) Winding characteristics: Winding vibration originates from the Lorentz force (related to current). When the winding clamping force is insufficient (loose), the natural frequency of the winding changes, usually triggering high-frequency resonance. If the core vibrates too strongly in the low-frequency range, it will "mask" the winding signal. Only in the mid-to-high frequencies, when the core sound is quieter, can the high-frequency "noise" caused by the loose winding be heard.

[0075] Constructing winding feature vectors (Regarding the first) (Sensors) From the FFT results, only the amplitude values ​​within the range of 600Hz to 2000Hz are retained, with a point taken every 100Hz. Set the rest to zero. Vibration components:

[0076] Current component (excitation):

[0077] Get Node Input features:

[0078] A total of 8 sensors obtained the core feature vector. This forms the first characteristic matrix.

[0079] A total of 8 sensors obtained the winding feature vector. This forms the second characteristic matrix.

[0080] Ultimately, two sets of clear characteristic data were obtained: one set representing the core state and the other set representing the winding state.

[0081] The above method decouples electrical features from mechanical vibration features through frequency domain transformation and frequency masking, filters out power frequency interference, and separates the characteristic frequency bands of the core and windings, significantly improving the purity of the input features and providing a high signal-to-noise ratio independent input for the dual-channel physical graph neural network.

[0082] Step 102: Input the first feature matrix and the second feature matrix into the first and second channels of the trained dual-channel physical graph neural network, respectively, to obtain the first predicted vibration amplitude matrix, the first adjacency matrix, the second predicted vibration amplitude matrix, and the second adjacency matrix; the first adjacency matrix and the second adjacency matrix are used to characterize the mechanical transmission relationship between the nodes of the sensor network; Specifically, the dual-channel physical graph neural network decouples the different vibration sources of the core and windings, thereby solving the problem of unclear fault location.

[0083] Optionally, the dual-channel physical graph neural network model includes: a physics-guided self-attention layer and a graph convolutional regression prediction layer; correspondingly, the first feature matrix and the second feature matrix are respectively input into the first and second channels of the trained dual-channel physical graph neural network to obtain a first predicted vibration amplitude matrix, a first adjacency matrix, a second predicted vibration amplitude matrix, and a second adjacency matrix, specifically including: Step 1021: Input the first feature matrix and the second feature matrix into the self-attention layer of the first channel and the second channel respectively, and calculate the dynamic mechanical transmission relationship between the sensor network nodes of the first channel and the second channel respectively to obtain the first adjacency matrix and the second adjacency matrix.

[0084] Step 1022: Input the first feature matrix and the first adjacency matrix into the physical graph convolutional regression prediction layer of the first channel to obtain the first predicted vibration amplitude matrix of the sensor network; input the second feature matrix and the second adjacency matrix into the physical graph convolutional regression prediction layer of the second channel to obtain the second predicted vibration amplitude matrix of the sensor network.

[0085] Specifically, the embodiments of the present invention will be described in detail using the winding fault diagnosis channel as an example (the implementation steps of the core channel are exactly the same, only the input data is different).

[0086] In step 1021 above, the eight sensors are considered as eight nodes in the diagram. The feature vector of each node is: The weights of the connections (edges) between nodes are not preset, but are learned dynamically by the network.

[0087] The method for calculating the dynamic mechanical transmission relationship between sensor network nodes also includes: (10) (11) in, Representative sensor node To the node Mechanical transmission weight; The attention coefficient represents the sensor's... For sensors The initial level of attention; Represents the feature vectors of the i-th and j-th nodes; Represents the attention weight vector; The weight matrix is ​​a learnable matrix; This represents a vector concatenation operation; r is the loop variable; m represents the number of nodes in the sensor network.

[0088] All that will be obtained The results are organized into a matrix containing the weights passed between all nodes. .

[0089] In step 1022 above, the graph convolutional regression prediction layer is used to predict vibrations based on the structure and input.

[0090] (12) in For activation functions, ReLU or Sigmoid functions are usually chosen. It was calculated in the previous step. matrix. This is the input matrix of the network; The regression layer weight matrix maps the aggregated features back to the vibration amplitude space. It is initially set to random numbers and is automatically updated during training. This is the final output, representing the predicted vibration amplitude of the eight sensors at various frequency points.

