Transformer fault diagnosis method and diagnosis system based on symplectic geometry modal decoupling

Abstract

The application discloses a transformer fault diagnosis method and system based on symplectic geometry modal decoupling, relates to the technical field of transformer fault diagnosis, and is used for providing an intelligent diagnosis system which can deeply embed a transformer physical mechanism, adaptively match vibration evolution rules and has high efficient global optimization capability.The diagnosis method comprises the following steps: S1, transformer signal modal decoupling and physical characteristic extraction based on symplectic geometry; S2, deep mapping of fault characteristics based on a symplectic manifold associated gate fusion network SMAG-Net; S3, feature adaptive fusion and decision based on symplectic space physical constraints; and S4, model optimization based on a chaotic ocean predator algorithm.The application makes up for the shortcomings of the prior art in a full range from signal processing, architecture design to global optimization, and significantly improves the reliability and accuracy of transformer fault diagnosis in a strong noise environment.

Description

Technical Field

[0001] This invention relates to the field of transformer fault diagnosis technology, specifically a transformer fault diagnosis method and system based on symplectic geometric mode decoupling. Background Technology

[0002] Currently, monitoring methods for the internal mechanical condition of transformers mainly include dissolved gas analysis in oil, partial discharge detection, and vibration methods. Among these, vibration methods, due to their advantages of being non-contact, having a fast response speed, and directly reflecting the physical characteristics of the mechanical structure, show great application potential in the fault diagnosis of transformer windings and cores. The vibration signals of a transformer mainly originate from the magnetostriction of the core under the influence of an electromagnetic field and the electrodynamic vibration of the windings under the influence of load current. These vibration signals are transmitted to sensors through the tank wall and contain rich information about the equipment's operating status.

[0003] However, in practical engineering applications, fault diagnosis using vibration signals faces significant challenges. Firstly, transformers are subjected to alternating effects of strong electromagnetic interference and complex environmental noise, resulting in raw vibration signals from sensors typically exhibiting extremely low signal-to-noise ratios. Subtle early fault characteristics are easily masked by environmental noise. Traditional signal processing methods, such as Fourier Transform (FFT) and Short-Time Fourier Transform (STFT), often suffer from limited time-frequency resolution when processing typical non-stationary and nonlinear signals like transformer vibration, making it difficult to effectively decouple complex vibration modes.

[0004] In recent years, artificial intelligence algorithms, represented by deep convolutional neural networks (CNNs), have made significant progress in the field of fault diagnosis. Researchers have attempted to input transformer vibration signals into deep learning models to automatically extract fault features through multi-layer nonlinear mapping. While these methods perform well under specific laboratory conditions, two core bottlenecks have been exposed in practical applications. First, there is a severe disconnect between physical mechanisms and data-driven approaches. Traditional deep learning models are often treated as black boxes, with the scale setting of their convolutional kernels and the selection of dilation factors relying heavily on human experience or random experiments. When processing transformer vibration signals, these models ignore the physical properties of the signal itself, such as sampling frequency, energy distribution of each component, and coupling logic of the internal mechanical structure. For example, when the transformer load fluctuates, the energy contrast between the fundamental frequency and harmonics of the vibration signal changes. If the model cannot identify these spurious features caused by load changes at the physical level, it is prone to generating false alarms. Furthermore, the self-attention mechanism in traditional models is usually based on mathematical correlation for weight allocation, which is inconsistent with the mechanical coupling paths between the various components inside the transformer that follow physical laws, resulting in extremely poor robustness of the model when dealing with complex variable load conditions. Second, there is the dilemma of optimizing model hyperparameters. The performance of deep learning models is highly dependent on the configuration of hyperparameters. Existing optimization strategies mostly employ traditional heuristic algorithms such as grid search, genetic algorithms, and particle swarm optimization. However, the objective function of transformer fault diagnosis problems typically has high non-convexity and complex constraint spaces. Traditional algorithms are prone to getting trapped in local optima in the later stages of the search and have slow convergence speeds. Especially when facing multi-dimensional, strongly coupled hyperparameter combinations, the lack of effective population initialization mechanisms and dynamic balancing exploration and development capabilities often makes it difficult for optimization algorithms to find the global optimum within a limited time, thus limiting further improvements in the performance of the diagnostic system.

[0005] In summary, existing transformer fault diagnosis technologies have significant shortcomings in areas such as preserving the structural characteristics of signal processing, the physical adaptability of feature extraction, the logical rigor of feature fusion, and the global search capability for parameter optimization. How to construct an intelligent diagnostic system that can deeply integrate with the physical mechanisms of transformers, adaptively match vibration evolution laws, and possess efficient global optimization capabilities has become a key bottleneck in the leapfrog development of power system operation and maintenance technologies towards intelligence and precision. Summary of the Invention

[0006] The purpose of this invention is to provide a transformer fault diagnosis method and system based on symplectic geometric mode decoupling, which provides an intelligent diagnosis system that can deeply integrate the physical mechanism of transformers, adaptively match the vibration evolution law, and has efficient global optimization capabilities.

