A transformer fault alarm positioning method and system

Abstract

This application provides a transformer fault alarm and location method and system, belonging to the field of intelligent monitoring and fault diagnosis of power equipment, to solve the problems of information loss from multidimensional heterogeneous data, delayed fault warning, and ambiguous physical location in related technologies. The method embeds the multidimensional data tensors acquired by a sensor array into a low-dimensional smooth manifold to construct a neural differential manifold model to characterize the continuous evolution of states, and identifies fault attractors at the macroscopic level based on causal emergence theory. It determines the key directions of state evolution by calculating the gradient of the holographic attention field on the manifold, and maps it back to the original data space to generate a difference field, thereby achieving accurate physical location and early alarm of the fault source. The system includes a holographic sensing array implementing the above method, a manifold computing processing unit, and an interactive terminal.

Description

Technical Field

[0001] This application relates to the field of intelligent monitoring and fault diagnosis of power equipment, and in particular to a method and system for locating transformer fault alarms. Background Technology

[0002] As a core piece of equipment in the power system, the operating status of transformers directly affects the safety and stability of the power grid. Real-time and accurate fault monitoring and location of transformers are crucial for preventing major accidents and achieving predictive maintenance.

[0003] Currently, mainstream transformer fault diagnosis technologies primarily rely on the analysis of multi-source heterogeneous monitoring data, including dissolved gases in oil, vibration signals, and partial discharges. Existing technologies typically combine feature engineering with machine learning models. For example, time-frequency domain features are extracted from vibration signals, gas ratios are calculated from chromatographic data, and these manually designed features are then input into classifiers (such as support vector machines or neural networks) for fault identification and classification. More advanced methods attempt to automatically extract features using deep learning models (such as convolutional neural networks) or utilize multimodal fusion techniques for comprehensive judgment.

[0004] However, these existing technologies still have significant shortcomings. First, high-dimensional, heterogeneous sensor data suffers information loss or distortion during feature extraction and fusion, making it difficult to preserve the data's inherent complete topology and evolutionary relationships. Second, most methods treat fault diagnosis as a static classification or regression problem, failing to fundamentally model the continuous dynamic evolution of equipment status over time, especially lacking the ability to characterize the entire process from fault initiation to development and outbreak, resulting in delayed early warnings and ambiguous localization. Finally, existing solutions typically separate the perception, diagnosis, and decision-making modules, lacking a unified and self-consistent mathematical framework to describe the complete cognitive loop from data to state to action. Summary of the Invention

[0005] This application provides a transformer fault alarm location method and system, which can realize early and accurate alarm and physical spatial location of transformer faults from a unified and dynamic geometric evolution perspective.

[0006] Firstly, this application provides a transformer fault alarm location method. The method includes the following steps: acquiring the original data tensor of a transformer multi-dimensional sensor array, the original data tensor containing data in time dimension, spatial location dimension, and physical quantity type dimension; embedding the original data tensor into a low-dimensional smooth manifold to obtain a manifold coordinate sequence characterizing the overall state of the transformer; wherein the embedding process is achieved through high-order tensor decomposition combined with spectral manifold regularization; constructing a dynamic manifold evolution model based on the manifold coordinate sequence, the dynamic manifold evolution model describing the continuous evolution process of the transformer state on the manifold; identifying fault attractors and their attraction domains on the manifold based on the dynamic manifold evolution model; calculating the distance between the current manifold coordinates and the identified fault attractor attraction domain; when the distance is less than a preset threshold; determining the key manifold direction that triggers the state evolution based on the current manifold coordinates and the dynamic manifold evolution model; and mapping the key manifold direction inversely to the original data tensor space to generate a difference field with the original data, thereby locating the fault source in the physical space and generating alarm information.

[0007] By adopting the above technical solution, this application unifies the mapping of high-dimensional heterogeneous sensor data onto a low-dimensional smooth manifold, constructing a geometric representation of the device state. Learning a dynamic evolution model on this manifold allows for a continuous and natural depiction of state change trajectories. By identifying fault attractors on the manifold, the system can achieve early warning based on the distance between the current state and the attractor's "potential well." Finally, by analyzing and reverse-mapping the key geometric directions of state evolution, abstract fault modes can be accurately located back to their specific physical locations, thus solving the problems of information fragmentation, delayed warnings, and inaccurate positioning inherent in traditional methods.