[0091] The training process of this physical graph neural network is as follows: Using historical health data of transformers after commissioning or maintenance, the weight parameters of the neural network are trained. ), and obtain the reference mechanical transmission matrix of the equipment in a healthy state ( ).

[0092] Collection time is (e.g., 1 month) of health operation data. The data is then segmented. Samples (e.g.) (Number of samples), each sample lasting 1 second. Frequency masking and feature separation are performed to obtain the training dataset for the winding channels: (13) Forward propagation will construct a collection of The total training set of each sample is randomly shuffled and divided into: One batch of samples is input into the neural network in parallel each time. (In this example, N is 64) sample matrices Input the network, using the current random parameters Calculate the predicted value Calculate the training loss to measure the accuracy of the predictions. At this point, only the fitting accuracy is considered.

[0093] (14) in These are the actual values ​​collected. This is the calculated predicted value. This is the average training loss for this batch. It is a scalar value. The smaller it is, the better the network's current parameters are and the more accurate the predictions are.

[0094] Backpropagation is the core of parameter acquisition, utilizing gradient descent algorithms (such as Adam Optimizer) based on... Automatic adjustment The value. If the prediction error is large, fine-tune the parameters until the error becomes smaller. Repeat the above process until... Convergence (approaching 0).

[0095] Save the trained This will not be changed in subsequent testing phases. The average of the adjacency matrices generated from all healthy samples will be saved as the baseline matrix. .

[0096] The above method, through independent dual-channel modeling, physically isolates the core and winding vibration sources, eliminating spectral aliasing interference. The dynamic adjacency matrix accurately maps the mechanical transmission paths between sensors, enabling the model to keenly capture minute structural deformations and significantly improving the physical interpretability and spatial accuracy of fault location.

[0097] Step 103: Based on the first predicted vibration amplitude matrix, the first adjacency matrix, the second predicted vibration amplitude matrix, and the second adjacency matrix, determine the fault type and locate the physical area of ​​the fault.

[0098] Specifically, determining the fault type and locating the physical area of ​​the fault includes the following steps: Construct a structural consistency loss function, which includes prediction error and structural discrepancy.

[0099] Based on the first predicted vibration amplitude matrix and the first adjacency matrix, the first loss value is calculated using the structural consistency loss function.

[0100] Based on the second predicted vibration amplitude matrix and the second adjacency matrix, the second loss value is calculated using the structural consistency loss function; The first loss value is compared with a set first threshold to determine the core state; The second loss value is compared with a set second threshold to determine the winding state.

[0101] Specifically, the core condition includes normal and faulty states; the winding condition also includes normal and faulty states. Prediction error reflects deviations in the amplitude of the vibration signal. Structural differences reflect deviations between the current mechanical transmission relationship and the health baseline. For example, a structural consistency loss function can be constructed as follows: (15) in, These are the actual values ​​collected. The calculated predicted value; This is the calculated current adjacency matrix; This is the baseline adjacency matrix saved after model training is completed; The Frobenius norm (the sum of squares of the differences between matrix elements) is used to derive a scalar value. The balance coefficient is usually taken as... This is used to balance the weights of the two items and prevent one item from becoming too large and dominating the result. The diagnostic index is a scalar value used for final alarm determination; in the iron core channel, it is... In the winding channel is .

[0102] First item For prediction error, the second term Due to structural differences. When the increase in vibration amplitude is solely caused by load changes, change, However, it remains essentially unchanged (because the mechanical structure remains the same), so the first term will increase, but the second term will be very small, and the total loss may not exceed the threshold. When a mechanical failure occurs (such as loosening), the local structural stiffness changes, causing a sudden change in the vibration transmission path between sensors. At this time, r will deviate significantly from Â_base, causing the second term to increase sharply, thereby triggering an alarm, thus reducing false alarms and eliminating operating condition interference.

[0103] Based on the first predicted vibration amplitude matrix and the first adjacency matrix, the first loss value is calculated using formula (15). Based on the second predicted vibration amplitude matrix and the second adjacency matrix, the second loss value is calculated using formula (15). .

[0104] After obtaining the loss value, the fault determination and location logic is as follows: The preset first threshold and the preset second threshold can be different or the same. For example, the first threshold and the second threshold can both be set to be the same. (Empirical or statistical value).