[0007] The technical solution adopted by the present invention to solve its technical problem is: a transformer fault diagnosis method based on symplectic geometric mode decoupling, which includes the following steps.

[0008] S1 Transformer signal mode decoupling and physical feature extraction based on symplectic geometry.

[0009] S1.1 Acquire the vibration signal sequence of the transformer under operating conditions. , where x n It is a vibration signal.

[0010] S1.2 Constructing the Hamiltonian matrix .

[0011] S1.3 Diagonalization is performed to solve the problem. symplectic eigenvalues Using symplectic orthogonal matrices For the original trajectory matrix By projection, the symplectic component matrix is ​​obtained. .

[0012] S1.4 symplectic component matrix Perform diagonal averaging. elements in Calculate the reconstructed first Each time point component Calculate the energy percentage of each component. and fourth-order kurtosis .

[0013] S1.5 Calculation of symplectic geometrical entropy .

[0014] S1.6 Finally, the energy percentage Fourth-order kurtosis And symplectic geometrical entropy Row vectors are formed by horizontal splicing. .

[0015] S2 is based on the deep mapping of fault features in the symplectic manifold relational gated fusion network SMAG-Net.

[0016] A dual-branch system with complementary functions was designed based on the data obtained after symplectic geometric processing.

[0017] S3 Feature Adaptive Fusion and Decision Based on Symplectic Space Physical Constraints

[0018] S3.1 Constructing the symplectic correlation matrix A global gating gain mechanism is introduced.

[0019] S3.2 Perform feature nonlinear correlation fusion to obtain feature vectors. ;

[0020] S3.3 uses feature splicing operations to... Indicators of static physical background Integration, output diagnostic results ;

[0021] S4 Model Optimization Based on Chaotic Ocean Predator Algorithm.

[0022] The parameter set of the model to be optimized is mapped to the predator position vector in the Chaotic Ocean Predator Algorithm (CMPA). We designed a comprehensive fitness function that integrates diagnostic accuracy and physical-logical constraints.

[0023] Further, step S1.2 specifically includes: using the symplectic geometric mode decomposition (SGMD) algorithm to... Reconstruct the trajectory matrix. (1); where, This is the total length of the signal; For the embedded dimension; Subsequently, a symmetric matrix is ​​constructed. and based on Constructing the Hamiltonian matrix : (2).

[0024] Furthermore, in step S1.3, the symplectic similarity transformation matrix is ​​used. (3) To Diagonalization is performed; among which, Let be the symplectic orthogonal matrix obtained through the symplectic similarity transformation, satisfying ; An identity symplectic matrix; It is a symplectic eigenvalue matrix. is the symplectic eigenvalue.

[0025] Furthermore, The calculation process is as follows: For each pair of symplectic eigenvalues Find its corresponding vector pair ( , Make it simultaneously satisfy both the mapping relation and the symplectic orthogonality relation: (4); will The symmetric orthogonal matrix is ​​obtained by horizontally concatenating the column vectors that have undergone symmetric orthogonalization. ;use For the original trajectory matrix Projecting, for the first Each mode has a symplectic component matrix after projection. (5); among which, It is extraction The column vectors corresponding to specific symplectic feature pairs are truncated, and their first two columns are extracted. Construct the projection vector from the dimensional components.

[0026] Furthermore, in step S1.3, for elements in The i-th component after reconstruction (6); among which, For the first The number of elements on each anti-diagonal line; for The order of the vector is determined; subsequently, the energy proportion of each component is calculated, and the energy proportion of the Kth order component is denoted as... , (7) Construct the energy distribution vector Calculate principal components Fourth-order kurtosis (8); among which, Main component exist Amplitude at any given moment; This is the arithmetic mean of the component over the entire time span; This represents the total number of sampling points.

[0027] Furthermore, in step S1.4, the symplectic eigenvalues ​​are used. calculate (9); among which, This represents the normalized symptotic eigenvalue.

[0028] Furthermore, in step S2, a residual multilayer perceptron (Res-MLP) is used. Transformed into a network-recognizable physical context tensor , (10); where σ() is a nonlinear activation function; A neural network whose hidden layers can be learned; These are the learnable bias terms for each layer; This is the weight matrix.

[0029] Furthermore, in step S3.1, the symplectic correlation matrix elements (11); where tanh is the hyperbolic tangent activation function; symplectic inner product term To capture the dynamic correlation of signals in phase space; a scale-aware correction operator is introduced. (12), where s min The minimum step size; s i，l To experience the wild stride length; For kurtosis; dynamic gain weights of the global gating gain mechanism. (13); among which, This is the global scaling factor; For bias terms; For smooth activation function, the expression is: ; The current order The average value of the component entropy spectrum; Std(H) sym )for The standard deviation.

[0030] Furthermore, in step S3.2, the fused feature vector (14); among which, For the first The depth feature vector output by the branch of the road mode evolution; This is for the Adama product operation.

[0031] Furthermore, in step S3.3, (15); where Prob is the final output probability; Softmax is the normalized activation function; W out b is the trainable weight matrix of the output layer; out This is the bias term for the classification output layer.

[0032] Furthermore, in step S4, a piecewise chaotic mapping is used to construct the initial parameter population: (16); These are mapping control parameters.