[0008] Furthermore, the step of embedding the original data tensor into a low-dimensional smooth manifold specifically includes: performing Tucker decomposition on the original data tensor to obtain a core tensor and corresponding factor matrices of multiple dimensions; introducing spectral manifold constraints into the objective function of the decomposition; the spectral manifold constraints are implemented by using the quadratic form of the Laplacian matrix constructed based on data similarity as a regularization term; performing singular value decomposition based on the expanded matrix of the core tensor; or performing a specific combination of the factor matrices to calculate the manifold coordinate sequence.

[0009] By employing the above technical solution, high-order tensor decomposition is used to directly process multidimensional data tensors, avoiding the loss of structural information caused by flattening the data. The introduction of spectral manifold regularization constraints forces the embedding process to maintain the local neighborhood relationships between samples in the original high-dimensional data space, thereby obtaining a low-dimensional manifold representation that better reflects the essential topological structure of the data, laying the foundation for subsequent accurate dynamic analysis.

[0010] Furthermore, the dynamic manifold evolution model is learned by a neural differential manifold defined on the low-dimensional smooth manifold. The neural differential manifold is constructed as a neural network operating on the tangent bundle of the low-dimensional smooth manifold to learn the drift term function and diffusion term function of the manifold coordinate sequence evolution. The drift term function and diffusion term function together constitute a stochastic differential equation on the manifold to describe the continuous evolution of the state.

[0011] By employing the above technical solution, and using a neural differential manifold defined on the tangent space of the manifold to learn the evolutionary laws, it is ensured that all operations are strictly performed within the geometric structure of the manifold, avoiding the errors introduced by projecting manifold data onto Euclidean space for modeling. Describing the evolution using stochastic differential equations can characterize both deterministic trends (drift terms) and accommodate uncertainties (diffusion terms), thus better reflecting the true continuous dynamics of the physical system.

[0012] Further, the step of identifying fault attractors and their attraction domains on the manifold includes: performing a coarse-grained mapping on the manifold coordinates to obtain macroscopic state variables; the coarse-grained mapping is obtained by solving a learning process that maximizes the difference between the mutual information between macroscopic state variables and the mutual information between macroscopic states and microscopic states; constructing macroscopic dynamic equations in the space formed by the macroscopic state variables; solving for the stable fixed points of the macroscopic dynamic equations; using the stable fixed points as fault attractors; calculating the Jacobian matrix of the macroscopic dynamic equations at the stable fixed points; and determining the boundary of the attraction domain of the fault attractor based on the eigenvalues ​​of the Jacobian matrix.

[0013] By adopting the above technical solution and introducing a coarse-grained process based on maximizing effective information, it is possible to automatically extract concise and highly predictive macroscopic variables from complex microscopic evolution. These macroscopic variables often correspond to understandable equipment functional states. Identifying fault attractors and their attraction domains at the macroscopic level gives fault early warnings clearer physical meaning and stronger robustness.

[0014] Furthermore, the macroscopic dynamic equation includes causal emergence terms, which are learned through a neural causal emergence network. The structure of the neural causal emergence network is constrained so that its output function cannot be represented by a linear combination of the input microscopic state variables.

[0015] By adopting the above technical solutions, "causal emergent terms" at the macroscopic level are clearly modeled. These terms represent new laws and patterns that do not exist at the microscopic component level but emerge at the system-wide level. This enables the system to capture and explain sudden failure modes caused by complex interactions that cannot be predicted by simple linear extrapolation, thus improving the foresight and depth of diagnosis.

[0016] Further, the step of determining the key manifold direction that triggers the state evolution based on the current manifold coordinates and the dynamic manifold evolution model includes: constructing a holographic attention field on the manifold; the value of the holographic attention field at any point is obtained by weighting and integrating all points on the manifold with a holographic kernel function to measure the similarity between the similarity and the fault attractor region pointed to by that point; calculating the covariant derivative of the holographic attention field at the current manifold coordinates; and taking the direction of the covariant derivative as the key manifold direction.

[0017] By employing the aforementioned technical solution and constructing a holographic attention field, the fault tracing problem is transformed into a vector field analysis on a manifold. The key direction given by the covariant derivative geometrically indicates the steepest sliding path from the current state point to the fault attractor. This localization method, based on the overall dynamic characteristics of the system, possesses profound mathematical interpretation and physical significance, resulting in more reliable localization results.

[0018] Furthermore, the method also includes constructing a state-action joint manifold based on the macroscopic state variables, the fault attractor information, and a predefined set of maintenance actions, defining a value function manifold on the state-action joint manifold, obtaining the optimal policy function from the macroscopic state to the optimal maintenance action by solving the Hamilton-Jacobi-Bellman equation on the state-action joint manifold, and generating a recommended maintenance strategy based on the current macroscopic state and the optimal policy function.