[0105] like Then it is judged as normal.

[0106] like This is then identified as a fault, and the system is started. An alarm indicates a core fault (such as multiple grounding points or inter-segment short circuits).

[0107] If Then it is judged as normal.

[0108] like This is then identified as a fault, and the system is started. An alarm indicates a winding fault (such as loose clamping force).

[0109] The above method decouples fault types, solving the problem of faults being "indistinguishable".

[0110] Optionally, when the core state and / or winding state are confirmed to be faulty, their respective difference matrices are calculated. The difference matrices are used to characterize the difference between the adjacency matrix and the reference adjacency matrix, which is determined after the model training is completed. Find the element with the largest value in the difference matrix, and determine the area between the sensor nodes represented by the element where the fault point is located.

[0111] Specifically, when the core is faulty, the winding is faulty, or both are faulty simultaneously, the fault location can be determined by the difference in the adjacency matrix. For example, the difference matrix can be calculated using the following formula: (16) Find the matrix The element with the largest value ,element Larger means sensor and sensors The mechanical transmission path between them changed abruptly. The fault point is located in the sensor. and The area between.

[0112] Using the above method, maintenance personnel can narrow the scope of maintenance from the entire transformer to this specific local area based on the actual installation locations of the i-th and j-th nodes in the sensor network, thus achieving precise location maintenance.

[0113] To facilitate understanding of the solutions in the embodiments of the present invention, such as Figure 4 As shown, a specific example of converter transformer fault diagnosis is used for illustration.

[0114] (1) First, a simulation model of the converter transformer is established, modal analysis is performed, and the mesh node information and modal matrix of the tank surface are extracted. This step aims to obtain the inherent vibration characteristics of the equipment structure, providing a physical basis for subsequent sensor layout and network input.

[0115] (2) An improved algorithm (a multi-strategy evolutionary algorithm MS-SSA combining a weighted roulette wheel strategy and mirror back learning) was used, along with the modal confidence criterion (MAC), to calculate the globally optimal sensor installation location. This ensures that the acquired vibration signal has the highest structural resolution and signal-to-noise ratio.

[0116] (3) Arrange sensors at the optimized positions, acquire vibration signals and secondary voltage and current signals through synchronous acquisition devices, construct an original data matrix containing excitation and response, and use it as the actual input of the model.

[0117] (4) Using the frequency masking mechanism, the original data is separated into two different frequency bands: the core-dominant frequency band, associated with voltage excitation, and the winding-dominant frequency band, associated with current excitation, to ensure the decoupling of frequency domain features. This step is the key preprocessing for realizing the separation of the vibration source.

[0118] (5) Input the decoupled features into the dual-channel graph neural network. The network uses a self-attention mechanism to invert the current dynamic adjacency matrix (i.e., the mechanical transfer matrix) and predict the theoretical vibration value. This step simultaneously completes the reconstruction of the physical topology and the fitting of the vibration amplitude.

[0119] (6) Calculate the structural consistency loss function (including fitting residuals and structural variation residuals), and determine the fault type based on the loss value (distinguishing between core faults and winding faults). Finally, visualize and locate the physical area where the fault occurred by analyzing the difference matrix (the difference between the predicted adjacency matrix and the actual adjacency matrix).

[0120] The following describes the converter transformer fault diagnosis device based on sensor position optimization and physical graph neural network provided by the present invention. The converter transformer fault diagnosis device based on sensor position optimization and physical graph neural network described below and the converter transformer fault diagnosis method based on sensor position optimization and physical graph neural network described above can be referred to in correspondence.

[0121] like Figure 5 As shown, this invention provides a converter transformer fault diagnosis device based on sensor position optimization and physical graph neural network, comprising the following modules: The acquisition module 500 is used to synchronously acquire multi-channel vibration signals of the converter transformer, as well as the corresponding voltage and current signals; the multi-channel vibration signals are obtained through a sensor network arranged at an optimal position on the surface of the converter transformer tank; the optimal position is obtained through an intelligent optimization algorithm. The feature separation module 510 is used to perform frequency domain transformation on the multi-channel vibration signal and the corresponding voltage and current signals, and to perform feature separation through frequency masking to obtain the first feature matrix and the second feature matrix of the converter transformer. The prediction module 520 is used to input the first feature matrix and the second feature matrix into the first channel and the second channel of the trained dual-channel physical graph neural network, respectively, to obtain a first predicted vibration amplitude matrix, a first adjacency matrix, a second predicted vibration amplitude matrix, and a second adjacency matrix; the first adjacency matrix and the second adjacency matrix are used to characterize the mechanical transmission relationship between the nodes of the sensor network; The fault diagnosis module 530 is used to determine the fault type and locate the physical area of ​​the fault based on the first predicted vibration amplitude matrix, the first adjacency matrix, the second predicted vibration amplitude matrix, and the second adjacency matrix.