[0033] Furthermore, the comprehensive fitness function (17); where N is the total number of samples; For indicator functions; For the first Each sample model predicts the fault category; For the first Each sample contains a true fault label; λ is the weighting coefficient. This is the symplectic geometric energy distribution vector extracted in the first stage.

[0034] The present invention also provides a transformer fault diagnosis system based on symplectic geometric mode decoupling, for executing the transformer fault diagnosis method based on symplectic geometric mode decoupling of the present invention, including a data acquisition and sensing module, a communication and data transmission module, a core computing and diagnostic analysis server, and a data storage and human-computer interaction module.

[0035] The data acquisition and sensing module includes a piezoelectric vibration sensor array, a signal conditioning circuit, and a high-frequency analog-to-digital converter. The vibration sensor array is installed on the outer wall of the transformer tank. The vibration sensor array is used to capture simulated mechanical vibration signals of the transformer under operating conditions in real time. The acquired simulated signals are filtered and amplified by the signal conditioning circuit for anti-aliasing, and then converted into discrete one-dimensional vibration digital sequences by the high-frequency analog-to-digital converter.

[0036] The communication and data transmission module is connected to the output of the data acquisition and sensing module. The communication and data transmission module includes an industrial Ethernet interface and a fiber optic transceiver terminal, which is used to transmit the acquired high-frequency vibration digital sequence to the core computing and diagnostic analysis server without loss and with low latency, so as to ensure the reliability and real-time performance of data transmission in the strong electromagnetic interference environment at the transformer site.

[0037] The core computing and diagnostic analysis server is equipped with a central processing unit and tensor acceleration hardware units. Internally, it contains and runs core diagnostic analysis logic. After receiving the vibration digital sequence from the communication link, the server first constructs a Hamiltonian matrix using a symplectic geometric mode decomposition algorithm and reconstructs the signal. It then extracts the energy proportion, fourth-order kurtosis, and symplectic geometric spectral entropy of each modal component, concatenating them to form a multi-dimensional physical indicator feature vector. Subsequently, the dual-branch feature depth mapping unit within the server processes the data in parallel. The modal evolution branch uses adaptive dilated convolution based on physical characteristics to extract temporal dynamic evolution features, while the physical indicator branch transforms the energy distribution into a physical background tensor through a residual multilayer perceptron. The server also includes constrained adaptive fusion decision logic and global optimization logic. The constrained adaptive fusion decision logic constructs a symplectic correlation matrix and introduces a global gated gain mechanism driven in real-time by symplectic geometric spectral entropy. This nonlinearly correlates and fuses the temporal dynamic evolution features with the physical background tensor, thereby outputting an accurate digital diagnostic signal for transformer fault classification. Meanwhile, the core computing and diagnostic analysis server adopts the ocean predator algorithm improved by piecewise chaotic mapping, and uses a comprehensive fitness function that integrates diagnostic accuracy and physical energy distribution topology consistency as the feedback target to perform closed-loop optimization and real-time iterative update of the network framework hyperparameters within the system.

[0038] The data storage and human-computer interaction module is signal-connected to the core computing and diagnostic analysis server, and includes a large-capacity database storage array and an interactive display terminal. The database storage array is used for persistent storage and management of massive amounts of historical vibration raw data, symplectic component feature libraries, and optimal network parameters after optimization iterations. The interactive display terminal is used to render the transformer health status topology in real time, dynamically display feature evolution curves, and ultimately output intuitive fault alarm diagnosis results and status assessment reports to on-site maintenance personnel.

[0039] The beneficial effects of this invention are as follows: By introducing symplectic geometric mode decomposition with structure-preserving properties, this invention effectively solves the problem in the background technology that traditional signal processing methods are difficult to decouple complex vibration modes and easily lose signal physical topology information; by constructing a SMAG-Net network based on energy-frequency adaptive mapping and a dynamic gating fusion mechanism, it eliminates the bottlenecks of insufficient feature extraction accuracy and poor robustness under variable load conditions caused by fixed convolution scale and lack of physical logic in traditional deep learning models, achieving deep integration of data-driven and mechanical vibration coupling laws; finally, by introducing a CMPA algorithm based on piecewise chaotic mapping, it overcomes the defects of existing optimization strategies that are prone to getting trapped in local optima and have low search efficiency, ensuring the global optimality of model hyperparameters in complex constraint space. In summary, this invention comprehensively makes up for the shortcomings of existing technologies from signal processing and architecture design to global optimization, significantly improving the reliability and accuracy of transformer fault diagnosis in strong noise environments. Attached Figure Description

[0040] Figure 1 Here is a flowchart of symplectic geometric data processing;

[0041] Figure 2 This is a diagram of a two-branch neural network processing method;

[0042] Figure 3 Flowchart for the Chaotic Ocean Predator CMPA. Detailed Implementation

[0043] This invention utilizes symplectic geometric mode decomposition (SGMD), which preserves data structure, to extract physical indicators. It then constructs a symplectic manifold correlation-gated fusion network, SMAG-Net, and extracts multi-scale features through dilated convolution based on energy-frequency adaptive mapping. Furthermore, it combines symplectic correlation matrices with spectral entropy-driven dynamic gating units (S-Gates) to achieve feature enhancement guided by physical laws. Finally, it introduces piecewise chaotic mapping to improve the marine predator algorithm (CMPA) for closed-loop parameter optimization. This invention achieves a deep integration of physical mechanisms and deep learning, significantly improving diagnostic accuracy and the model's anti-interference capability. The present invention provides a transformer fault diagnosis method based on symplectic geometric mode decoupling, comprising the following steps.