[0019] By adopting the above technical solution, the decision optimization problem is also incorporated into a unified manifold geometric framework. Solving the optimal control problem on the state-action joint manifold ensures that the generated maintenance strategy is mathematically tightly coupled with the dynamic evolution law of the system. This enables the automatic weighing of the impact of different actions on various aspects such as delaying failures, ensuring safety, and controlling costs, achieving an intelligent closed loop from diagnosis to decision-making.

[0020] Furthermore, the step of obtaining the optimal policy function by solving the Hamilton-Jacobi-Bellman equation is approximately implemented by a deep game network through adversarial training. The deep game network includes a policy network and a value network. The policy network is used to generate parameterized action policies on the state-action joint manifold, and the value network is used to evaluate the long-term expected return of the action policy. The two are jointly optimized through an adversarial learning process based on gradient descent on the manifold.

[0021] By employing the above technical solution, the adversarial training mechanism of deep game networks is used to approximate the solution of complex optimal control problems on manifolds. The policy network and value network co-evolve in the adversarial process, eventually approximating Nash equilibrium, i.e., the optimal policy. This method provides a feasible technical path for efficiently solving optimal policies in high-dimensional continuous state-action spaces.

[0022] Secondly, this application provides a transformer fault alarm and location system. The system is used to implement the transformer fault alarm and location method as described in any one of the first aspects. The system includes: a holographic sensing array comprising multiple sensor nodes integrating vibration, ultrasonic, and temperature sensing units; the sensor nodes achieve coherent sensing between the arrays via metasurface antennas for collaboratively acquiring multidimensional raw data tensors of the transformer; a manifold computation processing unit including a hardware acceleration unit for performing large-scale tensor shrinkage and manifold embedding computations, and a dedicated processing unit for processing graph structures and manifold topology operations; and an interactive terminal including a light field display device and a human-machine interface for visualizing the manifold state, evolution process, and alarm location results, and receiving interactive commands.

[0023] By adopting the above technical solutions, this application provides a dedicated hardware system that matches the aforementioned methods. The holographic sensing array natively supports multidimensional coherent data acquisition, providing a high-quality data foundation for manifold analysis. A dedicated manifold computing processing unit provides hardware acceleration for tensor operations and geometric operations in the core algorithms, ensuring the real-time performance of complex algorithms in practical engineering. The interactive terminal provides intuitive visualization and interaction methods, enabling maintenance personnel to effectively understand and utilize advanced diagnostic results.

[0024] In summary, this application has at least the following beneficial effects:

[0025] A new paradigm for transformer fault diagnosis based on geometric manifold and dynamic evolution is provided, which enables early and accurate alarm and physical location of faults.

[0026] By using manifold embedding and regularization techniques, the topological structure of multi-source heterogeneous data is fully preserved, laying the foundation for accurate modeling.

[0027] By identifying causal emergent fault attractors at the macroscopic level, it is possible to capture complex fault precursor patterns that are difficult to detect using traditional methods.

[0028] By unifying diagnosis and decision-making within a manifold geometry framework, a complete autonomous cognitive closed loop from state perception to the generation of optimal maintenance strategies is achieved.

[0029] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0030] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0031] Figure 1 A schematic diagram of a transformer fault alarm and location system according to an embodiment of this application is shown.

[0032] Figure 2 A flowchart of a transformer fault alarm location method according to an embodiment of this application is shown. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0035] This application provides a transformer fault alarm location method and system, which maps multi-source monitoring data to a unified geometric manifold and models its dynamic evolution. This fully preserves the inherent correlation of the data, enables early and accurate warning of faults, and can reverse the location of abstract fault modes to specific physical locations, fundamentally improving the timeliness, accuracy, and interpretability of diagnosis.

[0036] In a first aspect, embodiments of this application disclose a transformer fault alarm and location system.

[0037] Figure 1 A schematic diagram of a transformer fault alarm and location system according to an embodiment of this application is shown.

[0038] Reference Figure 1 This system is a dedicated intelligent monitoring and diagnostic device. Its overall architecture consists of three core components that are functionally tightly coupled: a holographic sensing array, a manifold computing processing unit, and an interactive terminal. These three parts are electrically connected and exchange information with the control network through a pre-set data bus, jointly realizing holographic perception, in-depth computational analysis, and intuitive interactive presentation of the transformer's operating status.