[0122] On the other hand, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the converter transformer fault diagnosis method based on sensor position optimization and physical graph neural network provided by the above methods.

[0123] In another aspect, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the converter transformer fault diagnosis method based on sensor position optimization and physical graph neural network provided by the above methods.

[0124] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0125] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A sensor location-based optimization and physical graph neural network-based converter transformer fault diagnosis method, characterized in that, include: Simultaneously acquire multi-channel vibration signals of the converter transformer, as well as the corresponding voltage and current signals; The multi-channel vibration signal is obtained through a sensor network arranged at an optimal position on the surface of the converter transformer tank; the optimal position is obtained through an intelligent optimization algorithm. The multi-channel vibration signal and the corresponding voltage and current signals are subjected to frequency domain transformation, and feature separation is performed by frequency masking to obtain the first feature matrix and the second feature matrix of the converter transformer. The first feature matrix and the second feature matrix are respectively input into the first and second channels of the trained dual-channel physical graph neural network to obtain the first predicted vibration amplitude matrix, the first adjacency matrix, the second predicted vibration amplitude matrix, and the second adjacency matrix; the first adjacency matrix and the second adjacency matrix are used to characterize the mechanical transmission relationship between the nodes of the sensor network; Based on the first predicted vibration amplitude matrix, the first adjacency matrix, the second predicted vibration amplitude matrix, and the second adjacency matrix, the fault type and the physical area of ​​the fault are determined.

2. The converter transformer fault diagnosis method based on sensor location optimization and physical graph neural network according to claim 1, characterized in that, The intelligent optimization algorithm includes: a multi-strategy improved sparrow search algorithm; Correspondingly, the optimal position is obtained through an intelligent optimization algorithm, specifically including: Obtain the modal kinetic energy distribution on the surface of the converter transformer tank; Based on the modal kinetic energy distribution, a weighted roulette wheel betting strategy is used to generate the initial population of the sensor network deployment scheme; Based on the composite fitness value of each placement scheme in the current population, the placement scheme is updated to obtain a new generation of population; for the globally optimal placement scheme of the new generation of population, a local lens strategy based on modal contribution is used for local update; the iteration continues until the maximum number of iterations is reached or the composite fitness value no longer decreases, and the placement scheme with the smallest fitness value is output, which is the globally optimal location set of the sensor network.

3. The converter transformer fault diagnosis method based on sensor location optimization and physical graph neural network according to claim 2, characterized in that, The globally optimal placement scheme for the new generation population employs a local lensing strategy based on modal contribution for local updates, specifically including: Based on the modal contribution, identify the weakest sensor network node; Local lens variations are applied to the nodes of the short-board sensor network to generate candidate placement schemes; If the composite fitness value of the candidate placement scheme is better than that of the original globally optimal placement scheme, then the globally optimal placement scheme is updated.

4. The converter transformer fault diagnosis method based on sensor location optimization and physical graph neural network according to claim 2, characterized in that, The composite fitness function Fit(X) is defined as follows: ； Where X represents the deployment scheme of a sensor network containing m nodes; modal matrix elements The i-th node represents the mode amplitude in the j-th mode; w1 and w2 represent the weighting coefficients. From the mode matrix The extracted first Modal row vectors corresponding to the positions of each measurement point; From the mode matrix The extracted first Modal row vectors corresponding to the positions of each measurement point; ) represents the maximum modal confidence criterion among all node pairs in the node placement scheme X.

5. The converter transformer fault diagnosis method based on sensor location optimization and physical graph neural network according to claim 3, characterized in that, modal contribution This can be represented by the following formula: ； in, For the first The total kinetic energy of a node in all modes; This represents the maximum kinetic energy value among all nodes in the sensor network. This represents the maximum similarity between the i-th node and all nodes except itself. This represents the weighting coefficient.