[0044] S1 Transformer signal mode decoupling and physical feature extraction based on symplectic geometry.

[0045] S1.1 Collects mechanical vibration signals of the transformer under operating conditions.

[0046] First, vibration sensors installed on the outer wall of the transformer tank are used to collect real-time mechanical vibration signal sequences of the transformer under operating conditions. , where x nThe signal is a vibration signal. The acquisition process needs to cover the complete electromagnetic cycle of the transformer, and the sampling frequency... The sampling theorem should be satisfied to ensure that the high-frequency transient characteristics of the winding under short-circuit current impact can be captured.

[0047] S1.2 Constructing the Hamiltonian matrix .

[0048] To address the non-stationary vibration characteristics under strong electromagnetic interference inside the transformer, such as Figure 1 As shown, the symplectic geometric mode decomposition (SGMD) algorithm is used to process the original one-dimensional vibration signal sequence. Reconstruct the trajectory matrix. . (1). Among them, This is the total length of the signal; For the embedded dimension; Subsequently, a symmetric matrix is ​​constructed. and based on Constructing the Hamiltonian matrix : (2) Hamiltonian matrix Protect the nonlinear manifold characteristics in the vibration signal of the transformer.

[0049] S1.3 Hamiltonian matrix Diagonalization is performed to solve the problem. symplectic eigenvalues Using symplectic orthogonal matrices For the original trajectory matrix By projection, the symplectic component matrix is ​​obtained. .

[0050] Vibration signal sequence After embedding into the symplectic space and constructing the Hamiltonian matrix, the symplectic similarity transformation matrix is ​​used. (3) For the Hamiltonian matrix Diagonalization is performed to solve the matrix. symplectic eigenvalues . Let be the symplectic orthogonal matrix obtained through the symplectic similarity transformation, satisfying . An identity symplectic matrix; It is a symplectic eigenvalue matrix, where each symplectic eigenvalue... Both represent a specific energy component in the vibration signal, and their magnitude directly reflects the proportion of that mode in the total vibration energy. The calculation process is as follows: after passing through the Hamiltonian matrix... After eigenvalue decomposition, due to the matrix Symmetric structure, Each eigenvalue always appears in pairs and its real part has the opposite sign; therefore, for each pair of symplectic eigenvalues... Find its corresponding vector pair ( , Make it simultaneously satisfy both the mapping relation and the symplectic orthogonality relation: (4). Ultimately, it will The symmetric orthogonal matrix is ​​obtained by horizontally concatenating the column vectors that have undergone symmetric orthogonalization. Calculate the cumulative contribution rate when When the energy of the first-order eigenvalues ​​reaches 95%, the remaining... The order is considered noise and discarded. Using a symplectic orthogonal matrix... For the original trajectory matrix Perform projection. For the first... Each mode has a symplectic component matrix after projection. for: (5). Among them, It is to extract the symplectic orthogonal matrix. The column vectors corresponding to specific symplectic feature pairs are truncated, and their first two columns are extracted. Construct the projection vector from the dimensional components.

[0051] S1.4 symplectic component matrix Perform diagonal averaging.

[0052] because It is The matrix is ​​, while subsequent processing and diagnosis require a matrix of length . The one-dimensional component, therefore, needs to be... Perform diagonal averaging. Diagonal averaging: For a matrix... elements in The i-th component after reconstruction for: (6). Among them, For the first The number of elements on each anti-diagonal line; for The order of the vector. Then, the individual components are calculated. Energy percentage , i=1,2…K. Construct the energy distribution vector. Let the energy percentage of the Kth order component be denoted as . , (7). Calculate the principal components. Fourth-order kurtosis : (8). Among them, Main component exist Amplitude at any given moment; The arithmetic mean of the principal component over the entire time span; This represents the total number of sampling points.

[0053] S1.5 Calculation of symplectic geometrical entropy .

[0054] Using symplectic eigenvalues The symplectic geometric order of the data is calculated and used as an adaptive feedback signal to drive subsequent optimization algorithms. (9). Among them, It is the symplectic geometrical entropy, which reflects the degree of order in the transformer vibration system. The more complex the signal and the greater the noise interference, the higher the entropy value. The symplectic eigenvalue is... The magnitude of the paired eigenvalues ​​extracted from the Hamiltonian matrix; This represents the normalized symptotic eigenvalue.

[0055] S1.6 Finally, the energy percentage Fourth-order kurtosis And symplectic geometrical entropy Row vectors are formed by horizontal splicing. .

[0056] S2 is based on the deep mapping of fault features in the symplectic manifold relational gated fusion network SMAG-Net.