[0039] The holographic sensing array serves as the system's data acquisition front-end, physically comprising a network of multiple distributed sensor nodes. These nodes are fixed to key monitoring locations on the transformer body via bolts or strong magnetic adsorption. Specific locations include the side walls and top cover of the main oil tank, the riser flanges of the high-voltage and low-voltage bushings, the inlet and outlet oil pipes of the cooling system, and the oil chamber housing of the on-load tap changer. The number of nodes is determined based on the transformer's capacity and physical dimensions, typically ranging from 8 to 32 sensor nodes in an array, preferably 16. Each sensor node structurally integrates three sensing units based on different physical principles: a broadband piezoelectric accelerometer for capturing mechanical vibrations, with a frequency response range covering 10 Hz to 10 kHz; a high-sensitivity ultrasonic sensor for detecting partial discharges or acoustic emissions from internal particle impacts, with a center frequency of 150 kHz; and a patch-type non-contact infrared thermometer for measuring the surface temperature of the mounting point, with a measurement accuracy of ±0.5 degrees Celsius. These heterogeneous sensing units are collectively encapsulated within a cylindrical metal protective housing, approximately 50 mm in diameter and 30 mm in height, with an IP67 protection rating, and share a unified power supply and communication module. Specifically, each node's housing integrates a metasurface antenna based on a dielectric metamaterial design. This antenna, composed of periodically arranged subwavelength microstructures, allows for precise modulation of the phase of incident or emitted electromagnetic waves. By pre-programming the radiation phase of each node's metasurface antenna, the array achieves "coherent sensing" between nodes. Specifically, a unit designated as the master node periodically broadcasts a radio frequency synchronization signal containing a precise timestamp. Upon receiving this signal, the remaining slave nodes use phase-locked loop technology and an internal high-stability crystal oscillator to synchronize their sampling clocks with the master node, achieving a synchronization accuracy better than 1 microsecond. All nodes acquire data under this unified clock, ensuring that the acquired vibration, ultrasonic, and temperature data are not only strictly aligned in time, but also, because the spatial coordinates of each node are precisely measured and entered into the system beforehand, the data acquired by the entire array also possesses a known coherent relationship in the spatial dimension defined by the phase distribution of the metasurface antenna. This natively constitutes a raw data tensor containing three dimensions: time, spatial location, and physical quantity type. For example, for a data block containing 16 nodes, each node acquiring 3 physical quantities, and continuously acquiring data for 1 second at a sampling frequency of 10 kHz, its raw data tensor has the following dimensions: 10,000 points in the time dimension, 16 points in the spatial location dimension, and 3 points in the physical quantity type dimension.

[0040] The manifold computing processing unit is the core computing brain of the system. It receives raw data tensors from the holographic sensing array via industrial Ethernet and executes complex manifold geometry algorithms. Physically, this unit is a standalone rack-mount server, and in hardware, it consists of two types of dedicated processor boards interconnected via a high-speed PCIe 4.0 bus. The first type of board is a hardware acceleration unit for performing large-scale tensor shrinking and manifold embedding computations. This unit is typically implemented based on an ASIC chip or a high-capacity FPGA chip containing a large number of dedicated tensor computing cores, with a peak computing power of at least 100 TOPS and equipped with high-bandwidth HBM2e memory. It is specifically optimized for efficiently performing high-order tensor decomposition operations, such as Tucker decomposition of the aforementioned raw data tensor, solving for the core tensor and multiple factor matrices through iterative algorithms, and efficiently handling large-scale matrix multiplications and singular value decompositions involved in the decomposition process. This is a key computational step in embedding high-dimensional data into a low-dimensional manifold, typically reducing thousands of dimensions of raw data to a 10- to 100-dimensional manifold space, with a preferred manifold dimension of 20. The second type of board consists of dedicated processing units for handling graph structures and manifold topology operations. These units typically employ a GPU that supports massively parallel threads and complex conditional branches, or a processor specifically designed for graph computation. They excel at performing sparse matrix operations related to spectral graph theory, such as constructing adjacency graphs based on data point similarity and calculating their Laplacian matrices. They also handle differential geometric operations defined on manifolds, such as calculating the covariant derivatives of scalar fields (e.g., attention fields) on manifolds. These two types of processing units exchange intermediate computation results through a shared unified memory space, enabling efficient data flow between them. Together, they complete a series of computational tasks, from data manifold construction and dynamic evolution model learning to fault attractor identification and critical direction analysis. For example, the hardware acceleration unit first performs manifold embedding, obtaining a manifold coordinate sequence; the dedicated processing unit then learns a neural differential manifold model based on this sequence and identifies attractors corresponding to typical faults such as "multiple grounding points in the iron core" and "severe winding deformation" in the macroscopic state space. The entire calculation process trains and updates the model based on a 10-minute historical dataset, and the processing delay for new data can be controlled within 5 seconds.