6. The converter transformer fault diagnosis method based on sensor location optimization and physical graph neural network according to claim 1, characterized in that, The dual-channel physical graph neural network model includes: a physics-guided self-attention layer and a graph convolutional regression prediction layer; Correspondingly, the first feature matrix and the second feature matrix are respectively input into the first and second channels of the trained dual-channel physical graph neural network to obtain the first predicted vibration amplitude matrix, the first adjacency matrix, the second predicted vibration amplitude matrix, and the second adjacency matrix, specifically including: The first feature matrix and the second feature matrix are respectively input into the self-attention layer of the first channel and the second channel, and the dynamic mechanical transmission relationship between the sensor network nodes of the first channel and the second channel is calculated to obtain the first adjacency matrix and the second adjacency matrix. The first feature matrix and the first adjacency matrix are input into the physical graph convolutional regression prediction layer of the first channel to obtain the first predicted vibration amplitude matrix of the sensor network; the second feature matrix and the second adjacency matrix are input into the physical graph convolutional regression prediction layer of the second channel to obtain the second predicted vibration amplitude matrix of the sensor network.

7. The converter transformer fault diagnosis method based on sensor location optimization and physical graph neural network according to claim 6, characterized in that, The method for calculating the dynamic mechanical transmission relationship between sensor network nodes also includes: ； ； in, Representative sensor node To the node Mechanical transmission weight; The attention coefficient represents the sensor's... For sensors The initial level of attention; Represents the feature vectors of the i-th and j-th nodes; Represents the attention weight vector; The weight matrix is ​​a learnable matrix; This represents a vector concatenation operation; r is the loop variable; m represents the number of nodes in the sensor network.

8. The converter transformer fault diagnosis method based on sensor location optimization and physical graph neural network according to claim 1, characterized in that, The step of determining the fault type and locating the physical area of ​​the fault based on the first predicted vibration amplitude matrix, the first adjacency matrix, the second predicted vibration amplitude matrix, and the second adjacency matrix specifically includes: Construct a structural consistency loss function, which includes: prediction error and structural difference; Based on the first predicted vibration amplitude matrix and the first adjacency matrix, the first loss value is calculated using the structural consistency loss function; Based on the second predicted vibration amplitude matrix and the second adjacency matrix, the second loss value is calculated using the structural consistency loss function; The first loss value is compared with a set first threshold to determine the core state; The second loss value is compared with a set second threshold to determine the winding state.

9. The converter transformer fault diagnosis method based on sensor position optimization and physical graph neural network according to claim 8, characterized in that, The physical area where the location fault occurred specifically includes: Once the core state and / or winding state are confirmed to be faulty, their respective difference matrices are calculated. The difference matrices are used to characterize the difference between the adjacency matrix and the reference adjacency matrix, which is determined after the model training is completed. Find the element with the largest value in the difference matrix, and determine the area between the sensor nodes represented by the element where the fault point is located.

10. A converter transformer fault diagnosis device based on sensor position optimization and physical graph neural network, characterized in that, include: The acquisition module is used to synchronously acquire multi-channel vibration signals of the converter transformer, as well as the corresponding voltage and current signals; The multi-channel vibration signal is obtained through a sensor network arranged at an optimal position on the surface of the converter transformer tank; the optimal position is obtained through an intelligent optimization algorithm. The feature separation module is used to perform frequency domain transformation on the multi-channel vibration signal and the corresponding voltage and current signals, and to perform feature separation through frequency masking to obtain the first feature matrix and the second feature matrix of the converter transformer. The prediction module is used to input the first feature matrix and the second feature matrix into the first channel and the second channel of the trained dual-channel physical graph neural network, respectively, to obtain a first predicted vibration amplitude matrix, a first adjacency matrix, a second predicted vibration amplitude matrix, and a second adjacency matrix; the first adjacency matrix and the second adjacency matrix are used to characterize the mechanical transmission relationship between the nodes of the sensor network; The fault diagnosis module is used to determine the fault type and locate the physical area of ​​the fault based on the first predicted vibration amplitude matrix, the first adjacency matrix, the second predicted vibration amplitude matrix, and the second adjacency matrix.

Images