[0057] Based on the data obtained after symplectic geometric processing, a complementary dual-branch system was designed, such as... Figure 2 As shown.

[0058] (1) Modal Evolution Branch: This invention employs a TCN network with dilated convolution characteristics. Unlike ordinary convolution, dilated convolution introduces a void ratio, allowing the convolution kernel to obtain a receptive field covering the entire vibration cycle while maintaining constant parameters. For low-order, high-energy components, the network uses a larger dilation factor to capture their macroscopic envelope; for high-order, detailed components, the network uses a smaller dilation factor to focus on weak discharge or impact pulses, thereby achieving full-process modeling of the transformer vibration evolution. Input is provided by... A tensor consisting of mutually orthogonal symplectic geometric components This branch employs an adaptive dilated convolution based on physical characteristics. For the ... Road components In the Layer expansion factor The calculation formula is: (10). Among them, The sampling frequency of the original signal; For the first Each symplectic geometric component The center frequency; This refers to the percentage of energy used. This is a floor operator. It utilizes the sampling frequency. With component center frequency The ratio relationship ensures that the sampling step size of the convolution kernel can span the physical vibration period of the signal, thereby accurately capturing the phase and envelope characteristics of the transformer modal components. An adjustment term is introduced. The observation granularity is automatically adjusted based on the energy proportion of the components. A suitable receptive field is maintained for the main vibrational components with concentrated energy to extract evolutionary details, while the receptive field span is appropriately increased for weak high-frequency impact components to enhance the ability to capture sparse fault pulses. This is combined with layer indexing. Achieving linearly increasing receptive fields allows shallow layers of the network to focus on local waveform distortions, while deeper layers perceive long-term performance degradation trends. This eliminates the blindness of the exponential growth method of traditional dilated convolutions, enhancing the interpretability and robustness of the diagnostic model. The calculation formula for dilated convolutional layers is as follows: (11). Among them, The output feature matrix of the convolutional layer contains the dynamic evolution of the signal in the time domain, including transformer vibration waveform information; ReLU is a nonlinear activation function. The kernel weight matrix; To dilate the convolution operator; This is a bias term.

[0059] (2) Physical Indicator Branch: The input of this branch is a feature row vector that condenses global information. Its function is to provide the physical context for the model. Through the Residual Multilayer Perceptron (Res-MLP), the energy distribution vector... Transformed into a network-recognizable physical context tensor This process ensures that, under different load conditions, the model can automatically calibrate the sensitivity threshold in subsequent fusion processes based on the current energy distribution. (12). Where σ() is a nonlinear activation function; A neural network whose hidden layers can be learned; These are the learnable bias terms for each layer; This is the weight matrix, used to connect the original input with the output of the inner hidden layer mapping, avoiding the problems of vanishing or exploding gradients.

[0060] S3 Feature Adaptive Fusion and Decision Based on Symplectic Space Physical Constraints

[0061] S3.1 Constructing the symplectic correlation matrix A global gating gain mechanism is introduced.

[0062] In transformer systems, the vibrations of the windings and core components influence each other through structural coupling. To quantify the physical coupling relationships within the transformer's internal mechanical structure, a symplectic correlation matrix is ​​constructed. Symplectic correlation matrix Using the topological structure-preserving properties of symplectic spaces, the energy transfer characteristics between different vibration modes are described: (13). Where tanh is the hyperbolic tangent activation function; the symplectic inner product term Used to capture the dynamic correlation of signals in phase space; R i，j Let be the element in the i-th row and j-th column of R. When a mechanical loosening or deformation fault occurs, the coupling strength between specific modes will be distorted. (Matrix) The numerical changes will guide the model to identify abnormal energy transfer paths, acting as a guide for the characteristic flow.

[0063] To address the limitation of traditional models in maintaining consistent fault sensitivity across different scales, this study comprehensively considers the receptive field step size s during feature extraction. i，l With kurtosis, which reflects the impact intensity Introducing a scale-aware correction operator (14). Among them, s min Minimum step size; for components exhibiting significant impact characteristics and high receptive field matching; This will result in a higher base score, giving a higher initial attention weight to the characteristics of sudden failures.

[0064] After obtaining the initial attention weights for each layer of features using the size-aware correction operator, a global gating gain mechanism is further introduced to avoid the mis-amplification of environmental noise under healthy operating conditions. Gating gain Subject to the global system disorder index symplectic geometrical entropy The real-time drive works similarly to a physical switch controlling the state of a transformer: when the transformer is in a healthy state, the signal exhibits high orderliness, and the spectral entropy deviates very little from the mean. At this time, the gated output approaches zero, effectively filtering random noise from the electromagnetic environment. When a slight internal degradation occurs, and the dynamic system evolves into a chaotic state, causing a surge in spectral entropy, the gate will... Guided by a specific scale, the algorithm rapidly activates, performing nonlinear gain amplification on specific fault modes. A Softplus activation function is employed to ensure the smoothness of the gain gradient, avoiding drastic oscillations in diagnostic conclusions at critical states. The dynamic gain weights... The triggering logic is as follows: (15). Among them, It serves as a global scaling factor, used to control sensitivity to anomalous responses through training. Bias term; smooth activation function The mathematical expression is It acts as a unidirectional amplification mechanism. Unlike ReLU, it is smooth near the zero point. When the component behaves normally, it outputs a small, stable gain; when the component behaves abnormally, the gain increases linearly. It ensures the weights... Constant is positive; The current order The average value of the component entropy spectrum; Std(H) sym )for The standard deviation.