[0041] The interactive terminal serves as the system's human-machine interface and decision support output, typically deployed in the substation's main control room or maintenance center. It includes a light field display device based on microlens array technology or parallax barrier technology, capable of projecting a 3D image with realistic depth information to the viewer without the need for special glasses. This display device has a horizontal viewing angle of at least 60 degrees and a resolution of at least 4K. The system utilizes this device to transform the abstract mathematical results generated by the manifold computation processing unit into intuitive, dynamic 3D visualizations. For example, the system can project a 20-dimensional manifold into a 3D space using principal component analysis to display its approximate topological shape and use a moving light spot trajectory to represent the evolution of the current equipment state on the manifold. When the system identifies a fault attractor, it marks it in the 3D manifold projection with a spherical region of a specific color, such as red. Most importantly, the system can reverse-map the calculated "key manifold direction" to generate a physical spatial difference field, which is then superimposed on the corresponding surface of the transformer's 3D digital model in the form of a heat map or a 3D vector field. This accurately indicates the physical location where fault vibration or discharge is most likely to occur, for example, displaying a high-temperature red area at the base of the A-phase bushing. The terminal also integrates multiple natural interaction interfaces, including a 21-inch multi-touch capacitive screen for basic menu operations; a set of depth-sensing cameras to capture the operator's gestures in front of the screen, such as grabbing, dragging, and rotating the 3D model; and a directional microphone and voice recognition module, supporting operation via voice commands such as "zoom in here" and "show historical curves." Maintenance personnel can freely explore the visualized content through these interfaces or input commands to trigger the system to generate a comprehensive diagnostic report containing detailed spectrum, trend analysis, and maintenance recommendations.

[0042] In summary, the system in this embodiment acquires native, spatiotemporally coherent multidimensional data tensors through a holographic sensing array. Two types of dedicated hardware in the manifold computing processing unit then collaborate to perform in-depth analysis and diagnosis from data to a geometric dynamic model. Finally, the complex analysis results are presented to the user in an immersive, interactive 3D visualization via an interactive terminal. Through clear functional division and efficient data pipeline connections, the various components collectively achieve intelligent processing of the entire process for early detection, precise location, and decision support for latent faults in transformers.

[0043] Secondly, this application discloses a method for locating transformer fault alarms.

[0044] Figure 2 A flowchart of a transformer fault alarm location method according to an embodiment of this application is shown.

[0045] Reference Figure 2This method transforms multi-source, heterogeneous transformer monitoring data into a profound understanding of equipment health status through a unified geometric and dynamic framework, ultimately achieving early fault warning, precise physical location, and intelligent maintenance decisions. The core execution logic of this method is a complete cognitive loop, moving from high-dimensional data to low-dimensional manifold geometry, and then from abstract geometric analysis back to concrete physical space.

[0046] First, the system acquires a raw data tensor that is strictly aligned in time, space, and physical quantity type from the sensor array deployed on the transformer itself. Specifically, this data tensor is denoted as... Among them, dimension This represents the number of time sampling points; for example, sampling at 10 kHz within 1 second. ; dimension This represents the total number of sensor nodes distributed in space, and its value can be between 8 and 32. A preferred implementation is... ; dimension The number representing the types of physical quantities, for example, when each node collects three signals: vibration, ultrasound, and temperature. Each element of this tensor It has a clear physical meaning, representing a point in time. Spatial nodes The first measured The numerical values ​​of physical quantities. These data are natively provided by hardware arrays with coherent sensing capabilities, ensuring the inherent correlation of the data in the temporal and spatial dimensions.

[0047] Next, the above high-dimensional data tensor Embedding into a low-dimensional smooth manifold is fundamental to all subsequent analyses. The core algorithm for this step is Tucker tensor decomposition, which incorporates spectral manifold regularization. The goal of Tucker decomposition is to approximate the original tensor as:

[0048]

[0049] in, This is called the core tensor, and its dimension is... It is a preset, relatively small value, usually much smaller than the corresponding value. For example, it is advisable .matrix , , These are the factor matrices, respectively representing time, space, and characteristic modes. Decomposition is achieved by minimizing the reconstruction error. To achieve this.

[0050] To ensure that the low-dimensional representation obtained from the decomposition preserves the local structure of the original high-dimensional data, this method introduces spectral manifold regularization. Using the time factor matrix... For example, the first step is to construct a similarity graph for the time series. This involves calculating the time points... and similarity For example, the cosine similarity of all corresponding sensor data vectors can be used. A similarity matrix can then be constructed based on this. Then calculate the graph Laplacian matrix. ,in It is an angle matrix, its diagonal elements The spectral manifold regularization term is defined as follows: This penalty penalizes the differences between the row vectors of the factor matrix at similar time points, thus forcing the manifold representation to be smooth. Ultimately, the optimization problem to be solved is:

[0051]

[0052] in It is a regularization coefficient used to balance reconstruction accuracy and manifold smoothness.

[0053] After decomposition, the manifold coordinate sequence It can be obtained in one of the following two ways: one is to obtain the core tensor Expanded into a matrix along the time dimension Then, perform singular value decomposition on the matrix, and take the first left singular vector. The column is the main component, in which The first option is to define the manifold dimension, which can be between 10 and 100, with 20 being the preferred value; the second option is to directly use the time factor matrix. As manifold coordinates, at this time Regardless of the method used, the final result is a sequence. Each of them These concisely represent the transformer at all times The overall state.

[0054] After obtaining the manifold coordinate sequence, a dynamic evolution model needs to be constructed on it. This method uses a neural differential manifold to learn this dynamic. The low-dimensional manifold is denoted as... The goal is to learn a stochastic differential equation on a manifold:

[0055]

[0056] here, It is a drift term, representing a deterministic evolutionary trend; This is the diffusion term, representing the intensity of random noise; It is a standard Wiener process; These are the parameters of the neural network. The key innovation of neural differential manifolds lies in the network... and It is designed to operate on the tangent bundle of the manifold. In its specific implementation, the network uses a logarithmic mapping of the current point. (A tangent vector) is used as part of the learning objective, and an exponential mapping is utilized. To ensure that the next predicted point is always on the manifold The model is trained by maximizing the likelihood function of the observed sequence, or minimizing the geodesic distance between the predicted and actual points. This model can accurately depict the continuous, stochastic evolution trajectory of states on the manifold, providing a foundation for prediction and diagnosis.

[0057] Based on the dynamic model, identifying fault attractors and their attraction domains is the core of early warning. First, coarsening is performed, transforming the microscopic manifold coordinates... Mapping to more interpretable macroeconomic variables ,in Define a coarse-grained function parameterized by the neural network. This function learns by maximizing effective information. Effective information Defined as:

[0058]

[0059] First item The first term is mutual information, which measures the predictive power of a macroeconomic variable over time. It can be approximated using a neural network-based mutual information estimator. The second term is also mutual information, measuring the degree to which a macroeconomic variable depends on microscopic details. Hyperparameters are used to balance simplicity and predictive power. Maximizing This means finding a macroscopic description that discards redundant microscopic noise while retaining the strongest causal predictive power.

[0060] In the obtained macroscopic state space In the middle, construct the macroscopic dynamic equations:

[0061]

[0062] in, It is the decisive flow field, which contains causal emergent terms. This is achieved through a structurally constrained neural causal emergent network. To ensure the emergent property, the network is constrained, for example, the rank of its weight matrix is ​​limited to be less than the input dimension. It includes a non-linear activation function to ensure its output. Cannot be represented as input micro variables Any linear combination of these. This forces the network to capture nonlinear interaction patterns that emerge only at the macroscopic level. The fault attractor corresponds to the equation stable fixed point These fixed points are solved using numerical methods (such as Newton's iteration method). For a obtained stable fixed point... Calculate its Jacobian matrix.

[0063]

[0064] like All eigenvalues ​​real parts ,but It is a stable attractor. The boundary of the region of attraction can be analyzed. The eigenvector direction, or by calculating a scalar potential field. (satisfy The contour lines are used to approximate the determination. The system calculates the current macroscopic state in real time. To the most recent fault attractor Distance of the attraction domain boundary This distance It is a direct criterion for early warning.