[0065] S3.2 Perform feature nonlinear correlation fusion.

[0066] During the fusion process, the temporal characteristics of each path from the evolutionary branch are... First, with the association matrix The accumulated values ​​of each row direction are multiplied. The physical significance of this step lies in aligning and correcting the features based on the mechanical coupling strength. Subsequently, the corrected feature flow is multiplied by the gate weights. The Hadamard product is performed to enhance fault features and suppress redundant features. The fused feature vector is then obtained. (16). Among them, For the first The depth feature vector output by the branch of the road mode evolution is derived from the first... Each symplectic geometric component The dimension of the hidden layer neurons is determined by the preset number of neurons and is extracted by the adaptive dilated convolution operator. It is used to characterize the deep nonlinear evolution characteristics of a specific physical mode in the time domain. This is the Hadamard product operation. This two-layer filtering mechanism ensures the quality of the fused feature vector. It preserves the evolutionary details of the time series and embeds topological structure-preserving logic of the symplectic space, enabling the model to extract weak fault indicators in the context of strong noise.

[0067] S3.3 uses a feature concatenation operation to fuse the feature vectors. Indicators of static physical background Integration, output diagnostic results : (17). Where Prob is the final output probability; Softmax is the normalized activation function; W out b is the trainable weight matrix of the output layer; out This is the bias term for the classification output layer. Physical background features are preserved through feature concatenation. The independence of the classification layer allows it to determine specific fault types, such as loose windings, overheating of the core, and mechanical impact, by considering not only dynamic vibration evolution information but also the current overall energy topology of the transformer, ultimately resulting in a higher probability distribution. It has high credibility.

[0068] S4 Model Optimization Based on Chaotic Ocean Predator Algorithm.

[0069] like Figure 3 As shown, the parameter set of the model to be optimized is mapped to the predator position vector in the Chaotic Ocean Predator Algorithm (CMPA). To overcome the uneven distribution of hyperparameters in the initial stage of hyperparameter search in the standard marine predator algorithm, this invention employs a piecewise chaotic mapping to construct the initial parameter population: (18). The mapping control parameter is set to 0.4. The initial parameter set generated through this mapping has higher ergodicity and randomness, ensuring that CMPA can cover the parameter search space of the transformer under different load conditions. The CMPA algorithm simulates the predation behavior of marine organisms at different speed ratios, performing multi-scale searches of network parameters. Exploration phase: Brownian motion is used to quickly locate potential optimal parameter intervals globally, preventing the model from getting trapped in local optima. Transition phase: Half of the individuals in the population perform Levi's leaps for exploration, while the other half perform Brownian motion for refined development. Convergence phase: Predators rapidly approach the optimal solution using Levi's leaps, targeting, for example,... The fine parameters of the gating threshold are finely adjusted at the centimeter level.

[0070] Based on the model construction of this invention, this invention designs a comprehensive fitness function that integrates diagnostic accuracy and physical-logical constraints: (19). Where N is the total number of samples; This is an indicator function that takes the value 1 when the prediction is correct. For the first Each sample model predicts the fault category; For the first Each sample contains a true fault label; λ is the weighting coefficient. This represents the symplectic geometric energy distribution vector extracted in the first stage. This function not only requires maximizing diagnostic accuracy but also demands that the predicted probability distribution output by the model be topologically consistent with the physical energy distribution of the original signal. Through this closed-loop feedback, the algorithm is forced to eliminate mathematically accurate but physically incompatible parameter combinations.

[0071] The present invention discloses a transformer fault diagnosis system based on symplectic geometric mode decoupling, which is used to execute the transformer fault diagnosis method based on symplectic geometric mode decoupling of the present invention. The system includes a data acquisition and sensing module, a communication and data transmission module, a core computing and diagnostic analysis server, and a data storage and human-computer interaction module.

[0072] The data acquisition and sensing module includes a piezoelectric vibration sensor array, a signal conditioning circuit, and a high-frequency analog-to-digital converter. The vibration sensor array is installed on the outer wall of the transformer tank. The vibration sensor array is used to capture simulated mechanical vibration signals of the transformer under operating conditions in real time. The acquired simulated signals are filtered and amplified by the signal conditioning circuit for anti-aliasing, and then converted into discrete one-dimensional vibration digital sequences by the high-frequency analog-to-digital converter.

[0073] The communication and data transmission module is connected to the output of the data acquisition and sensing module. The communication and data transmission module includes an industrial Ethernet interface and a fiber optic transceiver terminal, which is used to transmit the acquired high-frequency vibration digital sequence to the core computing and diagnostic analysis server without loss and with low latency, so as to ensure the reliability and real-time performance of data transmission in the strong electromagnetic interference environment at the transformer site.