[0065] When distance When the value falls below a preset threshold (e.g., entering the boundary of the attraction domain or falling below a certain empirical value), the system initiates a localization analysis. The key to localization lies in determining the current state point. The fastest glide direction towards the fault attractor on the manifold. Therefore, in the micromanifold Construct a holographic attention field First, identify the fault attractor. The corresponding region on the micromanifold This can be achieved by satisfying all of them. point The set is used to approximate the value. For any point on the manifold... Its holographic attention value is defined as:

[0066]

[0067] in, It is a kernel function defined on a manifold, such as a Gaussian kernel based on geodesic distance. It assigns a weight to the contribution of the global point to the current point. It is a point The similarity to the fault region can be defined as In practical calculations, this integral is approximated by a weighted summation over a set of sampled points on the manifold. The key manifold direction represents this scalar field. At the current point Covariant derivative at point This tangent vector indicates the path direction in manifold geometry where the state evolves towards the fault region with the least resistance.

[0068] Obtain key directions Then, it needs to be reverse-mapped back to the original physical sensor space to locate the fault source. This is achieved by solving an optimization problem: finding a tensor for the original data tensor Minimum perturbation This allows the perturbation to undergo the same manifold embedding process. Then, the direction of the displacement generated on the manifold is the same as... Consistent. That is:

[0069]

[0070] in, It is a coefficient that controls the mapping scale; It is a regularization term used to ensure The physical rationality, for example, the use of Norms promote sparsity in the spatial dimension (faults are often local), or a total variational regularization term promotes smoothness. Solving this optimization problem (e.g., using proximal gradient descent) yields... This is referred to as the "difference field." The analysis of this difference field in the spatial dimension... Energy distribution on The spatial location with the highest energy This corresponds to the point in the physical sensor array where the anomaly is most pronounced, thus achieving precise physical location of the fault source. The system generates an alarm containing specific coordinates based on this location information.

[0071] Furthermore, this method is extended to maintenance decision support. Decision-making is formalized as an optimal control problem on a joint state-action manifold. First, discrete maintenance actions (such as "reduce load to 70%" or "planned maintenance of phase A") are parameterized as continuous action vectors. Construct a joint manifold The points on it are represented as Define the value function on this manifold. , indicating from state Depart and execute the action The expected long-term return (negative cost) that can be obtained afterward. Optimal strategy. The Hamilton-Jacobi-Bellman equations should be satisfied:

[0072]

[0073] in It is the macro-dynamics influenced by action. It is an instantaneous cost function that comprehensively considers safety risks, power outage losses, and maintenance costs.

[0074] To solve this complex equation, this method employs a deep game theory network for approximation through adversarial training. This network consists of two parts: a policy network and a policy network. and value network The policy network proposes action plans, and the value network evaluates the long-term value of these plans. Both are jointly trained through an adversarial process: the policy network... Trained to make Maximize action; and value network The algorithm is trained to fit the true reward estimated under the current policy via temporal difference learning or Monte Carlo sampling. Training uses a manifold-constrained optimizer to ensure that gradient updates occur in the tangent space of the joint manifold. After training converges, the algorithm is applied to the currently diagnosed macroscopic state. The optimal maintenance action is Based on this, the system can automatically generate a recommended maintenance strategy report that includes specific measures, expected results, and execution time windows.

[0075] In summary, this method embodiment clearly reveals how to achieve intelligent perception, accurate diagnosis, and proactive maintenance of transformer faults through a complete, coherent, and detailed technical chain, from data tensor processing, manifold geometry embedding, neural differential dynamics modeling, causal emergence analysis, holographic attention localization, to game-theoretic optimization decision-making. Each algorithmic step, mathematical formula, and its parameters have clear technical definitions, physical origins, and functional roles, enabling those skilled in the art to implement the invention based on this description and necessary engineering knowledge.

[0076] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0077] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for locating transformer fault alarms, characterized in that, Includes the following steps, Obtain the original data tensor of the transformer multidimensional sensor array. The original data tensor contains data in the time dimension, spatial location dimension, and physical quantity type dimension. The original data tensor is embedded into a low-dimensional smooth manifold to obtain a manifold coordinate sequence characterizing the overall state of the transformer. This embedding process is achieved through higher-order tensor decomposition combined with spectral manifold regularization. Based on the aforementioned manifold coordinate sequence, a dynamic manifold evolution model is constructed. This model describes the continuous evolution of the transformer state along the manifold dimension. Based on the dynamic manifold evolution model, fault attractors and their attraction domains on the manifold are identified, and the distance between the current manifold coordinates and the attraction domain of the identified fault attractor is calculated. When the distance is less than a preset threshold, the key manifold direction that triggers the state evolution is determined based on the current manifold coordinates and the dynamic manifold evolution model. The key manifold direction is reverse-mapped to the original data tensor space to generate a difference field with the original data, so as to locate the fault source in the physical space and generate alarm information.

2. The transformer fault alarm location method according to claim 1, characterized in that, The step of embedding the original data tensor into a low-dimensional smooth manifold specifically includes, The original data tensor is subjected to Tucker decomposition to obtain the core tensor and the corresponding multi-dimensional factor matrix. A spectral manifold constraint is introduced into the objective function of the decomposition. This constraint is implemented by using the quadratic form term of the Laplacian matrix constructed based on data similarity as a regularization term. The manifold coordinate sequence is calculated by performing singular value decomposition on the expanded matrix of the core tensor, or by performing a specific combination of the factor matrices.

3. The transformer fault alarm location method according to claim 1, characterized in that, The dynamic manifold evolution model is obtained by learning a neural differential manifold defined on the low-dimensional smooth manifold. The neural differential manifold is constructed as a neural network operating on the tangent bundle of the low-dimensional smooth manifold, used to learn the drift and diffusion term functions of the manifold coordinate sequence evolution. The drift term function and the diffusion term function together constitute a stochastic differential equation on the manifold, which is used to describe the continuous evolution of the state.

4. The transformer fault alarm location method according to claim 1, characterized in that, The step of identifying fault attractors and their attraction regions on the manifold includes, The manifold coordinates are coarse-grained to obtain macroscopic state variables. This coarse-grained mapping is obtained by solving a learning process that maximizes the difference between the mutual information between macroscopic state variables and the mutual information between macroscopic states and microscopic states. In the space defined by the macroscopic state variables, macroscopic dynamic equations are constructed, and the stable fixed points of the macroscopic dynamic equations are solved. These stable fixed points are then used as fault attractors. Calculate the Jacobian matrix of the macroscopic dynamic equation at the stable fixed point, and determine the attraction domain boundary of the fault attractor based on the eigenvalues ​​of the Jacobian matrix.

5. The transformer fault alarm location method according to claim 4, characterized in that, The macroscopic dynamic equations contain causal emergence terms. The causal emergence term is learned through a neural causal emergence network whose structure is constrained so that its output function cannot be represented by a linear combination of the input microstate variables.

6. The transformer fault alarm location method according to claim 1, characterized in that, The step of determining the key manifold direction that triggers the state evolution based on the current manifold coordinates and the dynamic manifold evolution model includes: A holographic attention field is constructed on the manifold. The value of the holographic attention field at any point is obtained by weighting and integrating the similarity between all points on the manifold and the fault attractor region pointed to by that point using a holographic kernel function. Calculate the covariant derivative of the holographic attention field at the current manifold coordinates, and take the direction of the covariant derivative as the key manifold direction.

7. The transformer fault alarm location method according to claim 4, characterized in that, It also includes, Based on the macroscopic state variables, the fault attractor information, and the predefined set of maintenance actions, a state-action joint manifold is constructed, and a value function manifold is defined on the state-action joint manifold. By solving the Hamiltonian-Jacobi-Bellman equations on the state-action joint manifold, the optimal policy function from the macroscopic state to the optimal maintenance action is obtained. Based on the current macroscopic state and the optimal strategy function, a recommended maintenance strategy is generated.

8. The transformer fault alarm location method according to claim 7, characterized in that, The step of obtaining the optimal policy function by solving the Hamilton-Jacobi-Bellman equation is approximately implemented by a deep game theory network through adversarial training. The deep game theory network comprises a policy network and a value network. The policy network is used to generate parameterized action policies on the state-action joint manifold. The value network is used to evaluate the long-term expected return of the action strategy, and the two are jointly optimized through an adversarial learning process based on gradient descent on a manifold.

9. A transformer fault alarm and location system, characterized in that, The system for implementing the transformer fault alarm location method as described in any one of claims 1 to 8 includes, The holographic sensing array comprises multiple sensor nodes integrating vibration, ultrasonic, and temperature sensing units. These sensor nodes achieve coherent sensing between the arrays via metasurface antennas, enabling collaborative acquisition of the transformer's multidimensional raw data tensor. The manifold computation processing unit includes a hardware acceleration unit for performing large-scale tensor shrinking and manifold embedding computations, and a dedicated processing unit for processing graph structures and manifold topology operations. The interactive terminal includes a light field display device and a human-machine interface, used to visualize the manifold state, evolution process and alarm location results, and to receive interactive commands.

Images