[0074] The core computing and diagnostic analysis server is equipped with a central processing unit and tensor acceleration hardware units. Internally, it contains and runs core diagnostic analysis logic. After receiving the vibration digital sequence from the communication link, the server first constructs a Hamiltonian matrix using a symplectic geometric mode decomposition algorithm and reconstructs the signal. It then extracts the energy proportion, fourth-order kurtosis, and symplectic geometric spectral entropy of each modal component, concatenating them to form a multi-dimensional physical indicator feature vector. Subsequently, the dual-branch feature depth mapping unit within the server processes the data in parallel. The modal evolution branch uses adaptive dilated convolution based on physical characteristics to extract temporal dynamic evolution features, while the physical indicator branch transforms the energy distribution into a physical background tensor through a residual multilayer perceptron. The server also includes constrained adaptive fusion decision logic and global optimization logic. The constrained adaptive fusion decision logic constructs a symplectic correlation matrix and introduces a global gated gain mechanism driven in real-time by symplectic geometric spectral entropy. This nonlinearly correlates and fuses the temporal dynamic evolution features with the physical background tensor, thereby outputting an accurate digital diagnostic signal for transformer fault classification. Meanwhile, the core computing and diagnostic analysis server adopts the ocean predator algorithm improved by piecewise chaotic mapping, and uses a comprehensive fitness function that integrates diagnostic accuracy and physical energy distribution topology consistency as the feedback target to perform closed-loop optimization and real-time iterative update of the network framework hyperparameters within the system.

[0075] The data storage and human-computer interaction module is signal-connected to the core computing and diagnostic analysis server, and specifically includes a large-capacity database storage array and an interactive display terminal. The database storage array is used for persistent storage and management of massive amounts of historical vibration raw data, symplectic component feature libraries, and optimal network parameters after optimization iterations. The interactive display terminal is used to render the transformer health status topology in real time, dynamically display feature evolution curves, and ultimately output intuitive fault alarm diagnosis results and status assessment reports to on-site maintenance personnel.

[0076] This invention effectively solves the problem in the background technology that traditional signal processing methods are difficult to decouple complex vibration modes and easily lose signal physical topology information by introducing symplectic geometric mode decomposition with structure-preserving properties. By constructing a SMAG-Net network based on energy-frequency adaptive mapping and a dynamic gating fusion mechanism, it eliminates the bottlenecks of insufficient feature extraction accuracy and poor robustness under variable load conditions caused by fixed convolution scale and lack of physical logic in traditional deep learning models, achieving deep integration of data-driven and mechanical vibration coupling laws. Finally, by introducing a CMPA algorithm based on piecewise chaotic mapping, it overcomes the defects of existing optimization strategies that are prone to getting trapped in local optima and have low search efficiency, ensuring the global optimality of model hyperparameters in complex constraint space. In summary, this invention comprehensively makes up for the shortcomings of existing technologies from signal processing and architecture design to global optimization, significantly improving the reliability and accuracy of transformer fault diagnosis in high-noise environments.

Claims

1. A transformer fault diagnosis method based on symplectic geometric mode decoupling, characterized in that, Includes the following steps: S1 Transformer signal mode decoupling and physical feature extraction based on symplectic geometry; S1.1 Acquire vibration signal sequence of transformer under operating conditions , where x n It is a vibration signal; S1.2 Constructing the Hamiltonian matrix ; S1.3 Diagonalization is performed to solve the problem. symplectic eigenvalues Using symplectic orthogonal matrices For the original trajectory matrix By projection, the symplectic component matrix is ​​obtained. ; S1.4 pairs Perform diagonal averaging. elements in Calculate the reconstructed first Each time point component Calculate each component Energy percentage and fourth-order kurtosis ; S1.5 Calculation of symplectic geometrical entropy ; S1.6 will , and Row vectors are formed by horizontal splicing. ; S2 performs deep mapping of fault features based on the symplectic manifold correlation gated fusion network SMAG-Net; Through Residual Multilayer Perceptron (Res-MLP) Transformed into a network-recognizable physical context tensor , Where σ() is a nonlinear activation function; A neural network whose hidden layers can be learned; These are the learnable bias terms for each layer; This is the weight matrix; S3 Feature Adaptive Fusion and Decision Based on Symplectic Space Physical Constraints; S3.1 Constructing the symplectic correlation matrix A global gating gain mechanism is introduced; S3.2 Perform feature nonlinear correlation fusion to obtain feature vectors. ; S3.3 uses feature splicing operations to... Indicators of static physical background Integration, output diagnostic results ; S4 Model Optimization Based on Chaotic Oceanic Predator Algorithm; The parameter set of the model to be optimized is mapped to the predator position vector in the Chaotic Ocean Predator Algorithm (CMPA). Design a comprehensive fitness function that integrates diagnostic accuracy and physical-logical constraints; comprehensive fitness function Where N is the total number of samples; For indicator functions; For the first Each sample model predicts the fault category; For the first Each sample contains a true fault label; λ is the weighting coefficient. This refers to the symplectic geometric energy distribution vector extracted in the first stage; These are mapping control parameters.

2. The transformer fault diagnosis method based on symplectic geometric mode decoupling according to claim 1, characterized in that, Step S1.2 specifically includes: using the symplectic geometric mode decomposition (SGMD) algorithm to... Reconstruction ;in, This is the total length of the signal; For the embedded dimension; Subsequently, a symmetric matrix is ​​constructed. and based on structure .

3. The transformer fault diagnosis method based on symplectic geometric mode decoupling according to claim 2, characterized in that, In step S1.3, the symplectic similarity transformation matrix is ​​used. right Diagonalization is performed; among which, ; An identity symplectic matrix; It is a symplectic eigenvalue matrix. is the symplectic eigenvalue.

4. The transformer fault diagnosis method based on symplectic geometric mode decoupling according to claim 3, characterized in that, The calculation process is as follows: For each pair of symplectic eigenvalues Find its corresponding feature vector pair ( , Make it simultaneously satisfy both the mapping relation and the symplectic orthogonality relation: ; Among them, v i for The corresponding feature vector will The symmetric orthogonal matrix is ​​obtained by horizontally concatenating the column vectors that have undergone symmetric orthogonalization. ;use right Projecting, for the first Each mode has a symplectic component matrix after projection. ;in, It is extraction The column vectors corresponding to specific symplectic feature pairs are truncated, and their first two columns are extracted. Construct the projection vector from the dimensional components.

5. The transformer fault diagnosis method based on symplectic geometric mode decoupling according to claim 4, characterized in that, In step S1.3, for elements in The i-th component after reconstruction ;in, For the first The number of elements on each anti-diagonal line; for The order of the vector is determined; subsequently, the energy proportion of each component is calculated, and the energy proportion of the Kth order component is denoted as... , Construct energy distribution vector Calculate principal components Fourth-order kurtosis ;in, for exist Amplitude at any given moment; This is the arithmetic mean of the component over the entire time span; This represents the total number of sampling points.

6. The transformer fault diagnosis method based on symplectic geometric mode decoupling according to claim 5, characterized in that, In step S1.4, using calculate ;in, This represents the normalized symptotic eigenvalue.

7. The transformer fault diagnosis method based on symplectic geometric mode decoupling according to claim 1, characterized in that, In step S3.1, the symplectic correlation matrix elements Where tanh is the hyperbolic tangent activation function; symplectic inner product term To capture the dynamic correlation of signals in phase space; a scale-aware correction operator is introduced. , where s min The minimum step size; s i，l To experience the wild stride length; For kurtosis; dynamic gain weights of the global gating gain mechanism. ;in, This is the global scaling factor; For bias terms; For smooth activation functions; The current order The average value of the component entropy spectrum; Std(H) sym )for The standard deviation.

8. The transformer fault diagnosis method based on symplectic geometric mode decoupling according to claim 7, characterized in that, In step S3.2, the fused feature vector ;in, For the first The depth feature vector output by the branch of the road mode evolution; This is for the Adama product operation.

9. The transformer fault diagnosis method based on symplectic geometric mode decoupling according to claim 8, characterized in that, In step S3.3, Where Softmax is the normalized activation function; W out b is the trainable weight matrix of the output layer; out This is the bias term for the classification output layer.

10. The transformer fault diagnosis method based on symplectic geometric mode decoupling according to claim 9, characterized in that, In step S4, a piecewise chaotic mapping is used to construct the initial parameter population: .

11. A transformer fault diagnosis system based on symplectic geometric mode decoupling, used to execute the transformer fault diagnosis method based on symplectic geometric mode decoupling as described in claim 1, characterized in that, The system includes a data acquisition and sensing module, a communication and data transmission module, a core computing and diagnostic analysis server, and a data storage and human-computer interaction module. The data acquisition and sensing module captures simulated mechanical vibration signals of the transformer under operating conditions in real time. After anti-aliasing filtering and amplification, the acquired simulated signals are converted into discrete one-dimensional vibration digital sequences. The communication and data transmission module is connected to the output of the data acquisition and sensing module and transmits the vibration digital sequences to the core computing and diagnostic analysis server. Upon receiving the vibration digital sequences, the core computing and diagnostic analysis server constructs a Hamiltonian matrix using a symplectic geometric mode decomposition algorithm and reconstructs the signal, extracting the energy proportion of each modal component. The fourth-order kurtosis and symplectic geometric spectral entropy are concatenated to form a multi-dimensional physical indicator feature vector. Subsequently, an adaptive dilated convolution based on physical characteristics is used to extract the temporal dynamic evolution features. The energy distribution is transformed into a physical background tensor through a residual multilayer perceptron. By constructing a symplectic correlation matrix and introducing a global gated gain mechanism driven in real time by the symplectic geometric spectral entropy, the temporal dynamic evolution features and the physical background tensor are nonlinearly correlated and fused to output an accurate digital diagnostic signal for transformer fault classification. The data storage and human-computer interaction module is connected to the core computing and diagnostic analysis server for persistent storage and management of massive historical vibration raw data, symplectic component feature library, and optimal network parameters after optimization iteration.

Images